MPS issue job003486

Titlemv2test failure
Assigned userGareth Rees
DescriptionRunning mv2test produces the following assertion failure:

AnalysisThe backtrace looks like this:

    CBSEnter (cbs=0x103ffe1f0) at cbs.c:94
    CBSDelete (cbs=0x103ffe1f0, base=0x10409c000, limit=0x10409e000) at cbs.c:1095
    MVTReturnBlockSegs (mvt=0x103ffe160, block=0x103fef618, arena=0x1003a9000) at poolmv2.c:916
    MVTNoteNew (cbs=0x103ffe1f0, block=0x103ff2c08, oldSize=8312, newSize=16504) at poolmv2.c:953
    cbsBlockGrow (cbs=0x103ffe1f0, block=0x103ff2c08, oldSize=8312) at cbs.c:419
    cbsInsertIntoTree (baseReturn=0x7fff5fbfa900, limitReturn=0x7fff5fbfa8f8, cbs=0x103ffe1f0, base=0x101dfd000, limit=0x101dfe3a0) at cbs.c:523
    CBSInsertReturningRange (baseReturn=0x7fff5fbfa958, limitReturn=0x7fff5fbfa950, cbs=0x103ffe1f0, base=0x101dfd000, limit=0x101dfe3a0) at cbs.c:858
    CBSInsert (cbs=0x103ffe1f0, base=0x101dfd000, limit=0x101dfe3a0) at cbs.c:895
    MVTFree (pool=0x103ffe160, base=0x101dfd000, size=5024) at poolmv2.c:674
    PoolFree (pool=0x103ffe160, old=0x101dfd000, size=5021) at pool.c:331
    mps_free (pool=0x103ffe160, p=0x101dfd000, size=5021) at mpsi.c:706
    stress (class=0x1002c7ad8, arena=0x1003a9000, size=0x100002230 <randomSize>) at mv2test.c:261
    stress_with_arena_class (aclass=0x1002cb398) at mv2test.c:303
    main (argc=1, argv=0x7fff5fbffa90) at mv2test.c:322

CBS callbacks are only allowed to call back into the CBS in very limited ways, and the CBSEnter/CBSLeave protocol checks this. In particular, MVTNoteNew (which is a CBS callback) is wrong to call MVTReturnBlockSegs. Why does it call it? Because the ABQ is full.

The reason why CBSInsert and CBSDelete use callbacks (instead of returning the new or deleted block via a function parameters) is that each CBS maintains an "emergency list" of blocks which have been freed but for which it has not yet been able to allocate a CBSBlockStruct to refer to it (because of low memory). These emergency list blocks get merged back into the splay tree in an eager fashion, which might result in many calls to the callbacks.

It's not good for the CBS to have this emergency block structure. Better if that's implemented in a separate module, and available as a fallback strategy in any pool (for example MV might want to use it instead of just losing memory).

More detailed analysis and plan here [1].
How foundautomated_test
Evidence[1] <>
Observed in1.111.0
Created byGareth Rees
Created on2013-05-16 16:10:07
Last modified byGareth Rees
Last modified on2013-05-22 00:21:39
History2013-05-16 GDR Created.


Change Effect Date User Description
182032 closed 2013-05-21 13:35:41 Gareth Rees Refactor MVT so that blocks are pushed onto the ABQ in just one place.
Fix bug in MVTDelete: need to remove the whole of the old range from the ABQ, not just the range we're deleting.