Re: [PATCH 25/30] ACPI / hotplug / PCI: Check for new devices on enabled slots

[Alex Williamson <alex.williamson@redhat.com>]

On Thu, 2013-09-05 at 07:19 -0600, Alex Williamson wrote:
> On Thu, 2013-09-05 at 13:54 +0200, Rafael J. Wysocki wrote:
> > On Wednesday, September 04, 2013 10:06:02 PM Alex Williamson wrote:
> > > On Wed, 2013-09-04 at 21:37 -0600, Alex Williamson wrote:
> > > > On Thu, 2013-09-05 at 01:35 +0200, Rafael J. Wysocki wrote:
> > > > > On Wednesday, September 04, 2013 05:12:14 PM Alex Williamson wrote:
> > > > > > On Thu, 2013-09-05 at 00:54 +0200, Rafael J. Wysocki wrote:
> > > > > > > On Wednesday, September 04, 2013 02:36:34 PM Alex Williamson wrote:
> > > > > > > > On Thu, 2013-07-18 at 01:32 +0200, Rafael J. Wysocki wrote:
> > > > > > > > > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > > > > > > > > 
> > > > > > > > > The current implementation of acpiphp_check_bridge() is pretty dumb:
> > > > > > > > >  - It enables a slot if it's not enabled and the slot status is
> > > > > > > > >    ACPI_STA_ALL.
> > > > > > > > >  - It disables a slot if it's enabled and the slot status is not
> > > > > > > > >    ACPI_STA_ALL.
> > > > > > > > > 
> > > > > > > > > This behavior is not sufficient to handle the Thunderbolt daisy
> > > > > > > > > chaining case properly, however, because in that case the bus
> > > > > > > > > behind the already enabled slot needs to be rescanned for new
> > > > > > > > > devices.
> > > > > > > > > 
> > > > > > > > > For this reason, modify acpiphp_check_bridge() so that slots are
> > > > > > > > > disabled and stopped if they are not in the ACPI_STA_ALL state.
> > > > > > > > > 
> > > > > > > > > For slots in the ACPI_STA_ALL state, devices behind them that don't
> > > > > > > > > respond are trimmed using a new function, trim_stale_devices(),
> > > > > > > > > introduced specifically for this purpose.  That function walks
> > > > > > > > > the given bus and checks each device on it.  If the device doesn't
> > > > > > > > > respond, it is assumed to be gone and is removed.
> > > > > > > > > 
> > > > > > > > > Once all of the stale devices directy behind the slot have been
> > > > > > > > > removed, acpiphp_check_bridge() will start looking for new devices
> > > > > > > > > that might have appeared on the given bus.  It will do that even if
> > > > > > > > > the slot is already enabled (SLOT_ENABLED is set for it).
> > > > > > > > > 
> > > > > > > > > In addition to that, make the bus check notification ignore
> > > > > > > > > SLOT_ENABLED and go for enable_device() directly if bridge is NULL,
> > > > > > > > > so that devices behind the slot are re-enumerated in that case too.
> > > > > > > > > 
> > > > > > > > > This change is based on earlier patches from Kirill A Shutemov
> > > > > > > > > and Mika Westerberg.
> > > > > > > > > 
> > > > > > > > > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > > > > > > > > Tested-by: Mika Westerberg <mika.westerberg@linux.intel.com>
> > > > > > > > > ---
> > > > > > > > 
> > > > > > > > FYI, git bisect landed on this patch as the cause of my serial console
> > > > > > > > dying on current upstream.  Further debugging to come...  Thanks,
> > > > > > > 
> > > > > > > Well, sorry about that.
> > > > > > > 
> > > > > > > What exactly do you mean by "dying"?
> > > > > > 
> > > > > > Sorry, I was hoping to have more details quickly, but it's been a pain
> > > > > > to debug.  By dying I mean serial console output suddenly stops during
> > > > > > kernel boot and nothing more comes out of it until after the system is
> > > > > > rebooted.  The problem happens when acpiphp_check_bridge() calls
> > > > > > enable_slot().  The serial console dies somewhere down in
> > > > > > acpiphp_bus_trim().  I think this is happening on the 00:1f ISA bridge,
> > > > > > so there's a good chance the serial ports are described as somewhere
> > > > > > under there.
> > > > > 
> > > > > Can you please check if that is the acpiphp_bus_trim() called by
> > > > > acpiphp_bus_add() or the other one called from trim_stale_devices()?
> > > > > 
> > > > > Just add a dump_stack() or WARN_ON(1) to trim_stale_devices() next to
> > > > > the acpiphp_bus_trim() call and see if that triggers.  I *think* it's the one
> > > > > in acpiphp_bus_add(), but it won't hurt to verify that.
> > > > 
> > > > Here's the call path:
> > > > 
> > > > [   16.120824]  [<ffffffff81627e6c>] dump_stack+0x55/0x76
> > > > [   16.125979]  [<ffffffff8162132e>] enable_slot+0x4ee/0x5e0
> > > > [   16.131396]  [<ffffffff813418fb>] ? trim_stale_devices+0x5b/0xf0
> > > > [   16.137420]  [<ffffffff81341b35>] acpiphp_check_bridge+0xd5/0x110
> > > > [   16.143531]  [<ffffffff81342acb>] hotplug_event+0x16b/0x260
> > > > [   16.149115]  [<ffffffff81072cd9>] ? process_one_work+0x189/0x540
> > > > [   16.155136]  [<ffffffff81342bf0>] hotplug_event_work+0x30/0x70
> > > > [   16.160978]  [<ffffffff81072d3b>] process_one_work+0x1eb/0x540
> > > > [   16.166819]  [<ffffffff81072cd9>] ? process_one_work+0x189/0x540
> > > > [   16.172836]  [<ffffffff8107353c>] worker_thread+0x11c/0x370
> > > > [   16.178426]  [<ffffffff81073420>] ? rescuer_thread+0x350/0x350
> > > > [   16.184276]  [<ffffffff8107b0ea>] kthread+0xea/0xf0
> > > > [   16.189165]  [<ffffffff8107b000>] ? kthread_create_on_node+0x160/0x160
> > > > [   16.195700]  [<ffffffff816395dc>] ret_from_fork+0x7c/0xb0
> > > > [   16.201109]  [<ffffffff8107b000>] ? kthread_create_on_node+0x160/0x160
> > > > 
> > > > The actual death of the serial console occurs in acpi_device_set_power()
> > > > called from:
> > > > 
> > > > enable_slot()
> > > >  acpiphp_bus_add()
> > > >   acpiphp_bus_trim()
> > > >    acpi_bus_trim()
> > > >     acpi_walk_namespace()
> > > >      acpi_bus_remove()
> > > >       acpi_device_unregister()
> > > >        acpi_device_set_power()
> > > > 
> > > > I can't seem to get a path from the acpi devices in question there, so I
> > > > have no idea what's getting trimmed here.  It worries me quite a bit by
> > > > introducing this trimming that apparently wasn't happening before
> > > > though.  Thanks,
> > > 
> > > As suspected, the pnp.bus_id/id of the last device before the serial
> > > console dies is COM1/PNP0501.  I also see all of these being trimmed
> > > out:
> > > 
> > >  MBRD/PNP0C02
> > >  DMAC/PNP0200
> > >  MATH/PNP0C04
> > >  PIC/PNP0000
> > >  HPET/PNP0103
> > >  RTC/PNP0B00
> > >  SPKR/PNP0800
> > >  TIME/PNP0100
> > >  LNK{A-H}/PNP0C0F
> > > 
> > > This seems like a bad idea.  I forgot to mention, the original
> > > hotplug_event is called with a device check on \_SB_.PCI0.PEX2.  The box
> > > where I'm seeing this is a pretty generic X58 based Nehalem workstation
> > > (Lenovo S20).
> > 
> > Thanks for the info!
> > 
> > The acpiphp_bus_trim() in acpiphp_bus_add() is a leftover and a bad one.
> > I don't think it's actually necessary, at least Thunderbolt works without
> > it just fine on my Aspire S5.
> > 
> > The patch below should help, can you please test it?
> > 
> > Rafael
> > 
> > ---
> > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > Subject: ACPI / hotplug / PCI: Don't trim devices before scanning the namespace
> > 
> > In acpiphp_bus_add() we first remove device objects corresponding to
> > the given handle and the ACPI namespace branch below it which are
> > then re-created by acpi_bus_scan().  This used to be done to clean
> > up after surprise removals, but now we do the cleanup through
> > trim_stale_devices() which checks if the devices in question are
> > actually gone before removing them, so the device hierarchy trimming
> > in acpiphp_bus_add() is not necessary any more and, moreover, it may
> > lead to problems if it removes device objects corresponding to
> > devices that are actually present.
> > 
> > Reported-by: Alex Williamson <alex.williamson@redhat.com>
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > ---
> >  drivers/pci/hotplug/acpiphp_glue.c |    1 -
> >  1 file changed, 1 deletion(-)
> > 
> > Index: linux-pm/drivers/pci/hotplug/acpiphp_glue.c
> > ===================================================================
> > --- linux-pm.orig/drivers/pci/hotplug/acpiphp_glue.c
> > +++ linux-pm/drivers/pci/hotplug/acpiphp_glue.c
> > @@ -487,7 +487,6 @@ static void acpiphp_bus_add(acpi_handle
> >  {
> >  	struct acpi_device *adev = NULL;
> >  
> > -	acpiphp_bus_trim(handle);
> >  	acpi_bus_scan(handle);
> >  	acpi_bus_get_device(handle, &adev);
> >  	if (adev)
> > 
> > 
> 
> It boots again and the console works, but I see screens and screens of
> messages like this:
> 
> [   18.288122] pci 0000:00:00.0: no hotplug settings from platform
> [   18.288127] pcieport 0000:00:01.0: no hotplug settings from platform
> [   18.288142] pci 0000:01:00.0: no hotplug settings from platform
> [   18.288157] pci 0000:01:00.1: no hotplug settings from platform
> [   18.288162] pcieport 0000:00:03.0: no hotplug settings from platform
> [   18.288176] pci 0000:02:00.0: no hotplug settings from platform
> [   18.288190] pci 0000:02:00.1: no hotplug settings from platform
> [   18.288195] pcieport 0000:00:07.0: no hotplug settings from platform
> [   18.288209] pci 0000:03:00.0: no hotplug settings from platform
> [   18.288224] pci 0000:03:00.1: no hotplug settings from platform
> [   18.288228] pci 0000:00:14.0: no hotplug settings from platform
> [   18.288233] pci 0000:00:14.1: no hotplug settings from platform
> [   18.288237] pci 0000:00:14.2: no hotplug settings from platform
> [   18.288242] pci 0000:00:16.0: no hotplug settings from platform
> [   18.288247] pci 0000:00:16.1: no hotplug settings from platform
> [   18.288251] pci 0000:00:16.2: no hotplug settings from platform
> [   18.288256] pci 0000:00:16.3: no hotplug settings from platform
> [   18.288260] pci 0000:00:16.4: no hotplug settings from platform
> [   18.288265] pci 0000:00:16.5: no hotplug settings from platform
> [   18.288269] pci 0000:00:16.6: no hotplug settings from platform
> [   18.288274] pci 0000:00:16.7: no hotplug settings from platform
> [   18.288278] pci 0000:00:1a.0: no hotplug settings from platform
> [   18.288279] pci 0000:00:1a.0: using default PCI settings
> [   18.288292] pci 0000:00:1a.1: no hotplug settings from platform
> [   18.288293] pci 0000:00:1a.1: using default PCI settings
> [   18.288307] ehci-pci 0000:00:1a.7: no hotplug settings from platform
> [   18.288308] ehci-pci 0000:00:1a.7: using default PCI settings
> [   18.288322] pci 0000:00:1b.0: no hotplug settings from platform
> [   18.288327] pcieport 0000:00:1c.0: no hotplug settings from platform
> [   18.288332] pcieport 0000:00:1c.4: no hotplug settings from platform
> [   18.288344] pci 0000:05:00.0: no hotplug settings from platform
> [   18.288349] pci 0000:00:1d.0: no hotplug settings from platform
> [   18.288350] pci 0000:00:1d.0: using default PCI settings
> [   18.288360] pci 0000:00:1d.1: no hotplug settings from platform
> [   18.288361] pci 0000:00:1d.1: using default PCI settings
> [   18.288374] pci 0000:00:1d.2: no hotplug settings from platform
> [   18.288374] pci 0000:00:1d.2: using default PCI settings
> [   18.288387] pci 0000:00:1d.3: no hotplug settings from platform
> [   18.288387] pci 0000:00:1d.3: using default PCI settings
> 
> The boot is noticeably slower.  What's going to happen on systems that
> actually have a significant I/O topology vs my little workstation?

Just to give you an idea:

CONFIG_HOTPLUG_PCI_ACPI=y

$ dmesg | wc
  5697  49935 384368

$ dmesg | tail --lines=1
[   53.137123] Ebtables v2.0 registered

-- vs --

# CONFIG_HOTPLUG_PCI_ACPI is not set

$ dmesg | wc
 1053  9176 71652

$dmesg | tail --lines=1
[   28.917220] Ebtables v2.0 registered

So it spews out 5x more output with acpiphp enabled and takes and extra
24s to boot (nearly 2x).  Not good.  Thanks,

Alex