Hi, On Mon, Sep 1, 2014 at 9:14 PM, Bjorn Helgaas <bhelgaas@xxxxxxxxxx> wrote: > [+cc Linus (author of ad7edfe04908), Matthew (author of 07ff9220908c > from full history), Yinghai (author of 2f5d8e4ff947), Richard] > > On Thu, Aug 28, 2014 at 3:55 PM, Rajat Jain <rajatxjain@xxxxxxxxx> wrote: >> The PCIe root port of the Intel Haswell CPU, has a behavior to endlessly >> retry the configuration cycles, if an endpoint responds with a CRS >> (Configuration Request Retry Status), and the "CRS Software Visibility" >> flag is not set at the root port. This results in a CPU hang, when the >> kernel tries to enumerate the device that responds with CRS. > > Help me understand this. I thought a config read of Vendor ID would see: > > - 0xffff if no device is present, or > - 0x0001 if the device is present but not ready yet, and the OS > should retry the read. This only happens when Software Visibility is > enabled. > - Normal Vendor ID if device is present and ready. If Software > Visibility is disabled (as it is in Linux today) and the endpoint > responds with CRS, the Root Complex may retry the config read > invisibly to software. Yes, this is absolutely correct. However, note that if software visibility is disabled, the root complex may retry the config reads - but the PCIe spec leaves it open on whether or not it should retry or for how long it should retry. In this current case of Intel CPU, I confirmed with Intel that their CPU has chosen to retry indefinitely. In other words, the Intel root complex will keep on retrying (invisible to software) until the endpoint responds with a status other than CRS. > > I suppose these retries are causing your hang, but since the device is > actually present, I would think the CPU would see a read that took a > very long time to complete, but that did eventually complete. No, the read never completes as the root complex keeps on retrying since it keeps getting CRS. > Is this > causing a CPU timeout? Is the endpoint broken and returning CRS > endlessly? Er, yes the endpoint would keep on returning CRS for a veryyyy long time (until it is configured using I2C). However, the end point is not really broken because (in our case) the I2C configuration it is expecting, is to be done from user space (using i2c-dev). Since this enumeration (and this hang) happens at the kernel boot up time, we never really boot up and the userspace never gets a chance to do that I2C configuration. Thus we are stuck in an endless loop of CRS. I see reasons to enable CRS (without regard to the my specific HW) today because even in the case of a broken endpoint (that returns CRS indefinitely), the system should be able to ignore that device and move on. In other words a broken endpoint should not cause the entire PCI hierarchy to fail. The primary reason for introducing CRS in the PCIe spec, IMHO probably was to arm the OS so that it can deal with such broken devices? > >> Please note that this root port behavior (of endless retries) is still >> compliant with PCIe spec as the spec leaves the behavior open to >> implementation, on how many retries to do if "CRS visibility flag" is >> not enabled and it receives a CRS. (Intel has chosen to retry indefinitely) >> >> Ref1: >> https://www.pcisig.com/specifications/pciexpress/ECN_CRS_Software_Visibility_No27.pdf >> >> Ref2: >> PCIe spec V3.0, pg119, pg127 for "Configuration Request Retry Status" >> >> Following CPUs are affected: >> http://ark.intel.com/products/codename/42174/Haswell#@All >> >> Thus we need to enable the CRS visibility flag for such root ports. The >> commit ad7edfe04908 ("[PCI] Do not enable CRS Software Visibility by >> default") suggests to maintain a whitelist of the systems for which CRS >> should be enabled. This patch does the same. > > I'm not a fan of adding a whitelist for devices that work correctly. > I don't think that's a maintainable solution. Since we haven't had > many systems yet that care about CRS, some kind of "enable CRS on > machines newer than X" might work. I agree. I am willing to do whatever is suggested by the community. The only thing I was not sure is if it is possible to get the list of blacklisted devices, and the availability of those devices for testing. > > From Linus' email (https://lkml.org/lkml/2007/12/27/92), the only > detailed data I found was from Ciaran McCreesh > (https://lkml.org/lkml/2007/10/29/500). That case is interesting > because RootCap claims CRSVisible is not supported, but RootCtl claims > it is enabled: > > 00:04.0 PCI bridge: ATI Technologies Inc Unknown device 7934 > (prog-if 00 [Normal decode]) > Bus: primary=00, secondary=02, subordinate=02, sec-latency=0 > Capabilities: [58] Express (v1) Root Port (Slot+), MSI 00 > RootCtl: ErrCorrectable- ErrNon-Fatal- ErrFatal- PMEIntEna- > CRSVisible+ > RootCap: CRSVisible- > 02:00.0 Ethernet controller: Unknown device 0001:8168 (rev 01) > > The Linux code removed by ad7edfe04908 doesn't bother to check the > RootCap bit; it simply tries to set the RootCtl enable. Per spec, > that would be safe because the RootCtl enable bit should be hardwired > to 0 if the Root Port doesn't implement the capability, but I can > imagine a hardware bug where CRSVisible is partly implemented and the > designer assumes it won't be used because he doesn't advertise that > it's supported. > > I think you should add that RootCap test to your patch. Yes, I agree it is a good idea, I will add that. > It would be > very interesting to test that on some of these machines where we found > problems. > >> Note: Looking at the spec and reading about the CRS, IMHO the "CRS >> visibility" looks like a good thing to me that should always be enabled >> on the root ports that support it. And may be we should always enable >> it if supported and maintain a blacklist of devices on which should be >> disabled (because of known issues). >> >> How I stumbled upon this and tested the fix: >> >> Root port: PCI bridge: Intel Corporation Device 2f02 (rev 01) >> >> I have a PCIe endpoint (a PLX 8713 NT bridge) that will keep on responding >> with CRS for a long time when the kernel tries to enumerate the >> endpoint, trying to indicate that the device is not yet ready. This is >> because it needs some configuration over I2C in order to complete its >> reset sequence. This results in a CPU hang during enumeration. > > Is the I2C configuration something the OS needs to do? If so, I > suppose we will always get CRS after reset, and we'll always have to > wait for a timeout, then do that I2C configuration, then somehow try > to re-enumerate the device? The current problem is that we do not have drivers for I2C interface of such PCIe devices in the kernel. I have such a driver almost ready and I'm planning to send it out to the community. But until that happens, we're configuring it in user space using i2c-dev, and then will initiate hot-plug manually. However, with this patch I was hoping to solve CRS issue in general that other people might see if they have a broken endpoint (or something that takes really long to initialize itself). > That seems like a broken usage model. I > guess maybe the I2C configurator could initiate a rescan to sort of > patch things up, [A side comment: One question w.r.t this particular HW I have at my disposal (that requires an I2C configuration before it can be enumerated over PCI). Is there a way to ensure that the I2C subsystem and drivers get initialized BEFORE the PCI enumeration happens? IF so, we can just use that to ensure that the I2C config happens before the PCI enumeration and the endpoint will never respond with CRS. To the best of my understanding, I could only see that we cannot make any assumptions about (and we cannot also force) any ordering between different subsystems such as I2C and PCI.] Thanks, Rajat > but it still seems wrong that we have the timeout and > generate a warning. > >> I used this setup to fix and test this issue. After enabling the CRS >> visibility flag at the root port, I see that CPU moves on as expected >> declaring the following (instead of freezing): >> >> pci 0000:30:00.0 id reading try 50 times with interval 20 ms to get >> ffff0001 > > This message is from pci_bus_check_dev() (added by 2f5d8e4ff947). > pci_bus_read_dev_vendor_id() already has an internal timeout mechanism > with a message that makes sense. I don't understand why we need > another one with a different, unintelligible, message in > pci_bus_check_dev(). > >> Signed-off-by: Rajat Jain <rajatxjain@xxxxxxxxx> >> Signed-off-by: Rajat Jain <rajatjain@xxxxxxxxxxx> >> Signed-off-by: Guenter Roeck <groeck@xxxxxxxxxxx> >> --- >> Hi Bjorn / folks, >> >> I had also saught suggestions on how this patch should be modelled. >> Please find a suggestive alternative here: >> >> https://lkml.org/lkml/2014/8/1/186 >> >> Please let me know your thoughts. >> >> Thanks, >> >> Rajat >> >> >> >> >> drivers/pci/probe.c | 28 ++++++++++++++++++++++++++++ >> 1 file changed, 28 insertions(+) >> >> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c >> index e3cf8a2..909ca75 100644 >> --- a/drivers/pci/probe.c >> +++ b/drivers/pci/probe.c >> @@ -740,6 +740,32 @@ struct pci_bus *pci_add_new_bus(struct pci_bus *parent, struct pci_dev *dev, >> } >> EXPORT_SYMBOL(pci_add_new_bus); >> >> +static const struct pci_device_id crs_whitelist[] = { >> + { PCI_VDEVICE(INTEL, 0x2f00), }, >> + { PCI_VDEVICE(INTEL, 0x2f01), }, >> + { PCI_VDEVICE(INTEL, 0x2f02), }, >> + { PCI_VDEVICE(INTEL, 0x2f03), }, >> + { PCI_VDEVICE(INTEL, 0x2f04), }, >> + { PCI_VDEVICE(INTEL, 0x2f05), }, >> + { PCI_VDEVICE(INTEL, 0x2f06), }, >> + { PCI_VDEVICE(INTEL, 0x2f07), }, >> + { PCI_VDEVICE(INTEL, 0x2f08), }, >> + { PCI_VDEVICE(INTEL, 0x2f09), }, >> + { PCI_VDEVICE(INTEL, 0x2f0a), }, >> + { PCI_VDEVICE(INTEL, 0x2f0b), }, >> + { }, >> +}; >> + >> +static void pci_enable_crs(struct pci_dev *dev) >> +{ >> + /* Enable CRS Software visibility only for whitelisted systems */ >> + if (pci_is_pcie(dev) && >> + pci_pcie_type(dev) == PCI_EXP_TYPE_ROOT_PORT && >> + pci_match_id(crs_whitelist, dev)) >> + pcie_capability_set_word(dev, PCI_EXP_RTCTL, >> + PCI_EXP_RTCTL_CRSSVE); >> +} >> + >> /* >> * If it's a bridge, configure it and scan the bus behind it. >> * For CardBus bridges, we don't scan behind as the devices will >> @@ -787,6 +813,8 @@ int pci_scan_bridge(struct pci_bus *bus, struct pci_dev *dev, int max, int pass) >> pci_write_config_word(dev, PCI_BRIDGE_CONTROL, >> bctl & ~PCI_BRIDGE_CTL_MASTER_ABORT); >> >> + pci_enable_crs(dev); >> + >> if ((secondary || subordinate) && !pcibios_assign_all_busses() && >> !is_cardbus && !broken) { >> unsigned int cmax; >> -- >> 1.7.9.5 >> -- 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
