We observed a problem on at least two POWER machines with PCIe devices with very large SR-IOV BARs. The total size is larger than host bridge MMIO window. We expect normal BARs of this device get address successfuly, and SR-IOV BARs deserted. However, this only happens if PCI realloc is disabled (pci=realloc=off or not config at all). Otherwise we fail to assign both BARs. Here is what happened when realloc is enabled: * SR-IOV resources will NOT be considered as optional until the last round of realloc. Both normal BARs and SR-IOV BARs of this device will be sized and recorded in upstream bridge as initial size. * Realloc never works because this bridge is too big, at end of each round, the bridge is restored to the value before assigning, i.e. it keeps its size. * At last round, we finally consider SR-IOV resources optional and size them seperately. We should get a small bridge this time. * However, in pbus_size_mem() we call calculate_memsize() to calculate the final size. calculate_memsize() never shrinks window, i.e. if old size is larger than newly sized result, old size will be honored. Unfortunately, the old size at this time is the sum of both normal BARs and SR-IOV BARs when doing realloc. * The bridge window failed to get address just like previous rounds. Without PCI realloc, SR-IOV resources are considered as optional from the beginning and sized seperately. The old size would be 0. We get a small bridge window and assinged successfully. Everything work as expected. I actually don't quite understand what realloc does. Apparently this is not the right fix ... Cc: Yinghai Lu <yinghai@xxxxxxxxxx> Signed-off-by: Guo Chao <yan@xxxxxxxxxxxxxxxxxx> --- drivers/pci/setup-bus.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/drivers/pci/setup-bus.c b/drivers/pci/setup-bus.c index a5a63ec..39837d7 100644 --- a/drivers/pci/setup-bus.c +++ b/drivers/pci/setup-bus.c @@ -1623,7 +1623,7 @@ again: res->end = fail_res->end; res->flags = fail_res->flags; if (fail_res->dev->subordinate) - res->flags = 0; + res->flags = res->start = res->end = 0; } free_list(&fail_head); -- 1.9.1 -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
