On Tue, 2022-05-17 at 16:54 +0200, Rafael J. Wysocki wrote: > On Tue, May 17, 2022 at 4:48 PM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote: > > > > On Mon, May 16, 2022 at 10:59:32PM +0200, Rafael J. Wysocki wrote: > > > On Mon, May 16, 2022 at 10:09 PM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote: > > > > On Fri, May 13, 2022 at 10:00:48PM +0000, Jingar, Rajvi wrote: > > > > > > -----Original Message----- > > > > > > From: Bjorn Helgaas <helgaas@xxxxxxxxxx> > > > > > > Sent: Thursday, May 12, 2022 11:36 AM > > > > > > To: Rafael J. Wysocki <rafael@xxxxxxxxxx> > > > > > > Cc: Jingar, Rajvi <rajvi.jingar@xxxxxxxxx>; Wysocki, Rafael J > > > > > > <rafael.j.wysocki@xxxxxxxxx>; Bjorn Helgaas <bhelgaas@xxxxxxxxxx>; > > > > > > David Box > > > > > > <david.e.box@xxxxxxxxxxxxxxx>; Linux PCI <linux-pci@xxxxxxxxxxxxxxx>; > > > > > > Linux > > > > > > Kernel Mailing List <linux-kernel@xxxxxxxxxxxxxxx>; Linux PM <linux- > > > > > > pm@xxxxxxxxxxxxxxx> > > > > > > Subject: Re: [PATCH v5 2/2] PCI/PM: Fix pci_pm_suspend_noirq() to > > > > > > disable PTM > > > > > > > > > > > > On Thu, May 12, 2022 at 07:52:36PM +0200, Rafael J. Wysocki wrote: > > > > > > > On Thu, May 12, 2022 at 7:42 PM Bjorn Helgaas <helgaas@xxxxxxxxxx> > > > > > > > wrote: > > > > > > > > On Thu, May 12, 2022 at 03:49:18PM +0200, Rafael J. Wysocki wrote: > > > > > > > > > > > > > > > Something like this should suffice IMV: > > > > > > > > > > > > > > > > > > if (!dev_state_saved || pci_dev->current_state != PCI_D3cold) > > > > > > > > > > > > > > > > > > pci_disable_ptm(pci_dev); > > > > > > > > > > > > > > > > It makes sense to me that we needn't disable PTM if the device is > > > > > > > > in > > > > > > > > D3cold. But the "!dev_state_saved" condition depends on what the > > > > > > > > driver did. Why is that important? Why should we not do the > > > > > > > > following? > > > > > > > > > > > > > > > > if (pci_dev->current_state != PCI_D3cold) > > > > > > > > pci_disable_ptm(pci_dev); > > > > > > > > > > > > > > We can do this too. I thought we could skip the power state > > > > > > > check if dev_state_saved was unset, because then we would know > > > > > > > that the power state was not D3cold. It probably isn't worth > > > > > > > the hassle though. > > > > > > > > > > We see issue with certain platforms where only checking if device > > > > > power state in D3Cold is not enough and the !dev_state_saved check > > > > > is needed when disabling PTM. Device like nvme is relying on ASPM, > > > > > it stays in D0 but state is saved. Touching the config space wakes > > > > > up the device which prevents the system from entering into low power > > > > > state. > > > > > > > > Correct me if I'm wrong: for NVMe devices, nvme_suspend() has already > > > > saved state and put the device in some low-power state. Disabling PTM > > > > here is functionally OK but prevents a system low power state, so you > > > > want to leave PTM enabled. > > > > > > > > But I must be missing something because pci_prepare_to_sleep() > > > > currently disables PTM for Root Ports. If we leave PTM enabled on > > > > NVMe but disable it on the Root Port above it, any PTM Request from > > > > NVMe will cause an Unsupported Request error. > > > > > > > > Disabling PTM must be coordinated across PTM Requesters and PTM > > > > Responders. That means the decision to disable cannot depend on > > > > driver-specific things like whether the driver has saved state. > > > > > > Setting state_saved generally informs pci_pm_suspend_noirq() that the > > > device has already been handled and it doesn't need to do anything to > > > it. > > > > > > But you are right that PTM should be disabled on downstream devices as > > > well as on the ports that those devices are connected to and it can be > > > done even if the given device has already been handled, so the > > > state_saved value is technically irrelevant. > > > > > > That's why I suggested to check if the power state is between D0 and > > > D3cold (exclusive) and only disable PTM if that is the case. It is > > > pointless to disable PTM for devices in D3cold and it may be harmful > > > for devices that are left in D0. > > > > "... it may be harmful for devices that are left in D0" -- I want to > > understand this better. It sounds like nvme_suspend() leaves the > > device in some device-specific low-power flavor of D0, and subsequent > > config accesses take it out of that low-power situation? > This is exactly what we see. It's not all machines, but in our lab we've seen in it on 3 production systems out of about 20. And they were all different generations, a 7th, 8th, and 10th gen. nvme_suspend is relying on NVMe APST / PCIe ASPM to put the device in a low power state. The link state will be L1 or deeper while the device remains in D0. https://nvmexpress.org/resources/nvm-express-technology-features/nvme-technology-power-features/ > That's my understanding of it. > > > If that's the case, it sounds a little brittle. I don't think it's > > obvious that "pci_dev->state_saved was set by the driver" means "no > > config accesses allowed in pci_pm_suspend_noirq()." > > Well, yes and no. The device may be in D3cold then, so > pci_pm_suspend_noirq() should at least check that before accessing its > config space. > > > And pci_pm_suspend_noirq() calls quirks via pci_fixup_device(), which are > > very likely to do config accesses. > > > > Maybe PTM needs to be disabled earlier, e.g., in pci_pm_suspend()? I > > don't think PTM uses any interrupts, so there's probably no reason > > interrupts need to be disabled before disabling PTM. > > That certainly is worth investigation. For one, I don't see any > obvious downsides of doing so. We will look at this. David
