RFR(S): Use Vectored Exception Handling on Windows

[Thomas Stüfe]

Hi Ludovic,

sorry for the delay, and thanks for the extensive answer. Please find
remarks inline.

On Fri, Jun 26, 2020 at 12:11 AM Ludovic Henry <luhenry at microsoft.com>
wrote:

> Hi Thomas,
>
> It seems that the problem you're describing stems from the current
> exception handler treating two cases: 1. any exception knowingly triggered
> by Java code and treated by HotSpot (ex: safepoint-polling, arraycopy
> stubs, stackoverflow in Java code), and 2. exceptional cases leading to
> crashes (ex: uncaught C++ exception, an access violation in VM or
> native/external code, etc.). There is the same problem on Unix because
> there is only one system (signal handling) for both cases. Fortunately,
> Windows proposes different systems, each with its own advantages.
>
> The order in which Windows invokes each of these systems is the following:
>  1. Vectored Exception Handler registered with
> `AddVectoredExceptionHandler`
>  2. Structured Exception Handler
>  3. Vectored Exception Handler registered with `AddVectoredContinueHandler`
>  4. Unhandled Exception Handler
>
> Today, Hotspot on x86/x86_64 catches the exception at 2. via a handler
> registered with `RtlAddFunctionTable`. This handler does both the
> Java-triggered exceptions and any other exceptions.
>
> Now, from the point of view of an external library or application
> embedding the JVM inside their own process, they still have all the above
> options to register an exception handler, irrespective of how Hotspot does
> it. This creates the following cases:
>  - If the application uses VEH: they will (with Hotspot using SEH) be
> called _before_ Hotspot's exception handler and will then have to be aware
> that they may get exceptions unrelated to them and will have to ignore them
> accordingly
>  - If the application uses SEH: they will only get exceptions related to
> their code area
>
> If Hotspot is to use VEH, an exception would play as follow:
>  - If the application uses VEH and their registered handler executes
> _before_ Hotspot's one: same as above
>  - If the application uses VEH and their registered handler executes
> _after_ Hotspot's one: Hotspot has to make sure that the exception was
> triggered by Hotspot and ignore them otherwise (a range check on the PC can
> be used here to emulate how it's done with RltAddFunctionTable)
>  - If the application uses SEH: the same case as to where the
> application's handler executes _after_ Hotspot's one
>
> This all assumes that Hotspot's VEH handler doesn't trigger a crash report
> (VMError::report_and_die) on any exception it doesn't know how to handle.
> The simplest way to do that is simply _not_ to do it in Hotspot's VEH
> handler, and to do it by registering a Win32 Unhandled Exception Handler
> (with SetUnhandlerdExceptionFilter [1]). This handler is _only_ called when
> no other exception handler treated the exception (by returning
> EXCEPTION_CONTINUE_EXECUTION or EXCEPTION_EXECUTE_HANDLER). Invoking it
> means the application is "toast" and not in a runnable state anymore, which
> fits nicely with the purpose of the Hotspot crash report.
>
>
Okay, If I get this correctly:

Today:
  App uses VEH - they execute before us and have to handle this correctly
(->A)
  App uses SEH - no interaction

With proposed switch:
  App uses VEH - they may or may not execute before us. If they come before
us: (->A). If they come after us -> (B)
  App uses SEH -> (B)

A) this case exists today. An app getting signals via VEH would have to
willingly ignore signals for us to get them. This does not change, your
patch would mean this happens less often, so I do not see a backward
compatibility problem here.

B) this is a new case. We would have to ignore signals not meant for us.
Technically by just ignoring them. Distinguishing this is a bit difficult
though. Note the subtle difference to Unix: there we have signal chaining,
so an application which is really really interested in signals for its own
purposes uses it (e.g. by preloading libjsig) and then we know its handler
and hand over the signal.

On windows we do not know this (?), we only can distinguish our crashes
from their crashes via crash pc, rejecting any crash not in our code
(dynamic or static). Well, arguably this would be just how it is today with
our code scoped via SEH. With the added safety net of the unhandled
exception filter (what happens if multiple parties call this?).

Okay this seems safe enough to try it at least.

My only very small personal gripe would be that I always liked how I can
quickly use SEH to check if a pointer is valid without disturbing anyone.
But within the hotspot at least I can just as well use SafeFetch.

Thank you,

Thomas

I hope this sheds some light on possible solutions ahead of us.
>
> Thank you,
>
> --
> Ludovic
>
> [1]
> https://docs.microsoft.com/en-us/windows/win32/api/errhandlingapi/nf-errhandlingapi-setunhandledexceptionfilter
> ________________________________________
> From: Thomas Stüfe <thomas.stuefe at gmail.com>
> Sent: Sunday, June 21, 2020 05:55
> To: Ludovic Henry
> Cc: hotspot-runtime-dev at openjdk.java.net
> Subject: Re: RFR(S): Use Vectored Exception Handling on Windows
>
> Hi,
>
> We at SAP had used VEH in our own Windows Itanium port and I dimly
> remember it being a source of problems. That is many years ago and I
> realize that it is not worth much, but it makes me bit apprehensive of this
> change.
>
> The main problem I see is that this will be an observable change in
> behavior.
>
> We currently use SEH, so our error handler is guaranteed to be invoked
> only for exceptions from within our own code. With VEH we now follow the
> Unix way of things and suddenly our error handler becomes a global resource.
>
> We will suddenly be invoked for crashes outside the VM, e.g. in foreign
> launcher code atop of us or in non-java side threads, which will generate
> whole new classes of hs-err files for crashes the VM is not responsible
> for. Which are then perceived as VM crashes and sent to us vendors instead
> of going to the right people. This is the way it works on Unix today, and
> it is a constant annoyance and increases our support workload.
>
> We also may introduce new problems since suddenly we interfere with
> application exception handling. At the very least, we have to think up a
> scheme for signal chaining (both ways: VM->foreign code and foreign
> code->VM). For the first, we probably need some form of libjsig preloading,
> or some other way to divert signal handler instalment. That would also need
> cooperation from the application programmers and/or operators.
>
> Matters are even more complicated, since foreign code may use SEH instead
> of VEH, so what happens if a JNI library below me wants to use SEH, does
> that still work?
>
> I feel this should not be rushed. Even considered "brittle" SEH has served
> us well, I do not recall many problems in the past aside from having to add
> the occasional __try/__except. Are there actual bugs we have to solve?
>
> Lastly, personally I always found SEH quite a neat concept, and one of the
> few places where Windows was superior to Unix :)
>
> Thanks, Thomas
>
>
> On Fri, Jun 19, 2020 at 5:23 PM Ludovic Henry <luhenry at microsoft.com
> <mailto:luhenry at microsoft.com>> wrote:
> Hello,
>
> First, some context and definitions:
> - when talking about exception here, I'm talking about Win32 exception
> which are equivalent to signals on Linux and other Unix, I am _not_ talking
> about Java exceptions.
> - an explanation of an _exception filter_ can be found at
> https://docs.microsoft.com/en-us/cpp/cpp/writing-an-exception-filter?view=vs-2019
> <
> https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdocs.microsoft.com%2Fen-us%2Fcpp%2Fcpp%2Fwriting-an-exception-filter%3Fview%3Dvs-2019&data=02%7C01%7Cluhenry%40microsoft.com%7Cd552fedab47f45c6fe9808d815e2758f%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637283409665642403&sdata=fjcrwcQYAg3TstTSO2YHKziszwlusbYV6uUXINydD1E%3D&reserved=0>.
> There is only a limited concept of that in Java with type-based exception
> filter (ex: `try { ... } catch (IOException ioe) { ... } catch (Throwable
> t) { ... }`).
> - in Win32, there exist two exception handling mechanism:
>   - Structured Exception Handling: the historical one, based on `__try {}
> __except (...) {}`
>   - Vectored Exception Handling: introduced in Windows XP / Windows Server
> 2003, much more similar to signals on Linux
>
> These exception handling mechanisms are used to catch any exceptions like
> Access Violation, Stack Overflow, Divide by Zero, Overflow, and more. These
> exceptions are equivalent to signal on Linux and are then core to many
> mechanisms in the OpenJDK.
>
> Today, the OpenJDK uses Structured Exception Handling to catch such
> exceptions, creating several requirements. First, all code that might
> trigger an exception on purpose (like a Access Violation / SIGSEGV in the
> arraycopy stub), needs to be wrapped up in a __try / __except. Because it's
> not feasible to wrap every single instance of such code, these __try /
> __except are put at the top-level most function of any thread started by
> the runtime. Second, for code generated by Hotspot, `RtlAddFunctionTable`
> is used to simulate the use of __try / __except for a specific code area.
> This function needs platform specific code with the generation of  a
> trampoline that calls the exception filter declared in the runtime. It's
> also meant to be used as a one to one mapping with try / catch in user
> code, and not as a "catch all the exceptions in this code area". Third,
> Structured Exception Handling expects to be able to unwind the stack.
> However, because Hotspot doesn't guarantee the usage of the
> platform-specific ABI internally, the platform-specific unwinder might
> break. Hotspot's usage of `RtlAddFunctionTable` for the code cache relies
> on the assumption that Structured Exception Handling never tries to unwind
> the stack (which it would fail to do because of the different ABI) before
> calling the registered exception filter.
>
> Discussing that with Windows Kernel maintainers, this approach is highly
> discouraged, considered brittle, and the better solution is Vectored
> Exception Handling. Vectored Exception Handling is conceptually much more
> similar to signal / sigaction on Linux and other Unix systems. It will
> catch all exceptions happening across the process, and no __try / __except
> will be required. It also removes the requirement to call
> `RtlAddFunctionTable`.  The exception filter then behaves like a signal
> handler with the possibility to modify the registers at will, modifying the
> PC to step over an instruction after an expected Access Violation for
> example. Vectored Exception Handling is also already used for AOT code.
>
> The changes can be found at
> http://cr.openjdk.java.net/~burban/ludovic_vecexc/<
> https://nam06.safelinks.protection.outlook.com/?url=http:%2F%2Fcr.openjdk.java.net%2F~burban%2Fludovic_vecexc%2F&data=02%7C01%7Cluhenry%40microsoft.com%7Cd552fedab47f45c6fe9808d815e2758f%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637283409665652395&sdata=pTewy1%2BeB43HX4y0ypDwMDGRjBoNP6yBGrhRi7ncm1c%3D&reserved=0>.
> As I am not an author, I have not created a corresponding bug in JBS.
>
> Thank you, and looking forward for your feedback!
>
> --
> Ludovic
>
>
>