[PATCH v4 0/4] Miscellaneous fixes for pci subsystem

[PATCH v4 0/4] Miscellaneous fixes for pci subsystem

[Ziming Du]

Miscellaneous fixes for pci subsystem

Changes in v4:
- Remove the architecture-specific #ifdef to apply the alignment
  check for all platforms (including x86), as device registers are
  naturally aligned anyway.
- Fix a potential issue in proc_bus_pci_read() to make it consistent
  with proc_bus_pci_write(), as suggested by Ilpo Järvinen.
- Link to v3: https://lore.kernel.org/linux-pci/20260108015944.3520719-1-duziming2@huawei.com/

Changes in v3:
- Check *ppos before assign it to pos.
- Link to v2: https://lore.kernel.org/linux-pci/20251224092721.2034529-1-duziming2@huawei.com/

Changes in v2:
- Correct grammer and indentation.
- Remove unrelated stack traces from the commit message.
- Modify the handling of pos by adding a non-negative check to ensure
  that the input value is valid.
- Use the existing IS_ALIGNED macro and ensure that after modification,
  other cases still retuen -EINVAL as before.
- Link to v1: https://lore.kernel.org/linux-pci/20251216083912.758219-1-duziming2@huawei.com/

Yongqiang Liu (2):
  PCI: Prevent overflow in proc_bus_pci_write()
  PCI/sysfs: Prohibit unaligned access to I/O port
Ziming Du (2):
  PCI/sysfs: Fix null pointer dereference during hotplug
  PCI: Prevent overflow in proc_bus_pci_read()

 drivers/pci/pci-sysfs.c |  7 +++++++
 drivers/pci/proc.c      | 11 +++++++++--
 2 files changed, 16 insertions(+), 2 deletions(-)

-- 
2.43.0

[PATCH v4 1/4] PCI/sysfs: Prohibit unaligned access to I/O port

[Ziming Du]

Unaligned access is harmful for non-x86 archs such as arm64. When we
use pwrite or pread to access the I/O port resources with unaligned
offset, system will crash as follows:

Unable to handle kernel paging request at virtual address fffffbfffe8010c1
Internal error: Oops: 0000000096000061 [#1] SMP
Call trace:
 _outw include/asm-generic/io.h:594 [inline]
 logic_outw+0x54/0x218 lib/logic_pio.c:305
 pci_resource_io drivers/pci/pci-sysfs.c:1157 [inline]
 pci_write_resource_io drivers/pci/pci-sysfs.c:1191 [inline]
 pci_write_resource_io+0x208/0x260 drivers/pci/pci-sysfs.c:1181
 sysfs_kf_bin_write+0x188/0x210 fs/sysfs/file.c:158
 kernfs_fop_write_iter+0x2e8/0x4b0 fs/kernfs/file.c:338
 vfs_write+0x7bc/0xac8 fs/read_write.c:586
 ksys_write+0x12c/0x270 fs/read_write.c:639
 __arm64_sys_write+0x78/0xb8 fs/read_write.c:648

Although x86 might handle unaligned I/O accesses by splitting cycles,
this approach is still limited because PCI device registers typically
expect natural alignment. A global prohibition of unaligned accesses
ensures consistent behavior across all architectures and prevents
unexpected hardware side effects.

Fixes: 8633328be242 ("PCI: Allow read/write access to sysfs I/O port resources")
Signed-off-by: Yongqiang Liu <liuyongqiang13@huawei.com>
Signed-off-by: Ziming Du <duziming2@huawei.com>
Suggested-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
---
 drivers/pci/pci-sysfs.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/pci/pci-sysfs.c b/drivers/pci/pci-sysfs.c
index c2df915ad2d29..18e5d4603b472 100644
--- a/drivers/pci/pci-sysfs.c
+++ b/drivers/pci/pci-sysfs.c
@@ -31,6 +31,7 @@
 #include <linux/of.h>
 #include <linux/aperture.h>
 #include <linux/unaligned.h>
+#include <linux/align.h>
 #include "pci.h"
 
 #ifndef ARCH_PCI_DEV_GROUPS
@@ -1166,12 +1167,16 @@ static ssize_t pci_resource_io(struct file *filp, struct kobject *kobj,
 			*(u8 *)buf = inb(port);
 		return 1;
 	case 2:
+		if (!IS_ALIGNED(port, count))
+			return -EINVAL;
 		if (write)
 			outw(*(u16 *)buf, port);
 		else
 			*(u16 *)buf = inw(port);
 		return 2;
 	case 4:
+		if (!IS_ALIGNED(port, count))
+			return -EINVAL;
 		if (write)
 			outl(*(u32 *)buf, port);
 		else
-- 
2.43.0

[PATCH v4 2/4] PCI/sysfs: Fix null pointer dereference during hotplug

[Ziming Du]

During the concurrent process of creating and rescanning in VF, the
resource files for the same pci_dev may be created twice. The second
creation attempt fails, resulting the res_attr in pci_dev to kfree(),
but the pointer is not set to NULL. This will subsequently lead to
dereferencing a null pointer when removing the device.

When we perform the following operation:
  echo $sriov_totalvfs > /sys/class/net/"$pfname"/device/sriov_numvfs &
  sleep 0.5
  echo 1 > /sys/bus/pci/rescan
  pci_remove "$pfname"
system will crash as follows:

  Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
  Call trace:
   __pi_strlen+0x14/0x150
   kernfs_find_ns+0x54/0x120
   kernfs_remove_by_name_ns+0x58/0xf0
   sysfs_remove_bin_file+0x24/0x38
   pci_remove_resource_files+0x44/0x90
   pci_remove_sysfs_dev_files+0x28/0x40
   pci_stop_bus_device+0xb8/0x118
   pci_stop_and_remove_bus_device+0x20/0x40
   pci_iov_remove_virtfn+0xb8/0x138
   sriov_disable+0xbc/0x190
   pci_disable_sriov+0x30/0x48
   hinic_pci_sriov_disable+0x54/0x138 [hinic]
   hinic_remove+0x140/0x290 [hinic]
   pci_device_remove+0x4c/0xf8
   device_remove+0x54/0x90
   device_release_driver_internal+0x1d4/0x238
   device_release_driver+0x20/0x38
   pci_stop_bus_device+0xa8/0x118
   pci_stop_and_remove_bus_device_locked+0x28/0x50
   remove_store+0x128/0x208

Fix this by set the pointer to NULL after releasing 'res_attr' immediately.

Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Signed-off-by: Ziming Du <duziming2@huawei.com>
---
 drivers/pci/pci-sysfs.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/pci/pci-sysfs.c b/drivers/pci/pci-sysfs.c
index 18e5d4603b472..fbcbf39232732 100644
--- a/drivers/pci/pci-sysfs.c
+++ b/drivers/pci/pci-sysfs.c
@@ -1227,12 +1227,14 @@ static void pci_remove_resource_files(struct pci_dev *pdev)
 		if (res_attr) {
 			sysfs_remove_bin_file(&pdev->dev.kobj, res_attr);
 			kfree(res_attr);
+			pdev->res_attr[i] = NULL;
 		}
 
 		res_attr = pdev->res_attr_wc[i];
 		if (res_attr) {
 			sysfs_remove_bin_file(&pdev->dev.kobj, res_attr);
 			kfree(res_attr);
+			pdev->res_attr_wc[i] = NULL;
 		}
 	}
 }
-- 
2.43.0

[PATCH v4 3/4] PCI: Prevent overflow in proc_bus_pci_write()

[Ziming Du]

From: Yongqiang Liu <liuyongqiang13@huawei.com>

When the value of *ppos over the INT_MAX, the pos is over set to a
negative value which will be passed to get_user() or
pci_user_write_config_dword(). Unexpected behavior such as a soft lockup
will happen as follows:

 watchdog: BUG: soft lockup - CPU#0 stuck for 130s! [syz.3.109:3444]
 RIP: 0010:_raw_spin_unlock_irq+0x17/0x30
 Call Trace:
  <TASK>
  pci_user_write_config_dword+0x126/0x1f0
  proc_bus_pci_write+0x273/0x470
  proc_reg_write+0x1b6/0x280
  do_iter_write+0x48e/0x790
  vfs_writev+0x125/0x4a0
  __x64_sys_pwritev+0x1e2/0x2a0
  do_syscall_64+0x59/0x110
  entry_SYSCALL_64_after_hwframe+0x78/0xe2

Fix this by adding a non-negative check before assign *ppos to pos.

Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Signed-off-by: Yongqiang Liu <liuyongqiang13@huawei.com>
Signed-off-by: Ziming Du <duziming2@huawei.com>
Reviewed-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
---
 drivers/pci/proc.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/drivers/pci/proc.c b/drivers/pci/proc.c
index 9348a0fb80847..2d51b26edbe74 100644
--- a/drivers/pci/proc.c
+++ b/drivers/pci/proc.c
@@ -113,10 +113,14 @@ static ssize_t proc_bus_pci_write(struct file *file, const char __user *buf,
 {
 	struct inode *ino = file_inode(file);
 	struct pci_dev *dev = pde_data(ino);
-	int pos = *ppos;
+	int pos;
 	int size = dev->cfg_size;
 	int cnt, ret;
 
+	if (*ppos > INT_MAX)
+		return -EINVAL;
+	pos = *ppos;
+
 	ret = security_locked_down(LOCKDOWN_PCI_ACCESS);
 	if (ret)
 		return ret;
-- 
2.43.0

[PATCH v4 4/4] PCI: Prevent overflow in proc_bus_pci_read()

[Ziming Du]

proc_bus_pci_read() assigns *ppos directly to an unsigned integer variable.
For large offsets, this implicit conversion may truncate the value and
cause reads from an incorrect position.

proc_bus_pci_write() explicitly validates *ppos and rejects values larger
than INT_MAX, while proc_bus_pci_read() currently accepts them. This
difference in position handling is unjustified.

Fix this by validating *ppos in proc_bus_pci_read() and rejecting offsets
larger than INT_MAX before the assignment, matching proc_bus_pci_write().

Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Signed-off-by: Ziming Du <duziming2@huawei.com>
Suggested-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
---
 drivers/pci/proc.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/pci/proc.c b/drivers/pci/proc.c
index 2d51b26edbe74..f4ef7629dc78b 100644
--- a/drivers/pci/proc.c
+++ b/drivers/pci/proc.c
@@ -29,9 +29,12 @@ static ssize_t proc_bus_pci_read(struct file *file, char __user *buf,
 				 size_t nbytes, loff_t *ppos)
 {
 	struct pci_dev *dev = pde_data(file_inode(file));
-	unsigned int pos = *ppos;
+	int pos;
 	unsigned int cnt, size;
 
+	if (*ppos > INT_MAX)
+		return -EINVAL;
+	pos = *ppos;
 	/*
 	 * Normal users can read only the standardized portion of the
 	 * configuration space as several chips lock up when trying to read
-- 
2.43.0

Re: [PATCH v4 0/4] Miscellaneous fixes for pci subsystem

[duziming]

Hi all,

Gentle ping on this patchset. Any feedback would be greatly appreciated.

On 2026/1/15 15:52, Ziming Du wrote:

> Miscellaneous fixes for pci subsystem
>
> Changes in v4:
> - Remove the architecture-specific #ifdef to apply the alignment
>    check for all platforms (including x86), as device registers are
>    naturally aligned anyway.
> - Fix a potential issue in proc_bus_pci_read() to make it consistent
>    with proc_bus_pci_write(), as suggested by Ilpo Järvinen.
> - Link to v3:https://lore.kernel.org/linux-pci/20260108015944.3520719-1-duziming2@huawei.com/
>
> Changes in v3:
> - Check *ppos before assign it to pos.
> - Link to v2:https://lore.kernel.org/linux-pci/20251224092721.2034529-1-duziming2@huawei.com/
>
> Changes in v2:
> - Correct grammer and indentation.
> - Remove unrelated stack traces from the commit message.
> - Modify the handling of pos by adding a non-negative check to ensure
>    that the input value is valid.
> - Use the existing IS_ALIGNED macro and ensure that after modification,
>    other cases still retuen -EINVAL as before.
> - Link to v1:https://lore.kernel.org/linux-pci/20251216083912.758219-1-duziming2@huawei.com/
>
> Yongqiang Liu (2):
>    PCI: Prevent overflow in proc_bus_pci_write()
>    PCI/sysfs: Prohibit unaligned access to I/O port
> Ziming Du (2):
>    PCI/sysfs: Fix null pointer dereference during hotplug
>    PCI: Prevent overflow in proc_bus_pci_read()
>
>   drivers/pci/pci-sysfs.c |  7 +++++++
>   drivers/pci/proc.c      | 11 +++++++++--
>   2 files changed, 16 insertions(+), 2 deletions(-)

Thanx du!

Re: [PATCH v4 0/4] Miscellaneous fixes for pci subsystem

[Bjorn Helgaas]

On Fri, Jan 30, 2026 at 03:53:10PM +0800, duziming wrote:
> Hi all,
> 
> Gentle ping on this patchset. Any feedback would be greatly appreciated.

Sorry I didn't get to this.  Poke again after v6.20-rc1.

> On 2026/1/15 15:52, Ziming Du wrote:
> 
> > Miscellaneous fixes for pci subsystem
> > 
> > Changes in v4:
> > - Remove the architecture-specific #ifdef to apply the alignment
> >    check for all platforms (including x86), as device registers are
> >    naturally aligned anyway.
> > - Fix a potential issue in proc_bus_pci_read() to make it consistent
> >    with proc_bus_pci_write(), as suggested by Ilpo Järvinen.
> > - Link to v3:https://lore.kernel.org/linux-pci/20260108015944.3520719-1-duziming2@huawei.com/
> > 
> > Changes in v3:
> > - Check *ppos before assign it to pos.
> > - Link to v2:https://lore.kernel.org/linux-pci/20251224092721.2034529-1-duziming2@huawei.com/
> > 
> > Changes in v2:
> > - Correct grammer and indentation.
> > - Remove unrelated stack traces from the commit message.
> > - Modify the handling of pos by adding a non-negative check to ensure
> >    that the input value is valid.
> > - Use the existing IS_ALIGNED macro and ensure that after modification,
> >    other cases still retuen -EINVAL as before.
> > - Link to v1:https://lore.kernel.org/linux-pci/20251216083912.758219-1-duziming2@huawei.com/
> > 
> > Yongqiang Liu (2):
> >    PCI: Prevent overflow in proc_bus_pci_write()
> >    PCI/sysfs: Prohibit unaligned access to I/O port
> > Ziming Du (2):
> >    PCI/sysfs: Fix null pointer dereference during hotplug
> >    PCI: Prevent overflow in proc_bus_pci_read()
> > 
> >   drivers/pci/pci-sysfs.c |  7 +++++++
> >   drivers/pci/proc.c      | 11 +++++++++--
> >   2 files changed, 16 insertions(+), 2 deletions(-)
> 
> Thanx du!
> 

Re: [PATCH v4 0/4] Miscellaneous fixes for pci subsystem

[duziming]

在 2026/2/7 6:29, Bjorn Helgaas 写道:
> On Fri, Jan 30, 2026 at 03:53:10PM +0800, duziming wrote:
>> Hi all,
>>
>> Gentle ping on this patchset. Any feedback would be greatly appreciated.
> Sorry I didn't get to this.  Poke again after v6.20-rc1.
ping？
>> On 2026/1/15 15:52, Ziming Du wrote:
>>
>>> Miscellaneous fixes for pci subsystem
>>>
>>> Changes in v4:
>>> - Remove the architecture-specific #ifdef to apply the alignment
>>>     check for all platforms (including x86), as device registers are
>>>     naturally aligned anyway.
>>> - Fix a potential issue in proc_bus_pci_read() to make it consistent
>>>     with proc_bus_pci_write(), as suggested by Ilpo Järvinen.
>>> - Link to v3:https://lore.kernel.org/linux-pci/20260108015944.3520719-1-duziming2@huawei.com/
>>>
>>> Changes in v3:
>>> - Check *ppos before assign it to pos.
>>> - Link to v2:https://lore.kernel.org/linux-pci/20251224092721.2034529-1-duziming2@huawei.com/
>>>
>>> Changes in v2:
>>> - Correct grammer and indentation.
>>> - Remove unrelated stack traces from the commit message.
>>> - Modify the handling of pos by adding a non-negative check to ensure
>>>     that the input value is valid.
>>> - Use the existing IS_ALIGNED macro and ensure that after modification,
>>>     other cases still retuen -EINVAL as before.
>>> - Link to v1:https://lore.kernel.org/linux-pci/20251216083912.758219-1-duziming2@huawei.com/
>>>
>>> Yongqiang Liu (2):
>>>     PCI: Prevent overflow in proc_bus_pci_write()
>>>     PCI/sysfs: Prohibit unaligned access to I/O port
>>> Ziming Du (2):
>>>     PCI/sysfs: Fix null pointer dereference during hotplug
>>>     PCI: Prevent overflow in proc_bus_pci_read()
>>>
>>>    drivers/pci/pci-sysfs.c |  7 +++++++
>>>    drivers/pci/proc.c      | 11 +++++++++--
>>>    2 files changed, 16 insertions(+), 2 deletions(-)
>> Thanx du!
>>

Re: [PATCH v4 1/4] PCI/sysfs: Prohibit unaligned access to I/O port

[Bjorn Helgaas]

On Fri, Jan 16, 2026 at 04:17:18PM +0800, Ziming Du wrote:
> Unaligned access is harmful for non-x86 archs such as arm64. When we
> use pwrite or pread to access the I/O port resources with unaligned
> offset, system will crash as follows:

> @@ -1166,12 +1167,16 @@ static ssize_t pci_resource_io(struct file *filp, struct kobject *kobj,
>  			*(u8 *)buf = inb(port);
>  		return 1;
>  	case 2:
> +		if (!IS_ALIGNED(port, count))
> +			return -EINVAL;

I assume "IS_ALIGNED(port, 1)" is *always* true, so can we just do
this once before the switch instead of adding it to the "case 2" and
"case 4"?

>  		if (write)
>  			outw(*(u16 *)buf, port);
>  		else
>  			*(u16 *)buf = inw(port);
>  		return 2;
>  	case 4:
> +		if (!IS_ALIGNED(port, count))
> +			return -EINVAL;
>  		if (write)
>  			outl(*(u32 *)buf, port);
>  		else
> -- 
> 2.43.0
> 

Re: [PATCH v4 2/4] PCI/sysfs: Fix null pointer dereference during hotplug

[Bjorn Helgaas]

On Fri, Jan 16, 2026 at 04:17:19PM +0800, Ziming Du wrote:
> During the concurrent process of creating and rescanning in VF, the
> resource files for the same pci_dev may be created twice.

Where are the two resource file creations?  This will help review the
patch.

> The second
> creation attempt fails, resulting the res_attr in pci_dev to kfree(),
> but the pointer is not set to NULL. This will subsequently lead to
> dereferencing a null pointer when removing the device.
> 
> When we perform the following operation:
>   echo $sriov_totalvfs > /sys/class/net/"$pfname"/device/sriov_numvfs &

I think it would be more informative to include an actual sample here.
We can easily substitute the device names and numbers, given a
concrete example.  It's a little bit harder to intuit what $pfname and
$sriov_totalvfs should be.  E.g.,

  $ cat /sys/bus/pci/devices/0000:02:00.0/sriov_totalvfs
  128
  $ echo 128 > /sys/bus/pci/devices/0000:02:00.0/sriov_numvfs &

Unless it's important to use /sys/class/net/..., use
/sys/bus/pci/devices/... both places to make it simpler.

>   sleep 0.5
>   echo 1 > /sys/bus/pci/rescan

These look like shell commands ...

>   pci_remove "$pfname"

but what is "pci_remove"?  I guess it must be an echo into
/sys/bus/pci/devices/.../remove; expanding it here would be better.

> system will crash as follows:
> 
>   Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
>   Call trace:
>    __pi_strlen+0x14/0x150
>    kernfs_find_ns+0x54/0x120
>    kernfs_remove_by_name_ns+0x58/0xf0
>    sysfs_remove_bin_file+0x24/0x38
>    pci_remove_resource_files+0x44/0x90
>    pci_remove_sysfs_dev_files+0x28/0x40
>    pci_stop_bus_device+0xb8/0x118
>    pci_stop_and_remove_bus_device+0x20/0x40
>    pci_iov_remove_virtfn+0xb8/0x138
>    sriov_disable+0xbc/0x190
>    pci_disable_sriov+0x30/0x48
>    hinic_pci_sriov_disable+0x54/0x138 [hinic]
>    hinic_remove+0x140/0x290 [hinic]
>    pci_device_remove+0x4c/0xf8
>    device_remove+0x54/0x90
>    device_release_driver_internal+0x1d4/0x238
>    device_release_driver+0x20/0x38
>    pci_stop_bus_device+0xa8/0x118
>    pci_stop_and_remove_bus_device_locked+0x28/0x50
>    remove_store+0x128/0x208
> 
> Fix this by set the pointer to NULL after releasing 'res_attr' immediately.

This *sounds* like it would still be racy unless there's a lock around
this.  If there is a lock, please mention what it is and where it's
held.

> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> Signed-off-by: Ziming Du <duziming2@huawei.com>
> ---
>  drivers/pci/pci-sysfs.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/drivers/pci/pci-sysfs.c b/drivers/pci/pci-sysfs.c
> index 18e5d4603b472..fbcbf39232732 100644
> --- a/drivers/pci/pci-sysfs.c
> +++ b/drivers/pci/pci-sysfs.c
> @@ -1227,12 +1227,14 @@ static void pci_remove_resource_files(struct pci_dev *pdev)
>  		if (res_attr) {
>  			sysfs_remove_bin_file(&pdev->dev.kobj, res_attr);
>  			kfree(res_attr);
> +			pdev->res_attr[i] = NULL;
>  		}
>  
>  		res_attr = pdev->res_attr_wc[i];
>  		if (res_attr) {
>  			sysfs_remove_bin_file(&pdev->dev.kobj, res_attr);
>  			kfree(res_attr);
> +			pdev->res_attr_wc[i] = NULL;
>  		}
>  	}
>  }
> -- 
> 2.43.0
> 

Re: [PATCH v4 2/4] PCI/sysfs: Fix null pointer dereference during hotplug

[duziming]

在 2026/2/27 1:14, Bjorn Helgaas 写道:
> On Fri, Jan 16, 2026 at 04:17:19PM +0800, Ziming Du wrote:
>> During the concurrent process of creating and rescanning in VF, the
>> resource files for the same pci_dev may be created twice.
> Where are the two resource file creations?  This will help review the
> patch.

The process of creating VFs:

sriov_numvfs_store
     hinic_pci_sriov_configure
         hinic_pci_sriov_enable
             pci_enable_sriov
                 sriov_enable
                     sriov_add_vfs
                         pci_iov_add_virtfn
                             pci_bus_add_device
                                 pci_create_sysfs_dev_files
                                     pci_create_resource_files

The process of rescanning VFs:

rescan_store
     pci_rescan_bus
         pci_bus_add_devices
             pci_bus_add_device
                 pci_create_sysfs_dev_files
                     pci_create_resource_files

>> The second
>> creation attempt fails, resulting the res_attr in pci_dev to kfree(),
>> but the pointer is not set to NULL. This will subsequently lead to
>> dereferencing a null pointer when removing the device.
>>
>> When we perform the following operation:
>>    echo $sriov_totalvfs > /sys/class/net/"$pfname"/device/sriov_numvfs &
> I think it would be more informative to include an actual sample here.
> We can easily substitute the device names and numbers, given a
> concrete example.  It's a little bit harder to intuit what $pfname and
> $sriov_totalvfs should be.  E.g.,
>
>    $ cat /sys/bus/pci/devices/0000:02:00.0/sriov_totalvfs
>    128
>    $ echo 128 > /sys/bus/pci/devices/0000:02:00.0/sriov_numvfs &
>
> Unless it's important to use /sys/class/net/..., use
> /sys/bus/pci/devices/... both places to make it simpler.
>
>>    sleep 0.5
>>    echo 1 > /sys/bus/pci/rescan
> These look like shell commands ...
>
>>    pci_remove "$pfname"
> but what is "pci_remove"?  I guess it must be an echo into
> /sys/bus/pci/devices/.../remove; expanding it here would be better.

Yes, you are right, based on your comments, I have revised it as follows:

$ cat /sys/bus/pci/devices/0000:02:00.0/sriov_totalvfs
128
$ echo 128 > /sys/bus/pci/devices/0000:02:00.0/sriov_numvfs &
$ sleep 0.5
$ echo 1 > /sys/bus/pci/rescan
$ echo 1 > /sys/bus/pci/devices/0000:02:00.0/remove

>> system will crash as follows:
>>
>>    Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
>>    Call trace:
>>     __pi_strlen+0x14/0x150
>>     kernfs_find_ns+0x54/0x120
>>     kernfs_remove_by_name_ns+0x58/0xf0
>>     sysfs_remove_bin_file+0x24/0x38
>>     pci_remove_resource_files+0x44/0x90
>>     pci_remove_sysfs_dev_files+0x28/0x40
>>     pci_stop_bus_device+0xb8/0x118
>>     pci_stop_and_remove_bus_device+0x20/0x40
>>     pci_iov_remove_virtfn+0xb8/0x138
>>     sriov_disable+0xbc/0x190
>>     pci_disable_sriov+0x30/0x48
>>     hinic_pci_sriov_disable+0x54/0x138 [hinic]
>>     hinic_remove+0x140/0x290 [hinic]
>>     pci_device_remove+0x4c/0xf8
>>     device_remove+0x54/0x90
>>     device_release_driver_internal+0x1d4/0x238
>>     device_release_driver+0x20/0x38
>>     pci_stop_bus_device+0xa8/0x118
>>     pci_stop_and_remove_bus_device_locked+0x28/0x50
>>     remove_store+0x128/0x208
>>
>> Fix this by set the pointer to NULL after releasing 'res_attr' immediately.
> This *sounds* like it would still be racy unless there's a lock around
> this.  If there is a lock, please mention what it is and where it's
> held.

The rescan process holds pci_lock_rescan_remove(), while the creation 
process

holds "device_lock(&pdev->dev)"

>> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
>> Signed-off-by: Ziming Du <duziming2@huawei.com>
>> ---
>>   drivers/pci/pci-sysfs.c | 2 ++
>>   1 file changed, 2 insertions(+)
>>
>> diff --git a/drivers/pci/pci-sysfs.c b/drivers/pci/pci-sysfs.c
>> index 18e5d4603b472..fbcbf39232732 100644
>> --- a/drivers/pci/pci-sysfs.c
>> +++ b/drivers/pci/pci-sysfs.c
>> @@ -1227,12 +1227,14 @@ static void pci_remove_resource_files(struct pci_dev *pdev)
>>   		if (res_attr) {
>>   			sysfs_remove_bin_file(&pdev->dev.kobj, res_attr);
>>   			kfree(res_attr);
>> +			pdev->res_attr[i] = NULL;
>>   		}
>>   
>>   		res_attr = pdev->res_attr_wc[i];
>>   		if (res_attr) {
>>   			sysfs_remove_bin_file(&pdev->dev.kobj, res_attr);
>>   			kfree(res_attr);
>> +			pdev->res_attr_wc[i] = NULL;
>>   		}
>>   	}
>>   }
>> -- 
>> 2.43.0
>>

Re: [PATCH v4 3/4] PCI: Prevent overflow in proc_bus_pci_write()

[Bjorn Helgaas]

On Fri, Jan 16, 2026 at 04:17:20PM +0800, Ziming Du wrote:
> From: Yongqiang Liu <liuyongqiang13@huawei.com>
> 
> When the value of *ppos over the INT_MAX, the pos is over set to a
> negative value which will be passed to get_user() or
> pci_user_write_config_dword(). Unexpected behavior such as a soft lockup
> will happen as follows:

I think it's crazy to worry about offsets overflowing INT_MAX.  We're
doing PCI config accesses.  Config space is only 4K at most, so we
already have a much smaller upper bound on the offset.

The procfs proc_bus_pci_write() is essentially the same as the sysfs
pci_write_config().  They should share some common implementation.

>  watchdog: BUG: soft lockup - CPU#0 stuck for 130s! [syz.3.109:3444]
>  RIP: 0010:_raw_spin_unlock_irq+0x17/0x30
>  Call Trace:
>   <TASK>
>   pci_user_write_config_dword+0x126/0x1f0
>   proc_bus_pci_write+0x273/0x470
>   proc_reg_write+0x1b6/0x280
>   do_iter_write+0x48e/0x790
>   vfs_writev+0x125/0x4a0
>   __x64_sys_pwritev+0x1e2/0x2a0
>   do_syscall_64+0x59/0x110
>   entry_SYSCALL_64_after_hwframe+0x78/0xe2
> 
> Fix this by adding a non-negative check before assign *ppos to pos.
> 
> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> Signed-off-by: Yongqiang Liu <liuyongqiang13@huawei.com>
> Signed-off-by: Ziming Du <duziming2@huawei.com>
> Reviewed-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
> ---
>  drivers/pci/proc.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/pci/proc.c b/drivers/pci/proc.c
> index 9348a0fb80847..2d51b26edbe74 100644
> --- a/drivers/pci/proc.c
> +++ b/drivers/pci/proc.c
> @@ -113,10 +113,14 @@ static ssize_t proc_bus_pci_write(struct file *file, const char __user *buf,
>  {
>  	struct inode *ino = file_inode(file);
>  	struct pci_dev *dev = pde_data(ino);
> -	int pos = *ppos;
> +	int pos;
>  	int size = dev->cfg_size;
>  	int cnt, ret;
>  
> +	if (*ppos > INT_MAX)
> +		return -EINVAL;
> +	pos = *ppos;

The first issue here is that "*ppos" (loff_t) is long long, but "pos"
is int, so we're assigning a 64-bit value to a 32-bit container and
losing any high bits.  So an offset of 0x100000000 is incorrectly
treated as valid.

A change like this should fix that and bring this closer to the
pci_write_config() implementation:

  - int pos = *ppos;
  + loff_t pos = *ppos;

  - if (pos >= size)
  + if (pos > dev->cfg_size)
      return 0;

There's also a second issue here:

        if (pos + nbytes > size)
                nbytes = size - pos;

"pos" is a signed int, "nbytes" is size_t, which is an *unsigned* int,
so "pos" is implicitly converted to an unsigned value.  I think this
is what causes the soft lockup you reported because an offset like the
0x80800001 in your test case is converted from signed -2139095039 to
unsigned 2155872257.  "size" is dev->cfg_size, e.g., 4096, so
2155872257 + nbytes is certainly larger than 4096, so nbytes ends up
being set to some huge unsigned size_t value.

This issue would probably be avoided simply by returning early when
"pos" is out of range.  But mixing signed and unsigned in that
"pos + nbytes" expression is just asking for trouble and we should
avoid it as pci_write_config() does.

So I'd like to see something that makes the procfs accessor
validations look like the sysfs accessors.  It's a little messy
because they use different names, so the patches will be ugly.  But I
think it's worth it to make them work the same way so we don't have to
analyze them separately.

Maybe could be done in a couple steps, e.g., one to simply rename
things and a second to make the functional changes.

Bjorn

Re: [PATCH v4 2/4] PCI/sysfs: Fix null pointer dereference during hotplug

[duziming]

在 2026/2/27 1:14, Bjorn Helgaas 写道:
> On Fri, Jan 16, 2026 at 04:17:19PM +0800, Ziming Du wrote:
>> During the concurrent process of creating and rescanning in VF, the
>> resource files for the same pci_dev may be created twice.
> Where are the two resource file creations?  This will help review the
> patch.
>
>> The second
>> creation attempt fails, resulting the res_attr in pci_dev to kfree(),
>> but the pointer is not set to NULL. This will subsequently lead to
>> dereferencing a null pointer when removing the device.
>>
>> When we perform the following operation:
>>    echo $sriov_totalvfs > /sys/class/net/"$pfname"/device/sriov_numvfs &
> I think it would be more informative to include an actual sample here.
> We can easily substitute the device names and numbers, given a
> concrete example.  It's a little bit harder to intuit what $pfname and
> $sriov_totalvfs should be.  E.g.,
>
>    $ cat /sys/bus/pci/devices/0000:02:00.0/sriov_totalvfs
>    128
>    $ echo 128 > /sys/bus/pci/devices/0000:02:00.0/sriov_numvfs &
>
> Unless it's important to use /sys/class/net/..., use
> /sys/bus/pci/devices/... both places to make it simpler.
>
>>    sleep 0.5
>>    echo 1 > /sys/bus/pci/rescan
> These look like shell commands ...
>
>>    pci_remove "$pfname"
> but what is "pci_remove"?  I guess it must be an echo into
> /sys/bus/pci/devices/.../remove; expanding it here would be better.
>
>> system will crash as follows:
>>
>>    Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
>>    Call trace:
>>     __pi_strlen+0x14/0x150
>>     kernfs_find_ns+0x54/0x120
>>     kernfs_remove_by_name_ns+0x58/0xf0
>>     sysfs_remove_bin_file+0x24/0x38
>>     pci_remove_resource_files+0x44/0x90
>>     pci_remove_sysfs_dev_files+0x28/0x40
>>     pci_stop_bus_device+0xb8/0x118
>>     pci_stop_and_remove_bus_device+0x20/0x40
>>     pci_iov_remove_virtfn+0xb8/0x138
>>     sriov_disable+0xbc/0x190
>>     pci_disable_sriov+0x30/0x48
>>     hinic_pci_sriov_disable+0x54/0x138 [hinic]
>>     hinic_remove+0x140/0x290 [hinic]
>>     pci_device_remove+0x4c/0xf8
>>     device_remove+0x54/0x90
>>     device_release_driver_internal+0x1d4/0x238
>>     device_release_driver+0x20/0x38
>>     pci_stop_bus_device+0xa8/0x118
>>     pci_stop_and_remove_bus_device_locked+0x28/0x50
>>     remove_store+0x128/0x208
>>
>> Fix this by set the pointer to NULL after releasing 'res_attr' immediately.
> This *sounds* like it would still be racy unless there's a lock around
> this.  If there is a lock, please mention what it is and where it's
> held.

I found that the primary race condition between VF creation and PCI 
rescan has been

addressed by the pci_lock_rescan_remove() lock added in commit 
a5338e365c45.

Given this, would setting the pointer to NULL after kfree still be 
considered a worthwhile defensive measure?

>> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
>> Signed-off-by: Ziming Du <duziming2@huawei.com>
>> ---
>>   drivers/pci/pci-sysfs.c | 2 ++
>>   1 file changed, 2 insertions(+)
>>
>> diff --git a/drivers/pci/pci-sysfs.c b/drivers/pci/pci-sysfs.c
>> index 18e5d4603b472..fbcbf39232732 100644
>> --- a/drivers/pci/pci-sysfs.c
>> +++ b/drivers/pci/pci-sysfs.c
>> @@ -1227,12 +1227,14 @@ static void pci_remove_resource_files(struct pci_dev *pdev)
>>   		if (res_attr) {
>>   			sysfs_remove_bin_file(&pdev->dev.kobj, res_attr);
>>   			kfree(res_attr);
>> +			pdev->res_attr[i] = NULL;
>>   		}
>>   
>>   		res_attr = pdev->res_attr_wc[i];
>>   		if (res_attr) {
>>   			sysfs_remove_bin_file(&pdev->dev.kobj, res_attr);
>>   			kfree(res_attr);
>> +			pdev->res_attr_wc[i] = NULL;
>>   		}
>>   	}
>>   }
>> -- 
>> 2.43.0
>>