From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Date: Tue, 27 Mar 2018 17:13:17 -0500 From: Bjorn Helgaas To: Lukas Wunner Cc: Mika Westerberg , Bjorn Helgaas , "Rafael J. Wysocki" , Len Brown , Mario.Limonciello@dell.com, Michael Jamet , Yehezkel Bernat , Andy Shevchenko , linux-pci@vger.kernel.org, linux-acpi@vger.kernel.org, Greg Kroah-Hartman Subject: Re: [PATCH v3 5/5] ACPI / hotplug / PCI: Mark stale PCI devices disconnected Message-ID: <20180327221317.GH7759@bhelgaas-glaptop.roam.corp.google.com> References: <20180226132112.81447-1-mika.westerberg@linux.intel.com> <20180226132112.81447-6-mika.westerberg@linux.intel.com> <20180327194512.GD7759@bhelgaas-glaptop.roam.corp.google.com> <20180327205236.GA15613@wunner.de> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii In-Reply-To: <20180327205236.GA15613@wunner.de> Sender: linux-acpi-owner@vger.kernel.org List-ID: [+cc Greg] On Tue, Mar 27, 2018 at 10:52:36PM +0200, Lukas Wunner wrote: > On Tue, Mar 27, 2018 at 02:45:12PM -0500, Bjorn Helgaas wrote: > > I hate this pci_dev_set_disconnected() thing and hope to remove it > > someday. It's racy and adds additional states that are hard to > > manage, e.g., "device has been removed and will not respond to PCI > > accesses, but pci_dev_set_disconnected() hasn't been called yet". If > > we can handle that case (and we should), we shouldn't need > > pci_dev_set_disconnected(). > > Absent pci_dev_is_disconnected(), how can the PCI core detect if > a device is still there (versus has been hot-removed)? The only > valid method I know of is reading the vendor ID from config space. > However reading config spaces takes time. To properly handle > surprise removal, we ought to check presence of the device frequently > lest we clutter the logs with errors, unnecessarily occupy the CPU > and bus or in some cases hang the machine. Ideally, we should check > presence of the device every time we access it. I'm not proposing a change in pci_dev_is_disconnected() for *this* patch, so we can have a discussion here, but I don't think resolving this is a precondition for this patch. I was just expressing remorse, which I agree is not constructive, so I should have kept quiet. I know you understand the problem, but to try to make amends, I'll recap it here for any onlookers: - driver calls pci_dev_is_disconnected(), learns device is present - device is removed - driver reads register, believing device is still present Obviously the register read fails, because the device is gone. So the driver has to be prepared for that failure regardless of whether we have pci_dev_is_disconnected(). > Now, imagine every time we access config or mmio space, we read the > vendor ID from config space, so we essentially double the number of > accesses to the device. Does that make sense, performance-wise? No, but I wouldn't advocate that strategy. > If pci_dev_is_disconnected() returns true, the device is either there > or was until very recently, so in general checking that instead of > reading the vendor ID should be sufficient and should result in > acceptable performance. > > I keep hearing these complaints about the PCI_DEV_DISCONNECTED flag, > from Greg but now also from you, but I'm not seeing constructive > criticism how checking presence of a device should be handled instead, > in a way that doesn't negatively impact performance. > > IMO it was a mistake to constrain visibility of the flag to the PCI core > (at Greg's behest), it should have been made available to drivers so > they're afforded a better method to detect surprise removal than just > reading the vendor ID every time they access the device. To avoid the race, drivers have to check for a valid value anyway. If they *also* check pci_dev_is_disconnected(), that clutters the code and gives a false sense of security. When a read fails, it normally returns ~0 as the data to the driver. The driver knows the semantics of the registers it reads, and often it can tell immediately that a successful read of the register could never return that value, so it doesn't need to try any more accesses to Vendor ID or anything else. If ~0 *is* a possible valid value, the driver can read some other register (maybe Vendor ID, or maybe an MMIO register that can be read faster). Bjorn