Exception handler - catch bad memory accesses by the JIT #11795

hrydgard · 2019-02-12T20:00:31Z

NOTE: See 2020 Update below.

Installs an exception handler that filters out exceptions from the JIT and let the emulated executable continue running (or bail out cleanly, if set not to ignore bad memory accesses).

The main idea is to prevent game crashes from crashing the emulator, even with fast-memory enabled. This will hopefully get rid of a lot of bad callstacks from the Google Play crash reporting, and improve the user experience with bad ISOs and whatever.

~~This is the last new major feature that will make the cut for 1.8.0. Other than this, it's just bugfixes until its out.~~

A bit of the exception handling code is taken from Dolphin. It's also GPL though, and I originally wrote that code anyway so it's fine :)

A few things remain:

Does not seem to be catching exceptions on ARM 32-bit?? or my bad test write accidentally went through and overwrote something in the crowded memory space.
Bad reads should return 0. Or maybe some other value, not sure - right now they effectively just return whatever was in the destination register, as the instruction is simply skipped.
SIMD instructions are not handled in the x86 analyze function. It also is very ugly and buggy in general, needs a rewrite, or maybe we can use udis.

Executable that does a bad write:

crash.zip

2020 update

Rebased this on master, now it has the lovely bluescreens from #13092.

Not yet tested on ARM32.

Note that it's a little confusing when running from VS' debugger - first the debugger will catch the exception. Just press F5 and it will continue to the handler.

unknownbrackets

Bad reads returning 0 at least matches fast memory off, so I think it is a good plan.

Travis seems to have still failed with actual failures, like error: unknown type name 'kern_return_t' on iOS, No context definition for architecture on x86 Android?, and a link error on x86_64 (doesn't have the x86 disassembler - would be the old mk file.)

-[Unknown]

unknownbrackets · 2019-02-15T13:46:32Z

Core/System.cpp

 	if (coreParameter.updateRecent) {
 		g_Config.AddRecent(filename);
 	}
+
+	InstallExceptionHandler(&Memory::HandleFault);


Is it okay that this may occur on a different thread than the CPU thread (depending on Vulkan vs OpenGL, I think?)

-[Unknown]

I believe all these methods are process-global, I know the Windows path is, but hm, maybe a concern indeed.

I guess it's probably only a concern for thread_set_exception_ports. I think for Mac it'll always be on the CPU thread (I guess it's safe on Vulkan too anyway.)

Checked and the threaded CPU startup is after this line.

-[Unknown]

Common/ExceptionHandlerSetup.cpp

unknownbrackets · 2019-02-15T13:54:25Z

UI/EmuScreen.cpp

@@ -1347,6 +1376,11 @@ void EmuScreen::renderUI() {
 		DrawProfile(*ctx);
 	}
 #endif
+
+	if (coreState == CORE_ERROR) {


Hm, we currently also use CORE_ERROR for LoadExec changes (some games switch executable) that fail, and also game load fail. We should probably make sure these still do what they're supposed to do...

-[Unknown]

Alternatively we should assign a new coreState? CORE_RUNTIME_ERROR?

That makes sense. Or maybe even CORE_ERROR_CRASHDUMP or something to indicate plainly that it's the sort of error a crashdump makes sense for?

I guess that would be all runtime errors, though, except maybe LoadExec swaps of the binary mid-game.

-[Unknown]

Yeah, I think CORE_ERROR_RUNTIME would be specific enough for now. I'll fix it up in a bit. And if we want to split it further in the future we can.

UI/EmuScreen.cpp

Common/MachineContext.h

Core/MemMap.cpp

unknownbrackets · 2019-02-18T13:49:10Z

Looks like it's still gregs on Linux/x86_32. I guess iOS should use the USE_SIGACTION_ON_APPLE path?

-[Unknown]

hrydgard · 2019-02-25T21:27:42Z

Once I've gotten 1.8.0 out and stable, that's definitely one of the first things I want to look at. Got some ideas for a hyper-low-latency, so-so quality time stretcher that would probably be a better option for emulators than the usual libraries.

But please don't post such requests on completely different issues - post on the original one instead, I see them all.

Regarding this particular one, I think it may have to wait, too much work required to get it good enough to make 1.8.0.

Panderner · 2020-03-20T15:47:44Z

Does Tony Hawk's Underground 2 and MTX Mototrax helps this issue?

hrydgard · 2020-03-20T15:52:30Z

No.

This is so that we can catch game crashes gracefully and automatically report them, with as many details as we can get, in the future. However it is super complicated to get right, I'm gonna need to sit down a week or so someday and finish this...

hrydgard · 2020-07-04T13:06:55Z

Rebasing this, but haven't started digging again yet.

Note to self: I've been thinking more about this, and the case where we simply stop executing and don't offer a way to continue (but do for example offer approximate crash reporting and prevent the whole app from exiting) is actually rather easy. We can just set CoreState = ERROR and set PC directly to a suitable location in the dispatcher before returning from the handler, instead of carefully trying to disassemble and skip the offending instruction. We can thus fall out of execution safely (the state of the registers will be trash but that's just the way it has to be). This part is implemented.

This way we don't need to fully implement x64Analyzer etc support to get this to a useful state - we'll simply not offer a recovery attempt option for cases when a game crashes with an instruction it doesn't understand...

To get this super neat and clean though and to know exactly which MIPS instruction we crashed on, much more advanced register state handling will be necessary.

hrydgard · 2020-07-12T22:49:58Z

Alright, little more testing and it will finally be time to get this merged. Rebased and ready.

I've scaled back the plans a little - for now this will always kill the running game instead of trying to recover by returning 0 or something. It would be cool to keep stepping from a crash - but if you really need that, turn off Fast Memory and it will work.

Related to that, I want to kill or at least do something about bIgnoreBadMemAccess . Pretty sure its current form is a bad idea.

Exactly what kind of reports we will send to report.ppsspp.org remains to be designed. But there sure is plenty of information we could collect.

… the PSP memory space.

…BadMemAccess is false). Add setting for ignore bad memory accesses

hrydgard · 2020-07-14T11:35:13Z

Finally a green build. Here we go.

I'm gonna fix it up on Windows/ARM in a separate PR, just realized that it's probably broken.

unknownbrackets · 2020-07-17T02:14:43Z

Common/ExceptionHandlerSetup.cpp

+	int sicode = info->si_code;
+	if (sicode != SEGV_MAPERR && sicode != SEGV_ACCERR) {
+		// Huh? Return.
+		return;


Should we be calling the old handler in this case?

-[Unknown]

unknownbrackets · 2020-07-17T02:28:06Z

Core/MemMap.cpp

@@ -533,19 +533,16 @@ bool HandleFault(uintptr_t hostAddress, void *ctx) {
 			// It was a read. Fill the destination register with 0.
 			// TODO
 		}
-		// Move on to the next instruction.
+		// Move on to the next instruction. Note that handling bad accesses like this is pretty slow.


We could set a flag to rebuild jit with slow memory, either for this block or all blocks. Might be faster overall.

-[Unknown]

hrydgard added armjit Occurs with JIT on 32-bit ARM devices, but not another CPU backend. arm64jit Occurs with JIT on 64-bit ARM devices, but not another CPU backend. x86jit x86/x64 JIT bugs labels Feb 12, 2019

hrydgard added this to the v1.8.0 milestone Feb 12, 2019

hrydgard force-pushed the exception-handler branch from 6bd2a06 to 276a0de Compare February 12, 2019 20:01

unknownbrackets reviewed Feb 15, 2019

View reviewed changes

hrydgard modified the milestones: v1.8.0, v1.9.0 Feb 25, 2019

hrydgard modified the milestones: v1.9.0, v1.10.0 Aug 6, 2019

santiagomina83 approved these changes Jan 21, 2020

View reviewed changes

ghost mentioned this pull request Mar 20, 2020

[Android] Spider-Man 3 crashes unexpectedly #10196

Open

hrydgard modified the milestones: v1.10.0, v1.11.0 May 10, 2020

hrydgard force-pushed the exception-handler branch from 942f8c6 to 9341d6c Compare July 4, 2020 13:15

hrydgard mentioned this pull request Jul 4, 2020

Bad memory access handling improvements #13092

Merged

hrydgard force-pushed the exception-handler branch from 9341d6c to d3db7cc Compare July 12, 2020 22:44

hrydgard changed the title ~~WIP: Exception handler - catch bad memory access by the JIT~~ Exception handler - catch bad memory accesses by the JIT Jul 13, 2020

hrydgard added 6 commits July 14, 2020 09:25

Simple exception handler so we can ignore accesses that happen within…

1fce6de

… the PSP memory space.

Skip bad reads/writes by the guest executable.

aa802ec

ARM/ARM64 instruction analysis, hook up to handler

c988d42

Draw a minimal crash dump on the emu screen after a crash (if bIgnore…

cbc9095

…BadMemAccess is false). Add setting for ignore bad memory accesses

Mac/Linux buildfixes

c3016fe

Exceptions: A bit more consistency in callback registeration.

96a40bb

hrydgard added 9 commits July 14, 2020 09:25

More info on crash screen

236cb57

Add x64Analyzer to Android.mk. Some minor cleanup

fdcf4f0

MachineContext, Linux/Android: Try to support android-x86 (32-bit)

ac456c2

Try USE_SIGACTION_ON_APPLE on IOS.

15c7358

HandleFault: Add missing nullcheck

430c3ee

iOS buildfix attempt

465367b

Make fastmem memory exceptions report the exceptions to Core correctly.

a56f391

Various platform buildfixes

6f97c3d

Another iOS buildfix attempt

f6cc45a

hrydgard force-pushed the exception-handler branch from 55adff0 to f6cc45a Compare July 14, 2020 07:32

hrydgard added 2 commits July 14, 2020 10:34

Yet another iOS buildfix attempt

893aa29

There's no sigcontext on 64-bit iOS either, apparently.

f62df94

hrydgard merged commit fe91f62 into master Jul 14, 2020

hrydgard deleted the exception-handler branch July 14, 2020 11:35

unknownbrackets reviewed Jul 17, 2020

View reviewed changes

unknownbrackets mentioned this pull request Jan 30, 2021

Update README.md for 1.11 #14011

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Exception handler - catch bad memory accesses by the JIT #11795

Exception handler - catch bad memory accesses by the JIT #11795

hrydgard commented Feb 12, 2019 •

edited

Loading

unknownbrackets left a comment

unknownbrackets Feb 15, 2019

hrydgard Feb 17, 2019

unknownbrackets Feb 17, 2019 •

edited

Loading

unknownbrackets Feb 15, 2019

hrydgard Feb 17, 2019

unknownbrackets Feb 17, 2019

hrydgard Feb 18, 2019 •

edited

Loading

unknownbrackets commented Feb 18, 2019

hrydgard commented Feb 25, 2019 •

edited

Loading

Panderner commented Mar 20, 2020

hrydgard commented Mar 20, 2020

hrydgard commented Jul 4, 2020 •

edited

Loading

hrydgard commented Jul 12, 2020 •

edited

Loading

hrydgard commented Jul 14, 2020

unknownbrackets Jul 17, 2020

unknownbrackets Jul 17, 2020

Exception handler - catch bad memory accesses by the JIT #11795

Exception handler - catch bad memory accesses by the JIT #11795

Conversation

hrydgard commented Feb 12, 2019 • edited Loading

2020 update

unknownbrackets left a comment

Choose a reason for hiding this comment

unknownbrackets Feb 15, 2019

Choose a reason for hiding this comment

hrydgard Feb 17, 2019

Choose a reason for hiding this comment

unknownbrackets Feb 17, 2019 • edited Loading

Choose a reason for hiding this comment

unknownbrackets Feb 15, 2019

Choose a reason for hiding this comment

hrydgard Feb 17, 2019

Choose a reason for hiding this comment

unknownbrackets Feb 17, 2019

Choose a reason for hiding this comment

hrydgard Feb 18, 2019 • edited Loading

Choose a reason for hiding this comment

unknownbrackets commented Feb 18, 2019

hrydgard commented Feb 25, 2019 • edited Loading

Panderner commented Mar 20, 2020

hrydgard commented Mar 20, 2020

hrydgard commented Jul 4, 2020 • edited Loading

hrydgard commented Jul 12, 2020 • edited Loading

hrydgard commented Jul 14, 2020

unknownbrackets Jul 17, 2020

Choose a reason for hiding this comment

unknownbrackets Jul 17, 2020

Choose a reason for hiding this comment

hrydgard commented Feb 12, 2019 •

edited

Loading

unknownbrackets Feb 17, 2019 •

edited

Loading

hrydgard Feb 18, 2019 •

edited

Loading

hrydgard commented Feb 25, 2019 •

edited

Loading

hrydgard commented Jul 4, 2020 •

edited

Loading

hrydgard commented Jul 12, 2020 •

edited

Loading