memory: batch allocate ioeventfds[] in address_space_update_ioeventfds()
Reallocing the ioeventfds[] array each time an element is added is very
expensive as the number of ioeventfds increases. Batch allocate instead
to amortize the cost of realloc.
This patch reduces Linux guest boot times from 362s to 140s when there
are 2 virtio-blk devices with 1 virtqueue and 99 virtio-blk devices with
32 virtqueues.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20200218182226.913977-1-stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
diff --git a/memory.c b/memory.c
index aeaa8dc..09be40e 100644
--- a/memory.c
+++ b/memory.c
@@ -794,10 +794,19 @@
FlatView *view;
FlatRange *fr;
unsigned ioeventfd_nb = 0;
- MemoryRegionIoeventfd *ioeventfds = NULL;
+ unsigned ioeventfd_max;
+ MemoryRegionIoeventfd *ioeventfds;
AddrRange tmp;
unsigned i;
+ /*
+ * It is likely that the number of ioeventfds hasn't changed much, so use
+ * the previous size as the starting value, with some headroom to avoid
+ * gratuitous reallocations.
+ */
+ ioeventfd_max = QEMU_ALIGN_UP(as->ioeventfd_nb, 4);
+ ioeventfds = g_new(MemoryRegionIoeventfd, ioeventfd_max);
+
view = address_space_get_flatview(as);
FOR_EACH_FLAT_RANGE(fr, view) {
for (i = 0; i < fr->mr->ioeventfd_nb; ++i) {
@@ -806,8 +815,11 @@
int128_make64(fr->offset_in_region)));
if (addrrange_intersects(fr->addr, tmp)) {
++ioeventfd_nb;
- ioeventfds = g_realloc(ioeventfds,
- ioeventfd_nb * sizeof(*ioeventfds));
+ if (ioeventfd_nb > ioeventfd_max) {
+ ioeventfd_max = MAX(ioeventfd_max * 2, 4);
+ ioeventfds = g_realloc(ioeventfds,
+ ioeventfd_max * sizeof(*ioeventfds));
+ }
ioeventfds[ioeventfd_nb-1] = fr->mr->ioeventfds[i];
ioeventfds[ioeventfd_nb-1].addr = tmp;
}
