diff --git a/sys/kern/vfs_subr.c b/sys/kern/vfs_subr.c
--- a/sys/kern/vfs_subr.c
+++ b/sys/kern/vfs_subr.c
@@ -766,18 +766,24 @@
 	 * Not counting the 'maxproc' contribution, the (marginal) growth factor
 	 * of 'physvnodes' with respect to the physical memory size in MB is 64
 	 * until 'physvnodes' exceeds 98,304 (still not counting the 'maxproc'
-	 * contribution).  Thereafter, the marginal factor goes down to 16.
+	 * contribution).  Thereafter, the marginal factor goes down to 32.  32
+	 * has been chosen as it ensures that there are slightly more vnodes
+	 * than the default 'maxfiles' (whose marginal growth is 32, while here
+	 * it is slightly higher than 32 because of the 'maxproc' contribution,
+	 * but less than 33 because 'vm_cnt.v_page_count' is smaller than
+	 * 'physpages' used to compute 'maxfiles'), which avoids a deadlock
+	 * situation we do not currently handle.
 	 *
 	 * Separately, the memory required by vnodes and VM objects must not
 	 * exceed 1/10th of the kernel's heap size.  This last limit is computed
 	 * in 'virtvnodes'.
 	 *
 	 * With the current formulae, on 64-bit platforms, 'desiredvnodes' is
-	 * 'virtvnodes' up to a physical memory cutoff of ~1674MB, and then
+	 * 'virtvnodes' up to a physical memory cutoff of ~1722MB, and then
 	 * 'physvnodes' applies instead.
 	 */
-	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
-	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
+	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 32 +
+	    min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 32;
 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
 	desiredvnodes = min(physvnodes, virtvnodes);