diff --git a/usr.sbin/bhyve/pci_emul.c b/usr.sbin/bhyve/pci_emul.c
--- a/usr.sbin/bhyve/pci_emul.c
+++ b/usr.sbin/bhyve/pci_emul.c
@@ -106,6 +106,16 @@
 static uint64_t pci_emul_membase64;
 static uint64_t pci_emul_memlim64;
 
+struct pci_bar_allocation {
+	TAILQ_ENTRY(pci_bar_allocation) chain;
+	struct pci_devinst *pdi;
+	int idx;
+	enum pcibar_type type;
+	uint64_t size;
+};
+TAILQ_HEAD(pci_bar_list, pci_bar_allocation) pci_bars = TAILQ_HEAD_INITIALIZER(
+    pci_bars);
+
 #define	PCI_EMUL_IOBASE		0x2000
 #define	PCI_EMUL_IOLIMIT	0x10000
 
@@ -634,10 +644,6 @@
 pci_emul_alloc_bar(struct pci_devinst *pdi, int idx, enum pcibar_type type,
     uint64_t size)
 {
-	int error;
-	uint64_t *baseptr, limit, addr, mask, lobits, bar;
-	uint16_t cmd, enbit;
-
 	assert(idx >= 0 && idx <= PCI_BARMAX);
 
 	if ((size & (size - 1)) != 0)
@@ -652,6 +658,62 @@
 			size = 16;
 	}
 
+	/*
+	 * To reduce fragmentation of the MMIO space, we wanna allocate the BARs
+	 * by size. Therefore, don't allocate the BAR yet. We create a list of
+	 * all BAR allocation which is sorted by BAR size. When all PCI devices
+	 * are initialized, we will assign an address to the BARs.
+	 */
+
+	/* create a new list entry */
+	struct pci_bar_allocation *const new_bar = malloc(
+	    sizeof(struct pci_bar_allocation));
+	memset(new_bar, 0, sizeof(struct pci_bar_allocation));
+	new_bar->pdi = pdi;
+	new_bar->idx = idx;
+	new_bar->type = type;
+	new_bar->size = size;
+
+	/*
+	 * Search for a BAR which size is lower than the size of our newly
+	 * allocated BAR.
+	 */
+	struct pci_bar_allocation *bar = NULL;
+	TAILQ_FOREACH(bar, &pci_bars, chain) {
+		if (bar->size < size) {
+			break;
+		}
+	}
+
+	if (bar == NULL) {
+		/*
+		 * Either the list is empty or new BAR is the smallest BAR of
+		 * the list. Append it to the end of our list.
+		 */
+		TAILQ_INSERT_TAIL(&pci_bars, new_bar, chain);
+	} else {
+		/*
+		 * The found BAR is smaller than our new BAR. For that reason,
+		 * insert our new BAR before the found BAR.
+		 */
+		TAILQ_INSERT_BEFORE(bar, new_bar, chain);
+	}
+
+	return (0);
+}
+
+static int
+pci_emul_assign_bar(const struct pci_bar_allocation *const pci_bar)
+{
+	struct pci_devinst *const pdi = pci_bar->pdi;
+	const int idx = pci_bar->idx;
+	const enum pcibar_type type = pci_bar->type;
+	const uint64_t size = pci_bar->size;
+
+	int error;
+	uint64_t *baseptr, limit, addr, mask, lobits, bar;
+	uint16_t cmd, enbit;
+
 	switch (type) {
 	case PCIBAR_NONE:
 		baseptr = NULL;
@@ -1191,6 +1253,7 @@
 		bi->membase32 = pci_emul_membase32;
 		bi->membase64 = pci_emul_membase64;
 
+		/* first run: init devices */
 		for (slot = 0; slot < MAXSLOTS; slot++) {
 			si = &bi->slotinfo[slot];
 			for (func = 0; func < MAXFUNCS; func++) {
@@ -1230,6 +1293,15 @@
 			}
 		}
 
+		/* second run: assign BARs and free list */
+		struct pci_bar_allocation *bar;
+		struct pci_bar_allocation *bar_tmp;
+		TAILQ_FOREACH_SAFE(bar, &pci_bars, chain, bar_tmp) {
+			pci_emul_assign_bar(bar);
+			TAILQ_REMOVE(&pci_bars, bar, chain);
+			free(bar);
+		}
+
 		/*
 		 * Add some slop to the I/O and memory resources decoded by
 		 * this bus to give a guest some flexibility if it wants to