Date: Wed, 20 May 2015 14:47:49 -0500
From: Bjorn Helgaas <bhelgaas@google.com>
To: Yinghai Lu <yinghai@kernel.org>,
        Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: linux-pci@vger.kernel.org, linux-kernel@vger.kernel.org,
        netdev@vger.kernel.org, intel-wired-lan@lists.osuosl.org
Subject: e1000e pci_disable_link_state_locked() issues
Message-ID: <20150520194749.GA10210@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
User-Agent: Mutt/1.5.21 (2010-09-15)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2224
Lines: 48

I think we have some issues with the e1000e usage of
pci_disable_link_state_locked(), which Yinghai added with 9f728f53dd70
("PCI/e1000e: Add and use pci_disable_link_state_locked()").

That fixed an AER deadlock in the following path, where pci_bus_sem is held
by pci_walk_bus(), and we deadlocked when we tried to re-acquire it in
pci_disable_link_state():

  do_recovery
    broadcast_error_message(..., report_slot_reset)
      pci_walk_bus
        down_read(&pci_bus_sem)
          cb(...)                                       # report_slot_reset
            report_slot_reset
              dev->driver->err_handler->slot_reset      # e1000_io_slot_reset
                e1000_io_slot_reset
                  e1000e_disable_aspm
                    pci_disable_link_state
                      down_read(&pci_bus_sem)

9f728f53dd70 fixed that by changing e1000e_disable_aspm() to use
pci_disable_link_state_locked() instead, which assumes pci_bus_sem is
already held.

That's fine for the e1000_io_slot_reset() path, where pci_bus_sem really
*is* held.  But e1000e_disable_aspm() is also called from e1000_probe() and
__e1000_resume(), and in those paths, we *don't* hold pci_bus_sem.

In effect, the caller of pci_disable_link_state_locked() is promising that
pci_bus_sem is held, and __pci_disable_link_state() relies on that promise
for its locking.  But e1000e isn't upholding its end of the bargain.

I'm not 100% sure __pci_disable_link_state() actually *needs* that locking:
it is only called from a driver, and it should be impossible for a device
or any upstream bridge to go away while a driver is bound to it.  If
somebody wanted to analyze this further and propose a patch to remove the
locking (if it seems safe), that would be great.

But in any case, __pci_disable_link_state() should be able to rely on its
callers following the rules, so I'd like to see an e1000e change to use
pci_disable_link_state() from the paths where pci_bus_sem is not held.

Bjorn
--
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/