[SR-2655] in libdispatch/libpwq worker threads, too many signals are blocked on Linux

[weissi]

Previous ID	SR-2655
Radar	None
Original Reporter	@weissi
Type	Bug
Status	Resolved
Resolution	Done

Attachment: Download

Additional Detail from JIRA

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

md5: 071933e7988663468728927c7cfe2cd7

Issue Description:

Description

Certain signals are masked/blocked on libdispatch/libpwq worker threads.

On Darwin, the list of masked signals is the inverse of workq_threadmask:

#define sigcantmask (sigmask(SIGKILL) | sigmask(SIGSTOP))
#define threadmask (sigmask(SIGILL)|sigmask(SIGTRAP)|\
            sigmask(SIGIOT)|sigmask(SIGEMT)|\
            sigmask(SIGFPE)|sigmask(SIGBUS)|\
            sigmask(SIGSEGV)|sigmask(SIGSYS)|\
            sigmask(SIGPIPE)|sigmask(SIGKILL))

#define workq_threadmask (threadmask | sigcantmask)

There is actually a function _dispatch_pthread_sigmask which given a full sigmask calculates a similiar sigmask which is then used in certain situations.

However, the overcommit_worker_main function in libpwq chooses to mask all the signals which doesn't seem correct. To fix this issue, it seems enough to only block the signals _dispatch_pthread_sigmask elects instead of all.

There is also another place in libpwq where all signals are blocked on a thread. I'm unsure whether that is correct.

Repro

Running the attached test.c shows the problem:

Darwin

Darwin jwmbp.local 16.0.0 Darwin Kernel Version 16.0.0: Mon Aug 29 17:56:21 PDT 2016; root:xnu-3789.1.32~3/DEVELOPMENT_X86_64 x86_64
main
====
dispatch
=======
dispatch: SIGHUP is BLOCKED
dispatch: SIGINT is BLOCKED
dispatch: SIGQUIT is BLOCKED
dispatch: SIGALRM is BLOCKED
dispatch: SIGTERM is BLOCKED
dispatch: SIGURG is BLOCKED
dispatch: SIGTSTP is BLOCKED
dispatch: SIGCONT is BLOCKED
dispatch: SIGCHLD is BLOCKED
dispatch: SIGTTIN is BLOCKED
dispatch: SIGTTOU is BLOCKED
dispatch: SIGIO is BLOCKED
dispatch: SIGXCPU is BLOCKED
dispatch: SIGXFSZ is BLOCKED
dispatch: SIGVTALRM is BLOCKED
dispatch: SIGWINCH is BLOCKED
dispatch: SIGUSR1 is BLOCKED
dispatch: SIGUSR2 is BLOCKED
pthread
=======

Linux

libdispatch @ https://github.com/apple/swift-corelibs-libdispatch/tree/7fe8323838b0491691e515a7f826ffe508668856 ish

Linux build 4.4.0-36-generic #​55-Ubuntu SMP Thu Aug 11 18:01:55 UTC 2016 x86_64 x86_64 x86_64 GNU/Linux
main
====
dispatch
=======
dispatch: SIGHUP is BLOCKED
dispatch: SIGINT is BLOCKED
dispatch: SIGQUIT is BLOCKED
dispatch: SIGILL is BLOCKED
dispatch: SIGTRAP is BLOCKED
dispatch: SIGABRT is BLOCKED
dispatch: SIGIOT is BLOCKED
dispatch: SIGFPE is BLOCKED
dispatch: SIGBUS is BLOCKED
dispatch: SIGSEGV is BLOCKED
dispatch: SIGSYS is BLOCKED
dispatch: SIGPIPE is BLOCKED
dispatch: SIGALRM is BLOCKED
dispatch: SIGTERM is BLOCKED
dispatch: SIGURG is BLOCKED
dispatch: SIGTSTP is BLOCKED
dispatch: SIGCONT is BLOCKED
dispatch: SIGCHLD is BLOCKED
dispatch: SIGTTIN is BLOCKED
dispatch: SIGTTOU is BLOCKED
dispatch: SIGIO is BLOCKED
dispatch: SIGXCPU is BLOCKED
dispatch: SIGXFSZ is BLOCKED
dispatch: SIGVTALRM is BLOCKED
dispatch: SIGPROF is BLOCKED
dispatch: SIGWINCH is BLOCKED
dispatch: SIGUSR1 is BLOCKED
dispatch: SIGUSR2 is BLOCKED
pthread
=======

[swift-ci]

Comment by Daniel A. Steffen (JIRA)

as of Sierra, SIGPROF is also not masked by the Darwin kernel (needs to be fixed in _dispatch_pthread_sigmask still).

dgrove-oss (JIRA User) do you think this is something we could get into libpwq or should we fix it in libdispatch ? (we could change the function passed to pthread_workqueue_additem_np() to re-adjust the signal mask, a bit wasteful of course since libpwq just did that too)

[swift-ci]

Comment by David Grove (JIRA)

I think we should be able to modify libpwq as needed. frankeh (JIRA User), do you see any problems?

[swift-ci]

Comment by David Grove (JIRA)

Hubertus and I touched base this morning. Once the desired semantics are clear, it should be easy enough to make the changes to implement them.

It was unclear to us why libpwq is currently masking signals for the overcommit threads, but not for the regular worker threads. libpwq also masks signals on the manger thread.

From the dispatch perspective, what is the expected semantics of signal handling?

[swift-ci]

Comment by Daniel A. Steffen (JIRA)

dgrove-oss (JIRA User) we don't expect any difference between overcommit threads and non-overcommit (regular worker) threads as far as signals are concerned. I don't see any obvious reason why that is not the case in libpwq either, that seems wrong to me.

The expected semantics are that all non-synchronous signals (signals that are targeted at the process as a whole) are masked on worker threads.
The only signals that should be unmasked are the ones that are delivered to the generating thread itself (excluding the case of explicit calls to pthread_kill targeting the calling thread).

On Darwin kernel, the unmasked signals are SIGILL SIGTRAP SIGEMT SIGFPE SIGBUS SIGSEGV SIGSYS SIGPIPE SIGPROF (which is what _dispatch_pthread_sigmask emulates for threads manually pthread_create()d by the library)

[swift-ci]

Comment by David Grove (JIRA)

Opened pull request to libpwq (mheily/libpwq#21 Once it is merged, I'll put in a PR to libdispatch to update the libpwq version.

[swift-ci]

Comment by David Grove (JIRA)

libpwq changes merged upstream and dispatch updated in #179 and #180