mirror of
https://github.com/reactos/reactos.git
synced 2024-07-12 23:55:07 +00:00
7b28343c16
- The legacy ROS Mm uses structures known as page ops to track operations on a page such as page out, page in, and access fault. The idea is that each operation is forced to wait until the page has completed all other pending operations to start its work. The problem was that the page op's completion event was a NotificationEvent instead of a SynchronizationEvent. This caused all operations to proceed at the same time if they were waiting on a single page op to begin. Now that is fixed and page ops proceed one after another as intended. This bug has been around since r9077. - When waiting for sections to be unmapped or destroyed, the Mm waits for pending page operations to complete. The problem is that MmUnmapViewOfSection had a critical bug in which it forgot to dereference the page op it just retrieved. This caused zombie page ops to be stuck to that particular address if there were any pending operations at the time of the MmUnmapViewOfSection call. As a result, section destruction to bug check due to the hung page op after waiting 10 seconds for the operation to complete. This bug has been around since r18849. - Due to the combined effects of the above bugs, sometimes unmapping or freeing a section would hang the system or bug check the system (remember the "Failed to wait for page op" messages?). This was evident in smiley_'s theme work when many applications calling FreeLibrary at the same time would trigger this bug. There are likely many more cases that triggered this series of events which either could never be replicated or simply could not be explained. svn path=/trunk/; revision=56284 |
||
|---|---|---|
| reactos | ||
| rosapps | ||
| rostests | ||
| wallpaper | ||