RFR: 8234372: Investigate use of Thread::stack_base() and queries for "in stack"

[David Holmes]

On 12/02/2020 5:23 pm, Thomas Stüfe wrote:
> Tests ran through, no visible regressions so far.

Thanks Thomas!

As per my email response to Dan I made a further correction to the ARM code:

http://cr.openjdk.java.net/~dholmes/8234372/webrev.v3/

Thanks,
David


> Cheers, Thomas
> 
> On Tue, Feb 11, 2020 at 9:35 AM Thomas Stüfe <thomas.stuefe at gmail.com 
> <mailto:thomas.stuefe at gmail.com>> wrote:
> 
>     Hi David,
> 
>     I did not find anything wrong in your patch. Nice cleanup, and great
>     archaeological work :)
> 
>     Only small nits and some bike shedding:
> 
>     ---
> 
>     So, if I get this right:
> 
>     is_in_stack -> is in live stack (base ... sp]
>     on_local_stack  -> is in (base...start] includes guard pages
>     is_in_usable_stack -> is in (base...start - guard] excludes guard pages
> 
>     The naming is confusing but I saw you recommended renaming the
>     functions in the JBS comments, and I like all your suggestions
>     better than what we have now.
> 
>     -----
> 
>     http://cr.openjdk.java.net/~dholmes/8234372/webrev.v2/src/hotspot/cpu/aarch64/frame_aarch64.cpp.udiff.html
> 
>     -  if (locals > thread->stack_base() || locals < (address) fp())
>     return false;
>     +  if (locals >= thread->stack_base() || locals < (address) fp())
>     return false;
> 
>     This would be easier to read as a negated positive (also applies to
>     all other frame_xxx.cpp).
> 
>     Just an idea, maybe we could add a function
>     Thread::is_in_stack_limited_by(ptr, arbitrary_end_ptr) which could
>     compare that ptr is between (base .. arbitrary_end_ptr] and based on
>     that we could implement the other three stack functions.
> 
>     For cases like this we could then write:
>     if (!thread->is_in_stack_limited_by(locals, fp())
> 
>     But I am unsure, maybe I overthink things.
> 
>     -------
> 
>     http://cr.openjdk.java.net/~dholmes/8234372/webrev.v2/src/hotspot/os/linux/os_linux.cpp.udiff.html
> 
>     -  if (addr <  t->stack_base() && addr >=
>     t->stack_reserved_zone_base()) {
>     +  if (t->is_in_usable_stack(addr)) {
> 
>     First confused me but then I read Fredericks comment in JBS so I
>     think it is okay.
> 
>     But it would be nice to be able to remove this
>     manually-expand-stack-coding altogether :)
> 
>     -------
> 
>     Cheers, Thomas
> 
>     On Tue, Feb 11, 2020 at 6:40 AM David Holmes
>     <david.holmes at oracle.com <mailto:david.holmes at oracle.com>> wrote:
> 
>         Bug: https://bugs.openjdk.java.net/browse/JDK-8234372
>         webrev: http://cr.openjdk.java.net/~dholmes/8234372/webrev.v2/
> 
>         Following on from JDK-8215355 I checked all uses of
>         Thread::stack_base()
>         to watch for range tests that should be exclusive but are
>         inclusive, and
>         vice-versa. And in addition clarified and streamlined the
>         various "in
>         stack" checks that are made.
> 
>         Summary of changes:
> 
>         src/hotspot/cpu/aarch64/frame_aarch64.cpp
>         src/hotspot/cpu/arm/frame_arm.cpp
>         src/hotspot/cpu/ppc/frame_ppc.cpp
>         src/hotspot/cpu/s390/frame_s390.cpp
>         src/hotspot/cpu/sparc/frame_sparc.cpp
>         src/hotspot/cpu/x86/frame_x86.cpp
> 
>         In terms of actual bugs the implementation of
>         frame::safe_for_sender on
>         all platforms except x86 and aarch64 was using the wrong range
>         test in a
>         number of cases, so these are all now correct and consistent.
> 
>         All platforms had an incorrect range check in relation to the
>         "locals".
> 
>         All platforms now use is_in_usable_stack to check for a valid
>         sp, rather
>         than duplicating (sometimes incorrectly) that logic.
> 
>         --
> 
>         src/hotspot/os/linux/os_linux.cpp
>         src/hotspot/os/windows/os_windows.cpp
> 
>         Replaced explicit range check with is_in_usable_stack
> 
>         src/hotspot/os_cpu/linux_arm/os_linux_arm.cpp
>         src/hotspot/os_cpu/linux_s390/thread_linux_s390.cpp
> 
>         Replaced explicit range check with on_local_stack.
> 
>         ---
> 
>         src/hotspot/share/runtime/thread.?pp
> 
>         Moved is_in_usable_stack from Thread to JavaThread (guard
>         regions are
>         only relevant for JavaThreads).
> 
>         Clarified functionality and use of the three "in stack" variants.
> 
>         Removed redundant check from is_in_stack:
> 
>         !   // Allow non Java threads to call this without stack_base
>         !   if (_stack_base == NULL) return true;
> 
>         As this is executed by the current thread, and the very first
>         thing a
>         thread does is set its stack base and size, it is impossible to
>         find a
>         NULL stack_base (which is already asserted inside stack_base()). [I
>         tested this extensively just as a sanity check: tiers 1-5 plus
>         hotspot
>         runtime/serviceability/gc.]
> 
>         Misc cleanup to use stack_end() rather than recalulate it.
> 
>         ---
> 
>         There are some further possible cleanups here but I didn't want
>         to go
>         too far with things that would obscure the functional changes
>         too much.
>         As mentioned in the bug report the three "in stack" functions would
>         benefit from some minor renamings so that their relationship is
>         clearer.
>         But I can leave that to a follow on RFE. Further, it may be
>         possible to
>         replace a lot of the remaining uses of stack_base() with a more
>         constrained "in stack" function that takes a limit. For example,
>         rather
>         than something like:
> 
>         if (thread->stack_base() > fp && fp >= sp)
> 
>         have:
> 
>         if (thread->is_in_stack_range(fp /* addr*/, sp /*limit*/))
> 
>         which checks the given addr against stack_base and the limit,
>         and checks
>         the limit against stack_end(). The difficultly may lie in
>         determining
>         whether checking against the limit should be a > or >= test, as
>         it will
>         be dependent on the context. Again this seems like something for a
>         second RFE.
> 
>         ---
> 
>         Testing:
> 
>         Thanks to Andrew Haley for taking the frame changes for a spin on
>         ARM/PPC/Aarch64/S390(?).
> 
>         I also ran our tier 1 to 3 testing on x86 and sparc.
> 
>         Thanks,
>         David
>