Re: RFR: 8212995: Consider placing the Integer.IntegerCache and cached Integer objects in the closed archive heap region

[Jiangli Zhou]

Thanks!

Jiangli


On 11/2/18 5:15 PM, Ioi Lam wrote:

Hi Jiangli,

This looks good to me. Ship it!

Thanks

- Ioi


On 11/2/18 10:57 AM, Jiangli Zhou wrote:

Hi Ioi,

Here is the updated webrev with the warning message below for the Integer cache.

  http://cr.openjdk.java.net/~jiangli/8212995/webrev.03/

 995      *

 996      * WARNING: The cache is archived with CDS and reloaded from the shared  997      * archive at runtime. The archived cache (Integer[]) and Integer objects  998      * reside in the closed archive heap regions. Care should be taken when  999      * changing the implementation and the cache array should not be assigned

1000      * with new Integer object(s) after initialization.

Including core-lib-dev mailing list since the change now touches the core library file.

Thanks,
Jiangli


On 11/1/18 10:58 PM, Jiangli Zhou wrote:

Hi Ioi,

On Nov 1, 2018, at 9:37 PM, Ioi Lam <ioi....@oracle.com> wrote:

Hi Jiangli,

  576 void HeapShared::check_closed_archive_heap_region_object(InstanceKlass* k,

  577 Thread* THREAD) {
  578   // Check fields in the object
  579   for (JavaFieldStream fs(k); !fs.done(); fs.next()) {
  580     if (!fs.access_flags().is_static()) {
  581       BasicType ft = fs.field_descriptor().field_type();

  582       if (!fs.access_flags().is_final() && (ft == T_ARRAY || T_OBJECT)) {

  583         ResourceMark rm(THREAD);
  584         log_warning(cds, heap)(

  585           "Please check reference field in %s instance in closed archive heap region: %s %s",

  586           k->external_name(), (fs.name())->as_C_string(),
  587           (fs.signature())->as_C_string());
  588       }
  589     }
  590   }
  591 }

Checking that all static fields of the affected class (IntegerCache in this case) are final is not enough. The elements of a final array can be modified.

Just to clarify, the above checks all instance fields (non-static fields) in non-array objects. Static fields are not checked as mirrors are not in the closed archive heap region. In the IntegerCache subgraph case, it makes sure there is no non-final reference instance field in cached Integer objects (for future proof).

JLS requires that
https://docs.oracle.com/javase/specs/jls/se8/html/jls-5.html#jls-5.1.7

If the value p being boxed is an integer literal of type int between -128 and 127 inclusive (§3.10.1), or the boolean literal true or false (§3.10.3), or a character literal between '\u0000' and '\u007f' inclusive (§3.10.4), then let a and b be the results of any two boxing conversions of p. It is always the case that a == b. However, there's no requirement that all these special literal values must be created at system bootstrap time. So it's conceivable that IntegerCache may be rewritten to create the object dynamically:

   public static Integer valueOf(int i) {
       if (i >= IntegerCache.low && i <= IntegerCache.high) {
+       if (IntegerCache.cache[i + (-IntegerCache.low)] == null) {

+         IntegerCache.cache[i + (-IntegerCache.low)] = new Integer(i);

+       }
         return IntegerCache.cache[i + (-IntegerCache.low)];
       }
       return new Integer(i);
   }

Now, is this likely to happen? Probably not. However, in HotSpot, we should not assume that the library will always stay the same, and that the library writer knows what HotSpot requires.

So your suggestion was to add warnings specifically for Integer cache array. Looks like I’ve given it a much deeper interpretation.

I’ll add a warning about the cache. Will send new webrev tomorrow.

Thanks,
Jiangli

Thanks
- Ioi

On 11/1/18 3:47 PM, Jiangli Zhou wrote:
Hi Ioi,

Thanks for the review. I renamed both fields as suggested and added a warning for closed_archive_subgraph_entry_fields. A standalone warning message in Integer.java could be overlooked, so I added HeapShared::check_closed_archive_heap_region_object() for checking the instances being included in the closed archive heap regions at dump time.

http://cr.openjdk.java.net/~jiangli/8212995/webrev.02/

Thanks,

Jiangli

On 10/31/18 8:52 PM, Ioi Lam wrote:
Hi Jiangli,

static ArchivableStaticFieldInfo shareable_subgraph_entry_fields[] = {...}

static ArchivableStaticFieldInfo subgraph_entry_fields[] = {...}

Maybe these should be renamed to {open/closed}_archive_subgraph_entry_fields?

Also, I think we should add a strong warning about what objects can be placed in closed_archive_subgraph_entry_fields[]. Any references stored in these objects must not be modified at run time (or else we could have a pointer that from the closed region to the outside, violating the properties of the closed region.

Maybe we should also add a warning in Integer.java, something akin to "if you modify this class, check to see if it can still meet the requirements in heapShared.cpp"?

The rest of the code seems OK to me.

As a future improvement, for all the objects whose fields are all non-reference, final fields, maybe we can automatically put them in the closed archive region? For example, all archived primitive box objects are in this category.

Thanks
- Ioi

On 10/31/18 12:45 PM, Jiangli Zhou wrote:
On 10/31/18 12:08 PM, Jiangli Zhou wrote:

Hi Ioi,

Here is an updated webrev with renaming of the 'is_shared' argument. I decided to go with your suggestion, 'is_closed_archive'.

http://cr.openjdk.java.net/~jiangli/8212995/webrev.01/

BTW, in above webrev, I also included a typo fix for the following warning that Mandy found (thanks Mandy!)

@@ -1324,11 +1329,11 @@
    // header data

    const char* prop = Arguments::get_property("java.system.class.loader");

    if (prop != NULL) {

      warning("Archived non-system classes are disabled because the "               "java.system.class.loader property is specified (value = \"%s\"). " -            "To use archived non-system classes, this property must be not be set", prop); +            "To use archived non-system classes, this property must not be set", prop);

      _has_platform_or_app_classes = false;
    }

Thanks,
Jiangli

Thanks,

Jiangli

On 10/30/18 4:19 PM, Jiangli Zhou wrote:
Hi Ioi,

On 10/30/18 3:00 PM, Ioi Lam wrote:

Hi Jiangli,

This looks promising.

Now a full review yet, but I am wondering about the name of the is_shared parameter

   void add_subgraph_entry_field(int static_field_offset, oop v, bool is_shared);

Since this is part of "heapShared", everything is "shared" in some sense of the word. It could be confusing to say something is more shared than other things which also shared ...

How "is_closed_archive" instead?

Yes, our 'shared' has broader meaning. "is_closed_archive" or "is_closed_space" sounds good to me. I'll rename.

Thanks,
Jiangli

Thanks
- Ioi

On 10/30/2018 01:57 PM, Jiangli Zhou wrote:

Please review the following change for moving the archived Integer.IntegerCache and it's cached Integer objects (256) to the closed archiving heap region. The IntegerCache subgraph does not contain any reference that's changed at runtime (good candidate for sharing). Moving the whole subgraph into the closed archive heap region allows the memory to be shared by different JVM instances at runtime. The saving is 4K per JVM instance running the same or different java application simultaneously. Although 4K is not significant, in a larger picture the saving is much bigger (4k * (JVM_instance_num - 1) * host_num).

As part of the change, I also restructured the code to allow us to plug in more shareable subgraphs in the closed archive heap region for runtime footprint saving in the future.

The 'st' space is renamed to 'ca' (closed archive) space since it now contains other types of objects besides j.l.Strings.

webrev: http://cr.openjdk.java.net/~jiangli/8212995/webrev.00/
RFE: https://bugs.openjdk.java.net/browse/JDK-8212995

Before:

mc  space:      8416 [  0.0% of total] out of     12288 bytes [ 68.5% used] at 0x0000000800000000 rw  space:   3946640 [ 21.4% of total] out of 3948544 bytes [100.0% used] at 0x0000000800003000 ro  space:   7319328 [ 39.6% of total] out of 7319552 bytes [100.0% used] at 0x00000008003c7000 md  space:      2416 [  0.0% of total] out of      4096 bytes [ 59.0% used] at 0x0000000800ac2000 od  space:   6475944 [ 35.0% of total] out of 6479872 bytes [ 99.9% used] at 0x0000000800ac3000 st0 space:    438272 [  2.4% of total] out of 438272 bytes [100.0% used] at 0x00000007ffc00000 <<<<<<<<<< oa0 space:    282624 [  1.5% of total] out of 282624 bytes [100.0% used] at 0x00000007ff800000 <<<<<<<<<< total    :  18473640 [100.0% of total] out of 18485248 bytes [ 99.9% used]

After:

mc  space:      8416 [  0.0% of total] out of     12288 bytes [ 68.5% used] at 0x0000000800000000 rw  space:   3946640 [ 21.4% of total] out of 3948544 bytes [100.0% used] at 0x0000000800003000 ro  space:   7319304 [ 39.6% of total] out of 7319552 bytes [100.0% used] at 0x00000008003c7000 md  space:      2416 [  0.0% of total] out of      4096 bytes [ 59.0% used] at 0x0000000800ac2000 od  space:   6475920 [ 35.0% of total] out of 6479872 bytes [ 99.9% used] at 0x0000000800ac3000 ca0 space:    442368 [  2.4% of total] out of 442368 bytes [100.0% used] at 0x00000007ffc00000 <<<<<<<<<< oa0 space:    278528 [  1.5% of total] out of 278528 bytes [100.0% used] at 0x00000007ff800000 <<<<<<<<<< total    :  18473592 [100.0% of total] out of 18485248 bytes [ 99.9% used]

Tested with appcds tests on linux-x64 locally. Running tier1-teir4 tests.

Thanks,

Jiangli