[GR-49300] Break the initialization cycle in NIO/cgroups code. #7553
Conversation
First, we use a separate accessor for page-alignedness as it doesn't need the more sophisticated initialization of the directMemory field. Next, ensure PhysicalMemory initialization is serialized and when it is, set directMemory to a static value so that the container code can finish initialization without introducing a cyle. The final directMemory value based on the heap size is then published to JDK code by setting the VM init level to 1. Therefore, application code would use the non-static value as the upper bound. Closes: #556
oracle-contributor-agreement
bot
added
the
OCA Verified
There was a problem hiding this comment.
Thanks @christianhaeubl! This is sufficient for our use-case. +1 from me.
graalvmbot
force-pushed
the
chaeubl/GR-49300
branch
from
120ce63 to
06158a9
Compare
|
@jerboaa: I pushed another small fix, see the top commit. |
…ore thread-safe way.
graalvmbot
force-pushed
the
chaeubl/GR-49300
branch
from
d090636 to
a606940
Compare
| /** | ||
| * {@code java.nio.Bits} caches the max. direct memory size in the field {@code MAX_MEMORY}. We | ||
| * disable this cache and always call {@link DirectMemoryAccessors#getDirectMemory()} instead, which | ||
| * uses our own cache. Otherwise, it could happen that {@code MAX_MEMORY} caches a temporary value | ||
| * that is used during early VM startup, before {@link PhysicalMemory} is fully initialized. | ||
| */ | ||
| final class MaxMemoryAccessor { | ||
| // Checkstyle: stop | ||
|
|
||
| static long getMAX_MEMORY() { | ||
| return DirectMemoryAccessors.getDirectMemory(); | ||
| } | ||
|
|
||
| static void setMAX_MEMORY(long value) { | ||
| /* Nothing to do. */ | ||
| } | ||
|
|
||
| // Checkstyle: resume | ||
| } |
There was a problem hiding this comment.
Instead of going through the accessor an alternative could be to leaverage the init hooks in JDK code and actually set the init level to 1 once we know the real MaxDirectMemorySize (based on the heap size). For example here: https://github.com/oracle/graal/pull/7553/files#diff-db72af86eee732c8a7e3e53319b32fdd92ee0ec00735f413a02858c9dc0aff4dR126. That would mimic what HotSpot itself is doing. FWIW, that's what my original code was intending to do.
Both approaches work.
There was a problem hiding this comment.
The problem is that Native Image is already multi-threaded when this code is executed. Modifying the init level may have unexpected effects on other threads.
|
@christianhaeubl Thanks for getting this fix in! Much appreciated. |
No description provided.