Le Thursday 25 Jul 2013 à 12:11:15 (+0200), Andrea Arcangeli a écrit : > MADV_DONTFORK prevents fork to fail with -ENOMEM if the default > overcommit heuristics decides there's too much anonymous virtual > memory allocated. If the KVM secondary MMU is synchronized with MMU > notifiers or not, doesn't make a difference in that regard. > > Secondly it's always more efficient to avoid copying the guest > physical address space in the fork child (so we avoid to mark all the > guest memory readonly in the parent and so we skip the establishment > and teardown of lots of pagetables in the child). > > In the common case we can ignore the error if MADV_DONTFORK is not > available. Leave a second invocation that errors out in the KVM path > if MMU notifiers are missing and KVM is enabled, to abort in such > case. > > Signed-off-by: Andrea Arcangeli <aarca...@redhat.com> > --- > exec.c | 1 + > 1 file changed, 1 insertion(+) > > diff --git a/exec.c b/exec.c > index c99a883..d3bb58d 100644 > --- a/exec.c > +++ b/exec.c > @@ -1162,6 +1162,7 @@ ram_addr_t qemu_ram_alloc_from_ptr(ram_addr_t size, > void *host, > > qemu_ram_setup_dump(new_block->host, size); > qemu_madvise(new_block->host, size, QEMU_MADV_HUGEPAGE); > + qemu_madvise(new_block->host, size, QEMU_MADV_DONTFORK); > > if (kvm_enabled()) > kvm_setup_guest_memory(new_block->host, size); >
This patch solve a bug where the network down script of a regular tap interface is not executed because fork fail when a pci-assigment is done on a large guest. Tested-By: Benoit Canet <ben...@irqsave.net>