On Thu, May 7, 2020 at 7:36 AM Mika Westerberg <mika.westerberg@xxxxxxxxxxxxxxx> wrote: > > On Thu, May 07, 2020 at 07:24:37AM -0500, Bjorn Helgaas wrote: > > On Thu, May 7, 2020 at 6:45 AM Mika Westerberg > > <mika.westerberg@xxxxxxxxxxxxxxx> wrote: > > > > > > On Wed, May 06, 2020 at 05:42:28PM -0500, Bjorn Helgaas wrote: > > > > On Thu, Apr 16, 2020 at 11:32:45AM +0300, Mika Westerberg wrote: > > > > > Kai-Heng Feng reported that it takes long time (>1s) to resume > > > > > Thunderbolt connected PCIe devices from both runtime suspend and system > > > > > sleep (s2idle). > > > > > > > > > > These PCIe downstream ports the second link capability (PCI_EXP_LNKCAP2) > > > > > announces support for speeds > 5 GT/s but it is then capped by the > > > > > second link control (PCI_EXP_LNKCTL2) register to 2.5 GT/s. This > > > > > possiblity was not considered in pci_bridge_wait_for_secondary_bus() so > > > > > it ended up waiting for 1100 ms as these ports do not support active > > > > > link layer reporting either. > > > > > > > > > > PCIe spec 5.0 section 6.6.1 mandates that we must wait minimum of 100 ms > > > > > before sending configuration request to the device below, if the port > > > > > does not support speeds > 5 GT/s, and if it does we first need to wait > > > > > for the data link layer to become active before waiting for that 100 ms. > > > > > > > > > > PCIe spec 5.0 section 7.5.3.6 further says that all downstream ports > > > > > that support speeds > 5 GT/s must support active link layer reporting so > > > > > instead of looking for the speed we can check for the active link layer > > > > > reporting capability and determine how to wait based on that (as they go > > > > > hand in hand). > > > > > > > > I can't quite tell what the defect is here. > > > > > > > > I assume you're talking about this text from sec 6.6.1: > > > > > > > > - With a Downstream Port that does not support Link speeds greater > > > > than 5.0 GT/s, software must wait a minimum of 100 ms before > > > > sending a Configuration Request to the device immediately below > > > > that Port. > > > > > > > > - With a Downstream Port that supports Link speeds greater than 5.0 > > > > GT/s, software must wait a minimum of 100 ms after Link training > > > > completes before sending a Configuration Request to the device > > > > immediately below that Port. Software can determine when Link > > > > training completes by polling the Data Link Layer Link Active bit > > > > or by setting up an associated interrupt (see Section 6.7.3.3 ). > > > > > > > > I don't understand what Link Control 2 has to do with this. The spec > > > > talks about ports *supporting* certain link speeds, which sounds to me > > > > like the Link Capabilities. It doesn't say anything about the current > > > > or target link speed. > > > > > > PCIe 5.0 page 764 recommends using Supported Link Speeds Vector in Link > > > Capabilities 2 register and that's what pcie_get_speed_cap() is doing. > > > > > > However, we can avoid figuring the speed altogether by checking the > > > dev->link_active_reporting instead because that needs to be implemented > > > by all Downstream Ports that supports speeds > 5 GT/s (PCIe 5.0 page 735). > > > > I understand that part. But the code as-is matches sec 6.6.1, which > > makes it easy to understand. Checking link_active_reporting instead > > makes it harder to understand because it makes it one step removed > > from 6.6.1. And link_active_reporting must be set for ports that > > support > 5 GT/s, but it must *also* be set in some hotplug cases, so > > that further complicates the connection between it and 6.6.1. > > > > And apparently there's an actual defect, and I don't understand what > > that is yet. It sounds like we agree that pcie_get_speed_cap() is > > doing the right thing. Is there something in > > pci_bridge_wait_for_secondary_bus() that doesn't match sec 6.6.1? > > The defect is that some downstream PCIe ports on a system Kai-Heng is > using have Supported Link Speeds Vector with > 5GT/s and it does not > support active link reporting so the currect code ends up waiting 1100 ms > slowing down resume time. That sounds like a hardware defect that should be worked around with a quirk or something. If we just restructure the code to avoid the problem, we're likely to reintroduce it later because there's no hint in the code about this problem. > The Target Link Speed of this port (in Link Control 2) has the speed > capped to 2.5 GT/s. I saw you mentioned this before, but I don't understand yet why it's relevant. The spec says "supported,", not "current" or "target".
