RFR: 8308503: AArch64: SIGILL when running with -XX:UseBranchProtection=pac-ret on hardware without PAC feature [v2]

[Andrew Haley]

On Thu, 25 May 2023 08:29:17 GMT, Hao Sun <haosun at openjdk.org> wrote:

>> When revisiting the behavior of UseBranchProtection [1], we get one SIGILL error when running with -XX:UseBranchProtection=pac-ret on hardware without PAC.
>> 
>> Problem:
>> 
>> We build and run `java --version` with the following configuration matrix `Config X VMoption X Machine`.
>> 
>> 
>>   Config = {--enable-branch-protection, null}
>>   VMoption = {-XX:UseBranchProtection=pac-ret, -XX:UseBranchProtection=standard}
>>   Machine = {w/ PAC, w/o PAC}
>> 
>> 
>> VM crashes with SIGILL error for configure `Config=null, VMoption=pac-ret, Machine=w/o PAC`. The unrecognized instruction is `pacia x30, x29`, i.e. `pacia(lr, rfp)` generated by function `MacroAssembler::protect_return_address()`. [2]
>> 
>> Root cause:
>> 
>> 1. Instruction `pacia` is not in the NOP space. That's why `Config=null, VMoption=pac-ret` passes on `hardware w/ PAC`, but fails on `hardware w/o PAC`.
>> 
>> 2. -XX:UseBranchProtection=pac-ret behaves differently from the document [3], i.e.
>> 
>> 
>>   In order to use Branch Protection features in the VM,
>>   --enable-branch-protection must be used
>> 
>> 
>> `_rop_protection` is not turned off for `Config=null`. That's why `VMoption=pac-ret, Machine=w/o PAC` passes with
>> `Config=--enable-branch-protection` but fails with `Config=null`.
>> 
>> Fix:
>> 
>> This patch refines the parsing of -XX:UseBranchProtection=pac-ret:
>> 
>> 1. We handle "pac-ret" and "standard" in the same way, since only one type of branch protection is implemented for now, i.e. "pac-ret". We may update "standard" in the future if "bti" protection is added.
>> 
>> 2. `_rop_protection` is not turned on unless all the three conditions are satisfied [4]. Otherwise, it's kept off and one warning message is emitted.
>> 
>> 
>> // Enable PAC if this code has been built with branch-protection, the
>> // CPU/OS supports it, and incompatible preview features aren't enabled.
>> 
>> 
>> [1] https://bugs.openjdk.org/browse/JDK-8287325?focusedCommentId=14581099&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-14581099
>> [2] https://github.com/openjdk/jdk/blob/master/src/hotspot/cpu/aarch64/macroAssembler_aarch64.cpp#L5976
>> [3] https://github.com/openjdk/jdk/blob/master/doc/building.md#branch-protection
>> [4] https://github.com/openjdk/jdk/blob/master/src/hotspot/cpu/aarch64/vm_version_aarch64.cpp#L457
>
> Hao Sun has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Disable ROP-protection if VMContinuations is on
>   
>   Virtual threads are proposed to be a permanent feature in JDK 21, and
>   Arguments:enable_preview check no longer works. As an alternative, we
>   check if VMContinuations is on.
>   
>   In this way, ROP protection is enabled only with VM options
>   `-XX:UseBranchProtection=standard|pac-ret
>   -XX:+UnlockExperimentalVMOptions -XX:-VMContinuations` on hardware with
>   the support of PAC feature.

src/hotspot/cpu/aarch64/vm_version_aarch64.cpp line 468:

> 466:     } else if (VMContinuations) {
> 467:       // Not currently compatible with continuation freeze/thaw.
> 468:       warning("ROP-protection is incompatible with virtual threads feature. Disabling ROP-protection.");

Suggestion:

      warning("ROP-protection is incompatible with VMContinuations. Disabling ROP-protection.");

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

PR Review Comment: https://git.openjdk.org/jdk/pull/14095#discussion_r1205219492