[BUG] sysfs: duplicate resource file creation during PCIe rescan

[moubingquan]

Hi all,

When uninstalling the driver on the ARM64 platform and invoking `sriov_disable()`,
another thread simultaneously performed `echo 1 > /sys/bus/pci/rescan`,
which resulted in a call trace error (`sysfs: cannot create duplicate filename`) when subsequently loading the driver.

Under certain multi-threaded scenarios (e.g. VF teardown + PCI rescan triggered in parallel),
The following sequence may result in files appearing in sysfs that should not exist:

1. sriov_disable() uninstalls VFs and removes the corresponding VF files in sysfs.
2.At the same time, when rescan_store() rescan the entire PCI device tree,
there is a possibility that VF files that should have been deleted are added back,
resulting in VF files in sysfs that should have been removed but were not.

Tested on:
- Kernel version: Linux version 6.15.0-rc4+ (phisik3 at 10-50-163-153-ARM-openEuler22-3)
- Platform: ARM64 (Huawei Kunpeng920)
- Repro steps:
1.Thread A unloads the driver and VF (requires calling sriov_disable()).
2.Thread B calls `echo 1 > /sys/bus/pci/rescan `

The system will report a call trace as follows:

sysfs: cannot create duplicate filename '/devices/pci0000:3c/0000:3c:01.0/0000:3e:00.2/resource2'
CPU: 138 UID: 0 PID: 11067 Comm: sh Kdump: loaded Tainted: G      D W  O        6.15.0-rc4+ #1 PREEMPT
Tainted: [D]=DIE, [W]=WARN, [O]=OOT_MODULE
Call trace:
  show_stack+0x20/0x38 (C)
  dump_stack_lvl+0x80/0xf8
  dump_stack+0x18/0x28
  sysfs_warn_dup+0x6c/0x90
  sysfs_add_bin_file_mode_ns+0x12c/0x178
  sysfs_create_bin_file+0x7c/0xb8
  pci_create_attr+0x104/0x1b0
  pci_create_resource_files.part.0+0x50/0xd0
  pci_create_sysfs_dev_files+0x30/0x50
  pci_bus_add_device+0x40/0x120
  pci_bus_add_devices+0x40/0x98
  pci_bus_add_devices+0x6c/0x98
  pci_rescan_bus+0x38/0x58
  rescan_store+0x80/0xb0
  bus_attr_store+0x2c/0x48
  sysfs_kf_write+0x84/0xa8
  kernfs_fop_write_iter+0x120/0x1b8
  vfs_write+0x338/0x3f8
  ksys_write+0x70/0x110
  __arm64_sys_write+0x24/0x38
  invoke_syscall+0x50/0x120
  el0_svc_common.constprop.0+0xc8/0xf0
  do_el0_svc+0x24/0x38
  el0_svc+0x34/0xf0
  el0t_64_sync_handler+0xc8/0xd0
  el0t_64_sync+0x1ac/0x1b0

The general analysis and corresponding code are as follows:

drivers/pci/iov.c

736 static void sriov_disable(struct pci_dev *dev)
737 {
738         struct pci_sriov *iov = dev->sriov;
739
740         if (!iov->num_VFs)
741                 return;
742
743         sriov_del_vfs(dev);
744         iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE);
745         pci_cfg_access_lock(dev);
746         pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
747         ssleep(1);
748         pci_cfg_access_unlock(dev);
749
750         pcibios_sriov_disable(dev);
751
752         if (iov->link != dev->devfn)
753                 sysfs_remove_link(&dev->dev.kobj, "dep_link");
754
755         iov->num_VFs = 0;
756         pci_iov_set_numvfs(dev, 0);
757 }

sriov_disable() will unload the VF and remove its files from sysfs.

drivers/pci/pci-sysfs.c

  435 static ssize_t rescan_store(const struct bus_type *bus, const char *buf, size_t count)
  436 {
  437         unsigned long val;
  438         struct pci_bus *b = NULL;
  439
  440         if (kstrtoul(buf, 0, &val) < 0)
  441                 return -EINVAL;
  442
  443         if (val) {
  444                 pci_lock_rescan_remove();
  445                 while ((b = pci_find_next_bus(b)) != NULL)
  446                         pci_rescan_bus(b);
  447                 pci_unlock_rescan_remove();
  448         }
  449         return count;
  450 }
  451 static BUS_ATTR_WO(rescan);

The `rescan_store()` function will scan the entire PCI bus, including the relevant sysfs files for the aforementioned VFs.

Initially, it seemed that directly adding the pci_lock_rescan_remove() lock to sriov_disable() could solve the problem.
However, it was later discovered that this might lead to a deadlock issue (because the remove_store() function would call sriov_disable(), causing a deadlock).

drivers/pci/pci-sysfs.c

  487 static ssize_t remove_store(struct device *dev, struct device_attribute *attr,
  488                             const char *buf, size_t count)
  489 {
  490         unsigned long val;
  491
  492         if (kstrtoul(buf, 0, &val) < 0)
  493                 return -EINVAL;
  494
  495         if (val && device_remove_file_self(dev, attr))
  496
         //Subsequently, sriov_disable() will be invoked.
                 pci_stop_and_remove_bus_device_locked(to_pci_dev(dev));
  497         return count;
  498 }
  499 static DEVICE_ATTR_IGNORE_LOCKDEP(remove, 0220, NULL,
  500                                   remove_store);

The function `pci_stop_and_remove_bus_device_locked()` acquires the `pci_lock_rescan_remove()` lock and subsequently calls `sriov_disable()`.
If the lock is added within `sriov_disable()`, it could lead to a deadlock.

Should we add a new lock to address this issue, or are there any other ideas? I would like to consult and discuss this with everyone.

Thanks,
Bingquan Mou <moubingquan at h-partners.com>