Changeset 43235 in vbox for trunk/src/VBox/VMM/VMMR0

Timestamp:

Sep 6, 2012 11:53:40 PM (12 years ago)

Author:

vboxsync

Message:

GMMR0.cpp: Fixed bug in GMMR0CleanupVM/gmmR0CleanupVMScanChunk affecting bound mode on all 32-bit hosts + 64-bit darwin. Problem was caused by unncessary scanning of chunks bound to other VMs and accidentally relinking them into the set of the VM about to die. Once the GVM structure was finally fried, almost all pChunk->pSet members would point to the dead VMs GVM::gmm.s.Private member.

Also fixed a missing redo-from-start when someone else freed a chunk while we were scanning the list. Expecting this to only occure rarely, but should be reproducible when many VMs are doing cleanups at the same time in unbound mode.

File:

• trunk/src/VBox/VMM/VMMR0/GMMR0.cpp (modified) (4 diffs)

trunk/src/VBox/VMM/VMMR0/GMMR0.cpp

@@ -1215 +1215 @@
          * shared pages will be 'left behind'.)
          */
+        /** @todo r=bird: This scanning+freeing could be optimized in bound mode! */
         uint64_t    cPrivatePages = pGVM->gmm.s.Stats.cPrivatePages; /* save */
 
@@ -1226 +1227 @@
             {
                 uint32_t const cFreeChunksOld = pGMM->cFreedChunks;
-                if (gmmR0CleanupVMScanChunk(pGMM, pGVM, pChunk))
+                if (   (   !pGMM->fBoundMemoryMode
+                        || pChunk->hGVM == pGVM->hSelf)
+                    && gmmR0CleanupVMScanChunk(pGMM, pGVM, pChunk))
                 {
                     /* We left the giant mutex, so reset the yield counters. */
@@ -1241 +1244 @@
                 }
                 if (pGMM->cFreedChunks != cFreeChunksOld)
+                {
+                    fRedoFromStart = true;
                     break;
+                }
             }
         } while (fRedoFromStart);
@@ -1349 +1355 @@
 static bool gmmR0CleanupVMScanChunk(PGMM pGMM, PGVM pGVM, PGMMCHUNK pChunk)
 {
+    Assert(!pGMM->fBoundMemoryMode || pChunk->hGVM == pGVM->hSelf);
+
     /*
      * Look for pages belonging to the VM.