Disable backtraces on OSX

[alexcrichton]

This PR is a revert of #44251 which disables backtraces via RUST_BACKTRACE=1 on OSX. It was discovered in #45731 that OSX binaries generating a backtrace started generating a segfault in the range of 4279e2b...8493813. The most likely candidate PR in that range is #45400 which switched to compiling libtest with ThinLTO. Furthermore it's probably mostly likely that libbacktrace is segfaulting due to #45511, or otherwise bad DWARF data in libtest.

Now the bad DWARF information is certainly a bug with ThinLTO that needs to be fixed! Despite this, however, this is a presumed resurgence of #21889 where you can feed arbitrary data to libbacktrace effectively, and if it segfaults that's presumably a security vulnerability. In that sense a fix to DWARF + ThinLTO is not enough, but this indicates a deeper underlying problem that OSX binaries may be vulnerable to the situation described in #21889.

As a result, this PR reverts backtrace support for OSX, which can return once we're confident enough in the implementation of libbacktrace for OSX. I'm also going to nominate this for a beta backport as it's something we'll want to get out pretty soon.

Closes #45731

[alexcrichton]

cc @rust-lang/libs
cc @JohnColanduoni

[rust-highfive]

r? @sfackler

(rust_highfive has picked a reviewer for you, use r? to override)

[kennytm]

cc #24346

[sfackler]

Is there a way of doing this that's less invasive to the libbacktrace source?

[JohnColanduoni]

It's worth noting that if the vulnerability is indeed in the DWARF parser it is likely on other platforms as well; there is no OS X specific code in the DWARF parser and its interactions with it are minimal (it just searches for the appropriate named sections).

The stack trace from #45731 makes me doubt that this is the case unless the given crash is due to the DWARF parser trashing memory and the Mach-O code simply happens to be the one to step in it.

@sfackler The only this that needs to reverted to prevent the code from running is the changes to the cfg attributes in sys_common/mod.rs and sys/unix/backtrace/printing/mod.rs from the first commit. The Mach-O code will not be compiled into libbacktrace on non-Apple platforms.

[JohnColanduoni]

I tracked down the crash in #45731. It turns out the SIGBUS was due to an IO error (attempt to access page in empty mmaped file), not an errant memory access.

[alexcrichton]

Thanks for the investigation @JohnColanduoni! Do you have a fix available locally for this? If so I think we can probably just apply that and avoid reverting.

@sfackler probably, I just did the easiest thing of reverting the PR rather than trying to pick-and-choose what happened.

[sfackler]

Ah gotcha, I didn't realize the OSX stuff was already a patch on top of libbacktrace.

[JohnColanduoni]

@alexcrichton This should do it: #45866.

[alexcrichton]

Awesome! Closing in favor of that