On 11/13/2019 12:28 AM, Bjorn Helgaas wrote:
On Tue, Nov 12, 2019 at 11:29:55PM +0530, Vidya Sagar wrote:
On 11/12/2019 7:51 PM, Bjorn Helgaas wrote:
On Tue, Nov 12, 2019 at 01:59:23PM +0100, Thierry Reding wrote:
On Mon, Nov 11, 2019 at 04:32:35PM -0600, Bjorn Helgaas wrote:
On Mon, Nov 11, 2019 at 11:31:18AM +0530, Vidya Sagar wrote:
On 11/6/2019 10:11 PM, Bjorn Helgaas wrote:
Based on Vidya's backtrace, I think the resume path with problems
is this:
pci_pm_resume_noirq
pci_pm_default_resume_early
pci_power_up
if (platform_pci_power_manageable(dev))
platform_pci_set_power_state(dev, PCI_D0) # <-- FW delay here?
pci_raw_set_power_state
pci_update_current_state
pci_device_is_present # <-- config read returns CRS
So I think your suggestion is that Vidya's firmware should be
doing the delay inside platform_pci_set_power_state()?
Vidya, you typically work on Tegra, so I assume this is on an
arm64 system? Does it have ACPI? Do you have access to the
firmware developers to ask about who they expect to do the delays?
Yes. This is on arm64 (Tegra) and we don't have any ACPI or any
other firmware for that matter. PCIe is brought up directly in the
kernel.
I assume that your device is coming out of D3cold because apparently
you're seeing a CRS status from the config read when
pci_update_current_state() calls pci_device_is_present(). CRS status
should only happen after reset or power-on from D3cold, and you're not
doing a reset.
I'm pretty sure platform_pci_power_manageable() returns false on
your system (can you confirm that?) because the only scenarios with
platform power management are MID (Intel platform) and ACPI (which you
don't have).
Maybe you have some other platform-specific mechanism that controls
power to PCI devices, and it's not integrated into the
platform_pci_*() framework?
My understanding after reading the PCIe specification is that CRS is a
mechanism that allows an endpoint to signal that it isn't ready yet for
operation after reset or power-on from D3cold. There's nothing in there
that's platform specific. This is really only for specific endpoints.
I don't see how adding platform specific PM code would help in this
case. At a platform level we don't know if users are going to plug in a
PCI endpoint that needs a long delay before it's ready after reset and/
or exit from D3cold.
Right, see below.
I do understand that perhaps pci_device_is_present() is perhaps not the
best place to do complex CRS handling, but if a mechanism is clearly
described in the specification, isn't it something that should be dealt
with in the core? That way we don't have to quirk this for every device
and platform.
Definitely; we don't want quirks for endpoints (unless they're
actually broken) or for platforms (unless there's a platform hardware
or firmware defect).
There's no question that we need to delay and handle CRS after
power-on from D3cold. I'm trying to get at the point that PCI itself
doesn't tell us how to do that power-on. The mechanisms defined by
PCI rely on config space, which is only accessible in D0-D3hot, not
D3cold. Power-on from D3cold can only happen via ACPI methods or
other platform-specific mechanisms, and the current design abstracts
those via platform_pci_set_power_state(). This is partly based on
Rafael's response in [1].
The PCIe specification says that:
Software that intends to take advantage of this mechanism must
ensure that the first access made to a device following a valid
reset condition is a Configuration Read Request accessing both
bytes of the Vendor ID field in the device's Configuration Space
header.
So doesn't that mean that pci_device_is_present() is already much too
late because we've potentially made other configuration read requests in
the meantime?
Wouldn't it make more sense to push the CRS handling up a bit? The
existing pci_power_up() function seems like it would be a good place.
For example, adding code to deal with CRS right after the platform PCI
PM calls but before pci_raw_set_power_state() seems like it would fit
the restrictions given in the above quote from the specification.
Yep, I think that's the right point. I'm trying to figure out how to
integrate it. Rafael suggests that delays may be platform-specific
and should be in platform_pci_set_power_state(), but the CRS handling
itself isn't platform-specific and maybe could be higher up.
I'm fishing to see if Tegra has some kind of power control for
endpoints that is not related to the platform_pci_*() framework. How
did the endpoint get put in D3cold in the first place? I assume
something in the suspend path did that? Maybe this happens when we
suspend the Tegra RC itself, e.g., tegra_pcie_pm_suspend()?
tegra_pcie_pm_suspend
tegra_pcie_phy_power_off
tegra_pcie_power_off
tegra_pcie_pm_resume
tegra_pcie_power_on
tegra_pcie_phy_power_on
If a path like tegra_pcie_pm_resume() is causing the D3cold -> D0
transition for the endpoint, I don't think we want to do CRS handling
there because that path shouldn't be touching the endpoint. But maybe
it should be doing the delays required by PCIe r5.0, sec 6.6.1, before
any config accesses are issued to devices.
Here, I'm exercising suspend-to-RAM sequence and the PCIe endpoint of
concern is Intel 750 NVMe drive.
PCIe host controller driver already has 100ms of delay as per the spec,
To make this more concrete, where specifically is this delay?
It is after PERST is deasserted and before making the first check for DLActive+
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/pci/controller/dwc/pcie-tegra194.c?h=v5.4-rc7#n816
and dw_pcie_wait_for_link() has maximum timeout of 1sec before bailing out
Anyway, in this case, link is coming up within 100ms but config space is not accessible as device returns CRS.
but this particular device is taking 1023ms to get ready to respond to
configuration space requests (till that time, it responds with
configuration request retry statuses)
I've put a dump_stack () to see the path resume sequence takes and here it is
Call trace:
dump_backtrace+0x0/0x158
show_stack+0x14/0x20
dump_stack+0xb0/0xf4
pci_bus_generic_read_dev_vendor_id+0x19c/0x1a0
pci_bus_read_dev_vendor_id+0x48/0x70
pci_update_current_state+0x68/0xd8
pci_power_up+0x40/0x50
pci_pm_resume_noirq+0x7c/0x138
dpm_run_callback.isra.16+0x20/0x70
device_resume_noirq+0x120/0x238
async_resume_noirq+0x24/0x58
async_run_entry_fn+0x40/0x148
process_one_work+0x1e8/0x360
worker_thread+0x40/0x488
kthread+0x118/0x120
ret_from_fork+0x10/0x1c
pci 0005:01:00.0: ready after 1023ms
Spec also mentions the following
Unless Readiness Notifications mechanisms are used (see Section
6.23), the Root Complex and/or system software must allow at
least 1.0 s after a Conventional Reset of a device, before it
may determine that a device which fails to return a Successful
Completion status for a valid Configuration Request is a broken
device. This period is independent of how quickly Link training
completes.
My understanding is that this 1sec waiting is supposed to be done by
the PCIe sub-system and not the host controller driver.
That doesn't say we must wait 1s; it only says we can't decide the
device is broken before 1s. But regardless, I agree that the CRS
handling doesn't belong in the driver for either the endpoint or the
PCIe host controller.
My question is whether this wait should be connected somehow with
platform_pci_set_power_state(). It sounds like the tegra host
controller driver already does the platform-specific delays, and I'm
not sure it's reasonable for platform_pci_set_power_state() to do the
CRS polling. Maybe something like this? I'd really like to get
Rafael's thinking here.
diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index e7982af9a5d8..052fa316c917 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -964,9 +964,14 @@ void pci_refresh_power_state(struct pci_dev *dev)
*/
void pci_power_up(struct pci_dev *dev)
{
+ pci_power_state_t prev_state = dev->current_state;
+
if (platform_pci_power_manageable(dev))
platform_pci_set_power_state(dev, PCI_D0);
+ if (prev_state == PCI_D3cold)
+ pci_dev_wait(dev, "D3cold->D0", PCIE_RESET_READY_POLL_MS);
+
pci_raw_set_power_state(dev, PCI_D0);
pci_update_current_state(dev, PCI_D0);
}
FWIW, this particular device is taking a little more time than 1 sec
(i.e. 1023 ms) I'm now wondering why is it that the CRS has a
timeout of 60 secs than just 1 sec?
pci_bus_wait_crs() does an exponential backoff, i.e., it does reads
after 0ms, 2ms, 4ms, 8ms, ..., 512ms, 1024ms, so we don't know
*exactly* when the device became ready. All we know is that it
became ready somewhere between 512ms and 1024ms.
But 821cdad5c46c ("PCI: Wait up to 60 seconds for device to become
ready after FLR") does suggest that the Intel 750 NVMe may require
more than 1s. I think the 60s timeout is just a convenient large
number and I'm not sure it's derived from anything in the spec.
I tried out the above patch that you came up with.
Couple of points here.
In Tegra194, we put the endpoint to D0 explicitly (instead of leaving it in D3Hot)
and then go for link transition to L2 (because of Tegra specific reasons).
Even if I comment out the API in pcie-tegra194.c which puts the devices back into D0
(from D3Hot) just before going for L2 transition, device is in D3Hot but not in D3Cold
So, if (prev_state == PCI_D3cold) didn't work really.
Also, I'm wondering why should there be a check in the first place anyway?
why can't pci_dev_wait() be called always?
[1] https://lore.kernel.org/r/11429373.7ySiFsEkgL@kreacher