Stack probe

[antoyo]

Hi.
To have better safety, could we use guard pages and stack probes in may?
I'm not sure it's possible, just asking.
Thanks.

[Xudong-Huang]

thanks for the link!
currently May doesn't use guard page for coroutines stacks, this need alloc extra one more page for each coroutine. Now we only has a very unsafe way to probe the stack overflow implemented by checking the footprint of the coroutine stack. So this is why exceed stack could cause undefined behavior.

I check the __rust_probestack src code and find that this function needs to be called in the user's code, not in the May library, because it only used for one big single function frame to detect stack overflow, and for libary design we don't have any information about how the stack frames would grow. If users forget call this for a large stack frame (bigger than one page) function it still triggers undefined behavior even we has a stack gurad page here.

[vitiral]

I check the __rust_probestack src code and find that this function needs to be called in the user's code, not in the May library, because it only used for one big single function frame to detect stack overflow, and for libary design we don't have any information about how the stack frames would grow.

I may not be understanding, but if it has to be called "in the user's code" then couldn't the go! macro call it?

[Xudong-Huang]

@vitiral: go! macro is just a thin wrapper, which doesn't need stack probe at all. only those leaf functions that have very big stack frame would be benefit from __rust_probestack when they are running with a guard page on the stack. But may doesn't have guard pages for coroutines now, so call __rust_probestack is no help.

[vitiral]

I feel the big issue with may is that it does not guard against undefined behavior. I think most use cases don't care that they have to limit their stack size (most co-routines will spawn a single function call!) and wouldn't mind if the coroutine paniced in that case. Undefined behavior, however, is a huge problem.

Is it possible to prevent undefined behavior and instead panic... or even abort?

[Xudong-Huang]

yes, it's possible to add a stack guard page for each coroutine. And this would need an extra one page memory allocated. for most cases this is not a big issue which I preferred to add the behavior in future. if stack overflow detected, OS will kill the program with a segment fault(maybe with better output by rust, this is not a UB, I think) just like normal thread stack overflow implemented by rust.

I don't think we can have a panic if stack overflow happened, when panic happened, it will need more stack which would make things worse. If we are using guard page, there is no place to put the panic code, rust run-time takes over the action for stack overflow.

[vitiral]

cool, thanks for the response!

I think it makes sense for go! to use the stack probe, but allow users to access a coroutine without probed memory (through an unsafe function) if they really want to squeeze performance.

Thanks!

[killme2008]

I think this is a serious issue that prevent people to use may in production.People are not afraid of panic but undefined. 😄

Do you have any plan to fix this issue?

[Xudong-Huang]

Sorry about this delay. I have a solution. But not implement it. You are right, we should have it as soon as posdible🙀 发自我的 iPhone 在 2019年4月23日，10:36，dennis zhuang <[email protected]<mailto:[email protected]>> 写道： I think this is a serious issue that prevent people to use may in production.People are not afraid of panic but undefined. 😄 Do you have any plan to fix this issue? — You are receiving this because you commented. Reply to this email directly, view it on GitHub<#29 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/ADLGBK6FH5O2QTGQ777CJ7LPRZY3NANCNFSM4EMWIPNA>.

[Xudong-Huang]

now the generator lib use guarded stack, may will automatically has the desired behavior.

ref Xudong-Huang/generator-rs#12