How often does this happen? Is there any way we could reproduce?
If reproducible, we could turn off the sync.Pool usage and see if anything changes.

@catenacyber says it is still happening multiple times a day on OSS-Fuzz.

Philippe, do you have any hints about an easy way to get a reproduction case on our own machines?
I looked briefly at the instructions for running oss-fuzz itself and they were a bit daunting.

How often does this happen?

Around 50 times a day on oss-fuzz

Is there any way we could reproduce?

I did not manage to reproduce it myself, did not try very hard though...

I looked briefly at the instructions for running oss-fuzz itself and they were a bit daunting.

Well, the hard thing is that this bug does not reproduce for a specific input.
But running it should be ok cf https://google.github.io/oss-fuzz/getting-started/new-project-guide/#testing-locally
That is

• install docker
• cd /path/to/oss-fuzz
• python infra/helper.py build_image gonids
• python infra/helper.py build_fuzzers gonids
• python infra/helper.py run_fuzzer --corpus-dir=<path-to-temp-corpus-dir> gonids fuzz_parserule

Then, I guess you need to wait one hour, and relaunch the fuzzer if it did not trigger the bug, until it does

we could turn off the sync.Pool usage and see if anything changes.

Is there some environment variable to do so ?

Maybe oss-fuzz uses -fork=2 as an extra argument to run_fuzzer

we could turn off the sync.Pool usage and see if anything changes.

Is there some environment variable to do so ?

No, you'd have to edit the code to replace pool allocations with new or make.

Is there some environment variable to do so ?

No, you'd have to edit the code to replace pool allocations with new or make.

so rebuild the standard library ?

Just edit it, rebuild will be automatic.

So, I did google/oss-fuzz#6623 with regex.go not using sync package

google/oss-fuzz#6623 looks worth a shot. Thanks.

It looks like the bug is still happening but much less often.

One last stack trace is

panic: runtime error: slice bounds out of range [:107271103185152] with length 45246

goroutine 17 [running, locked to thread]:
regexp.(*Regexp).Split(0x10c0000b2640, {0x10c00010801f, 0x76dbc0}, 0xffffffffffffffff)
	regexp/regexp.go:1260 +0x61c
github.com/google/gonids.(*Rule).option(0x10c00034a180, {0x100c0007c4350, {0x10c000108016, 0x6}}, 0x10c000334080)
	github.com/google/gonids/parser.go:675 +0x36cf
github.com/google/gonids.parseRuleAux({0x10c000108000, 0x62e0000d8400}, 0x0)
	github.com/google/gonids/parser.go:943 +0x6b3
github.com/google/gonids.ParseRule(...)
	github.com/google/gonids/parser.go:972
github.com/google/gonids.FuzzParseRule({0x62e0000d8400, 0x0, 0x10c000000601})
	github.com/google/gonids/fuzz.go:20 +0x54
main.LLVMFuzzerTestOneInput(...)
	./main.3230035416.go:21
AddressSanitizer:DEADLYSIGNAL

regexp.go:1260 seems to prove that this is the modified regex.go file without sync right ?

Any more clues ?
Any more debug assertions to insert in regexp.go ?

This started on August 16th, so is likely a Go 1.17 issue.

Could you bisect to a particular commit that introduced the corruption?

Could you bisect to a particular commit that introduced the corruption?

oss-fuzz uses latest Golang release, so they switched from 1.16.x to 1.17 on August 16th, but we do not know which commit exactly in this major release induced the buggy behavior...

I have another possible theory:

regexp.(*Regexp).Split(0x10c0000b2640, {0x10c0001c801f, 0x76dbc0}, 0xffffffffffffffff)
	regexp/regexp.go:1266 +0x61c

Is there any way you can get at the regexp and the contents of the string? In this case, 0x10c0000b2640 is the address of a Regexp, the first field of which is a string I'd like to see. Also, we'd need the very long string at address 0x10c0001c801f for 0x76dbc0 bytes (7MB+!).
If you have any way for us to reproduce that execution state, we could grab those values ourselves.

My theory is that regexp is using the code in internal/bytealg pretty hard, and maybe it's tripping up on some weird corner case. The string pointers always end in 0x1f, which is maybe one of those corner cases. That might explain the intermittency - if the same test case gets allocated in a different place, it might not trigger the bug.
Not sure why we would hit it only for 1.17. https://go-review.googlesource.com/c/go/+/310184 seems to be the only CL of note in internal/bytealg for 1.17, and it looks reasonable to me.

Put the hex encoded version of one string here :
https://catenacyber.fr/stringhex.txt

If it's an alignment issue, could we try to reproduce by mmap'ing a large block and creating a string with all the different possible alignments / ending pointer bytes?

What is the regexp that is being used?

regexp.MustCompile(,\s*).Split(fuzzedinput, -1)

So, there are kinds of a lot of answers to split....

Nothing obvious with the string+regexp you posted. Putting it at different alignments doesn't seem to trigger anything.
The string you posted doesn't seem long enough, though. It is only 227927 bytes, and all the failure traces above have a string that is 0x76dbc0 = 7789504 bytes.

This string posted comes with stack trace

panic: runtime error: slice bounds out of range [699494522:64250]
--
  |  
  | goroutine 17 [running, locked to thread]:
  | regexp.(*Regexp).Split(0x10c0000b2640, {0x10c000ac201f, 0x962c00}, 0xffffffffffffffff)
  | regexp/regexp.go:1266 +0x617
  | github.com/google/gonids.(*Rule).option(0x10c000066000, {0x0, {0x10c000ac2016, 0x10c000044000}}, 0x10c0001dc000)
  | github.com/google/gonids/parser.go:675 +0x36c5
  | github.com/google/gonids.parseRuleAux({0x10c000ac2000, 0x37a78}, 0x0)
  | github.com/google/gonids/parser.go:942 +0x6b3
  | github.com/google/gonids.ParseRule(...)
  | github.com/google/gonids/parser.go:971
  | github.com/google/gonids.FuzzParseRule({0x7f369cb75800, 0x0, 0x10c000000601})
  | github.com/google/gonids/fuzz.go:20 +0x54
  | main.LLVMFuzzerTestOneInput(...)
  | ./main.3953748960.go:21

When running manually, it seems my string is always aligned on 16 bytes.like 0x10c0001185a0

That doesn't seem right. The string posted is only 227927 bytes, but the stack trace shows that is is 9841664 bytes.

To unalign a string, do:

s2 := (strings.Repeat("*", i) + s)[i:]

So, the string must have been corrupted before the call to regexp.(*Regexp).Split

I'm not sure I understand. How did you get the string, if not just dumping the 0x962c00 bytes starting at address 0x10c000ac201f ?

I'm not sure I understand. How did you get the string, if not just dumping the 0x962c00 bytes starting at address 0x10c000ac201f ?

oss-fuzz provides the stack trace, and the input that triggered it.
The input that triggers it, when run over gonids.FuzzParseRule ends up once at gonids.(*Rule).option to call regexp.(*Regexp).Split and that is where I got the string

Hm, then I guess the values in the stack traces are incorrect. Which can happen, especially with regabi introduced in 1.17.
I don't have any ideas on how to figure out the problem without bisecting through the 1.17 commits.
I am unable to reproduce on my laptop. It never crashes for me...
Speaking of which, it's always possible that it is a bad machine. Have you reproduced on different machines?

Another thing to try, run with GODEBUG=cpu.all=off. That will disable most of the complicated bytealg code.

I am unable to reproduce on my laptop. It never crashes for me...

FWIW, I was not able to reproduce either using the oss-fuzz local execution steps on a clean Linux VM (amd64, Ubuntu 20.04) following the directions outlined above in #49075 (comment), with 3 separate runs that totaled about 24 hours.

I also ran for about 24 hours using dvyukov/go-fuzz against the same gonids fuzz target, which also did not crash (although I don't think there is an enormous amount of signal from that result, given libFuzzer and go-fuzz have different process restart strategies, different mutations, different propensity for creating large inputs, and so on).

@catenacyber some questions for you:

1. I am curious if you are able to reproduce in a clean environment using those steps you outlined for local execution?

2. I wonder if some additional steps might be needed, such as perhaps downloading a snapshot of the oss-fuzz working corpus (vs. if I followed, the oss-fuzz local execution steps you outlined above results in only using a seed corpus downloaded from https://rules.emergingthreats.net?).

3. I wonder if -max_len needs to be set, or alternatively, perhaps if the live oss-fuzz working corpus has larger inputs that push libFuzzer to use larger examples? Following the steps above, my local execution currently reports:

INFO: -max_len is not provided; libFuzzer will not generate inputs larger 
than 4096 bytes
INFO: seed corpus: files: 67174 min: 1b max: 3228b total: 19520296b rss: 57Mb

That said, it might be that size of inputs in the corpus is not meaningful if this turns out to be due to some corruption.

4. If I followed, it looks like the live gonids Dockerfile on the oss-fuzz repo is using the attempted workaround of removing use of sync.Pool, which I think you said resulted in "It looks like the bug is still happening but much less often." If Keith or others are trying to reproduce, it might make sense to not use that attempted workaround when trying to reproduce locally in order to increase frequency?

In any event, sorry if anything I wrote is off base, but curious for your thoughts.

Do I understand correctly that the libfuzzer signal handler overflows the golang stack and overflows unrelated nearby memory ?

Yes.

Would having a guard page after a golang stack help debug this ? (by making reproducing easier)

Yes. The hacking I described has a similar effect.

Or setting up another signal handler to backtrace the second SEGV ?

I'm not sure what the second SEGV is here. There is a SIGALRM, and handling that silently overwrites the heap. When the kinda-guard-page is hacked in, handling the SIGALRM instead triggers a SIGSEGV.

cf https://github.com/catenacyber/nallocfuzz/blob/main/nallocfuzz.c#L112 for an example to do this without recompiling all llvm

I'm not sure what that is showing me. I can change the fuzz sources no problem, I just don't know how to include those changes when building the reproducer. aka, how do I replace the built-in libFuzzer.a with my own libFuzzer.a? Just putting a -L early in the command line didn't seem to work.

I have LLVM built now, but for some reason it isn't built with the sanitizers (fuzzer, asan, ubsan, ...), so the repro complains that those libraries are missing. Trying to figure that out now. Currently trying this: https://stackoverflow.com/questions/56847345/how-to-compile-clang-with-sanitizers

Would having a guard page after a golang stack help debug this ? (by making reproducing easier)

Yes. The hacking I described has a similar effect.

Having a guard page after stack allows earlier catching, right ?

Or setting up another signal handler to backtrace the second SEGV ?

I'm not sure what the second SEGV is here. There is a SIGALRM, and handling that silently overwrites the heap. When the kinda-guard-page is hacked in, handling the SIGALRM instead triggers a SIGSEGV.

The heap or the stack ?

cf https://github.com/catenacyber/nallocfuzz/blob/main/nallocfuzz.c#L112 for an example to do this without recompiling all llvm

I'm not sure what that is showing me. I can change the fuzz sources no problem, I just don't know how to include those changes when building the reproducer. aka, how do I replace the built-in libFuzzer.a with my own libFuzzer.a? Just putting a -L early in the command line didn't seem to work.

Change -fsanitize=fuzzer with your own libFuzzer.a ;-)

I have LLVM built now, but for some reason it isn't built with the sanitizers (fuzzer, asan, ubsan, ...), so the repro complains that those libraries are missing. Trying to figure that out now. Currently trying this: https://stackoverflow.com/questions/56847345/how-to-compile-clang-with-sanitizers

I use oss-fuzz Docker images to do this cf
https://github.com/google/oss-fuzz/blob/master/infra/base-images/base-clang/checkout_build_install_llvm.sh

I would use a second signal handler to gather more debug information : this signal was triggered with this backtrace.

So far so good, with this patch to libFuzzer:

--- a/compiler-rt/lib/fuzzer/FuzzerUtilPosix.cpp
+++ b/compiler-rt/lib/fuzzer/FuzzerUtilPosix.cpp
@@ -82,6 +82,7 @@ static void SetSigaction(int signum,
   // dedicated stack) in order to be able to detect stack overflows; keep the
   // flag if it's set.
   new_sigact.sa_flags = SA_SIGINFO | (sigact.sa_flags & SA_ONSTACK);
+  new_sigact.sa_flags |= SA_ONSTACK;
   new_sigact.sa_sigaction = callback;
   if (sigaction(signum, &new_sigact, nullptr)) {
     Printf("libFuzzer: sigaction failed with %d\n", errno);

I'm not sure that's a general fix, but for Go it's enough because Go does its own sigaltstack. Maybe everyone who cares does their own sigaltstack? In that case, this may be all that's required.

Having a guard page after stack allows earlier catching, right ?

Yes, almost certainly. Having a mode where the Go runtime keeps guard pages at the end of goroutine stacks would be awesome.

The heap or the stack ?

The heap. It writes off the end of the stack into an earlier page, which is probably a general heap page. Could also be a different goroutine's stack, I suppose. All equally bad.

Yes, almost certainly. Having a mode where the Go runtime keeps guard pages at the end of goroutine stacks would be awesome.

Does it not exist already ?
Looks like it would be useful...

So far so good, with this patch to libFuzzer:

It should then be :

--- a/compiler-rt/lib/fuzzer/FuzzerUtilPosix.cpp
+++ b/compiler-rt/lib/fuzzer/FuzzerUtilPosix.cpp
@@ -82,6 +82,7 @@ static void SetSigaction(int signum,
   // dedicated stack) in order to be able to detect stack overflows; keep the
   // flag if it's set.
-   new_sigact.sa_flags = SA_SIGINFO | (sigact.sa_flags & SA_ONSTACK);
+  new_sigact.sa_flags = SA_SIGINFO | SA_ONSTACK;
   new_sigact.sa_sigaction = callback;
   if (sigaction(signum, &new_sigact, nullptr)) {
     Printf("libFuzzer: sigaction failed with %d\n", errno);

But that looks dubious, like it will break another case.
Do you know why sigact.sa_flags & SA_ONSTACK is false ?

Does it not exist already ?

It does not.

Do you know why sigact.sa_flags & SA_ONSTACK is false ?

It's the default. I think if we redesigned unix from scratch we'd always force signal handlers to have their own stack.

A bit of background: when running with the Go runtime, all signals must be installed with the SA_ONSTACK flag set. This is documented at https://pkg.go.dev/os/signal#hdr-Go_programs_that_use_cgo_or_SWIG . For a Go program (not -buildmode=c-archive) the Go runtime will normally install a signal handler for SIGALRM with SA_ONSTACK set. In which case there shouldn't be a problem here.

But if this program is using -buildmode=c-archive to build the Go program, then the Go runtime will not set a signal handler for SIGALRM. If there is no signal handler when the Go runtime is initialized, it will be left alone. If there is a signal handler, then the Go runtime will set the SA_ONSTACK flag for it.

So while I'm sure I'm missing something, right now the only way I can see that a problem can arise is if we are using -buildmode=c-archive, and the Go runtime is initialized, and then after that happens the fuzzer code installs a SIGALRM signal handler.

As a side note, setting SA_ONSTACK does no harm, as it only affects code if something calls sigaltstack for the thread. So the libfuzzer code may as well always set it.

Change https://go.dev/cl/494117 mentions this issue: runtime: fix misaligned SP for libfuzzer entry

I'm somewhat sure my patch fixes the issue, although hard to be 100% sure with the low failure rate. If anyone else wants to run some overnight jobs, that would be great. Especially on 1.19/1.20, as I've been working at tip.

Patch for libFuzzer has been mailed: https://reviews.llvm.org/D150231

So while I'm sure I'm missing something, right now the only way I can see that a problem can arise is if we are using -buildmode=c-archive, and the Go runtime is initialized, and then after that happens the fuzzer code installs a SIGALRM signal handler.

This is indeed the situation, Go initializes first.

Thanks for these inputs.

As I understand it, the problem lies rather in https://github.com/mdempsky/go114-fuzz-build ,the framework to link libFuzzer and golang: it should set a signaler handler that sets SA_ONSTACK
I am testing this : https://github.com/catenacyber/gonids/tree/repro/mainc

By the way, will the feature "guard page after a stack" be implemented ? It looks like ASAN for golang to me...
It would have been useful here, and may be in the future...

As I understand it, the problem lies rather in https://github.com/mdempsky/go114-fuzz-build ,the framework to link libFuzzer and golang: it should set a signaler handler that sets SA_ONSTACK
I am testing this : https://github.com/catenacyber/gonids/tree/repro/mainc

I don't think that is the problem. libfuzzer would still be setting a SA_ONSTACK-less handler even if other handlers are registered correctly.

By the way, will the feature "guard page after a stack" be implemented ? It looks like ASAN for golang to me...
It would have been useful here, and may be in the future...

There is no plan for that at the moment. I'm happy to review the code if someone wants to take it on...

Normally the compiler+linker enforces that Go can't overflow the stack. Signals not on an alternate stack are the only way I know of that the Go stack can overflow.
Go has asan already. Not for Go stack overflows though, it is intended for the C parts of cgo.

I don't think that is the problem. libfuzzer would still be setting a SA_ONSTACK-less handler even if other handlers are registered correctly.

I think the chaining of signal handlers solves the issue.
I will run some more tests, but setting a golang signal handler solves (or hides) the issue so far...

I think the chaining of signal handlers solves the issue.

I can see that preventing the crash, if your signal.Notify call happens after the libFuzzer's sigaction call.

But won't that prevent delivery of the signals that libFuzzer wants? I presume signal.Notify eats the signal and libFuzzer just won't see that signal raised at all any more. So I guess the question is, does the timeout flag still work with your additional code? I think that's what libFuzzer is using signals for.

I'm going to close this issue (and the backports). This probably needs a fix somewhere but it isn't in the Go repository.
We can reopen if we figure out some workaround that would require stdlib code.

your signal.Notify call happens after the libFuzzer's sigaction call.

I think it does

But won't that prevent delivery of the signals that libFuzzer wants?

Oh, is there a way for golang to pass signal to the next handler ?
Otherwise, I need to set up, instead of a golang signal handler, a new C signal handler with SA_ONSTACK and passing to the old signal handler (ie libFuzzer)

Cf https://github.com/catenacyber/gonids/tree/repro/pocsig
Build with

CC=clang go build -o fuzz.a -buildmode c-archive cfuzz.go
clang -fsanitize=address -c sigfuzz.c
clang++ -fsanitize=address sigfuzz.o fuzz.a -o fuzz_sig

Running ./fuzz_sig -> looks like C signal handlers reuse golang stack (0x10c000020090 versus 0x10c00006a680)

Running C_SA_ONSTACK=1 ./fuzz_sig -> looks like C signal handlers have a different stack than golang (0x10c00001a0a0 versus 0x7fb151cb2a00)

Running GO_SIGHANDLER=1 shows that golang does not relay the signal to the previous handler

TL;DR @randall77 should go114-fuzz-build be fixed by adding a C signal handler with SA_ONSTACK and relaying to the next signal handler ?

You don't need to install a new signal handler, you just need to reinstall the old signal handler with the SA_ONSTACK flag set. Call sigaction to get the old one, set SA_ONSTACK, and pass it back to sigaction.

Indeed Ian, thanks.

So, you confirm that I cannot do that in pure Golang, and need to resort to C ?

And I cannot build a static library containing both C and go with only one go command, right ?
That is, in

CC=clang go build -o sigfuzz.a -buildmode c-archive sigfuzz.go
clang -fsanitize=address -c sigfuzz.c
clang++ -fsanitize=address sigfuzz.o sigfuzz.a -o fuzz_sig

I cannot compile sigfuzz.c and sigfuzz.go into sigfuzz.a in one command ?

So, you confirm that I cannot do that in pure Golang, and need to resort to C ?

You can do it in pure Go using syscall.Syscall and defining struct sigaction yourself.

Or, just to be clear, you can do it using cgo without any supporting C files, although I don't know if that meets your requirements.

And I cannot build a static library containing both C and go with only one go command, right ?

If you put your Go and C files into a single directory, then you can run go build -buildmode=c-archive in that directory and get a single Go archive that includes both the Go and the C files.

defining struct sigaction yourself.

Not sure I see how to do that...

Thanks anyways :-)

To define struct sigaction yourself in Go, you could copy the various definitions of sigactiont in the runtime package.