On 6/23/20 2:44 PM, Daniel D. Daugherty wrote:
On 6/18/20 8:54 PM, Chris Plummer wrote:
[I've added runtime-dev to this SA review since understanding
interpreter invokes (code generated by
TemplateInterpreterGenerator::generate_normal_entry()) and stack
walking is probably more important than understanding SA.]
Hello,
Please help review the following:
https://bugs.openjdk.java.net/browse/JDK-8244383
http://cr.openjdk.java.net/~cjplummer/8244383/webrev.00/index.html
Thanks for helping!
Sorry for the delay in reviewing this one. I've come back to it a couple
of times because code like this is very hard to review.
General comment:
This fix reminds of the crazy things that AsyncGetCallTrace has to
do in order to gather call trace data. I'm guessing that SA is
attaching to the VM in an asynchronous manner and that's why it
can observe things like partially constructed frames. If that's a
correct guess, then how is SA stopping/suspending the threads?
I'm just curious here.
On linux SA uses ptrace. I'm not familiar with the details of how it
works. I'm not sure where ptrace allows suspends to happen, but
certainly it has no knowledge of JVM safepoints or other synchronization
that the JVM does. So from the JVM and SA point of view the suspend can
happen at any arbitrary JVM instruction.
From what I can gather, PTRACE_ATTACH suspends the entire process, so
that means all threads are suspended once you attach. However,
PTRACE_GETREGS can be called on individual threads (LWPs), but I don't
see any indication in the SA code that you need to attach to each LWP first.
Or this might be a case where SA is examining a core file in
which case the various threads stacks are not necessarily at
good/safepoint-safe pause points.
For this bug and test it's a live process, but I think the bug being
addressed here can happen just as well with a core file. Unfortunately
we have very little core file testing support. I'm actually in the
middle of addressing that right now.
src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/runtime/x86/X86Frame.java
No comments.
src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/debugger/bsd/amd64/BsdAMD64CFrame.java
No comments.
src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/runtime/amd64/AMD64CurrentFrameGuess.java
L104: // two locations, then we canot determine the frame.
typo: s/canot/cannot/
ok
L127: // it's validity will help us determine the state of the
new frame push.
typo: s/it's/its/
ok
L148: System.out.println("CurrentFrameGuess: frame pushed
but not initaliazed.");
typo: s/initaliazed/initialized/
ok
L220: System.out.println("CurrentFrameGuess:
choosing interpreter frame: sp = " +
L221: spFound + ", fpFound = " +
fp + ", pcFound = " + pc);
This debug output doesn't make sense to me:
"sp = " label and 'spFound' value
"fpFound = " label and 'fp' value
"pcFound = " label and 'pc' value
but I may not have enough context...
From the point of view of the person reading the output, they want to
know the values for sp, fp, and pc. But within the code these values are
stored in the "found" variables.
With code like this, it's really hard to figure out if you've covered
all the cases unless you've been in the observer seat yourself and
even then your test runs may not hit all the possible cases. All you
can really do is start with a set of adaptive changes, run with those
for a while and tweak them as you gather more observations.
Yes, and I know there is still a very tiny gap or two in coverage that
are maybe one or two instructions long, but they aren't worth dealing
with. This bug was already very rare, and with the fixes I've done I
don't see any issues now. SA is a debugger, so perfection in this regard
is not expected.
Chris, nice job with this bit of insanity!
Thanks! I mostly stuck with this one to help with my SA expertise.
Otherwise it wouldn't have been worth the time.
Chris
Thumbs up!
Dan
The crux of the bug is when doing stack walking the topmost frame is
in an inconsistent state because we are in the middle of pushing a
new interpreter frame. Basically we are executing code generated by
TemplateInterpreterGenerator::generate_normal_entry(). Since the PC
register is in this code, SA assumes the topmost frame is an
interpreter frame.
The first issue with this interpreter frame assumption is if we
haven't actually pushed the frame yet, then the current frame is the
caller's frame, and could be compiled. But since SA thinks it's
interpreted, later on it tries to convert the frame->bcp to a BCI,
but frame->bcp is only valid for interpreter frames. Thus the
"illegal BCI" failures. If the previous frame happened to be
interpreted, then the existing SA code works fine.
The other state of frame pushing that was problematic was when the
new frame had been pushed, but frame->method and frame->bcp were not
setup yet. This also would lead to "illegal BCI" later on because
garbage would be stored in these locations.
Fixing the above problems requires trying to determine the state of
the frame push through a series of checks, and then adapting what is
considered to be the current frame based on the outcome of the
checks. The first things checked is that frame->method is valid (we
can successfully instantiate a wrapper for the Method* without
failure) and that frame->bcp is within the method. If both these pass
then we can use the frame as-is.
If the above checks fail, then we try to determine whether the issue
is that the frame is not yet pushed and the current frame is actually
compiled, or the frame has been pushed but not yet initialized. This
is done by first getting the return address from the stack or RAX
(it's location depends on how far along we are in the entry code) and
comparing this to what is stored in frame->return_addr. If they are
the same, then we have pushed the frame but not yet initialized it.
In this case we use the previous frame (senderSP() and senderFP()) as
the current frame since the current frame is not yet initialized. If
the return address check fails, then we assume the new frame is not
yet pushed, and and treat the current frame as compiled, even though
PC points into the interpreter (we replace PC with RAX in this case).
Comments in the code pretty well explain all the above, so it is
probably easier to follow the logic in the code along with the
comments rather than apply my above description to the code.
I should add that it's very rare that we ever get into this special
error handling code. This bug was very hard to reproduce initially. I
was only able to make progress with reproducing and debugging by
inserting delay loops in various spots in the code generated by
TemplateInterpreterGenerator::generate_normal_entry(). By doing this
I was able to reproduce the issue quite easily and hit all the logic
in the new code I've added.
The fix is basically entirely contained within
AMD64CurrentFrameGuess.java. The rest of the changes are minor:
src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/runtime/amd64/AMD64CurrentFrameGuess.java
-Main fix for CR
src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/runtime/x86/X86Frame.java
-Added getInterpreterFrameBCP(), which is now needed by
AMD64CurrentFrameGuess.java
-I also simplified some code by using the existing
getInterpreterFrameMethod()
rather than replicating inline what it does.
src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/debugger/bsd/amd64/BsdAMD64CFrame.java
-I noticed the windows version of this code had some extra checks
that were missing
from the bsd version. I then looked at the linux version, but it had
been heavily modified
a short while back to leverage DWARF info to determine frames. So I
looked at the previous
rev and it too had these extra checks. I decided to add them to the
BSD port. I'm not sure
if it helps at all, but it certainly doesn't seem to do any harm.
thanks,
Chris
