On Mon, 16 Nov 2026 09:14:31 GMT, Ashay Rane <[email protected]> wrote:

>> This patch enables PAC and AUT instructions for pointer signing and
>> authentication in JIT code on Windows/ARM64 when "UseBranchProtection"
>> flag is set and when the processor supports the feature.
>> 
>> Specifically, JVM detects hardware support by reading ID_AA64ISAR1_EL1
>> and ID_AA64ISAR2_EL1 registers via CP4031 and CP4032 keys in the Windows
>> registry.  Then, if the "UseBranchProtection" flag is set, then HotSpot
>> emits the signing and authenticating instructions.  Windows uses
>> AArch64's "B" key, so this patch selects between "paciaz" and "pacibz"
>> instructions (and similarly, between "autiaz" and "autibz" instructions)
>> depending on whether HotSpot is being compiled for Linux versus Windows.
>> Both the MacroAssembler and the Windows' PAC helpers need to stay in
>> sync, so this patch needs changes to both macroAssembler_aarch64.cpp as
>> well as pauth_windows_aarch64.inline.hpp.
>> 
>> Since Linux supports both "A" as well as "B" keys, it's possible to use
>> "B" keys for both Linux and Windows, although I fear doing so may
>> introduce more churn in the code.
>> 
>> This patch also introduces a cross-platform (Linux and Windows) test
>> that compiles and runs a method under both C1 and C2 before checking
>> whether the generated instructions include the "paci[ab]z" and the
>> "auti[ab]z" instructions.
>> 
>> CI currently doesn't test with branch protection enabled, so I manually
>> changed the default value of "UseBranchProtection" in
>> globals_aarch64.cpp to be "pac-ret" before running tier 1 through tier 3
>> tests in the HotSpot, JDK, and LangTools test groups to ensure that this
>> patch does not introduce any new failures.  I also validated this patch
>> by running the same tests with the default value for
>> "UseBranchProtection".
>> 
>> Going forward, to catch return address corruption bugs, this patch
>> enables branch protection in CI for Windows/ARM64.  We deliberately do
>> not enable branch protection in CI for Linux/AArch64 so that if/when a
>> failure occurs on Windows/ARM64 and if the failure seems unrelated to
>> differences in the OS, then we have a better chance of knowing whether
>> stricter return address protection caused the failure.
>> 
>> ---------
>> - [x] I confirm that I make this contribution in accordance with the 
>> [OpenJDK Interim AI Policy](https://openjdk.org/legal/ai).
>
> Ashay Rane has updated the pull request incrementally with one additional 
> commit since the last revision:
> 
>   Make IO.java less brittle by permitting HotSpot warnings
>   
>   Also update DataDumpDcmdTest.java and LoadAgentDcmdTest.java to use
>   `OutputAnalyzer.stderrShouldBeEmptyIgnoreVMWarnings()`.

Ran through tier 1 through 3 tests on a machine with hardware support (for 
paca) and on a machine without hardware support, both with 
UseBranchProtection=standard versus UseBranchProtection=none, and I didn't see 
any new failures.  The Windows/x64 CI failure in the last run was strange since 
this patch doesn't change anything for the x64 port.  When I re-ran that job, 
it passed (see 
https://github.com/raneashay/openjdk-jdk/actions/runs/43837074309).

Merged the master branch to trigger CI.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/33595#issuecomment-6939764798

Reply via email to