| Title | Assertion failure in amssshe |
| Status | closed |
| Priority | essential |
| Assigned user | Gareth Rees |
| Organization | Ravenbrook |
| Description | See for example [1]: amssshe failed: log follows -------------------------------------------------------------------------------- lii6gc/cool/amssshe: randomize(): choosing initial state (v3): 303772349. ........trace.c:417: MPS ASSERTION FAILED: !TraceSetIsMember(SegWhite(seg), trace) |
| Analysis | The issue is that AMSWhiten adds the segment to the white set for the trace, even if there were no objects in the segment and so nothing was added to trace->condemned. If this is the case for all AMS segments in the condemned set, then trace->condemned will be zero, and so TraceIsEmpty() will return TRUE, and so TraceCondemnZones or traceCondemnAll will return ResFAIL, and the trace will be destroyed via TraceDestroyInit without going through reclaim. But this leaves the AMS segments white even though the trace is finished. The same issue is present in AWL and LO. AMC is safe because if the segment is unbuffered then the whole segment is accounted as condemned; if the segment is buffered and there is nothing in the buffer then the segment is not condemned; otherwise there is something in the buffer and the condemned bytes are given by SegSize(seg) - AddrOffset(BufferScanLimit(buffer), BufferLimit(buffer)). AMS and AWL were accidentally fixed by RB in change 190607. LO was accidentally fixed by GDR in change 192195. |
| How found | automated_test |
| Evidence | [1] https://travis-ci.org/Ravenbrook/mps/jobs/400813949 |
| Created by | Gareth Rees |
| Created on | 2018-07-06 13:08:40 |
| Last modified by | Gareth Rees |
| Last modified on | 2018-07-06 14:26:54 |
| History | 2018-07-06 GDR Created. |
| Change | Effect | Date | User | Description |
|---|---|---|---|---|
| 194475 | closed | 2018-07-06 14:26:54 | Gareth Rees | Check that traces with no condemned objects can be started and finished without error in each automatic pool class. Regression test for job004086. |
| 192195 | open | 2016-09-05 16:23:15 | Gareth Rees | Merge branch/2016-04-13/mortality. |
| 190607 | open | 2016-03-31 23:43:21 | Richard Brooksby | Don't allow pools to whiten segments without condemning objects, so that a condemned size of zero implies no white segments, allowing quick trace destruction. Imported from Git Author: Richard Brooksby <rb@ravenbrook.com> 1459464106 +0100 Committer: Richard Brooksby <rb@ravenbrook.com> 1459464106 +0100 sha1: f8f87f053ba174f9fd2bcf58831312fbe7db145d push-state: complete parent-changes: 0c4e6bef63ad2da89e93de328f1b824cf6a87912=[190579] |