Quite some time ago there was a thread on qemu-devel, started by Andrea, about modifying qemu to better use transparent huge pages:
http://lists.gnu.org/archive/html/qemu-devel/2010-03/msg01250.html That thread hasn't reached any conclusion, but some time after that Avi implemented a similar change: commit 36b586284e678da28df3af9fd0907d2b16f9311c Author: Avi Kivity <a...@redhat.com> Date: Mon Sep 5 11:07:05 2011 +0300 qemu_vmalloc: align properly for transparent hugepages and KVM To make good use of transparent hugepages, KVM requires that guest-physical and host-virtual addresses share the low 21 bits (as opposed to just the low 12 bits normally required). Adjust qemu_vmalloc() to honor that requirement. Ignore it for small region to avoid fragmentation. Signed-off-by: Avi Kivity <a...@redhat.com> Signed-off-by: Anthony Liguori <aligu...@us.ibm.com> diff --git a/oslib-posix.c b/oslib-posix.c index 196099c..a304fb0 100644 --- a/oslib-posix.c +++ b/oslib-posix.c @@ -35,6 +35,13 @@ extern int daemon(int, int); #endif +#if defined(__linux__) && defined(__x86_64__) + /* Use 2MB alignment so transparent hugepages can be used by KVM */ +# define QEMU_VMALLOC_ALIGN (512 * 4096) +#else +# define QEMU_VMALLOC_ALIGN getpagesize() +#endif + #include "config-host.h" #include "sysemu.h" #include "trace.h" @@ -80,7 +87,12 @@ void *qemu_memalign(size_t alignment, size_t size) void *qemu_vmalloc(size_t size) { void *ptr; - ptr = qemu_memalign(getpagesize(), size); + size_t align = QEMU_VMALLOC_ALIGN; + + if (size < align) { + align = getpagesize(); + } + ptr = qemu_memalign(align, size); trace_qemu_vmalloc(size, ptr); return ptr; } (why it is 64bit-only is a different, unrelated question). But apparently, THP does not work still, even with 2Mb alignment: when running a guest, AnonHugePages in /proc/meminfo stays at 0 - either in kvm mode or in tcg mode. Any idea why? What else is needed for THP to work? This is quite a frequent question in #kvm IRC channel, and I always suggested using -mem-path for this, but I'm curios why it doesn't work automatically when it probably should? Thanks, /mjt
