lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date:   Wed, 22 Nov 2017 01:26:02 +0100
From:   "Rafael J. Wysocki" <rjw@...ysocki.net>
To:     Brian Norris <briannorris@...omium.org>
Cc:     Jeffy Chen <jeffy.chen@...k-chips.com>,
        linux-kernel@...r.kernel.org, bhelgaas@...gle.com,
        linux-pm@...r.kernel.org, tony@...mide.com,
        shawn.lin@...k-chips.com, dianders@...omium.org,
        linux-pci@...r.kernel.org, linux-acpi@...r.kernel.org,
        Len Brown <lenb@...nel.org>
Subject: Re: [RFC PATCH v10 6/7] PCI / PM: Move acpi wakeup code to pci core

On Tuesday, November 14, 2017 3:51:11 AM CET Brian Norris wrote:
> Hi Rafael,
> 
> I'll answer some of it from my perspective, though Jeffy might have had
> different ideas (and answers) when he implemented this.
> 
> On Wed, Nov 08, 2017 at 11:32:20PM +0100, Rafael J. Wysocki wrote:
> > On Friday, October 27, 2017 9:26:11 AM CET Jeffy Chen wrote:
> > > Move acpi wakeup code to pci core as pci_set_wakeup(), so that other
> > > platforms could reuse it.
> > 
> > What exactly do you want to reuse?
> > 
> > It looks like that's just several lines of code in acpi_pci_wakeup()
> > and acpi_pci_propagate_wakeup() which invoke ACPI-specific lower-level
> > functions, so IMO not worth it at all.
> 
> The important part he's sharing here is the walking of the tree
> structure, in which it's possible for some parent along the way to
> handle wakeup for its children. I'm not sure how valuable nor how
> reusable that is.

ACPI very well may be the only case needing to walk the hierarchy for that.

> In this case (the Rockchip platforms Jeffy and I are working on), I
> think we really want to just support a single WAKE# pin for the whole
> system, so maybe the complexity isn't needed.

So unless I know you'll need it, please don't add it.

> The spec does describe that there are good reasons for supporting more than
> 1 WAKE# pin though (e.g., 1 per device), so it doesn't seem really wise to
> shoehorn oursleves into a single setup.

One WAKE# pin per device is reasonable enough, but WAKE# is specifically
defined as out-of-band and "orthogonal" to the PCIe hierarchy.  What you
need is a way to program WAKE# and (possibly) wakeup power on the given
platform for each device having a WAKE# pin individually.
 
The main reason why ACPI needs to walk the hierarchy is that on some systems
the firmware takes over the handling of the native PME mechanism which then is
taken care of by the AML (and the kernel is not granted control of the
relevant PCIe registers).  I don't think you'll ever see anything like that
on non-ACPI systems.

> But that can be implemented either via copying the "few" lines of
> tree-walking logic, or by trying to share them.
> 
> > The structure for other platform code may be the same or similar, but
> > the details will almost certainly be different and I don't think that
> > having more callback pointers in pci_platform_pm_ops is necessarily better.
> 
> I suppose that's reasonable.
> 
> > > Also add .setup_dev() / .setup_host_bridge() / .cleanup() platform pm
> > > ops's callbacks to setup and cleanup pci devices and host bridge for
> > > wakeup.
> > 
> > Why are they needed?
> 
> The implementation is in patch 7, if you really needed more info about
> why, or provide alternatives.

Well, that's quite questionable.

At least defining ->dev_wakeup to do device_set_wakeup_capable() doesn't
really make sense to me.

> The current set of hooks assumes that there is no state information or
> initialization needed for tracking actions of these platform PM hooks on
> a device. For ACPI this works, because devices have "companion"
> acpi_dev's to handle everything, and the ACPI framework generally
> prepares GPE's for you, IIUC. For 'pci-mid', the operations happen to be
> trivial (and arguably wrong -- several are no-ops, where we might expect
> the platform to tell us whether the operation was actually supported or
> not).
> 
> For device tree, there isn't really a canonical place to store this
> information, nor to initialize something like wakeup interrupts.

The only thing you need IMO is ->set_wakeup which also is what ACPI needs
and that enables or disables wakeup for the device.  It doesn't actually
have to track anything (other than, possibly, whether or not wakeup has
been enabled for it already).

And since the core takes care of native PMEs for you, the only thing
this needs to cover is the WAKE# pin programming AFAICS.

The setup part, in turn, in the ACPI case is done via acpi_platform_notify()
and, analogously, acpi_platform_notify_remove() does the cleanup.  You seem
to be trying to add something like it via pci_platform_pm_ops, but is it
really the most suitable place for that?

> Technically, we could shoehorn this into the .set_wakeup() call, but
> we'd probably rather not do things like request_irq() on every attempt
> to suspend/resume the system (among other reasons, we'd lose information
> that we might otherwise track in /proc/ or /sys/).
> 
> Or the inverse of the above: where would you suggest initializing or
> tearing down the wakeirq?

Again, ACPI does that via acpi_platform_notify()/acpi_platform_notify_remove(),
so maybe something similar?

In any case, you need a way to associate DT-provided data with PCI devices and,
ideally, that should be done at the enumeration time.

> An alternative could be to include any necessary state into the
> pci_host_bridge or pci_dev and just inline the setup code into
> pci.c/remove.c (e.g., pci_register_host_bridge()) and pci-driver.c
> (pci_device_{probe,remove}()). But I'm not sure that's much more
> beautiful.

Well, that's not the way it is done in the ACPI case anyway.

Thanks,
Rafael

Powered by blists - more mailing lists