linux-stable.git/drivers/pci, branch v5.4.244

PCI: pciehp: Fix AB-BA deadlock between reset_lock and device_lock

2023-05-17T09:36:04+00:00

commit f5eff5591b8f9c5effd25c92c758a127765f74c1 upstream.

In 2013, commits

  2e35afaefe64 ("PCI: pciehp: Add reset_slot() method")
  608c388122c7 ("PCI: Add slot reset option to pci_dev_reset()")

amended PCIe hotplug to mask Presence Detect Changed events during a
Secondary Bus Reset.  The reset thus no longer causes gratuitous slot
bringdown and bringup.

However the commits neglected to serialize reset with code paths reading
slot registers.  For instance, a slot bringup due to an earlier hotplug
event may see the Presence Detect State bit cleared during a concurrent
Secondary Bus Reset.

In 2018, commit

  5b3f7b7d062b ("PCI: pciehp: Avoid slot access during reset")

retrofitted the missing locking.  It introduced a reset_lock which
serializes a Secondary Bus Reset with other parts of pciehp.

Unfortunately the locking turns out to be overzealous:  reset_lock is
held for the entire enumeration and de-enumeration of hotplugged devices,
including driver binding and unbinding.

Driver binding and unbinding acquires device_lock while the reset_lock
of the ancestral hotplug port is held.  A concurrent Secondary Bus Reset
acquires the ancestral reset_lock while already holding the device_lock.
The asymmetric locking order in the two code paths can lead to AB-BA
deadlocks.

Michael Haeuptle reports such deadlocks on simultaneous hot-removal and
vfio release (the latter implies a Secondary Bus Reset):

  pciehp_ist()                                    # down_read(reset_lock)
    pciehp_handle_presence_or_link_change()
      pciehp_disable_slot()
        __pciehp_disable_slot()
          remove_board()
            pciehp_unconfigure_device()
              pci_stop_and_remove_bus_device()
                pci_stop_bus_device()
                  pci_stop_dev()
                    device_release_driver()
                      device_release_driver_internal()
                        __device_driver_lock()    # device_lock()

  SYS_munmap()
    vfio_device_fops_release()
      vfio_device_group_close()
        vfio_device_close()
          vfio_device_last_close()
            vfio_pci_core_close_device()
              vfio_pci_core_disable()             # device_lock()
                __pci_reset_function_locked()
                  pci_reset_bus_function()
                    pci_dev_reset_slot_function()
                      pci_reset_hotplug_slot()
                        pciehp_reset_slot()       # down_write(reset_lock)

Ian May reports the same deadlock on simultaneous hot-removal and an
AER-induced Secondary Bus Reset:

  aer_recover_work_func()
    pcie_do_recovery()
      aer_root_reset()
        pci_bus_error_reset()
          pci_slot_reset()
            pci_slot_lock()                       # device_lock()
            pci_reset_hotplug_slot()
              pciehp_reset_slot()                 # down_write(reset_lock)

Fix by releasing the reset_lock during driver binding and unbinding,
thereby splitting and shrinking the critical section.

Driver binding and unbinding is protected by the device_lock() and thus
serialized with a Secondary Bus Reset.  There's no need to additionally
protect it with the reset_lock.  However, pciehp does not bind and
unbind devices directly, but rather invokes PCI core functions which
also perform certain enumeration and de-enumeration steps.

The reset_lock's purpose is to protect slot registers, not enumeration
and de-enumeration of hotplugged devices.  That would arguably be the
job of the PCI core, not the PCIe hotplug driver.  After all, an
AER-induced Secondary Bus Reset may as well happen during boot-time
enumeration of the PCI hierarchy and there's no locking to prevent that
either.

Exempting *de-enumeration* from the reset_lock is relatively harmless:
A concurrent Secondary Bus Reset may foil config space accesses such as
PME interrupt disablement.  But if the device is physically gone, those
accesses are pointless anyway.  If the device is physically present and
only logically removed through an Attention Button press or the sysfs
"power" attribute, PME interrupts as well as DMA cannot come through
because pciehp_unconfigure_device() disables INTx and Bus Master bits.
That's still protected by the reset_lock in the present commit.

Exempting *enumeration* from the reset_lock also has limited impact:
The exempted call to pci_bus_add_device() may perform device accesses
through pcibios_bus_add_device() and pci_fixup_device() which are now
no longer protected from a concurrent Secondary Bus Reset.  Otherwise
there should be no impact.

In essence, the present commit seeks to fix the AB-BA deadlocks while
still retaining a best-effort reset protection for enumeration and
de-enumeration of hotplugged devices -- until a general solution is
implemented in the PCI core.

Link: https://lore.kernel.org/linux-pci/CS1PR8401MB0728FC6FDAB8A35C22BD90EC95F10@CS1PR8401MB0728.NAMPRD84.PROD.OUTLOOK.COM
Link: https://lore.kernel.org/linux-pci/20200615143250.438252-1-ian.may@canonical.com
Link: https://lore.kernel.org/linux-pci/ce878dab-c0c4-5bd0-a725-9805a075682d@amd.com
Link: https://lore.kernel.org/linux-pci/ed831249-384a-6d35-0831-70af191e9bce@huawei.com
Link: https://bugzilla.kernel.org/show_bug.cgi?id=215590
Fixes: 5b3f7b7d062b ("PCI: pciehp: Avoid slot access during reset")
Link: https://lore.kernel.org/r/fef2b2e9edf245c049a8c5b94743c0f74ff5008a.1681191902.git.lukas@wunner.de
Reported-by: Michael Haeuptle 
Reported-by: Ian May 
Reported-by: Andrey Grodzovsky 
Reported-by: Rahul Kumar 
Reported-by: Jialin Zhang 
Tested-by: Anatoli Antonovitch 
Signed-off-by: Lukas Wunner 
Signed-off-by: Bjorn Helgaas 
Cc: stable@vger.kernel.org # v4.19+
Cc: Dan Stein 
Cc: Ashok Raj 
Cc: Alex Michon 
Cc: Xiongfeng Wang 
Cc: Alex Williamson 
Cc: Mika Westerberg 
Cc: Sathyanarayanan Kuppuswamy 
Signed-off-by: Greg Kroah-Hartman

PCI: pciehp: Use down_read/write_nested(reset_lock) to fix lockdep errors

2023-05-17T09:36:04+00:00

commit 085a9f43433f30cbe8a1ade62d9d7827c3217f4d upstream.

Use down_read_nested() and down_write_nested() when taking the
ctrl->reset_lock rw-sem, passing the number of PCIe hotplug controllers in
the path to the PCI root bus as lock subclass parameter.

This fixes the following false-positive lockdep report when unplugging a
Lenovo X1C8 from a Lenovo 2nd gen TB3 dock:

  pcieport 0000:06:01.0: pciehp: Slot(1): Link Down
  pcieport 0000:06:01.0: pciehp: Slot(1): Card not present
  ============================================
  WARNING: possible recursive locking detected
  5.16.0-rc2+ #621 Not tainted
  --------------------------------------------
  irq/124-pciehp/86 is trying to acquire lock:
  ffff8e5ac4299ef8 (&ctrl->reset_lock){.+.+}-{3:3}, at: pciehp_check_presence+0x23/0x80

  but task is already holding lock:
  ffff8e5ac4298af8 (&ctrl->reset_lock){.+.+}-{3:3}, at: pciehp_ist+0xf3/0x180

   other info that might help us debug this:
   Possible unsafe locking scenario:

	 CPU0
	 ----
    lock(&ctrl->reset_lock);
    lock(&ctrl->reset_lock);

   *** DEADLOCK ***

   May be due to missing lock nesting notation

  3 locks held by irq/124-pciehp/86:
   #0: ffff8e5ac4298af8 (&ctrl->reset_lock){.+.+}-{3:3}, at: pciehp_ist+0xf3/0x180
   #1: ffffffffa3b024e8 (pci_rescan_remove_lock){+.+.}-{3:3}, at: pciehp_unconfigure_device+0x31/0x110
   #2: ffff8e5ac1ee2248 (&dev->mutex){....}-{3:3}, at: device_release_driver+0x1c/0x40

  stack backtrace:
  CPU: 4 PID: 86 Comm: irq/124-pciehp Not tainted 5.16.0-rc2+ #621
  Hardware name: LENOVO 20U90SIT19/20U90SIT19, BIOS N2WET30W (1.20 ) 08/26/2021
  Call Trace:
   
   dump_stack_lvl+0x59/0x73
   __lock_acquire.cold+0xc5/0x2c6
   lock_acquire+0xb5/0x2b0
   down_read+0x3e/0x50
   pciehp_check_presence+0x23/0x80
   pciehp_runtime_resume+0x5c/0xa0
   device_for_each_child+0x45/0x70
   pcie_port_device_runtime_resume+0x20/0x30
   pci_pm_runtime_resume+0xa7/0xc0
   __rpm_callback+0x41/0x110
   rpm_callback+0x59/0x70
   rpm_resume+0x512/0x7b0
   __pm_runtime_resume+0x4a/0x90
   __device_release_driver+0x28/0x240
   device_release_driver+0x26/0x40
   pci_stop_bus_device+0x68/0x90
   pci_stop_bus_device+0x2c/0x90
   pci_stop_and_remove_bus_device+0xe/0x20
   pciehp_unconfigure_device+0x6c/0x110
   pciehp_disable_slot+0x5b/0xe0
   pciehp_handle_presence_or_link_change+0xc3/0x2f0
   pciehp_ist+0x179/0x180

This lockdep warning is triggered because with Thunderbolt, hotplug ports
are nested. When removing multiple devices in a daisy-chain, each hotplug
port's reset_lock may be acquired recursively. It's never the same lock, so
the lockdep splat is a false positive.

Because locks at the same hierarchy level are never acquired recursively, a
per-level lockdep class is sufficient to fix the lockdep warning.

The choice to use one lockdep subclass per pcie-hotplug controller in the
path to the root-bus was made to conserve class keys because their number
is limited and the complexity grows quadratically with number of keys
according to Documentation/locking/lockdep-design.rst.

Link: https://lore.kernel.org/linux-pci/20190402021933.GA2966@mit.edu/
Link: https://lore.kernel.org/linux-pci/de684a28-9038-8fc6-27ca-3f6f2f6400d7@redhat.com/
Link: https://lore.kernel.org/r/20211217141709.379663-1-hdegoede@redhat.com
Link: https://bugzilla.kernel.org/show_bug.cgi?id=208855
Reported-by: "Theodore Ts'o" 
Signed-off-by: Hans de Goede 
Signed-off-by: Bjorn Helgaas 
Reviewed-by: Lukas Wunner 
Cc: stable@vger.kernel.org
[lukas: backport to v5.4-stable]
Signed-off-by: Lukas Wunner 
Signed-off-by: Greg Kroah-Hartman

PCI: imx6: Install the fault handler only on compatible match

2023-05-17T09:35:46+00:00

[ Upstream commit 5f5ac460dfe7f4e11f99de9870f240e39189cf72 ]

commit bb38919ec56e ("PCI: imx6: Add support for i.MX6 PCIe controller")
added a fault hook to this driver in the probe function. So it was only
installed if needed.

commit bde4a5a00e76 ("PCI: imx6: Allow probe deferral by reset GPIO")
moved it from probe to driver init which installs the hook unconditionally
as soon as the driver is compiled into a kernel.

When this driver is compiled as a module, the hook is not registered
until after the driver has been matched with a .compatible and
loaded.

commit 415b6185c541 ("PCI: imx6: Fix config read timeout handling")
extended the fault handling code.

commit 2d8ed461dbc9 ("PCI: imx6: Add support for i.MX8MQ")
added some protection for non-ARM architectures, but this does not
protect non-i.MX ARM architectures.

Since fault handlers can be triggered on any architecture for different
reasons, there is no guarantee that they will be triggered only for the
assumed situation, leading to improper error handling (i.MX6-specific
imx6q_pcie_abort_handler) on foreign systems.

I had seen strange L3 imprecise external abort messages several times on
OMAP4 and OMAP5 devices and couldn't make sense of them until I realized
they were related to this unused imx6q driver because I had
CONFIG_PCI_IMX6=y.

Note that CONFIG_PCI_IMX6=y is useful for kernel binaries that are designed
to run on different ARM SoC and be differentiated only by device tree
binaries. So turning off CONFIG_PCI_IMX6 is not a solution.

Therefore we check the compatible in the init function before registering
the fault handler.

Link: https://lore.kernel.org/r/e1bcfc3078c82b53aa9b78077a89955abe4ea009.1678380991.git.hns@goldelico.com
Fixes: bde4a5a00e76 ("PCI: imx6: Allow probe deferral by reset GPIO")
Fixes: 415b6185c541 ("PCI: imx6: Fix config read timeout handling")
Fixes: 2d8ed461dbc9 ("PCI: imx6: Add support for i.MX8MQ")
Signed-off-by: H. Nikolaus Schaller
Signed-off-by: Lorenzo Pieralisi
Reviewed-by: Richard Zhu
Signed-off-by: Sasha Levin

PCI: Unify delay handling for reset and resume

2023-03-22T12:28:09+00:00

commit ac91e6980563ed53afadd925fa6585ffd2bc4a2c upstream.

Sheng Bi reports that pci_bridge_secondary_bus_reset() may fail to wait
for devices on the secondary bus to become accessible after reset:

Although it does call pci_dev_wait(), it erroneously passes the bridge's
pci_dev rather than that of a child.  The bridge of course is always
accessible while its secondary bus is reset, so pci_dev_wait() returns
immediately.

Sheng Bi proposes introducing a new pci_bridge_secondary_bus_wait()
function which is called from pci_bridge_secondary_bus_reset():

https://lore.kernel.org/linux-pci/20220523171517.32407-1-windy.bi.enflame@gmail.com/

However we already have pci_bridge_wait_for_secondary_bus() which does
almost exactly what we need.  So far it's only called on resume from
D3cold (which implies a Fundamental Reset per PCIe r6.0 sec 5.8).
Re-using it for Secondary Bus Resets is a leaner and more rational
approach than introducing a new function.

That only requires a few minor tweaks:

- Amend pci_bridge_wait_for_secondary_bus() to await accessibility of
  the first device on the secondary bus by calling pci_dev_wait() after
  performing the prescribed delays.  pci_dev_wait() needs two parameters,
  a reset reason and a timeout, which callers must now pass to
  pci_bridge_wait_for_secondary_bus().  The timeout is 1 sec for resume
  (PCIe r6.0 sec 6.6.1) and 60 sec for reset (commit 821cdad5c46c ("PCI:
  Wait up to 60 seconds for device to become ready after FLR")).
  Introduce a PCI_RESET_WAIT macro for the 1 sec timeout.

- Amend pci_bridge_wait_for_secondary_bus() to return 0 on success or
  -ENOTTY on error for consumption by pci_bridge_secondary_bus_reset().

- Drop an unnecessary 1 sec delay from pci_reset_secondary_bus() which
  is now performed by pci_bridge_wait_for_secondary_bus().  A static
  delay this long is only necessary for Conventional PCI, so modern
  PCIe systems benefit from shorter reset times as a side effect.

Fixes: 6b2f1351af56 ("PCI: Wait for device to become ready after secondary bus reset")
Link: https://lore.kernel.org/r/da77c92796b99ec568bd070cbe4725074a117038.1673769517.git.lukas@wunner.de
Reported-by: Sheng Bi 
Tested-by: Ravi Kishore Koppuravuri 
Signed-off-by: Lukas Wunner 
Signed-off-by: Bjorn Helgaas 
Reviewed-by: Mika Westerberg 
Reviewed-by: Kuppuswamy Sathyanarayanan 
Cc: stable@vger.kernel.org # v4.17+
Signed-off-by: Greg Kroah-Hartman

PCI: Add ACS quirk for Wangxun NICs

2023-03-11T15:44:15+00:00

[ Upstream commit a2b9b123ccac913e9f9b80337d687a2fe786a634 ]

Wangxun has verified there is no peer-to-peer between functions for the
below selection of SFxxx, RP1000 and RP2000 NICS.  They may be
multi-function devices, but the hardware does not advertise ACS capability.

Add an ACS quirk for these devices so the functions can be in independent
IOMMU groups.

Link: https://lore.kernel.org/r/20230207102419.44326-1-mengyuanlou@net-swift.com
Signed-off-by: Mengyuan Lou 
Signed-off-by: Bjorn Helgaas 
Signed-off-by: Sasha Levin

PCI: Avoid FLR for AMD FCH AHCI adapters

2023-03-11T15:44:07+00:00

commit 63ba51db24ed1b8f8088a897290eb6c036c5435d upstream.

PCI passthrough to VMs does not work with AMD FCH AHCI adapters: the guest
OS fails to correctly probe devices attached to the controller due to FIS
communication failures:

  ata4: softreset failed (1st FIS failed)
  ...
  ata4.00: qc timeout after 5000 msecs (cmd 0xec)
  ata4.00: failed to IDENTIFY (I/O error, err_mask=0x4)

Forcing the "bus" reset method before unbinding & binding the adapter to
the vfio-pci driver solves this issue, e.g.:

  echo "bus" > /sys/bus/pci/devices//reset_method

gives a working guest OS, indicating that the default FLR reset method
doesn't work correctly.

Apply quirk_no_flr() to AMD FCH AHCI devices to work around this issue.

Link: https://lore.kernel.org/r/20230128013951.523247-1-damien.lemoal@opensource.wdc.com
Reported-by: Niklas Cassel 
Signed-off-by: Damien Le Moal 
Signed-off-by: Bjorn Helgaas 
Cc: stable@vger.kernel.org
Signed-off-by: Greg Kroah-Hartman

PCI: hotplug: Allow marking devices as disconnected during bind/unbind

2023-03-11T15:44:07+00:00

commit 74ff8864cc842be994853095dba6db48e716400a upstream.

On surprise removal, pciehp_unconfigure_device() and acpiphp's
trim_stale_devices() call pci_dev_set_disconnected() to mark removed
devices as permanently offline.  Thereby, the PCI core and drivers know
to skip device accesses.

However pci_dev_set_disconnected() takes the device_lock and thus waits for
a concurrent driver bind or unbind to complete.  As a result, the driver's
->probe and ->remove hooks have no chance to learn that the device is gone.

That doesn't make any sense, so drop the device_lock and instead use atomic
xchg() and cmpxchg() operations to update the device state.

As a byproduct, an AB-BA deadlock reported by Anatoli is fixed which occurs
on surprise removal with AER concurrently performing a bus reset.

AER bus reset:

  INFO: task irq/26-aerdrv:95 blocked for more than 120 seconds.
  Tainted: G        W          6.2.0-rc3-custom-norework-jan11+
  schedule
  rwsem_down_write_slowpath
  down_write_nested
  pciehp_reset_slot                      # acquires reset_lock
  pci_reset_hotplug_slot
  pci_slot_reset                         # acquires device_lock
  pci_bus_error_reset
  aer_root_reset
  pcie_do_recovery
  aer_process_err_devices
  aer_isr

pciehp surprise removal:

  INFO: task irq/26-pciehp:96 blocked for more than 120 seconds.
  Tainted: G        W          6.2.0-rc3-custom-norework-jan11+
  schedule_preempt_disabled
  __mutex_lock
  mutex_lock_nested
  pci_dev_set_disconnected               # acquires device_lock
  pci_walk_bus
  pciehp_unconfigure_device
  pciehp_disable_slot
  pciehp_handle_presence_or_link_change
  pciehp_ist                             # acquires reset_lock

Link: https://bugzilla.kernel.org/show_bug.cgi?id=215590
Fixes: a6bd101b8f84 ("PCI: Unify device inaccessible")
Link: https://lore.kernel.org/r/3dc88ea82bdc0e37d9000e413d5ebce481cbd629.1674205689.git.lukas@wunner.de
Reported-by: Anatoli Antonovitch 
Signed-off-by: Lukas Wunner 
Signed-off-by: Bjorn Helgaas 
Cc: stable@vger.kernel.org # v4.20+
Cc: Keith Busch 
Signed-off-by: Greg Kroah-Hartman

PCI/PM: Observe reset delay irrespective of bridge_d3

2023-03-11T15:44:06+00:00

commit 8ef0217227b42e2c34a18de316cee3da16c9bf1e upstream.

If a PCI bridge is suspended to D3cold upon entering system sleep,
resuming it entails a Fundamental Reset per PCIe r6.0 sec 5.8.

The delay prescribed after a Fundamental Reset in PCIe r6.0 sec 6.6.1
is sought to be observed by:

  pci_pm_resume_noirq()
    pci_pm_bridge_power_up_actions()
      pci_bridge_wait_for_secondary_bus()

However, pci_bridge_wait_for_secondary_bus() bails out if the bridge_d3
flag is not set.  That flag indicates whether a bridge is allowed to
suspend to D3cold at *runtime*.

Hence *no* delay is observed on resume from system sleep if runtime
D3cold is forbidden.  That doesn't make any sense, so drop the bridge_d3
check from pci_bridge_wait_for_secondary_bus().

The purpose of the bridge_d3 check was probably to avoid delays if a
bridge remained in D0 during suspend.  However the sole caller of
pci_bridge_wait_for_secondary_bus(), pci_pm_bridge_power_up_actions(),
is only invoked if the previous power state was D3cold.  Hence the
additional bridge_d3 check seems superfluous.

Fixes: ad9001f2f411 ("PCI/PM: Add missing link delays required by the PCIe spec")
Link: https://lore.kernel.org/r/eb37fa345285ec8bacabbf06b020b803f77bdd3d.1673769517.git.lukas@wunner.de
Tested-by: Ravi Kishore Koppuravuri 
Signed-off-by: Lukas Wunner 
Signed-off-by: Bjorn Helgaas 
Reviewed-by: Mika Westerberg 
Reviewed-by: Kuppuswamy Sathyanarayanan 
Cc: stable@vger.kernel.org # v5.5+
Signed-off-by: Greg Kroah-Hartman

PCI/sysfs: Fix double free in error path

2023-01-18T10:41:49+00:00

commit aa382ffa705bea9931ec92b6f3c70e1fdb372195 upstream.

When pci_create_attr() fails, pci_remove_resource_files() is called which
will iterate over the res_attr[_wc] arrays and frees every non NULL entry.
To avoid a double free here set the array entry only after it's clear we
successfully initialized it.

Fixes: b562ec8f74e4 ("PCI: Don't leak memory if sysfs_create_bin_file() fails")
Link: https://lore.kernel.org/r/20221007070735.GX986@pengutronix.de/
Signed-off-by: Sascha Hauer 
Signed-off-by: Bjorn Helgaas 
Cc: stable@vger.kernel.org
Signed-off-by: Greg Kroah-Hartman

PCI: Fix pci_device_is_present() for VFs by checking PF

2023-01-18T10:41:49+00:00

commit 98b04dd0b4577894520493d96bc4623387767445 upstream.

pci_device_is_present() previously didn't work for VFs because it reads the
Vendor and Device ID, which are 0xffff for VFs, which looks like they
aren't present.  Check the PF instead.

Wei Gong reported that if virtio I/O is in progress when the driver is
unbound or "0" is written to /sys/.../sriov_numvfs, the virtio I/O
operation hangs, which may result in output like this:

  task:bash state:D stack:    0 pid: 1773 ppid:  1241 flags:0x00004002
  Call Trace:
   schedule+0x4f/0xc0
   blk_mq_freeze_queue_wait+0x69/0xa0
   blk_mq_freeze_queue+0x1b/0x20
   blk_cleanup_queue+0x3d/0xd0
   virtblk_remove+0x3c/0xb0 [virtio_blk]
   virtio_dev_remove+0x4b/0x80
   ...
   device_unregister+0x1b/0x60
   unregister_virtio_device+0x18/0x30
   virtio_pci_remove+0x41/0x80
   pci_device_remove+0x3e/0xb0

This happened because pci_device_is_present(VF) returned "false" in
virtio_pci_remove(), so it called virtio_break_device().  The broken vq
meant that vring_interrupt() skipped the vq.callback() that would have
completed the virtio I/O operation via virtblk_done().

[bhelgaas: commit log, simplify to always use pci_physfn(), add stable tag]
Link: https://lore.kernel.org/r/20221026060912.173250-1-mst@redhat.com
Reported-by: Wei Gong 
Tested-by: Wei Gong 
Signed-off-by: Michael S. Tsirkin 
Signed-off-by: Bjorn Helgaas 
Cc: stable@vger.kernel.org
Signed-off-by: Greg Kroah-Hartman