Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

c-deps: bump CryptoPP to avoid SIGTRAP on macOS #31516

Merged
merged 1 commit into from
Oct 16, 2018

Conversation

benesch
Copy link
Contributor

@benesch benesch commented Oct 16, 2018

Bump CryptoPP to pick up a fix for #31380.
Details reproduced below.

Fix #31380.


As part of its CPU feature detection, CryptoPP installs a SIGILL signal
handler before issuing the cpuid instruction. The intent is to
gracefully degrade on CPUs that don't support the cpuid instruction.

The problem is that it is impossible to safely overwrite a signal
handler installed by the Go runtime in go1.10 on macOS
(golang/go#22805). This causes CockroachDB 2.0 to crash on macOS Mojave:
#31380.

The situation has improved on the Go front, as go1.11 makes it possible
to safely save and restore signal handlers installed by the Go runtime
on macOS.

Still, we can do better and support go1.10. There is no need to bother
installing a SIGILL handler, as the cpuid instruction is supported by
every x86-64 CPU. We can instead use conditional compilation to make
sure that we never execute a cpuid instruction on a non x86-64 CPU.

Note that CPU feature detection is performed at executable load time
(see the attribute(constructor) on DetectX86Features); therefore any
reference to function which calls DetectX86Features (notably HasAESNI)
corrupts the signal handler. It's not entirely clear why this corruption
later leads to the SIGTRAP seen in #31380--is
something in macOS or the Go runtime generating a SIGILL and trying to
handle it gracefully?--but regardless, not mucking with the signal
handler fixes the issue.

Release note (bug fix): CockroachDB no longer crashes due to a SIGTRAP error
soon after startup on macOS Mojave (#31380).

Bump CryptoPP to pick up a fix for cockroachdb#31380.
Details reproduced below.

Fix cockroachdb#31380.

---

As part of its CPU feature detection, CryptoPP installs a SIGILL signal
handler before issuing the cpuid instruction. The intent is to
gracefully degrade on CPUs that don't support the cpuid instruction.

The problem is that it is impossible to safely overwrite a signal
handler installed by the Go runtime in go1.10 on macOS
(golang/go#22805). This causes CockroachDB 2.0 to crash on macOS Mojave:
cockroachdb#31380.

The situation has improved on the Go front, as go1.11 makes it possible
to safely save and restore signal handlers installed by the Go runtime
on macOS.

Still, we can do better and support go1.10. There is no need to bother
installing a SIGILL handler, as the cpuid instruction is supported by
every x86-64 CPU. We can instead use conditional compilation to make
sure that we never execute a cpuid instruction on a non x86-64 CPU.

Note that CPU feature detection is performed at executable load time
(see the attribute(constructor) on DetectX86Features); therefore any
reference to function which calls DetectX86Features (notably HasAESNI)
corrupts the signal handler. It's not entirely clear why this corruption
later leads to the SIGTRAP seen in cockroachdb#31380--is
something in macOS or the Go runtime generating a SIGILL and trying to
handle it gracefully?--but regardless, not mucking with the signal
handler fixes the issue.

Release note (bug fix): CockroachDB no longer crashes due to a SIGTRAP error
soon after startup on macOS Mojave (cockroachdb#31380).
@benesch benesch requested review from tbg, mberhault and a team October 16, 2018 19:30
@cockroach-teamcity
Copy link
Member

This change is Reviewable

@mberhault
Copy link
Contributor

LGTM

@benesch
Copy link
Contributor Author

benesch commented Oct 16, 2018

Thanks! bors r=mberhault

@benesch
Copy link
Contributor Author

benesch commented Oct 16, 2018

Urgh.

bors r=mberhault

craig bot pushed a commit that referenced this pull request Oct 16, 2018
31516: c-deps: bump CryptoPP to avoid SIGTRAP on macOS r=mberhault a=benesch

Bump CryptoPP to pick up a fix for #31380.
Details reproduced below.

Fix #31380.

---

As part of its CPU feature detection, CryptoPP installs a SIGILL signal
handler before issuing the cpuid instruction. The intent is to
gracefully degrade on CPUs that don't support the cpuid instruction.

The problem is that it is impossible to safely overwrite a signal
handler installed by the Go runtime in go1.10 on macOS
(golang/go#22805). This causes CockroachDB 2.0 to crash on macOS Mojave:
#31380.

The situation has improved on the Go front, as go1.11 makes it possible
to safely save and restore signal handlers installed by the Go runtime
on macOS.

Still, we can do better and support go1.10. There is no need to bother
installing a SIGILL handler, as the cpuid instruction is supported by
every x86-64 CPU. We can instead use conditional compilation to make
sure that we never execute a cpuid instruction on a non x86-64 CPU.

Note that CPU feature detection is performed at executable load time
(see the attribute(constructor) on DetectX86Features); therefore any
reference to function which calls DetectX86Features (notably HasAESNI)
corrupts the signal handler. It's not entirely clear why this corruption
later leads to the SIGTRAP seen in #31380--is
something in macOS or the Go runtime generating a SIGILL and trying to
handle it gracefully?--but regardless, not mucking with the signal
handler fixes the issue.

Release note (bug fix): CockroachDB no longer crashes due to a SIGTRAP error
soon after startup on macOS Mojave (#31380).

31517: roachtest: deflake acceptance/bank/zerosum-splits r=andreimatei a=benesch

This test requires that the experimental_force_split_at session var be
set to force ALTER ... SPLIT AT to work even with the merge queue
enabled. gosql.DB's connection pool will occasionally open a new
connection which does not have the var set. Set the session var in the
same batch of statements as the ALTER ... SPLIT AT command so that the
session var is always set in the session that executes the ALTER ...
SPLIT AT command.

Fix #31510.

Release note: None

Co-authored-by: Nikhil Benesch <[email protected]>
@craig
Copy link
Contributor

craig bot commented Oct 16, 2018

Build succeeded

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

SIGINT crash during graceful shutdown with memory storage on macOS Mojave
3 participants