Re: [musl] Re: Pending patches for MT-fork stuff

[Rich Felker <dalias@libc.org>]

On Tue, Sep 29, 2020 at 04:34:07PM -0400, Rich Felker wrote:
> On Tue, Sep 29, 2020 at 01:51:30PM -0500, Jesse Hathaway wrote:
> > On Tue, Sep 29, 2020 at 1:36 PM Rich Felker <dalias@libc.org> wrote:
> > > Can you provide an strace (with -f) showing the hang? It's probably
> > > not related to this since fork does not seem to be involved. Depending
> > > on how you're using Go, it may just be Go bypassing libc then trying
> > > to use libc functions, which at least used to be a big problem; I
> > > don't know if it's fixed nowadays or not.
> > 
> > Thanks Rich, for taking a look, I have attached an strace of the
> > program compiled against musl & glibc. The first call to setreuid
> > succeeds in both, but the second call fails under musl. Jesse
> 
> The problem is this line:
> 
> > 8238  rt_sigprocmask(SIG_SETMASK, ~[HUP INT QUIT ILL TRAP ABRT BUS FPE SEGV TERM STKFLT CHLD PROF SYS RTMIN RT_1],  <unfinished ...>
> 
> Something broken in the Go runtime is bypassing libc and either
> calling SYS_rt_sigprocmask itself, or calling the libc sigprocmask
> function with a sigset_t it produced itself, blocking a libc-internal
> signal. This makes it invalid to make any further use of libc.
> 
> Either it (the Go runtime) needs to manipulate sigset_t objects via
> the public APIs for them (sigfillset, sigaddset, etc.) or its wrapper
> for sigprocmask needs to convert the Go-manipulated sigset_t to one
> valid for libc by iterating over the bits and using sigaddset, so that
> invalid bits don't end up in the one passed to libc.

Here is the offending code:

https://github.com/golang/go/blob/a413908dd064de6e3ea5b8d95d707a532bd3f4c8/src/runtime/signal_unix.go#L866

It should be calling sigfillset() (from libc) to get the starting
sigset_t rather than using its own all-one-bits initializer.

There may be other places in the runtime where the same error is made.
It looks like blockableSig (line 1132) is intended to do something
here, but has hard-coded (somewhere else) glibc knowledge rather than
probing via (libc's) sigaddset whether the signal number is valid.
This might be a preferred point to fix it at.

Rich