Hi Alan, 在 2024-10-09星期三的 11:50 -0400，Alan Stern写道： > On Wed, Oct 09, 2024 at 10:35:05AM +0800, duanchenghao wrote: > > Hi Alan, > > > > These are two patches, each addressing the same issue. The current > > patch is a direct solution to the problem of the interrupt bottom > > half > > being frozen. The patch you replied with is another, alternative > > solution to the same problem. Please review which solution is more > > suitable, or if there are any other revised proposals. > > > > > > Please review the patch I mentioned: > > https://lore.kernel.org/all/0a4dc46ae767c28dd207ae29511ede747f05539a.camel@xxxxxxxxxx/ > > I still think your whole approach is wrong. There is no need to > change > the way the HCD_FLAG_WAKEUP_PENDING flag gets set or cleared; that's > not > the reason for the problem you're seeing. > Thank you very much for your evaluation of the scheme. I have a question regarding why the set_bit operation for the HCD_FLAG_WAKEUP_PENDING flag is performed in the top half of an interrupt handler, while the clear_bit operation is done in the bottom half. This seems to contradict conventional practices. The issue is not limited to S4; if other processes freeze the work queue in the bottom half, the same problem may arise. The solution you described below should be able to resolve the current S4 issue, but for now, we haven't identified any other scenarios that require the same operation. Based on my understanding, the USB device is woken up in the bottom half of the interrupt, and both the set_bit and clear_bit operations for the HCD_FLAG_WAKEUP_PENDING flag should be executed within the same thread in the bottom half. May I ask if there are any other reasons why the set_bit is executed in the top half? > The problem occurs because when suspend_common() in > drivers/usb/core/hcd-pci.c sets do_wakeup, it does so by calling > device_may_wakeup(), and the value that function returns is not what > we > want. The value is based on whether the controller's power/wakeup > attribute is set, but we also have to take into account the type of > suspend that's happening. > > Basically, if msg is one of the suspend types for which wakeups > should > always be disabled, then do_wakeup should be set to false. There > isn't > a macro to test for these things, but there should be. Something > like > PMSG_IS_AUTO() in include/linux/pm.h; maybe call it > PMSG_NO_WAKEUP(). > This macro should return true if the PM_EVENT_FREEZE or > PM_EVENT_QUIESCE > bits are set in msg.event. > > Rafael, please correct me if I got any of this wrong. > > So the right way to fix the problem is to add to pm.h: > > #define PMSG_NO_WAKEUP(msg) (((msg.event) & \ > (PM_EVENT_FREEZE | PM_EVENT_QUIESCE)) != 0) > > and in suspend_common(): > > if (PMSG_IS_AUTO(msg)) > do_wakeup = true; > else if (PMSG_NO_WAKEUP(msg)) > do_wakeup = false; > else > do_wakeup = device_may_wakeup(dev); > > Then with do_wakeup set to false, none of the HCD_WAKEUP_PENDING() > tests > in the following code will be executed, so the routine won't fail > during > the freeze or restore phase with -EBUSY. > > You'll also have to define an hcd_pci_freeze() routine, just > like hcd_pci_suspend() except that it uses PMSG_FREEZE instead of > PMSG_SUSPEND. And the .freeze callback in usb_hcd_pci_pm_ops should > be > changed to hcd_pci_freeze. > > In fact, it looks like choose_wakeup() in drivers/usb/core/driver.c > could also use the new macro, instead of checking for FREEZE or > QUIESCE > directly the way it does now. > > Alan Stern
