[9] RFR(L) 8158168: SIGSEGV: CollectedHeap::fill_with_objects(HeapWord*, unsigned long, bool)+0xa8 (original) (raw)

dean.long at oracle.com dean.long at oracle.com
Thu Mar 23 18:25:47 UTC 2017

On 3/22/17 1:49 PM, Vladimir Ivanov wrote:

Also, it looks like the changes I made to ASB.appendChars(char[] s, int off, int end) are not needed. Agree.

Vladimir, don't you need to replace checkIndex with checkOffset in indexOf and lastIndexOf, so that we allow count == length? Yes, my bad. Good catch. Updated webrev in place. FTR I haven't done any extensive testing of the minimized fix. If we agree to proceed with it, the regression test should be updated as well. I think the viable solution would be to construct broken SBs (using reflection) and invoke affected methods on them.

We can construct broken SBs using the Helper class that gets patched into java.lang. I'll work on that.

dl

Best regards, Vladimir Ivanov

On 3/22/17 8:35 AM, Vladimir Ivanov wrote:

So are we convinced that the proposed changes will never lead to a crash due to a missing or incorrect bounds check, due to a racy use of an unsynchronized ASB instance e.g. StringBuilder?

If only we had a static analysis tool that could tell us if the code is safe. Because we don't, in my initial changeset, we always take a snapshot of the ASB fields by passing those field values to StringUTF16 before doing checks on them. And I wrote a test to make sure that those StringUTF16 interfaces are catching all the underflows and overflows I could imagine, and I added verification code to detect when a check was missed. However, all the reviewers have requested to minimize the amount of changes. In Vladimir's version, if there is a missing check somewhere, then yes it could lead to a crash. I'd like to point out that asserts and verification code are disabled by default. They are invaluable during problem diagnosis, but don't help at all from defence-in-depth perspective. But I agree that it's easier to reason about and test the initial version of the fix. I wonder if the reviewers have fully realized the potential impact here? This has exposed a flaw in the way intrinsics are used from core classes. FTR here are the checks I omitted in the minimized version (modulo separation of indexOf/lastIndexOf for trusted/non-trusted callers): http://cr.openjdk.java.net/~vlivanov/dlong/8158168/redundantchecks/ Other than that, the difference is mainly about undoing refactorings and removing verification logic (asserts + checks in the JVM). There are still unsafe accesses which are considered safe in both versions (see StringUTF16.Trusted usages in the initial version [1]). We used to provide safe wrappers for unsafe intrinsics which makes it much easier to reason about code correctness. I'd like to see compact string code refactored that way and IMO the initial version by Dean is a big step in the right direction. I still prefer to see a point fix in 9 and major refactoring happening in 10, but I'll leave the decision on how to proceed with the fix to core-libs folks. After finishing the exercise minimizing the fix, I'm much more comfortable with the initial fix [1] (though there are changes I consider excessive). Best regards, Vladimir Ivanov [1] http://cr.openjdk.java.net/~dlong/8158168/webrev.0

Some clarifications:

============ src/java.base/share/classes/java/lang/String.java: The bounds check is needed only in String.nonSyncContentEquals when it extracts info from AbstractStringBuilder. I don't see how out of bounds access can happen in String.contentEquals: if (n != length()) { return false; } ... for (int i = 0; i < n; i++) { if (StringUTF16.getChar(val, i) != cs.charAt(i)) {

OK. ============ src/java.base/share/classes/java/lang/StringConcatHelper.java: I think bounds checks in StringConcatHelper.prepend() are skipped intentionally, since java.lang.invoke.StringConcatFactory constructs method handle chains which already contain bounds checks: array length is precomputed based on argument values and all accesses are guaranteed to be in bounds. This is calling the trusted version of getChars() with no bounds checks. It was a little more obvious when I had the Trusted inner class. ============ src/java.base/share/classes/java/lang/StringUTF16.java: + static void putChar(byte[] val, int index, int c) { + assert index >= 0 && index < length(val) : "Trusted_ _caller_ _missed bounds check";_ _Unfortunately, asserts can affect inlining decisions (since they_ _increase bytecode size). In order to minimize possible_ _performance_ _impact, I suggest to remove them from the fix targeting 9._ _Sure._ _============_ _private static int indexOfSupplementary(byte[] value, int_ _ch, int_ _fromIndex, int max) {_ _if (Character.isValidCodePoint(ch)) {_ _final char hi = Character.highSurrogate(ch);_ _final char lo = Character.lowSurrogate(ch);_ _+ checkBoundsBeginEnd(fromIndex, max, value);_ _The check is redundant here. fromIndex & max are always_ _inbounds by_ _construction:_ _public static int indexOf(byte[] value, int ch, int_ _fromIndex) {_ _int max = value.length >> 1; if (fromIndex < 0) {_ _fromIndex = 0;_ _} else if (fromIndex >= max) { // Note: fromIndex might be near -1>>>1. return -1; } ... return indexOfSupplementary(value, ch, fromIndex, max); OK. ============ I moved bounds checks from StringUTF16.lastIndexOf/indexOf to ABS.indexOf/lastIndexOf. I think it's enough to do range check on ABS.value & ABS.count. After that, all accesses should be inbounds by construction (in String.indexOf/lastIndexOf): jdk/src/java.base/share/classes/java/lang/StringUTF16.java: static int lastIndexOf(byte[] src, byte srcCoder, int srcCount, String tgtStr, int fromIndex) { int rightIndex = srcCount - tgtCount; if (fromIndex > rightIndex) { fromIndex = rightIndex; } if (fromIndex < 0) {_ _return -1;_ _}_ _jdk/src/java.base/share/classes/java/lang/StringUTF16.java:_ _public static int lastIndexOf(byte[] src, int srcCount,_ _byte[] tgt, int tgtCount, int_ _fromIndex) {_ _int min = tgtCount - 1;_ _int i = min + fromIndex;_ _int strLastIndex = tgtCount - 1;_ _char strLastChar = getChar(tgt, strLastIndex);_ _startSearchForLastChar:_ _while (true) {_ _while (i >= min && getChar(src, i) != strLastChar) { There are 2 places: * getChar(tgt, strLastIndex) => getChar(tgt, tgtCount-1) - inbound * getChar(src, i); i in [ min; min+fromIndex ] min = tgtCount - 1 rightIndex = srcCount - tgtCount fromIndex <= rightIndex_ _0 <= min + fromIndex <= min + rightIndex == (tgtCount_ _- 1)_ _+ (srcCount - tgtCount) == srcCount - 1_ _Hence, should be covered by the check on count & value:_ _public int lastIndexOf(String str, int fromIndex) {_ _+ byte[] value = this.value;_ _+ int count = this.count;_ _+ byte coder = this.coder;_ _+ checkIndex(count, value.length >> coder); return String.lastIndexOf(value, coder, count, str, fromIndex); } OK, I will go with your version if it's OK with Sherman. dl Best regards, Vladimir Ivanov On 3/17/17 5:58 AM, Vladimir Ivanov wrote: I have the same concern. Can we fix the immediate problem in 9 and integrate verification logic in 10?

OK, Tobias is suggesting having verification logic only inside the intrinsics. Are you suggesting removing that as well? Yes and put them back in 10. I'm OK with removing all the verification, but that won't reduce the library changes much. I could undo the renaming to Trusted.getChar, but we would still have the bounds checks moved into StringUTF16. I suggest to go with a point fix for 9: just add missing range checks.

More information about the hotspot-dev mailing list