[Ran-Mewo]

Growing a RawMemoryFreeList up to its limit panics. raise_high_water subtracted the wrong way round when clamping to the limit, and current_capacity ignored the last block when the limit cut it short. Easy to hit with a small fixed heap, e.g. 24 MB.

[wks]

It looks like the code in the Rust MMTk core is ported from this piece of code in JikesRVM:

  /**
   * Raise the high water mark by requesting more pages from the OS
   *
   * @param blocks block count to raise by
   */
  private void raiseHighWater(int blocks) {
    Extent growExtent = Conversions.pagesToBytes(pagesPerBlock * blocks);
    if (highWater.EQ(limit)) {
      Log.write("limit=", limit);
      Log.write(", highWater=", highWater);
      Log.writeln(", growExtent=", growExtent);
      VM.assertions.fail("Attempt to grow FreeList beyond limit");
    }
    if (highWater.plus(growExtent).GT(limit)) {
      growExtent = highWater.diff(limit).toWord().toExtent();
    }
    mmap(highWater, growExtent);
    highWater = highWater.plus(growExtent);
  }

highWater.diff(limit) means highWater - limit. The highWater should not be higher than than limit, and the result of the diff shall be negative. It is a bug in JikesRVM, and I am surprised it has existed since 2016 and never gets found.

[k-sareen]

Looks like this PR and #1625 are both trying to fix the same issue? Or at least very similar issues

[qinsoon]

Looks like this PR and #1625 are both trying to fix the same issue? Or at least very similar issues

This PR seems to be a subset of #1625 with one additional test.