Hi Peter, Is it that sending zero page mostly service the first iteration, ie bluk-stage? And for the subsequent iteration, dirty pages are normally not zero.
Thanks On Wed, Jun 5, 2013 at 2:09 PM, Peter Lieven <p...@kamp.de> wrote: > > Am 05.06.2013 um 05:37 schrieb Alexey Kardashevskiy <a...@ozlabs.ru>: > >> On 06/05/2013 05:15 AM, Peter Lieven wrote: >>> >>> Am 04.06.2013 um 17:17 schrieb Paolo Bonzini <pbonz...@redhat.com>: >>> >>>> Il 04/06/2013 16:48, Peter Lieven ha scritto: >>>>> Still using 3.2, but strange enough the above example is also not >>>>> increasing RSS size for me. >>>>> >>>>> Can you try the following: >>>>> qemu git master with 1G of memory (hanging in bios with no boot device) >>>>> and migrate it. Before migration RSS Size os somewhat >>>>> around 16MB. After migration its RSS size is in the order of 1G. >>>> >>>> That may be a kernel bug. The kernel did not do the copy-on-write trick >>>> on huge zero pages. It was fixed last year, maybe 3.2 is not enough. >>>> Try adding a MADV_HUGEPAGE madvise to the testcase and see if it >>>> reproduces. >>> >>> that's it. thanks for the pointer. the huge zero page was introduced in 3.8. >>> >>> paolo, alexey: can you please verify the following works for you: >>> https://github.com/plieven/qemu/tree/fix-migration >> >> These two? >> 848b796 Tue Jun 4 14:43:04 2013 +0200 Peter Lieven migration: do not >> overwrite zero pages >> 2206ac8 Tue Jun 4 14:25:33 2013 +0200 Peter Lieven Revert "migration: do >> not sent zero pages in bulk stage" > > Yes, sorry forgot to mention this. > >> >> That works for me (qemu 1.5, kernel 3.9-rc2). >> Tested-by: Alexey Kardashevskiy <a...@ozlabs.ru> > > Thank you, > Peter
