Date: Thu, 11 Dec 2014 08:38:31 -0800
From: Guenter Roeck <groeck@juniper.net>
To: Bjorn Helgaas <bhelgaas@google.com>
CC: Rajat Jain <rajatxjain@gmail.com>, <linux-kernel@vger.kernel.org>,
        <linux-pci@vger.kernel.org>, Rajat Jain <rajatjain@juniper.net>
Subject: Re: [PATCH v2] PCI: pciehp: Check link state before accessing device
 during removal
Message-ID: <20141211163831.GA2845@svl-evodev-groeck.juniper.net>
References: <546E7120.5080505@gmail.com>
 <20141211002630.GC22886@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-Disposition: inline
In-Reply-To: <20141211002630.GC22886@google.com>
User-Agent: Mutt/1.5.21 (2010-09-15)
Sender: linux-kernel-owner@vger.kernel.org

On Wed, Dec 10, 2014 at 05:26:30PM -0700, Bjorn Helgaas wrote:
> On Thu, Nov 20, 2014 at 02:54:24PM -0800, Rajat Jain wrote:
> > While removing a card, we can't assume the presence to mean that the
> > access to card is OK. That is because the cause of removal may be a
> > link down event, and the card may still be physically present. Thus,
> > instead of presence, use the link state to decide whether or not it is
> > OK to access the card devices.
> > 
> > Here are the problem symptoms:
> > During the removal of a card due to link down, sometimes the following
> > error is seen (because pciehp_unconfigure_device() reads 0xFF from
> > bridge control register as the link is down, which cause it to assume
> > that the VGA bit is set):
> > 
> > pciehp 0000:21:05.0:pcie24: pcie_isr: intr_loc 100
> > pciehp 0000:21:05.0:pcie24: Data Link Layer State change
> > pciehp 0000:21:05.0:pcie24: slot(5): Link Down event
> > pciehp 0000:21:05.0:pcie24: Disabling domain:bus:device=0000:60:00
> > pciehp 0000:21:05.0:pcie24: pciehp_unconfigure_device: domain:bus:dev = 0000:60:00
> > pciehp 0000:21:05.0:pcie24: Cannot remove display device 0000:60:00.0
> > 
> > Ofcourse, when the link comes back up, the device addition fails too:
> > 
> > pciehp 0000:21:05.0:pcie24: pcie_isr: intr_loc 100
> > pciehp 0000:21:05.0:pcie24: Data Link Layer State change
> > pciehp 0000:21:05.0:pcie24: pciehp_check_link_active: lnk_status = 6011
> > pciehp 0000:21:05.0:pcie24: slot(5): Link Up event
> > pciehp 0000:21:05.0:pcie24: Enabling domain:bus:device=0000:60:00
> > pciehp 0000:21:05.0:pcie24: pciehp_check_link_active: lnk_status = 6011
> > pciehp 0000:21:05.0:pcie24: pciehp_check_link_status: lnk_status = 6011
> > pciehp 0000:21:05.0:pcie24: Device 0000:60:00.0 already exists at 0000:60:00, cannot hot-add
> > pciehp 0000:21:05.0:pcie24: Cannot add device at 0000:60:00
> > 
> > The problem is not seen with this patch applied. The device removal and
> > insertion works as expected.
> > 
> > Signed-off-by: Rajat Jain <rajatxjain@gmail.com>
> > Signed-off-by: Rajat Jain <rajatjain@juniper.net>
> > Signed-off-by: Guenter Roeck <groeck@juniper.net>
> > ---
> > v2: Use the already initialized "ctrl" instead of "p_slot->ctrl"
> > 
> >  drivers/pci/hotplug/pciehp_pci.c |    8 ++++----
> >  1 file changed, 4 insertions(+), 4 deletions(-)
> > 
> > diff --git a/drivers/pci/hotplug/pciehp_pci.c b/drivers/pci/hotplug/pciehp_pci.c
> > index 9e69403..911f85b 100644
> > --- a/drivers/pci/hotplug/pciehp_pci.c
> > +++ b/drivers/pci/hotplug/pciehp_pci.c
> > @@ -77,7 +77,7 @@ int pciehp_unconfigure_device(struct slot *p_slot)
> >  {
> >  	int rc = 0;
> >  	u8 bctl = 0;
> > -	u8 presence = 0;
> > +	bool link_active = false;
> >  	struct pci_dev *dev, *temp;
> >  	struct pci_bus *parent = p_slot->ctrl->pcie->port->subordinate;
> >  	u16 command;
> > @@ -85,7 +85,7 @@ int pciehp_unconfigure_device(struct slot *p_slot)
> >  
> >  	ctrl_dbg(ctrl, "%s: domain:bus:dev = %04x:%02x:00\n",
> >  		 __func__, pci_domain_nr(parent), parent->number);
> > -	pciehp_get_adapter_status(p_slot, &presence);
> > +	link_active = pciehp_check_link_active(ctrl);
> >  
> >  	pci_lock_rescan_remove();
> >  
> > @@ -98,7 +98,7 @@ int pciehp_unconfigure_device(struct slot *p_slot)
> >  	list_for_each_entry_safe_reverse(dev, temp, &parent->devices,
> >  					 bus_list) {
> >  		pci_dev_get(dev);
> > -		if (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE && presence) {
> > +		if (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE && link_active) {
> >  			pci_read_config_byte(dev, PCI_BRIDGE_CONTROL, &bctl);
> >  			if (bctl & PCI_BRIDGE_CTL_VGA) {
> >  				ctrl_err(ctrl,
> 
> Why do we even have this code to check for VGA devices?  I looked (briefly)
> and couldn't find anything in the spec that prohibits removal of VGA
> devices.
> 
For my part I don't know. I only know that I had to integrate the patch into
our images since I hit the problem repeatedly. Usually I wait with integrating
Rajat's patches until you accept them, but this one was too disruptive.

I would argue that while the patch may not be perfect, at least it improves
the situation substantially.

> If we do need it (and it looks like most or all hotplug drivers copied it),
> isn't there still a race?  Can't we have the following sequence?
> 
>   - pciehp_check_link_active()		# returns true
>   - Link goes down
>   - pci_read_config_byte()		# fails because link is down
> 
I would guess so. Question is how to address it. Read the configuration byte
first, then check if the link is down ? Check if link is still up after reading
the configuration byte ? Add a note that there may be a potential race condition
and do nothing until it is actually seen ?

Guenter
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/