Messages in this thread | | | Date | Thu, 11 Jun 2009 22:22:39 +0800 | From | Wu Fengguang <> | Subject | [PATCH 0/5] [RFC] HWPOISON incremental fixes |
| |
Hi all,
Here are the hwpoison fixes that aims to address Nick and Hugh's concerns. Note that - the early kill option is dropped for .31. It's obscure option and complex code and is not must have for .31. Maybe Andi also aims this option for notifying KVM, but right now KVM is not ready to handle that. - It seems that even fsync() processes are not easy to catch, so I abandoned the SIGKILL on fsync() idea. Instead, I choose to fail any attempt to populate the poisoned file with new pages, so that the corrupted page offset won't be repopulated with outdated data. This seems to be a safe way to allow the process to continue running while still be able to promise good (but not complete) data consistency. - I didn't implement the PANIC-on-corrupted-data option. Instead, I guess sending uevent notification to user space will be a more flexible scheme?
Thanks, Fengguang --
| |