Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Fri, 7 Apr 2023 18:58:29 -0500
From:   Bjorn Helgaas <helgaas@kernel.org>
To:     "You-Sheng Yang (vicamo)" <vicamo.yang@canonical.com>
Cc:     Nirmal Patel <nirmal.patel@linux.intel.com>,
        Jonathan Derrick <jonathan.derrick@linux.dev>,
        Lorenzo Pieralisi <lpieralisi@kernel.org>,
        Rob Herring <robh@kernel.org>,
        Krzysztof =?utf-8?Q?Wilczy=C5=84ski?= <kw@linux.com>,
        Bjorn Helgaas <bhelgaas@google.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Keith Busch <kbusch@kernel.org>, linux-pci@vger.kernel.org,
        linux-kernel@vger.kernel.org,
        Korneliusz Osmenda <korneliuszo@gmail.com>,
        Alexander Stein <alexander.stein@ew.tq-group.com>
Subject: Re: [PATCH] PCI: vmd: guard device addition and removal
Message-ID: <20230407235829.GA3834716@bhelgaas>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20230313173733.1815277-1-vicamo.yang@canonical.com>
Precedence: bulk

[+cc Korneliusz, Alexander]

On Tue, Mar 14, 2023 at 01:37:33AM +0800, You-Sheng Yang (vicamo) wrote:
> VMD may fail to create sysfs entries while `pci_rescan_bus()` called in
> some other drivers like t7xx wwan driver:
> 
>   sysfs: cannot create duplicate filename '/devices/.../resource0'
>   Call Trace:
>    <TASK>
>    sysfs_warn_dup.cold+0x17/0x34
>    sysfs_add_bin_file_mode_ns+0xc0/0xf0
>    sysfs_create_bin_file+0x6d/0xb0
>    pci_create_attr+0x117/0x260
>    pci_create_resource_files+0x6b/0x150
>    pci_create_sysfs_dev_files+0x18/0x30
>    pci_bus_add_device+0x30/0x80
>    pci_bus_add_devices+0x31/0x80
>    pci_bus_add_devices+0x5b/0x80
>    vmd_enable_domain.constprop.0+0x6b7/0x880 [vmd]
>    vmd_probe+0x16d/0x193 [vmd]

This is a long-standing issue, and I would *love* to nail it, but this
doesn't feel like the right solution to me.  What's unique about vmd
here?

I guess maybe it's similar to the situation Korneliusz and Alexander
ran into at [1]?

And why is t7xx called out specifically here?  This is a pretty
generic sysfs attribute issue and it doesn't *seem* like t7xx should
be special in that respect.  Oooh, maybe it's the fact that there's a
t7xx patch [2] coming that adds pci_rescan_bus() there?

Krzysztof has converted a lot of the sysfs files to static attributes,
e.g.,

  506140f9c06b ("PCI/sysfs: Convert "index", "acpi_index", "label" to static attributes")
  d93f8399053d ("PCI/sysfs: Convert "vpd" to static attribute")
  f42c35ea3b13 ("PCI/sysfs: Convert "reset" to static attribute")
  527139d738d7 ("PCI/sysfs: Convert "rom" to static attribute")
  e1d3f3268b0e ("PCI/sysfs: Convert "config" to static attribute")

I think that's the *best* approach since the sysfs infrastructure
already prevents races here.  But these last few files are more
difficult to convert, so we've been kind of stalled on them.

Bjorn

[1] https://lore.kernel.org/r/20230316091540.494366-1-alexander.stein@ew.tq-group.com
[2] https://lore.kernel.org/linux-pci/20230124204543.550d88e3@kernel.org/

> Fixes: 185a383ada2e ("x86/PCI: Add driver for Intel Volume Management Device (VMD)")
> Signed-off-by: You-Sheng Yang (vicamo) <vicamo.yang@canonical.com>
> ---
>  drivers/pci/controller/vmd.c | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/drivers/pci/controller/vmd.c b/drivers/pci/controller/vmd.c
> index 769eedeb8802..f050991bd1e9 100644
> --- a/drivers/pci/controller/vmd.c
> +++ b/drivers/pci/controller/vmd.c
> @@ -838,9 +838,13 @@ static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
>  	pci_add_resource_offset(&resources, &vmd->resources[1], offset[0]);
>  	pci_add_resource_offset(&resources, &vmd->resources[2], offset[1]);
>  
> +	pci_lock_rescan_remove();
> +
>  	vmd->bus = pci_create_root_bus(&vmd->dev->dev, vmd->busn_start,
>  				       &vmd_ops, sd, &resources);
>  	if (!vmd->bus) {
> +		pci_unlock_rescan_remove();
> +
>  		pci_free_resource_list(&resources);
>  		vmd_remove_irq_domain(vmd);
>  		return -ENODEV;
> @@ -893,6 +897,8 @@ static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
>  
>  	vmd_acpi_end();
>  
> +	pci_unlock_rescan_remove();
> +
>  	WARN(sysfs_create_link(&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,
>  			       "domain"), "Can't create symlink to domain\n");
>  	return 0;
> -- 
> 2.39.2
>