On 13/07/2012 21:20, "Olaf Hering" <olaf at aepfle.de> wrote: > On Tue, Jul 10, Keir Fraser wrote: > >> On 10/07/2012 19:09, "Olaf Hering" <olaf at aepfle.de> wrote: >>> I'm not sure, most likely the gfn will just disappear from the guest, >>> like a ballooned page disappears. Accessing it will likely cause a >>> crash. >> >> Best thing to do, is possible, is map the shared-info page in the >> xen-platform pci device's BAR memory range. Then it will not conflict with >> any RAM. >> >> If you do map it over the top of an existing RAM page, you will have to >> repopulate that RAM page before kexec, using populate_physmap hypercall. The >> good news is that the populate_physmap hypercall will have the side effect >> of unmapping the shared-info page, reayd to be mapped wherever the new >> kernel would like it to reside :) > > Keir, > > is this a safe thing to do in a SMP guest? > If arch/x86/xen/enlighten.c:xen_hvm_init_shared_info() allocates a page > (backed by mfn M and pfn A) and assigns *HYPERVISOR_shared_info and > *xen_vcpu then everything will reference these pointers. So pfn A now points at shared_info, and mfn M is lost (freed back to Xen). Xen_vcpu doesn't come into it, you'd have that mapped at yet another pfn. > If drivers/xen/platform-pci.c:platform_pci_init would also do a > XENMAPSPACE_shared_info call with pfn B, isnt there a small window where > pfn A is not backed by a mfn because mfn M is now connected to pfn C? As > a result other code paths which access *HYPERVISOR_shared_info and > *xen_vcpu between the hypercall and the update of the pointers will read > 0xff. Don't really understand this. After the XENMAPSPACE_shared_info_call: * PFN B points at shared_info, mfn M_B it previously mapped is lost (freed back to Xen). * PFN A maps nothing, reads return all-1s. Yes, obviously you can't atomically update the mapping of shinfo from A->B, ad update your pointer in the kernel at exactly the same time. Presumably you do this early during boot, or late during kexec, or otherwise at a time when other processors are not expected to touch shinfo. > > If I read the hypercall code of XENMEM_add_to_physmap correctly the mfn > backing *HYPERVISOR_shared_info will remain the same, so there is no need > to copy data from the old to the new *HYPERVISOR_shared_info. That is correct. > What do you think, is that race real? I suppose it is. I didn't imagine it would be a troublesome one though. -- Keir > Olaf
