CCed KVM guys. On 05/10/2013 01:11 PM, Stefan Hajnoczi wrote: > On Fri, May 10, 2013 at 4:28 AM, wenchao <wenchaolinux@xxxxxxxxx> wrote: >> 于 2013-5-9 22:13, Mel Gorman 写道: >> >>> On Thu, May 09, 2013 at 05:50:05PM +0800, wenchaolinux@xxxxxxxxx wrote: >>>> >>>> From: Wenchao Xia <wenchaolinux@xxxxxxxxx> >>>> >>>> This serial try to enable mremap syscall to cow some private memory >>>> region, >>>> just like what fork() did. As a result, user space application would got >>>> a >>>> mirror of those region, and it can be used as a snapshot for further >>>> processing. >>>> >>> >>> What not just fork()? Even if the application was threaded it should be >>> managable to handle fork just for processing the private memory region >>> in question. I'm having trouble figuring out what sort of application >>> would require an interface like this. >>> >> It have some troubles: parent - child communication, sometimes >> page copy. >> I'd like to snapshot qemu guest's RAM, currently solution is: >> 1) fork() >> 2) pipe guest RAM data from child to parent. >> 3) parent write down the contents. >> >> To avoid complex communication for data control, and file content >> protecting, So let parent instead of child handling the data with >> a pipe, but this brings additional copy(). I think an explicit API >> cow mapping an memory region inside one process, could avoid it, >> and faster and cow less pages, also make user space code nicer. > > A new Linux-specific API is not portable and not available on existing > hosts. Since QEMU supports non-Linux host operating systems the > fork() approach is preferable. > > If you're worried about the memory copy - which should be benchmarked > - then vmsplice(2) can be used in the child process and splice(2) can > be used in the parent. It probably doesn't help though since QEMU > scans RAM pages to find all-zero pages before sending them over the > socket, and at that point the memory copy might not make much > difference. > > Perhaps other applications can use this new flag better, but for QEMU > I think fork()'s portability is more important than the convenience of > accessing the CoW pages in the same process. Yup, I agree with you that the new syscall sometimes is not a good solution. Currently, we're working on live-update[1] that will be enabled on Qemu firstly, this feature let the guest run on the new Qemu binary smoothly without restart, it's good for us to do security-update. In this case, we need to move the guest memory on old qemu instance to the new one, fork() can not help because we need to exec() a new instance, after that all memory mapping will be destroyed. We tried to enable SPLICE_F_MOVE[2] for vmsplice() to move the memory without memory-copy but the performance isn't so good as we expected: it's due to some limitations: the page-size, lock, message-size limitation on pipe, etc. Of course, we will continue to improve this, but wenchao's patch seems a new direction for us. To coordinate with your fork() approach, maybe we can introduce a new flag for VMA, something like: VM_KEEP_ONEXEC, to tell exec() to do not destroy this VMA. How about this or you guy have new idea? Really appreciate for your suggestion. [1] http://marc.info/?l=qemu-devel&m=138597598700844&w=2 [2] https://lkml.org/lkml/2013/10/25/285 -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
