pcib: Assume a window where both the base and limit are 0 is uninitialized

Since the low bits of a window's limit are hardwired to 1, this
configuration looks like a minimally sized window at address 0.
However, PCI resources are not generally at address 0 (see the
__PCI_BAR_ZERO_VALID macro that was only defined on sparc64), and some
PCI-PCI bridges report these register values after a reset.  The
result today is a lot of spam in dmesg as the minimally-sized windows
fail to allocate.  By ignoring these windows and treating them as
closed the end result is the same, but there is less spam during boot.

Reported by:	jrtc27
Differential Revision:	https://reviews.freebsd.org/D43922

1 files changed, 18 insertions, 0 deletions

diff --git a/sys/dev/pci/pci_pci.c b/sys/dev/pci/pci_pci.c
index 788af7339418..5cdb1de71c3f 100644
--- a/sys/dev/pci/pci_pci.c
+++ b/sys/dev/pci/pci_pci.c
@@ -437,6 +437,24 @@ pcib_alloc_window(struct pcib_softc *sc, struct pcib_window *w, int type,
 	if (!pcib_is_window_open(w))
 		return;
 
+	/*
+	 * Assume that a window where both the base and limit read as
+	 * zero is not really open, or at least not assigned a valid
+	 * range by the firmware.  This can happen if a bridge device
+	 * is never initialized by firmware, or if a platform driver
+	 * resets the bridge.
+	 *
+	 * If devices behind this bridge have firmware-assigned
+	 * resources in this range then the window will be reallocated
+	 * on-demand.
+	 */
+	if (w->base == 0 && w->limit == ((pci_addr_t)1 << w->step) - 1) {
+		w->base = max_address;
+		w->limit = 0;
+		pcib_write_windows(sc, w->mask);
+		return;
+	}
+
 	if (w->base > max_address || w->limit > max_address) {
 		device_printf(sc->dev,
 		    "initial %s window has too many bits, ignoring\n", w->name);