8296507: GCM using more memory than necessary with in-place operations #11121

ascarpino · 2022-11-13T02:54:10Z

I would like a review of an update to the GCM code. A recent report showed that GCM memory usage for TLS was very large. This was a result of in-place buffers, which TLS uses, and how the code handled the combined intrinsic method during decryption. A temporary buffer was used because the combined intrinsic does gctr before ghash which results in a bad tag. The fix is to not use the combined intrinsic during in-place decryption and depend on the individual GHASH and CounterMode intrinsics. Direct ByteBuffers are not affected as they are not used by the intrinsics directly.

The reduction in the memory usage boosted performance back to where it was before despite using slower intrinsics (gctr & ghash individually). The extra memory allocation for the temporary buffer out-weighted the faster intrinsic.

    JDK 17:   122913.554 ops/sec
    JDK 19:    94885.008 ops/sec
    Post fix: 122735.804 ops/sec

There is no regression test because this is a memory change and test coverage already existing.

Progress

Change must be properly reviewed (1 review required, with at least 1 Reviewer)
Change must not contain extraneous whitespace
Commit message must refer to an issue

Issue

JDK-8296507: GCM using more memory than necessary with in-place operations

Reviewers

Jamil Nimeh (@jnimeh - Reviewer) ⚠️ Review applies to 340ab22f

Reviewing

Using git

Checkout this PR locally:
$ git fetch https://git.openjdk.org/jdk pull/11121/head:pull/11121
$ git checkout pull/11121

Update a local copy of the PR:
$ git checkout pull/11121
$ git pull https://git.openjdk.org/jdk pull/11121/head

Using Skara CLI tools

Checkout this PR locally:
$ git pr checkout 11121

View PR using the GUI difftool:
$ git pr show -t 11121

Using diff file

Download this PR as a diff file:
https://git.openjdk.org/jdk/pull/11121.diff

bridgekeeper · 2022-11-13T02:55:25Z

👋 Welcome back ascarpino! A progress list of the required criteria for merging this PR into master will be added to the body of your pull request. There are additional pull request commands available for use with this pull request.

openjdk · 2022-11-13T02:57:45Z

@ascarpino The following label will be automatically applied to this pull request:

security

When this pull request is ready to be reviewed, an "RFR" email will be sent to the corresponding mailing list. If you would like to change these labels, use the /label pull request command.

mlbridge · 2022-11-15T22:48:44Z

Webrevs

j-baker · 2022-11-16T16:18:15Z

It's possibly worth noting that while this is merely fixing a regression for x86, it's very likely a decent sized performance improvement on arm64, where intrinsics for AES-GCM (depending on JVM vendor) aren't added until after Java 17.

carterkozak · 2022-11-16T16:33:06Z

Thanks for looking into this, @ascarpino!

In testing this using a local build, it improves performance in cases using heap buffers (a super-set of the socket case), however servers which use direct byte-buffers still exhibit a similar performance regression (heavy allocation compared to jdk17, ~10% slower TLS performance in HTTP+TLS benchmarks). It's possible that has a different root cause, but the outcome is strikingly similar.

j-baker · 2022-11-16T16:44:20Z

Carter, when I looked at this a few months back (admittedly I'm a fairly careless profiler and didn't fully dig down to a root cause) I felt as though direct bytebuffers were possibly getting compromised round about here:

jdk/src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

Line 722 in 3416bfa

int implGCMCrypt(ByteBuffer src, ByteBuffer dst) {

, where... essentially it's easier for intrinsics to operate on byte arrays and so any direct data passed in gets copied into a new byte array which is then passed into the intrinsic.

It's possible that I'm misunderstanding, however. I think one could test this hypothesis by adjusting the size of PARALLEL_LEN. Halving it will lead to a less efficient intrinsic usage but correspondingly halve the allocation rate.

mcpowers · 2022-11-16T16:55:39Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

+     * large chunks of data into 1MB sized chunks. This is to place
+     * an upper limit on the number of blocks encrypted in the intrinsic.
+     *
+     * For decrypting in-place byte[], calling methods must ct must set to null


end of sentence mangled

"must ct must set to null" => "must set ct to null"?

mcpowers · 2022-11-16T16:57:14Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

    private static int implGCMCrypt(byte[] in, int inOfs, int inLen, byte[] ct,
                                    int ctOfs, byte[] out, int outOfs,
                                    GCTR gctr, GHASH ghash) {

        int len = 0;
-        if (inLen > SPLIT_LEN) {
+        // Loop if input length is greater than the SPLIT_LEN


comment doesn't add anything not already obvious from the code

yeah.. probably right

mcpowers · 2022-11-16T17:00:37Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

            while (inLen >= SPLIT_LEN) {
-                int partlen = implGCMCrypt0(in, inOfs + len, SPLIT_LEN, ct,
+                partlen = implGCMCrypt0(in, inOfs + len, SPLIT_LEN, ct,


why not int partlen and get rid of line 594

can reuse the same partlen for all loops through the while

mcpowers · 2022-11-16T17:01:50Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

                    ctOfs + len, out, outOfs + len, gctr, ghash);
                len += partlen;
                inLen -= partlen;
            }
        }
+
+        // Finish any remaining data


comment doesn't add anything special

mcpowers · 2022-11-16T17:21:25Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

@@ -666,6 +691,11 @@ abstract class GCMEngine {
        byte[] originalOut = null;
        int originalOutOfs = 0;

+        // True if op is in-place array decryption with the input & output


// Setting inPlaceArray to true turns off combined intrinsic processing.

yeah that's better

Actually the replacement isn't entirely accurate. This only applies to decryption and for buffers that don't overlap where input is ahead of output. That's why the comment is so wordy

mcpowers · 2022-11-16T17:38:19Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

-                    gctr, ghash);
+                byte[] array;
+                if (encryption) {
+                    array = dst.array();


You could factor out lines 764 and 770 by changing line 762 to
byte[] array = encryption ? dst.array() : src.array();

That was intentional since 763 checks the encryption boolean, I can define 'array' in that condition instead of having two conditions for the same thing

mcpowers · 2022-11-16T17:47:38Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

-                } else {
-                    Unsafe.getUnsafe().setMemory(((DirectBuffer)dst).address(),
-                        len + dst.position(), (byte)0);
+                // If this is an in-place array, don't zero the src


The comment doesn't jive with the line of code on the next line. It is the inverse of the comment.

ascarpino · 2022-11-16T18:19:30Z

It's possibly worth noting that while this is merely fixing a regression for x86, it's very likely a decent sized performance improvement on arm64, where intrinsics for AES-GCM (depending on JVM vendor) aren't added until after Java 17.

That is why it refers to "combined intrinsic" more than spelling out AVX512. The change affects all platforms

ascarpino · 2022-11-16T18:48:58Z

Thanks for looking into this, @ascarpino!

In testing this using a local build, it improves performance in cases using heap buffers (a super-set of the socket case), however servers which use direct byte-buffers still exhibit a similar performance regression (heavy allocation compared to jdk17, ~10% slower TLS performance in HTTP+TLS benchmarks). It's possible that has a different root cause, but the outcome is strikingly similar.

Well the provided test ran with heap bytebuffers and direct bytebuffers are handled differently because it has to copy data for the intrinsic. But that data allocation is pretty low and I believe was the same in 17. So I'm not aware of a direct bytebuffer slowdown as you now report

carterkozak · 2022-11-16T22:43:16Z

Great point, I neglected to add benchmark coverage for the direct buffer case. I've updated my benchmark repository with a server using direct buffers:
https://github.com/carterkozak/java-crypto-allocation-performance/blob/develop/java-crypto-allocation-performance/src/main/java/com/palantir/java/crypto/allocations/DirectBufferTransportLayerSecurityBenchmark.java

ascarpino · 2022-11-17T18:24:07Z

Thanks for looking into this, @ascarpino!

In testing this using a local build, it improves performance in cases using heap buffers (a super-set of the socket case), however servers which use direct byte-buffers still exhibit a similar performance regression (heavy allocation compared to jdk17, ~10% slower TLS performance in HTTP+TLS benchmarks). It's possible that has a different root cause, but the outcome is strikingly similar.

Looking at this, it's not related to the same in-place issues. This is a result of the combined intrinsics requirement . Maybe some better tuning can be done, but I think this is unavoidable. I can consider this in a future PR

carterkozak · 2022-11-17T19:01:50Z

Looking at this, it's not related to the same in-place issues. This is a result of the combined intrinsics requirement . Maybe some better tuning can be done, but I think this is unavoidable. I can consider this in a future PR

That makes sense, perhaps we could document the finding in a new jira issue for posterity in case it impacts other folks as well? I can't overstate my appreciation for your work, thank you!

ascarpino · 2022-11-21T18:16:09Z

btw. that last commit comment is wrong.. it's cleaning up from mcpowers's comments.

valeriepeng · 2022-11-30T23:49:22Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

+     * For decrypting in-place byte[], calling methods must ct must set to null
+     * to avoid combined intrinsic, call GHASH directly before GCTR to avoid
+     * a bad tag exception. This check is not performed here because it would
+     * impose a check every operation which is less efficient.


Missing "for" after "check"?

valeriepeng · 2022-12-01T02:10:01Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

+                            null, 0, array, dst.arrayOffset() + dst.position(),
+                            gctr, ghash);
+                    } else {
+                        int ofs = src.arrayOffset() + src.position();


Isn't this also used on line 774? Why not move this up and directly refer to it for both places?

The line 774 case only uses the calculated value once. I'm avoiding the unnecessary store & load operations when it is set the value to a variable. I see it when I run javap -c to view the bytecode. It's purely an optimization for performance. I do set it to the variable in the line 778 case because calculating the value twice is probably more expensive than once with the store & load operations.

jnimeh

This looks good, I only have nit comments.

jnimeh · 2022-11-15T23:03:02Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

+     * large chunks of data into 1MB sized chunks. This is to place
+     * an upper limit on the number of blocks encrypted in the intrinsic.
+     *
+     * For decrypting in-place byte[], calling methods must ct must set to null


Typo nit? Should it be "calling methods must set ct to null"

jnimeh · 2022-12-01T18:55:35Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

-                } else {
-                    Unsafe.getUnsafe().setMemory(((DirectBuffer)dst).address(),
-                        len + dst.position(), (byte)0);
+                // If this is no an in-place array, zero the dst buffer


nit: no -> not

openjdk · 2022-12-01T19:12:13Z

@ascarpino This change now passes all automated pre-integration checks.

ℹ️ This project also has non-automated pre-integration requirements. Please see the file CONTRIBUTING.md for details.

After integration, the commit message for the final commit will be:

8296507: GCM using more memory than necessary with in-place operations

Reviewed-by: jnimeh

You can use pull request commands such as /summary, /contributor and /issue to adjust it as needed.

At the time when this comment was updated there had been no new commits pushed to the master branch. If another commit should be pushed before you perform the /integrate command, your PR will be automatically rebased. If you prefer to avoid any potential automatic rebasing, please check the documentation for the /integrate command for further details.

➡️ To integrate this PR with the above commit message to the master branch, type /integrate in a new comment.

valeriepeng · 2022-12-01T23:44:45Z

src/java.base/share/classes/com/sun/crypto/provider/GaloisCounterMode.java

+                    originalOutOfs = outOfs;
+                    return new byte[out.length];
+                }
+                inPlaceArray = (!encryption);


Is the "inPlaceArray" reset somewhere? When inOfs >= outOfs and the function will return on line 1051, the inPlaceArray value will not be set on line 1053. Is this intentional? My vacation is coming up and I can't finish off this review before I leave. I see that Jamil has approved it. No need to hold up this for me. Thanks.

Reviewed-by: sviswanathan, vlivanov

Reviewed-by: dholmes

…tOutputStream0 when no content-length in response Reviewed-by: simonis, dfuchs

Reviewed-by: weijun

Reviewed-by: alanb, jpai

…en fastdebug is used Reviewed-by: stuefe, serb

Reviewed-by: prr

…erOfTrailingZeros/numberOfLeadingZeros()` Reviewed-by: kvn, thartmann

Reviewed-by: kbarrett, sjohanss

Reviewed-by: tschatzl

…ion reported in the error message Reviewed-by: jpai

…g pattern matching Reviewed-by: jlahoda

Reviewed-by: sviswanathan, ascarpino, jnimeh

Reviewed-by: sundar, jlahoda

openjdk · 2022-12-06T19:39:03Z

@ascarpino this pull request can not be integrated into master due to one or more merge conflicts. To resolve these merge conflicts and update this pull request you can run the following commands in the local repository for your personal fork:

git checkout gcm
git fetch https://git.openjdk.org/jdk master
git merge FETCH_HEAD
# resolve conflicts and follow the instructions given by git merge
git commit -m "Merge master"
git push

ascarpino · 2022-12-06T21:35:40Z

/integrate

openjdk · 2022-12-06T21:37:32Z

Going to push as commit b4da0ee.
Since your change was applied there have been 3 commits pushed to the master branch:

cd2182a: 8295724: VirtualMachineError: Out of space in CodeCache for method handle intrinsic
2cdc019: 8298178: Update to use jtreg 7.1.1
79d163d: 8293412: Remove unnecessary java.security.egd overrides

Your commit was automatically rebased without conflicts.

openjdk · 2022-12-06T21:37:59Z

@ascarpino Pushed as commit b4da0ee.

💡 You may see a message that your pull request was closed with unmerged commits. This can be safely ignored.

theRealAph · 2022-12-08T18:24:03Z

What benchmark was this? How large were the buffers?

carterkozak · 2022-12-08T18:34:26Z

@theRealAph I reported this oddity to the mailing list including a benchmark which I later updated to include additional coverage for direct buffers

openjdk bot added the security security-dev@openjdk.org label Nov 13, 2022

ascarpino changed the title ~~GCM using more memory than necessary with in-place operations~~ 8296507: GCM using more memory than necessary with in-place operations Nov 14, 2022

ascarpino marked this pull request as ready for review November 15, 2022 22:41

openjdk bot added the rfr Pull request is ready for review label Nov 15, 2022

mcpowers reviewed Nov 16, 2022

View reviewed changes

valeriepeng reviewed Nov 30, 2022

View reviewed changes

valeriepeng reviewed Dec 1, 2022

View reviewed changes

jnimeh approved these changes Dec 1, 2022

View reviewed changes

openjdk bot added the ready Pull request is ready to be integrated label Dec 1, 2022

valeriepeng reviewed Dec 1, 2022

View reviewed changes

vpaprotsk and others added 7 commits December 6, 2022 11:27

8288047: Accelerate Poly1305 on x86_64 using AVX512 instructions

0f21512

Reviewed-by: sviswanathan, vlivanov

8297164: Update troff man pages and CheckManPageOptions.java

5882910

Reviewed-by: dholmes

8297211: Expensive fillInStackTrace operation in HttpURLConnection.ge…

1b78fab

…tOutputStream0 when no content-length in response Reviewed-by: simonis, dfuchs

8297301: Cleanup unused methods in JavaUtilJarAccess

7274f2e

Reviewed-by: weijun

8297299: SequenceInputStream should not use Vector

2fdbeea

Reviewed-by: alanb, jpai

8297147: UnexpectedSourceImageSize test times out on slow machines wh…

2696b30

…en fastdebug is used Reviewed-by: stuefe, serb

8252713: jtreg time out of CtrlASCII.java seems to hang the Xserver.

ea11fbb

Reviewed-by: prr

Fei Gao and others added 9 commits December 6, 2022 11:27

8297172: Fix some issues of auto-vectorization of `Long.bitCount/numb…

2a410bf

…erOfTrailingZeros/numberOfLeadingZeros()` Reviewed-by: kvn, thartmann

8297186: G1 triggers unnecessary full GCs when heap utilization is low

d150b0e

Reviewed-by: kbarrett, sjohanss

merge with direct

d1b9946

8298145: Remove ContiguousSpace::capacity

67a7221

Reviewed-by: tschatzl

8297687: new URI(S,S,S,S) throws exception with incorrect index posit…

3803da6

…ion reported in the error message Reviewed-by: jpai

8297602: Compiler crash with type annotation and generic record durin…

3474f9a

…g pattern matching Reviewed-by: jlahoda

8297379: Enable the ByteBuffer path of Poly1305 optimizations

ea14fd8

Reviewed-by: sviswanathan, ascarpino, jnimeh

8298142: Update internal comment on language features in SourceVersion

1ae8bb5

Reviewed-by: sundar, jlahoda

merge with direct change

07a73bd

openjdk bot added merge-conflict Pull request has merge conflict with target branch and removed ready Pull request is ready to be integrated labels Dec 6, 2022

Merge branch 'master' into gcm

c3d9467

openjdk bot removed merge-conflict Pull request has merge conflict with target branch rfr Pull request is ready for review labels Dec 6, 2022

ascarpino added 3 commits December 6, 2022 12:25

why won't this merge hell stop 3

030c77a

why won't this merge hell stop 2

0e54b32

editor screwed up

a07c08c

openjdk bot added ready Pull request is ready to be integrated rfr Pull request is ready for review labels Dec 6, 2022

openjdk bot added the integrated Pull request has been integrated label Dec 6, 2022

openjdk bot closed this Dec 6, 2022

openjdk bot removed ready Pull request is ready to be integrated rfr Pull request is ready for review labels Dec 6, 2022

ascarpino deleted the gcm branch February 14, 2025 21:15

8296507: GCM using more memory than necessary with in-place operations #11121

8296507: GCM using more memory than necessary with in-place operations #11121

Conversation

ascarpino commented Nov 13, 2022 • edited by openjdk bot Loading

Progress

Issue

Reviewers

Reviewing

bridgekeeper bot commented Nov 13, 2022

openjdk bot commented Nov 13, 2022

mlbridge bot commented Nov 15, 2022 • edited Loading

Webrevs

j-baker commented Nov 16, 2022 • edited by bridgekeeper bot Loading

carterkozak commented Nov 16, 2022

j-baker commented Nov 16, 2022 • edited by bridgekeeper bot Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ascarpino commented Nov 16, 2022

ascarpino commented Nov 16, 2022

carterkozak commented Nov 16, 2022

ascarpino commented Nov 17, 2022

carterkozak commented Nov 17, 2022

ascarpino commented Nov 21, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jnimeh left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

openjdk bot commented Dec 1, 2022 • edited Loading

Choose a reason for hiding this comment

openjdk bot commented Dec 6, 2022

ascarpino commented Dec 6, 2022

openjdk bot commented Dec 6, 2022

openjdk bot commented Dec 6, 2022

theRealAph commented Dec 8, 2022

carterkozak commented Dec 8, 2022

ascarpino commented Nov 13, 2022 •

edited by openjdk bot

Loading

mlbridge bot commented Nov 15, 2022 •

edited

Loading

j-baker commented Nov 16, 2022 •

edited by bridgekeeper bot

Loading

j-baker commented Nov 16, 2022 •

edited by bridgekeeper bot

Loading

openjdk bot commented Dec 1, 2022 •

edited

Loading