[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20220511222243.GA822874@bhelgaas>
Date: Wed, 11 May 2022 17:22:43 -0500
From: Bjorn Helgaas <helgaas@...nel.org>
To: Jay Zhou <jianjay.zhou@...wei.com>
Cc: linux-pci@...r.kernel.org, linux-kernel@...r.kernel.org,
bhelgaas@...gle.com, alex.williamson@...hat.com,
weidong.huang@...wei.com, Donald Dutile <ddutile@...hat.com>,
Jason Gunthorpe <jgg@...pe.ca>,
Leon Romanovsky <leon@...nel.org>
Subject: Re: [PATCH] pci: avoid dead lock between device reset and sriov
disable
[+cc Don, Jason, Leon, just FYI]
On Mon, Apr 04, 2022 at 02:25:39PM +0800, Jay Zhou wrote:
> Call trace of PF SRIOV disable:
> sriov_numvfs_store
> device_lock <----------------- (1) get the device lock
> ->sriov_configure # e.g. vfio_pci_sriov_configure
> sriov_disable
> pci_cfg_access_lock <--- (4) wait dev->block_cfg_access to be 0
>
> Call trace of PF reset:
> reset_store
> pci_reset_function
> pci_dev_lock
> pci_cfg_access_lock <----- (2) set dev->block_cfg_access = 1
> device_lock <------------- (3) want to get the device lock
>
> These two oprations would wait for each other forever if the
> code execution sequence is (1)(2)(3)(4).
>
> Let's get the device lock and then the config access lock in
> pci_dev_lock().
>
> Signed-off-by: Jay Zhou <jianjay.zhou@...wei.com>
Thanks, I (finally) applied Yicong's identical patch from a year ago
to pci/virtualization for v5.19, with a commit log combined and
adapted from both:
PCI: Avoid pci_dev_lock() AB/BA deadlock with sriov_numvfs_store()
The sysfs sriov_numvfs_store() path acquires the device lock before the
config space access lock:
sriov_numvfs_store
device_lock # A (1) acquire device lock
sriov_configure
vfio_pci_sriov_configure # (for example)
vfio_pci_core_sriov_configure
pci_disable_sriov
sriov_disable
pci_cfg_access_lock
pci_wait_cfg # B (4) wait for dev->block_cfg_access == 0
Previously, pci_dev_lock() acquired the config space access lock before the
device lock:
pci_dev_lock
pci_cfg_access_lock
dev->block_cfg_access = 1 # B (2) set dev->block_cfg_access = 1
device_lock # A (3) wait for device lock
Any path that uses pci_dev_lock(), e.g., pci_reset_function(), may
deadlock with sriov_numvfs_store() if the operations occur in the sequence
(1) (2) (3) (4).
Avoid the deadlock by reversing the order in pci_dev_lock() so it acquires
the device lock before the config space access lock, the same as the
sriov_numvfs_store() path.
[bhelgaas: combined and adapted commit log from Jay Zhou's independent
subsequent posting:
https://lore.kernel.org/r/20220404062539.1710-1-jianjay.zhou@huawei.com]
Link: https://lore.kernel.org/linux-pci/1583489997-17156-1-git-send-email-yangyicong@hisilicon.com/
Also-posted-by: Jay Zhou <jianjay.zhou@...wei.com>
Signed-off-by: Yicong Yang <yangyicong@...ilicon.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@...gle.com>
> ---
> drivers/pci/pci.c | 10 +++++-----
> 1 file changed, 5 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> index 9ecce435fb3f..61a6db1d21f6 100644
> --- a/drivers/pci/pci.c
> +++ b/drivers/pci/pci.c
> @@ -5103,19 +5103,19 @@ static int pci_reset_bus_function(struct pci_dev *dev, bool probe)
>
> void pci_dev_lock(struct pci_dev *dev)
> {
> - pci_cfg_access_lock(dev);
> /* block PM suspend, driver probe, etc. */
> device_lock(&dev->dev);
> + pci_cfg_access_lock(dev);
> }
> EXPORT_SYMBOL_GPL(pci_dev_lock);
>
> /* Return 1 on successful lock, 0 on contention */
> int pci_dev_trylock(struct pci_dev *dev)
> {
> - if (pci_cfg_access_trylock(dev)) {
> - if (device_trylock(&dev->dev))
> + if (device_trylock(&dev->dev)) {
> + if (pci_cfg_access_trylock(dev))
> return 1;
> - pci_cfg_access_unlock(dev);
> + device_unlock(&dev->dev);
> }
>
> return 0;
> @@ -5124,8 +5124,8 @@ EXPORT_SYMBOL_GPL(pci_dev_trylock);
>
> void pci_dev_unlock(struct pci_dev *dev)
> {
> - device_unlock(&dev->dev);
> pci_cfg_access_unlock(dev);
> + device_unlock(&dev->dev);
> }
> EXPORT_SYMBOL_GPL(pci_dev_unlock);
>
> --
> 2.27.0
>
Powered by blists - more mailing lists