Debugging an OpenJDK crash on SPARC

I had to spend a little time recently fixing a crash in OpenJDK on Solaris SPARC.

What we're seeing is, from the hs_err file:

# A fatal error has been detected by the Java Runtime Environment:
#
#  SIGSEGV (0xb) at pc=0xffffffff57c745a8, pid=18442, tid=37
...
# Problematic frame:
# V  [libjvm.so+0x7745a8]  G1CollectedHeap::allocate_new_tlab(unsigned long, unsigned long, unsigned long*)+0xb8

Well that's odd. I only see this on SPARC, and I've seen it sporadically on Tribblix during the process of continually building OpenJDK on SPARC, but haven't seen it on Solaris. Until a customer hit it in production, which is rather a painful place to find a reproducer.

In terms of source, this is located in the file src/hotspot/share/gc/g1/g1CollectedHeap.cpp (all future source references will be relative to that directory), and looks like:

HeapWord* G1CollectedHeap::allocate_new_tlab(size_t min_size,
                                             size_t requested_size,
                                             size_t* actual_size) {
  assert_heap_not_locked_and_not_at_safepoint();
  assert(!is_humongous(requested_size), "we do not allow humongous TLABs");

  return attempt_allocation(min_size, requested_size, actual_size);
}

That's incredibly simple. There's not much that can go wrong there, is there?

The complexity here is that a whole load of functions get inlined. So what does it call? You find yourself in a twisty maze of passages, all alike. But anyway, the next one down is

inline HeapWord* G1CollectedHeap::attempt_allocation(size_t min_word_size,
                                                     size_t desired_word_size,
                                                     size_t* actual_word_size) {
  assert_heap_not_locked_and_not_at_safepoint();
  assert(!is_humongous(desired_word_size), "attempt_allocation() should not "
         "be called for humongous allocation requests");

  HeapWord* result = _allocator->attempt_allocation(min_word_size, desired_word_size, actual_word_size);

  if (result == NULL) {
    *actual_word_size = desired_word_size;
    result = attempt_allocation_slow(desired_word_size);
  }

  assert_heap_not_locked();
  if (result != NULL) {
    assert(*actual_word_size != 0, "Actual size must have been set here");
    dirty_young_block(result, *actual_word_size);
  } else {
    *actual_word_size = 0;
  }

  return result;
}

That then calls an inlined G1Allocator::attempt_allocation() in g1Allocator.hpp. That calls current_node_index(), which looks safe and then there are a couple of calls to mutator_alloc_region()->attempt_retained_allocation() and mutator_alloc_region()->attempt_allocation(), which come from g1AllocRegion.inline.hpp and both ultimately call a local par_allocate(), which then calls par_allocate_impl() or par_allocate() in heapRegion.inline.hpp.

Now, mostly all these are doing is calling something else. The one really complex piece of code is in par_allocate_impl() which contains

...
  do {
    HeapWord* obj = top();
    size_t available = pointer_delta(end(), obj);
    size_t want_to_allocate = MIN2(available, desired_word_size);
    if (want_to_allocate >= min_word_size) {
      HeapWord* new_top = obj + want_to_allocate;
      HeapWord* result = Atomic::cmpxchg(&_top, obj, new_top);
      // result can be one of two:
      //  the old top value: the exchange succeeded
      //  otherwise: the new value of the top is returned.
      if (result == obj) {
        assert(is_object_aligned(obj) && is_object_aligned(new_top), "checking alignment");
        *actual_size = want_to_allocate;
        return obj;
      }
    } else {
      return NULL;
    }
  } while (true);
}

Right, let's go back to the crash. We can open up the core file in
mdb, and look at the stack with $C

ffffffff7f39d751 libjvm.so`_ZN7VMError14report_and_dieEP6ThreadjPhPvS3_+0x3c(
    101cbb1d0?, b?, fffffffcb45dea7c?, ffffffff7f39ecb0?, ffffffff7f39e9a0?, 0?)
ffffffff7f39d811 libjvm.so`JVM_handle_solaris_signal+0x1d4(b?,
    ffffffff7f39ecb0?, ffffffff7f39e9a0?, 0?, ffffffff7f39e178?, 101cbb1d0?)
ffffffff7f39dde1 libjvm.so`_ZL17javaSignalHandleriP7siginfoPv+0x20(b?,
    ffffffff7f39ecb0?, ffffffff7f39e9a0?, 0?, 0?, ffffffff7e7dd370?)
ffffffff7f39de91 libc.so.1`__sighndlr+0xc(b?, ffffffff7f39ecb0?,
    ffffffff7f39e9a0?, fffffffcb4b38afc?, 0?, ffffffff7f20c7e8?)
ffffffff7f39df41 libc.so.1`call_user_handler+0x400((int) -1?,
    (siginfo_t *) 0xffffffff7f39ecb0?, (ucontext_t *) 0xc?)
ffffffff7f39e031 libc.so.1`sigacthandler+0xa0((int) 11?,
    (siginfo_t *) 0xffffffff7f39ecb0?, (void *) 0xffffffff7f39e9a0?)
ffffffff7f39e5b1 libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm+0xb8(
    10013d030?, 100?, 520?, ffffffff7f39f000?, 0?, 0?)

What you see here is the allocate_new_tlab() at the botton, it throws a signal, the signal handler catches it, passes it ultimately to JVM_handle_solaris_signal() which bails, and the JVM exits.

We can look at the signal. It's at address 0xffffffff7f39ecb0 and is of type siginfo_t, so we can just print it

java:core> ffffffff7f39ecb0::print -t siginfo_t

and we first see

siginfo_t {
    int si_signo = 0t11 (0xb)
    int si_code = 1
    int si_errno = 0
...

OK, the signal was indeed 11 = SIGSEGV. The interesting thing is the si_code of 1, which is defined as

#define SEGV_MAPERR     1       /* address not mapped to object */

Ah. Now, in the jvm you actually see a lot of SIGSEGV, but a lot of them are handled by that mysterious JVM_handle_solaris_signal(). In particular, it'll handle anything with SEGV_ACCERR which is basically something running off the end of an array.

Further down, you can see the fault address

struct  __fault = {
            void *__addr = 0x10
            int __trapno = 0
            caddr_t __pc = 0
            int __adivers = 0
        }

So, we're faulting on address 0x10. Yes, you try messing around down there and you will fault.

That confirms the crash is a SEGV. What are we actually trying to do? We can disassemble the allocate_new_tlab() function and see what's happening - remember the crash was at offset 0xb8

java:core> libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm::dis
...
 libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm+0xb8:

       ldx       [%i4 + 0x10], %i5

That's interesting, 0x10 was the fault address. What's %i4 then?

java:core> ::regs
%i4 = 0x0000000000000000

Yep. Given that, we'll try and read 0x10, giving the SEGV we see.

There's a little more context around that call site. A slightly
expanded view is

 libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm+0xa0:        nop
 libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm+0xa4:        add       %
i5, %g1, %g1
 libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm+0xa8:        casx      [
%g3], %i5, %g1
 libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm+0xac:        cmp       %
i5, %g1
 libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm+0xb0:        be,pn     %
xcc, +0x160  <libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm+0x210>
 libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm+0xb4:        nop
 libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm+0xb8:        ldx       [
%i4 + 0x10], %i5

Now, the interesting thing here is the casx (compare and swap) instruction. That lines up with the Atomic::cmpxchg() in par_allocate_impl() that we were suspecting above. So the crash is somewhere around there.

It turns out there's another way to approach this. If we compile without optimization then effectively we turn off the inlining. The way to do this is to add an entry to the jvm Makefile via make/hotspot/lib/JvmOverrideFiles.gmk

...
else ifeq ($(call isTargetOs, solaris), true)
    ifeq ($(call isTargetCpuArch, sparc), true)
      # ptribble port tweaks
      BUILD_LIBJVM_g1CollectedHeap.cpp_CXXFLAGS += -O0
    endif
endif

If we rebuild (having touched all the files in the directory to force
make to rebuild everything correctly), and run again, we get the full
call stack:

Now the crash is

# V  [libjvm.so+0x80cc48]  HeapRegion::top() const+0xc

which we can expand to the following stack leading up to where it goes
into the signal handler.:

ffffffff7f39dff1 libjvm.so`_ZNK10HeapRegion3topEv+0xc(0?, ffffffff7f39ef40?,
    101583e38?, ffffffff7f39f020?, fffffffa46de8038?, 10000?)
ffffffff7f39e0a1 libjvm.so`_ZN10HeapRegion17par_allocate_implEmmPm+0x18(0?,
    100?, 10000?, ffffffff7f39ef60?, ffffffff7f39ef40?, 8f00?)
ffffffff7f39e181                     
libjvm.so`_ZN10HeapRegion27par_allocate_no_bot_updatesEmmPm+0x24(0?, 100?,
    10000?, ffffffff7f39ef60?, 566c?, 200031?)
ffffffff7f39e231                     
libjvm.so`_ZN13G1AllocRegion12par_allocateEP10HeapRegionmmPm+0x44(100145440?,
    0?, 100?, 10000?, ffffffff7f39ef60?, 0?)
ffffffff7f39e2e1 libjvm.so`_ZN13G1AllocRegion18attempt_allocationEmmPm+0x48(
    100145440?, 100?, 10000?, ffffffff7f39ef60?, 3?, fffffffa46ceff48?)
ffffffff7f39e3a1 libjvm.so`_ZN11G1Allocator18attempt_allocationEmmPm+0xa4(
    1001453b0?, 100?, 10000?, ffffffff7f39ef60?, 7c0007410?, ffffffff7f39ea41?)
ffffffff7f39e461 libjvm.so`_ZN15G1CollectedHeap18attempt_allocationEmmPm+0x2c(
    10013d030?, 100?, 10000?, ffffffff7f39ef60?, 7c01b15e8?, 0?)
ffffffff7f39e521 libjvm.so`_ZN15G1CollectedHeap17allocate_new_tlabEmmPm+0x24(
    10013d030?, 100?, 10000?, ffffffff7f39ef60?, 0?, 0?)

So yes, this confirms that we are indeed in par_allocate_impl() and
it's crashing on the very first line of the code segment I showed
above, where it calls top(). All top() does is return the _top member
of a HeapRegion.

So the only thing that can happen here is that the HeapRegion itself
is NULL. Then the _top member is presumably at offset 0x10, and trying
to access it gives the SIGSEGV.

Now, in G1AllocRegion::attempt_allocation() there's an assert:

  HeapRegion* alloc_region = _alloc_region;
  assert_alloc_region(alloc_region != NULL, "not initialized properly");

However, asserts aren't compiled into production builds.

But the fix here is to fail if we've got NULL and let the caller
retry. There are a lot of calls here, and the general approach is to
return NULL if anything goes wrong, so I do the same for this extra
failure case, adding the following:

  if (alloc_region == NULL) {
    return NULL;
  }

With that, no more of those pesky crashes. (There might be others
lurking elsewhere, of course.)

Of course, what this doesn't explain is why the HeapRegion wasn't
correctly initialized in the first place. But that's another problem
entirely.

What's a decent password length?

What's a decent length for a password?

I think it's pretty much agreed by now that longer passwords are, in general, better. And fortunately stupid complexity requirements are on the way out.

Reading the NIST password rules gives the following:

• User chosen passwords must be at least 8 characters
• Machine chosen passwords must be at least 6 characters
• You must allow passwords to be at least 64 characters

Say what? A 6 character password is secure?

Initially, that seems way off, but it depends on your threat model. If you have a mechanism to block the really bad commonly used passwords, then 6 characters gives you a billion choices. Not many, but you should also be implementing technical measures such as rate limiting.

With that, if the only attack vector is brute force over the network, trying a billion passwords is simply impractical. Even with just passive rate limiting (limited by cpu power and network latency) an attacker will struggle; with active limiting they'll be trying for decades.

That's with just 6 random characters. Go to 8 and you're out of sight. And for this attack vector, no quantum computing developments will make any difference whatsoever.

But what if the user database itself is compromised?

Of course, if the passwords are in cleartext then no amount of fancy rules or length requirements is going to help you at all.

But if an attacker gets encrypted passwords then they can simply brute force them many orders of magnitude faster. Or use rainbow tables. And that's a whole different threat model.

Realistically, protecting against brute force or rainbow table attacks probably needs a 16 character password (or passphrase), and that requirement could get longer over time.

A corollary to this is that there isn't actually much to be gained to requiring password lengths between 8 and 16 characters.

In illumos, the default minimum password length is 6 characters. I recently increased the default in Tribblix to 8, which aligns with the user chosen limit that NIST give.

Tribblix image structural changes

The Tribblix live ISO and related images are put together every so slightly differently in the latest m34 release.

All along, there's been an overlay (think a group package) called base-iso that lists the packages that are present in the live image. On installation, this is augmented with a few extra packages that you would expect to be present in a running system but which don't make much sense in a live image, to construct the base system.

You can add additional software, but the base is assumed to be present.

The snag with this is that base-iso is very much a single-purpose generic concept. By its very nature it has to be minimal enough to not be overly bloated, yet contain as many drivers as necessary to handle the majority of systems.

As such, the regular ISO image has fallen between 2 stools - it doesn't have every single driver, so some systems won't work, while it has a lot of unnecessary drivers for a lot of common use cases.

So what I've done is split base-iso into 2 layers. There's a new core-tribblix overlay, which is the common packages, and then base-iso adds all the extra drivers. By and large, the regular live image for m34 isn't really any different to what was present before.

But the concepts of "what packages do I need for applications to work" and "what packages do I want to load on a given downloadable ISO" have now been split.

What this allows is to easily create other images with different rules. As of m34, for example, the "minimal" image is actually created from a new base-server overlay, which again sits atop core-tribblix and differs from base-iso in that it has all the FC drivers. If you're installing on a fibre-channel connected system then using the minimal image will work better (and if you're SAN-booted, it will work where the regular ISO won't).

The next use case is that images for cloud or virtual systems simply don't need most of the drivers. This cuts out a lot of packages (although it doesn't actually save that much space).

The standard Tribblix base system now depends on core-tribblix, not base-iso or any of the specific image layers. This is as it should be - userland and applications really shouldn't care what drivers are present.

One side-effect of this change is that it makes minimising zones easier, because what gets installed in a zone can be based on that stripped-down core-tribblix overlay.

The SunOS JDK builder

I've been building OpenJDK on Solaris and illumos for a while.

This has been moderately successful; illumos distributions now have access to up to date LTS releases, most of which work well. (At least 11 and 17 are fine; 21 isn't quite right.)

There are even some third-party collections of my patches, primarily for Solaris (as opposed to illumos) builds.

I've added another tool. The SunOS jdk builder.

The aim here is to be able to build every single jdk tag, rather than going to one of the existing repos which only have the current builds. And, yes, you could grope through the git history to get to older builds, but one problem with that is that you can't actually fix problems with past builds.

Most of the content is in the jdk-sunos-patches repository. Here there are patches for both illumos and Solaris (they're ever so slightly different) for every tag I've built.

(That's almost every jdk tag since the Solaris/SPARC/Studio removal, and a few before that. Every so often I find I missed one. And there's been the odd bad patch along the way.)

The idea here is to make it easy to build every tag, and to do so on a current system. I've had to add new patches to get some of the older builds to work. The world has changed, we have newer compilers and other tools, and the OS we're building on has evolved. So if someone wanted to start building the jdk from scratch (and remember that you have to build all the versions in sequence) then this would be useful.

I'm using it for a couple of other things.

One is to put back SPARC support on illumos and Solaris. The initial port I did was on x86 only, so I'm walking through older builds and getting them to work on SPARC. We'll almost certainly not get to jdk21, but 17 seems a reasonable target.

The other thing is to enable the test suites, and then run them, and hopefully get them clean. At the moment they aren't, but a lot of that is because many tests are OS-specific and they don't know what Solaris is so get confused. With all the tags, I can bisect on failures and (hopefully) fix them.

Building up networks of zones on Tribblix

With OpenSolaris and derivatives such as illumos, we gained the ability to build a whole IT infrastructure in a single box, using virtualized networking (crossbow) to build the underlying network and then attaching virtualized systems (zones) atop virtualized storage (zfs).

Some of this was present in Solaris 10, but it didn't have crossbow so the networking piece was a bit tricky (although I did manage to get surprisingly far by abusing the loopback interface).

In Tribblix, I've long had the notion of a router or proxy zone, which acts as a bridge between the outside world and a local virtual subnet. For the next release I've been expanding that into something much more flexible and capable.

What did I need to put this together?

The first thing is a virtual network. You use dladm to create an etherstub. Think of that as a virtual switch you can connect network links to.

To connect that to the world, a zone is created with 2 network interfaces (vnics). One over the system interface so it can connect to the outside world, and one over the etherstub.

That special router zone is a little bit more than that. It runs NAT to allow any traffic on the internal subnet - simple NAT, nothing complicated here. In order to do that the zone has to have IPFilter installed, and the zone creation script creates the right ipnat configuration file and ensures that IPFilter is started.

You also need to have IPFilter installed in the global zone. It doesn't have to be running there, but the installation is required to create the IPFilter devices. Those IPFilter devices are then exposed to the zone, and for that to work the zone needs to use exclusive-ip networking rather than shared-ip (and would need to do so anyway for packet forwarding to work).

One thing I learnt was that you can't lock the router zone's networking down with allowed-address. The anti-spoofing protection that allowed-address gives you prevents forwarding and breaks NAT.

The router zone also has a couple of extra pieces of software installed. The first is haproxy, which is intended as an ingress controller. That's not currently used, and could be replaced by something else. The second is dnsmasq, which is used as a dhcp server to configure any zones that get connected to the subnet.

With a network segment in place, and a router zone for management, you can then create extra zones.

The way this works in Tribblix is that if you tell zap to create a zone with an IP address that is part of a private subnet, it will attach its network to the corresponding etherstub. That works fine for an exclusive-ip zone, where the vnic can be created directly over the etherstub.

For shared-ip zones it's a bit trickier. The etherstub isn't a real network device, although for some purposes (like creating a vnic) it looks like one. To allow shared-ip, I create a dedicated shared vnic over the etherstub, and the virtual addresses for shared-ip zones are associated with that vnic. For this to work, it has to be plumbed in the global zone, but doesn't need an address there. The downside to the shared-ip setup (or it might be an upside, depending on what the zone's going to be used for) is that in this configuration it doesn't get a network route; normally this would be inherited off the parent interface, but there isn't an IP configuration associated with the vnic in the global zone.

The shared-ip zone is handed its IP address. For exclusive-ip zones, the right configuration fragment is poked into dnsmasq on the router zone, so that if the zone asks via dhcp it will get the answer you configured. Generally, though, if I can directly configure the zone I will. And that's either by putting the right configuration into the files in a zone so it implements the right networking at boot, or via cloud-init. (Or, in the case of a solaris10 zone, I populate sysidcfg.)

There's actually a lot of steps here, and doing it by hand would be rather (ahem, very) tedious. So it's all automated by zap, the package and system administration tool in Tribblix. The user asks for a router zone, and all it needs to be given is the zone's name, the public IP address, and the subnet address, and all the work will be done automatically. It saves all the required details so that they can be picked up later. Likewise for a regular zone, it will do all the configuration based on the IP address you specify, with no extra input required from the user.

The whole aim here is to make building zones, and whole systems of zones, much easier and more reliable. And there's still a lot more capability to add.

Keeping python modules in check

Any operating system distribution - and Tribblix is no different - will have a bunch of packages for python modules.

And one thing about python modules is that they tend to depend on other python modules. Sometimes a lot of python modules. Not only that, the dependency will be on a specific version - or range of versions - of particular modules.

Which opens up the possibility that two different modules might require incompatible versions of a module they both depend on.

For a long time, I was a bit lax about this. Most of the time you can get away with it (often because module writers are excessively cautious about newer versions of their dependencies). But occasionally I got bitten by upgrading a module and breaking something that used it, or breaking it because a dependency hadn't been updated to match.

So now I always check that I've got all the dependencies listed in packaging with

pip3 show modulename

and every time I update a module I check the dependencies aren't broken with

pip3 check

Of course, this relies on the machine having all the (interesting) modules installed, but on my main build machine that is generally true.

If an incompatibility is picked up by pip3 check then I'll either not do the update, or update any other modules to keep in sync. If an update is impossible, I'll take a note of which modules are blockers, and wait until they get an update to unjam the process.

A case in point was that urllib3 went to version 2.x recently. At first, nothing would allow that, so I couldn't update urllib3 at all. Now we're in a situation where I have one module I use that won't allow me to update urllib3, and am starting to see a few modules requiring urllib3 to be updated, so those are held downrev for the time being.

The package dependencies I declare tend to be the explicit module dependencies (as shown by pip3 show). Occasionally I'll declare some or all of the optional dependencies in packaging, if the standard use case suggests it. And there's no obvious easy way to emulate the notion of extras in package dependencies. But that can be handled in package overlays, which is the safest way in any case.

Something else the checking can pick up is when a dependency is removed, which is something that can be easily missed.

Doing all the checking adds a little extra work up front, but should help remove one class of package breakage.

It seemed like a simple problem to fix

While a bit under the weather last week, I decided to try and fix what at first glance appears to be a simple problem:

need to ship the manpage with exa

Now, exa is a modern file lister, and the package on Tribblix doesn't ship a man page. The reason for that, it turns out, is that there isn't a man page in the source, but you can generate one.

To build the man page requires pandoc. OK, so how to get pandoc, which wasn't available on Tribblix? It's written in Haskell, and I did have a Haskell package.

Only my version of Haskell was a bit old, and wouldn't build pandoc. The build complains that it's too old and unsupported. You can't even build an old version of pandoc, which is a little peculiar.

Off to upgrade Haskell then. You need Haskell to build Haskell, and it has some specific requirements about precisely which versions of Haskell work. I wanted to get to 9.4, which is the last version of Haskell that builds using make (and I'll leave Hadrian for another day). You can't build Haskell 9.4 with 9.2 which it claims to be too new, you have to go back to 9.0.

Fortunately we do have some bootstrap kits for illumos available, so I pulled 9.0 from there, successfully built Haskell, then cabal, and finally pandoc.

Back to exa. At which point you notice that it's been deprecated and replaced by eza. (This is a snag with modern point tools. They can disappear on a whim.)

So let's build eza. At which point I find that the MSRV (Minimum Supported Rust Version) has been bumped to 1.70, and I only had 1.69. Another update required. Rust is actually quite simple to package, you can just download the stable version and package it.

After all this, exa still doesn't have a man page, because it's deprecated (if you run man exa you get something completely different from X.Org). But I did manage to upgrade Haskell and Cabal, I managed to package pandoc, I updated rust, and I added a replacement utility - eza - which does now come with a man page.