RFR: 8340184: Bug in CompressedKlassPointers::is_in_encodable_range

[Thomas Stuefe]

On Mon, 16 Sep 2024 10:37:55 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

> Since JDK-8338526, we keep only those Klass in the encoding range that need a narrowKlass id.
> 
> To check whether a Klass has a narrowKlass id, we call `CompressedKlassPointers::is_in_encodable_range()`. There is a small bug that results from the confusion around "encoding range" vs "klass range".
> 
> The **"Encoding Range"** is the range that can be encoded with the current encoding base, encoding shift, and (implicitly) the bit size of the narrowKlass. The encoding range reaches from `[ <encoding base> ... <encoding base> + 1 << (32 + <encoding shift>) ).`. Its size is either 4G (shift=0) or 32G (shift=3).
> 
> The **"Klass Range"** is the range that actually holds Klass structures. It is part of the Encoding Range, but usually much smaller:
> - with zero-based encoding, the encoding base is zero, so it precedes the start of the Klass range
> - The encoding range can reach far beyond the end of the Klass range.
> 
> For a more detailed explanation, including pleasing ASCII art, please refer to `compressedKlass.hpp` in this patch.
> 
> ----
> 
> The error in this case, introduced with 8338526, was that we use the range `[<encoding base> ... <klass range end>)` for `is_in_encodable_range()`. That can lead to false positives since `<encoding base>` can be zero. In a highly contrived theoretical case, we could mis-classify a Klass as being encodable if it lives in metaspace outside class space, but its metaspace region happens to be located below the class space start.
> 
> The error is extremely unlikely because:
> - non-class Metaspace regions - freely allocated by mmap - typically live in high-address ranges
> - we only allow zero-based encoding with Xshare:off, itself a very unusual setting nowadays
> 
> But the error highlights misuse of the term "encoding range", so it should be fixed.
> 
> ----
> 
> Note that most of the patch has been selectively copied from Lilliput. Lilliput, with its non-22-bit narrowKlass, had needed to straighten out this code a while ago, and it is better for it.
> 
> The fix:
> - we now use the actual *Klass Range* for the "is encodable" check. We don't use the encoding base. That is the real fix.
> - To do that, we need to keep track of the Klass Range inside `CompressedKlassPointers`. That is simple, since we are given this range during initialization.
> 
> Minor cleanups:
> - I also removed the confusingly named `CompressedKlassPointers::range()`. This was a strange animal ( the distance between *klass range end* and *encoding range start* ) and was only used in a single place in t...

src/hotspot/cpu/aarch64/macroAssembler_aarch64.cpp line 5084:

> 5082: 
> 5083:   if (operand_valid_for_logical_immediate(
> 5084:         /*is32*/false, (uint64_t)CompressedKlassPointers::base())) {

Reviewer hint:

The before and after code should be exactly identical. What we do here is we calculate the max. value a left-shifted nKlass can have, and compute a mask over it.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/21015#discussion_r1761042543