On 05/30/2013 11:00 PM, Paolo Bonzini wrote: > Il 30/05/2013 11:08, Peter Lieven ha scritto: >> Am 30.05.2013 10:18, schrieb Alexey Kardashevskiy: >>> On 05/30/2013 05:49 PM, Paolo Bonzini wrote: >>>> Il 30/05/2013 09:44, Alexey Kardashevskiy ha scritto: >>>>> Hi! >>>>> >>>>> I found the migration broken on pseries platform, specifically, this patch >>>>> broke it: >>>>> >>>>> f1c72795af573b24a7da5eb52375c9aba8a37972 >>>>> migration: do not sent zero pages in bulk stage >>>>> >>>>> The idea is not to send zero pages to the destination guest which is >>>>> expected to have 100% empty RAM. >>>>> >>>>> However on pseries plaftorm the guest always has some stuff in the RAM as >>>>> a >>>>> part of initialization (device tree, system firmware and rtas (?)) so it >>>>> is >>>>> not completely empty. As the source guest cannot detect this, it skips >>>>> some >>>>> pages during migration and we get a broken destination guest. Bug. >>>>> >>>>> While the idea is ok in general, I do not see any easy way to fix it as >>>>> neither QEMUMachine::init nor QEMUMachine::reset callbacks has information >>>>> about whether we are about to receive a migration or not (-incoming >>>>> parameter) and we cannot move device-tree and system firmware >>>>> initialization anywhere else. >>>>> >>>>> ram_bulk_stage is static and cannot be disabled from the platform >>>>> initialization code. >>>>> >>>>> So what would the community suggest? >>>> Revert the patch. :) >>> I'll wait for 24 hours (forgot to cc: the author) and then post a revert >>> patch :) >>> >> does this problem only occur on pseries emulation? > > Probably not. On a PC, it would occur if you had 4K of zeros in the > source BIOS but not in the destination BIOS. When you reboot, the BIOS > image is wrong. > >> not sending zero pages is not only a performance benefit it also makes >> overcomitted memory usable. the madv_dontneed seems to kick in asynchronously >> and memory is not available immediately. > > You could also scan the page for nonzero values before writing it.
Scan where? On the source? It is there already. On the destination? Won't just trying to read a page kill all effect from MADV_DONTNEED? -- Alexey