[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20240104053803.GA16954@linuxonhyperv3.guj3yctzbm1etfxqx2vob5hsef.xx.internal.cloudapp.net>
Date: Wed, 3 Jan 2024 21:38:03 -0800
From: Saurabh Singh Sengar <ssengar@...ux.microsoft.com>
To: Alexander Stein <alexander.stein@...tq-group.com>
Cc: bhelgaas@...gle.com, linux-pci@...r.kernel.org,
linux-kernel@...r.kernel.org, decui@...rosoft.com
Subject: Re: [PATCH] PCI/sysfs: Fix race in pci sysfs creation
On Tue, Dec 12, 2023 at 12:28:05AM -0800, Saurabh Singh Sengar wrote:
> On Tue, Dec 12, 2023 at 08:19:11AM +0100, Alexander Stein wrote:
> > Hi Saurabh,
> >
> > thanks for the patch.
> >
> > Am Samstag, 9. Dezember 2023, 04:46:16 CET schrieb Saurabh Sengar:
> > > Currently there is a race in calling pci_create_resource_files function
> > > from two different therads, first therad is triggered by pci_sysfs_init
> > > from the late initcall where as the second thread is initiated by
> > > pci_bus_add_devices from the respective PCI drivers probe.
> > >
> > > The synchronization between these threads relies on the sysfs_initialized
> > > flag. However, in pci_sysfs_init, sysfs_initialized is set right before
> > > calling pci_create_resource_files which is wrong as it can create race
> > > condition with pci_bus_add_devices threads. Fix this by setting
> > > sysfs_initialized flag at the end of pci_sysfs_init and direecly call the
> >
> > Small typo here: direecly -> directly
> >
> > > pci_create_resource_files function from it.
> > >
> > > There can be an additional case where driver probe is so delayed that
> > > pci_bus_add_devices is called after the sysfs is created by pci_sysfs_init.
> > > In such cases, attempting to access already existing sysfs resources is
> > > unnecessary. Fix this by adding a check for sysfs attributes and return
> > > if they are already allocated.
> > >
> > > In both cases, the consequence will be the removal of sysfs resources that
> > > were appropriately allocated by pci_sysfs_init following the warning below.
> >
> > I'm not sure if this is the way to go. Unfortunately I can't trigger this
> > error on my imx6 platform at the moment (apparently timing is off).
> > But reading [1] again, the most expressive way is that pci_bus_add_devices()
> > needs to wait until pci_sysfs_init() has passed.
>
> (I correct my self a bit in my earlier reply)
> The problem with waiting is that sysfs entries will be created by pci_sysfs_init
> already and when pci_bus_add_devices try to create it will observe that the
> entries are already existing and in such case PCI code will remove the sysfs
> entries created by pci_sysfs_init. Resulting system will be having no sysfs
> entries.
Hi Alexander,
Have you got time to check this ? Please let me know if you think there is any
concern left with this patch.
- Saurabh
Powered by blists - more mailing lists