Implement dtrace_getupcstack in ARM64

[wma]

Summary

This allows using DTRACE for performance analysis of userspace
applications - the function call stack can be captured.
This is almost an exact copy of AMD64 solution.

Example of FlameGraph for iperf3
https://drive.google.com/file/d/0B-7yTLrPxaWtYUJDc2FaVWRJeDA/view?usp=sharing

[gnn]

Only one requested change from me. I have not tested this directly, only read the code.

sys/cddl/dev/dtrace/aarch64/dtrace_isa.c

66 ↗

(On Diff #14733)

I think this ought to be a tunable.

163 ↗

(On Diff #14733)

If this is related to x86 do we need this at all?

[emaste]

sys/cddl/dev/dtrace/aarch64/dtrace_isa.c

66 ↗

(On Diff #14733)

I'm not sure; in hwpmc we have a similar setting kern.hwpmc.callchaindepth that is a tunable, but it defaults to 32. In this case it's just safety belt and should be higher than any user stack depth we support.

163 ↗

(On Diff #14733)

This came in rS179237 for x86, and it looks like mips and powerpc have the same behaviour (and powerpc has the same comment). Perhaps @jhibbits has some insight on the latter case.

It might be sensible to keep the code (I'm not sure), but the comment definitely needs to be updated.

[jhibbits]

sys/cddl/dev/dtrace/aarch64/dtrace_isa.c

163 ↗

(On Diff #14733)

I think part of this is that mips, arm*, and powerpc are all copied from x86. I chose not to remove the comment, thinking I would revisit it later with someone who actually knows dtrace. So, my answer for is it related only to x86 is... no idea. It might be x86 specific, in which case it can be removed from every other arch.

[emaste]

sys/cddl/dev/dtrace/aarch64/dtrace_isa.c

163 ↗

(On Diff #14733)

This has been refactored somewhat in MIPS, but on a quick look there seems to be an equivalent:

                        if (*flags & CPU_DTRACE_FAULT)
                                goto fault;
...
fault:
        /*
         * We just got lost in backtrace, no big deal
         */
        *flags &= ~CPU_DTRACE_FAULT;
        return (-1);

@markj do you have any insight?

[wma]

sys/cddl/dev/dtrace/aarch64/dtrace_isa.c

163 ↗

(On Diff #14733)

In that case I'll leave the code as-is but update the comment with ARM64TODO to remember that this thing needs to be checked. Either way, removing this stuff from other architectures should be made in a separate commit with thorough testing, which I don't feel competent enough to perform...

[jhibbits]

sys/cddl/dev/dtrace/aarch64/dtrace_isa.c

163 ↗

(On Diff #14733)

Agreed. We can revisit it separately for all architectures.

[markj]

sys/cddl/dev/dtrace/aarch64/dtrace_isa.c

163 ↗

(On Diff #14733)

I'm not sure what the exact problem scenario is here - this code was in the initial OpenSolaris release. I wonder if this code was hitting a fault while trying to unwind past a thread trampoline or something like that.

[wma]

Changed ustack_max to be a define, I'm not convinced if anyone should be allowed to change it so didn't export that one to sysctl.

Added some more meaningful info to the x86 JVM comment.

[wma]

Is there anything else that should be changed here or can I commit the latest version?