Re: [PATCH v6 4/5] xhci: introduce xhci->lost_power flag
From: Mathias Nyman
Date: Wed Jan 08 2025 - 13:43:13 EST
On 8.1.2025 12.59, Théo Lebrun wrote:
On Wed Dec 18, 2024 at 6:49 PM CET, Théo Lebrun wrote:
On Tue Dec 17, 2024 at 10:00 PM CET, Roger Quadros wrote:
On 13/12/2024 18:03, Théo Lebrun wrote:
On Thu Dec 12, 2024 at 1:37 PM CET, Roger Quadros wrote:
On 10/12/2024 19:13, Théo Lebrun wrote:
The XHCI_RESET_ON_RESUME quirk allows wrappers to signal that they
expect a reset after resume. It is also used by some to enforce a XHCI
reset on resume (see needs-reset-on-resume DT prop).
Some wrappers are unsure beforehands if they will reset. Add a mechanism
to signal *at resume* if power has been lost. Parent devices can set
this flag, that defaults to false.
The XHCI_RESET_ON_RESUME quirk still triggers a runtime_pm_get() on the
controller. This is required as we do not know if a suspend will
trigger a reset, so the best guess is to avoid runtime PM.
Signed-off-by: Théo Lebrun <theo.lebrun@xxxxxxxxxxx>
---
drivers/usb/host/xhci.c | 3 ++-
drivers/usb/host/xhci.h | 6 ++++++
2 files changed, 8 insertions(+), 1 deletion(-)
diff --git a/drivers/usb/host/xhci.c b/drivers/usb/host/xhci.c
index 5ebde8cae4fc44cdb997b0f61314e309bda56c0d..ae2c8daa206a87da24d58a62b0a0485ebf68cdd6 100644
--- a/drivers/usb/host/xhci.c
+++ b/drivers/usb/host/xhci.c
@@ -1017,7 +1017,8 @@ int xhci_resume(struct xhci_hcd *xhci, pm_message_t msg)
spin_lock_irq(&xhci->lock);
- if (hibernated || xhci->quirks & XHCI_RESET_ON_RESUME || xhci->broken_suspend)
+ if (hibernated || xhci->quirks & XHCI_RESET_ON_RESUME ||
+ xhci->broken_suspend || xhci->lost_power)
reinit_xhc = true;
if (!reinit_xhc) {
diff --git a/drivers/usb/host/xhci.h b/drivers/usb/host/xhci.h
index 4914f0a10cff42dbc1448dcf7908534d582c848e..32526df75925989d40cbe7d59a187c945f498a30 100644
--- a/drivers/usb/host/xhci.h
+++ b/drivers/usb/host/xhci.h
@@ -1645,6 +1645,12 @@ struct xhci_hcd {
unsigned broken_suspend:1;
/* Indicates that omitting hcd is supported if root hub has no ports */
unsigned allow_single_roothub:1;
+ /*
+ * Signal from upper stacks that we lost power during system-wide
+ * suspend. Its default value is based on XHCI_RESET_ON_RESUME, meaning
+ * it is safe for wrappers to not modify lost_power at resume.
+ */
+ unsigned lost_power:1;
I suppose this is private to XHCI driver and not legitimate to be accessed
by another driver after HCD is instantiated?
Yes it is private.
Doesn't access to xhci_hcd need to be serialized via xhci->lock?
Good question. In theory maybe. In practice I don't see how
cdns_host_resume(), called by cdns_resume(), could clash with anything
else. I'll add that to be safe.
Just curious, what happens if you don't include patch 4 and 5?
Is USB functionality still broken for you?
No it works fine. Patches 4+5 are only there to avoid the below warning.
Logging "xHC error in resume" is a lie, so I want to avoid it.
How is it a lie?
The XHCI controller did loose its save/restore state during a PM operation.
As far as XHCI is concerned this is an error. no?
The `xhci->quirks & XHCI_RESET_ON_RESUME` is exactly the same thing;
both the quirk and the flag we add have for purpose:
1. skipping this block
if (!reinit_xhc) {
retval = xhci_handshake(&xhci->op_regs->status,
STS_CNR, 0, 10 * 1000 * 1000);
// ...
xhci_restore_registers(xhci);
xhci_set_cmd_ring_deq(xhci);
command = readl(&xhci->op_regs->command);
command |= CMD_CRS;
writel(command, &xhci->op_regs->command);
if (xhci_handshake(&xhci->op_regs->status,
STS_RESTORE, 0, 100 * 1000)) {
// ...
}
}
2. avoiding this warning:
xhci_warn(xhci, "xHC error in resume, USBSTS 0x%x, Reinit\n", temp);
I don't think the block skipped is important in resume duration (to be
confirmed). But the xhci_warn() is not desired: we do not want to log
warnings if we know it is expected.
I'll think some more about it.
About this series, there were two discussions:
- The desire to avoid putting the hardware init sequence of cdns3-ti
inside runtime_resume() as we don't do runtime PM.
*That is fine and will be fixed for the next revision.*
See [PATCH V6 2/5]: https://lore.kernel.org/lkml/8a1ed4be-c41c-46b6-ae25-33a6035b8c8d@xxxxxxxxxx/
- [PATCH V6 4/5] and [PATCH V6 5/5] are dedicated to avoiding a warning
at XHCI resume on J7200:
xhci_warn(xhci, "xHC error in resume, USBSTS 0x%x, Reinit\n", temp);
Adding a new quirk or private xhci_cd meme
https://lore.kernel.org/lkml/20241210-s2r-cdns-v6-4-28a17f9715a2@xxxxxxxxxxx/
https://lore.kernel.org/lkml/20241210-s2r-cdns-v6-5-28a17f9715a2@xxxxxxxxxxx/
Roger Quadros asked if we should not instead keep it, as there is
indeed a reinit of the xHC block. I don't think we do: we don't want
a warning at resume; IMO it would imply the reinit was unexpected.
Proof is there is already a platform with a ->broken_suspend boolean
that disables the warning even though there is a reinit. It doesn't
log a warning as the reinit was expected.
So we currently have:
- xhci->broken_suspend: set at probe & implies the resume sequence
won't work.
- xhci->quirks & XHCI_RESET_ON_RESUME: set at probe & implies the
controller reset during suspend.
IIUC xhci->broken_suspend is NOT equivalent to "the controller reset
during suspend". Else we wouldn't have both the broken_suspend flag
and the XHCI_RESET_ON_RESUME quirk.
In our case we want exactly the same thing as the
XHCI_RESET_ON_RESUME quirk but updated at resume depending on what
the wrapper driver detects.
We could either:
1. Update xhci->quirks at resume from upper layers.
2. Introduce a xhci->lost_power flag. It would be strictly equivalent
to the XHCI_RESET_ON_RESUME quirk BUT updated at resume from
upper layers.
@Mathias Nyman: what is your thought on the matter? Option (2) was
the one taken in this series. Is there another option I am missing?
This would be a fourth way the upper layers inform xhci_resume() that the
xHC host should be reset instead of restoring the registers.
option 1 creates the first dynamic xhci->quirk flag,
option 2 adds a xhci->lost_power flag that is touched outside of xhci driver.
Neither seem like a good idea just to get rid of a dev_warn() message.
Maybe its time to split xhci_resume() into xhci_reset_resume()
and xhci_restore_resume(), and let those upper layers decide what they need.
Doesn't cdns driver already have a xhci_plat_priv resume_quirk() function
called during xhci_plat_resume(), before xhci_resume()?
Can that be used?
Thanks
Mathias