Linus, Keith, and Paul, Can you try the attached patch and see if it helps your issues with enabling runtime PM for the NEC USB 3.0 host controller? You may have to increase the timeout for your system. Hi Rafael, I mentioned I've been seeing failures when PCI runtime PM was enabled on the xHCI host controller in my Lenovo x220. It turns out that several people have the same NEC chipset in other Lenovo models, and are also experiencing similar issues. When runtime PM is enabled for the host, and you plug in a device, we get a steady stream of: Oct 31 16:11:31 puck kernel: [ 5196.504007] xhci_hcd 0000:0e:00.0: PME# enabled Oct 31 16:11:40 puck kernel: [ 5205.503185] xhci_hcd 0000:0e:00.0: PME# disabled Oct 31 16:11:40 puck kernel: [ 5205.503202] xhci_hcd 0000:0e:00.0: setting latency timer to 64 Those three lines are repeated, but the new USB device never enumerates. I've discovered that if I add a small delay in the xhci_resume function (which is invoked by the dev_pm_ops resume method), then the host controller has time to issue an interrupt about the port status change before the runtime PM core puts it back into PCI suspend. I'm running Linus' latest tree as of yesterday (commit 839d8810), which includes your commit 379021d "PCI / PM: Extend PME polling to all PCI devices". Without any delays: Oct 31 16:11:31 puck kernel: [ 5196.484738] xhci_suspend started Oct 31 16:11:31 puck kernel: [ 5196.503891] xhci_hcd 0000:0e:00.0: hcd_pci_runtime_suspend: 0 Oct 31 16:11:31 puck kernel: [ 5196.504007] xhci_hcd 0000:0e:00.0: PME# enabled Oct 31 16:11:40 puck kernel: [ 5205.503185] xhci_hcd 0000:0e:00.0: PME# disabled Oct 31 16:11:40 puck kernel: [ 5205.503202] xhci_hcd 0000:0e:00.0: setting latency timer to 64 Oct 31 16:11:40 puck kernel: [ 5205.503207] xhci_resume started Oct 31 16:11:40 puck kernel: [ 5205.503247] xhci_resume done Oct 31 16:11:40 puck kernel: [ 5205.503248] xhci_hcd 0000:0e:00.0: hcd_pci_runtime_resume: 0 Oct 31 16:11:40 puck kernel: [ 5205.503255] xhci_suspend started Oct 31 16:11:40 puck kernel: [ 5205.509289] xhci_hcd 0000:0e:00.0: hcd_pci_runtime_suspend: 0 Oct 31 16:11:40 puck kernel: [ 5205.509359] xhci_hcd 0000:0e:00.0: PME# enabled Note the very small amount of time between xhci_resume being done and xhci_suspend being started. If I put a delay at the end of xhci_resume (as small as 30us) before returning success, the interrupt for the port status change has time to come in: Nov 1 06:45:19 puck kernel: [57557.041234] xhci_suspend started Nov 1 06:45:19 puck kernel: [57557.059419] xhci_hcd 0000:0e:00.0: hcd_pci_runtime_suspend: 0 Nov 1 06:45:19 puck kernel: [57557.059513] xhci_hcd 0000:0e:00.0: PME# enabled Nov 1 06:45:27 puck kernel: [57565.060976] xhci_hcd 0000:0e:00.0: PME# disabled Nov 1 06:45:27 puck kernel: [57565.060994] xhci_hcd 0000:0e:00.0: setting latency timer to 64 Nov 1 06:45:27 puck kernel: [57565.060999] xhci_resume started Nov 1 06:45:27 puck kernel: [57565.061039] xhci_resume done Nov 1 06:45:27 puck kernel: [57565.061071] xhci_hcd 0000:0e:00.0: hcd_pci_runtime_resume: 0 Nov 1 06:45:27 puck kernel: [57565.061079] xhci_suspend started Nov 1 06:45:27 puck kernel: [57565.061082] xhci_irq Nov 1 06:45:27 puck kernel: [57565.061085] Port Status Change Event for port 3 Nov 1 06:45:27 puck kernel: [57565.067076] xhci_resume started Nov 1 06:45:27 puck kernel: [57565.067119] xhci_resume done Nov 1 06:45:27 puck kernel: [57565.067150] xhci_hcd 0000:0e:00.0: hcd_pci_runtime_suspend: -16 Nov 1 06:45:27 puck kernel: [57565.067158] pci_pm_runtime_suspend(): hcd_pci_runtime_suspend+0x0/0x50 returns -16 Maybe resume takes too long, and the runtime PM system needs to mark the device as busy when the runtime_resume function returns? Catting the power/autosuspend_delay_ms file returns an I/O error: root@puck:/sys/bus/pci/devices/0000:0e:00.0# cat power/autosuspend_delay_ms cat: power/autosuspend_delay_ms: Input/output error Any thoughts on this? I can certainly add the attached patch to the xHCI driver, but it seems like there is something fundamentally broken in the runtime PM core if it's suspending devices without giving them a chance to interrupt. Sarah Sharp
diff --git a/drivers/usb/host/xhci.c b/drivers/usb/host/xhci.c index 1ff95a0..e238360 100644 --- a/drivers/usb/host/xhci.c +++ b/drivers/usb/host/xhci.c @@ -914,6 +914,9 @@ failed_restart: set_bit(HCD_FLAG_HW_ACCESSIBLE, &xhci->shared_hcd->flags); spin_unlock_irq(&xhci->lock); + + if (xhci->quirks & XHCI_NEC_HOST) + udelay(30); return 0; } #endif /* CONFIG_PM */
