[PATCH 0/2] ata: pata_parport: fix UAF on protocol module unload

[PATCH 0/2] ata: pata_parport: fix UAF on protocol module unload

[Pei Xiao]

This series fixes use-after-free issues in pata_parport when a protocol
module goes away while pi_adapter devices created by it are still
attached.

Patch 1 pins the protocol module before the device becomes visible.
Previously try_module_get() ran after device_register(), so a forced
module unload in between left pi->proto dangling from the moment the
device appeared on the bus.

Patch 2 makes pata_parport_unregister_driver() tear down all adapters
using the protocol. Without this, the rollback path of a multi-protocol
module init (e.g. kbic registering k951 then k971) left the devices of
the already-registered protocol alive while the module loader freed the
module memory; removing such a dangling device later crashed in
pi_disconnect() dereferencing pi->proto->disconnect.

Pei Xiao (2):
  ata: pata_parport: pin the protocol module before device_register()
  ata: pata_parport: unregister devices on protocol unregister

 drivers/ata/pata_parport/pata_parport.c | 27 +++++++++++++++++++++----
 1 file changed, 23 insertions(+), 4 deletions(-)

-- 
2.25.1

[PATCH 1/2] ata: pata_parport: pin the protocol module before device_register()

[Pei Xiao]

pi_init_one() calls device_register() before try_module_get(). Between
these two calls the device is already visible but the module is not
pinned yet, so an unload in this window leaves pi->proto dangling:

    pi_init_one()                       rmmod -f <proto>
    --------------------------------------------------------
    device_register(&pi->dev)
         device visible on the bus
                                           module memory freed
         pi->proto = pr        <- writes into freed memory / dangles
         try_module_get(...)   <- too late, module already gone

Take the module reference before registering the device, and drop it
on the device_register() failure path.

Fixes: 246a1c4c6b7f ("ata: pata_parport: add driver (PARIDE replacement)")
Signed-off-by: Pei Xiao <xiaopei01@kylinos.cn>
---
 drivers/ata/pata_parport/pata_parport.c | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/drivers/ata/pata_parport/pata_parport.c b/drivers/ata/pata_parport/pata_parport.c
index cf81a6128f55..7462f9b1acc5 100644
--- a/drivers/ata/pata_parport/pata_parport.c
+++ b/drivers/ata/pata_parport/pata_parport.c
@@ -509,6 +509,14 @@ static struct pi_adapter *pi_init_one(struct parport *parport,
 		return NULL;
 	}
 
+	pi->proto = pr;
+
+	if (!try_module_get(pi->proto->owner)) {
+		kfree(pi);
+		ida_free(&pata_parport_bus_dev_ids, id);
+		return NULL;
+	}
+
 	/* set up pi->dev before pi_probe_unit() so it can use dev_printk() */
 	pi->dev.parent = pata_parport_bus;
 	pi->dev.bus = &pata_parport_bus_type;
@@ -517,15 +525,12 @@ static struct pi_adapter *pi_init_one(struct parport *parport,
 	pi->dev.id = id;
 	dev_set_name(&pi->dev, "pata_parport.%u", pi->dev.id);
 	if (device_register(&pi->dev)) {
+		module_put(pi->proto->owner);
 		put_device(&pi->dev);
 		/* pata_parport_dev_release will do ida_free(dev->id) and kfree(pi) */
 		return NULL;
 	}
 
-	pi->proto = pr;
-
-	if (!try_module_get(pi->proto->owner))
-		goto out_unreg_dev;
 	if (pi->proto->init_proto && pi->proto->init_proto(pi) < 0)
 		goto out_module_put;
 
-- 
2.25.1

[PATCH 2/2] ata: pata_parport: unregister devices on protocol unregister

[Pei Xiao]

When a protocol module registers multiple protocols and a later
registration fails (e.g. kbic_init registering k951 then k971), the
rollback path calls pata_parport_unregister_driver() on the already
registered protocol. This removes the protocol from the IDR and
unregisters the driver, but leaves the dynamically created pi_adapter
devices untouched.

Since the module init then fails, the module loader frees the module
memory, bypassing the references held by the devices. Any later removal
of these dangling devices (e.g. via sysfs delete_device or parport
detach) hits pi_remove_one() -> pi_disconnect() -> pi->proto->disconnect,
dereferencing the freed module memory and crashing the kernel.

Tear down all pi_adapters using the protocol in
pata_parport_unregister_driver(), before driver_unregister(), while the
module is still alive so the ->disconnect / ->release_proto callbacks
are safe to call.

Fixes: 246a1c4c6b7f ("ata: pata_parport: add driver (PARIDE replacement)")
Signed-off-by: Pei Xiao <xiaopei01@kylinos.cn>
---
 drivers/ata/pata_parport/pata_parport.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/drivers/ata/pata_parport/pata_parport.c b/drivers/ata/pata_parport/pata_parport.c
index 7462f9b1acc5..4ee518807031 100644
--- a/drivers/ata/pata_parport/pata_parport.c
+++ b/drivers/ata/pata_parport/pata_parport.c
@@ -612,6 +612,18 @@ int pata_parport_register_driver(struct pi_protocol *pr)
 }
 EXPORT_SYMBOL_GPL(pata_parport_register_driver);
 
+static int pi_remove_by_proto(struct device *dev, void *data)
+{
+	struct pi_protocol *pr = data;
+	struct ata_host *host = dev_get_drvdata(dev);
+	struct pi_adapter *pi = host->private_data;
+
+	if (pi->proto == pr)
+		pi_remove_one(dev);
+
+	return 0;
+}
+
 void pata_parport_unregister_driver(struct pi_protocol *pr)
 {
 	struct pi_protocol *pr_iter;
@@ -623,6 +635,8 @@ void pata_parport_unregister_driver(struct pi_protocol *pr)
 			break;
 	}
 	idr_remove(&protocols, id);
+	/* remove adapters using this protocol while the module is still alive */
+	bus_for_each_dev(&pata_parport_bus_type, NULL, pr, pi_remove_by_proto);
 	driver_unregister(&pr->driver);
 	mutex_unlock(&pi_mutex);
 
-- 
2.25.1

Re: [PATCH 1/2] ata: pata_parport: pin the protocol module before device_register()

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Low] Unused label `out_unreg_dev` causes compiler warning.
--

commit 99f7607699410f6446da6e37e3bed84251a51bd8
Author: Pei Xiao <xiaopei01@kylinos.cn>

ata: pata_parport: pin the protocol module before device_register()

Take the module reference before registering the device to prevent a race
condition where the device is visible but the protocol module is not yet
pinned. Also ensure the reference is dropped on the device_register()
failure path.

> diff --git a/drivers/ata/pata_parport/pata_parport.c b/drivers/ata/pata_parport/pata_parport.c
> index cf81a6128f552..7462f9b1acc57 100644
> --- a/drivers/ata/pata_parport/pata_parport.c
> +++ b/drivers/ata/pata_parport/pata_parport.c

[ ... ]

> @@ -517,15 +525,12 @@ static struct pi_adapter *pi_init_one(struct parport *parport,
>  	pi->dev.id = id;
>  	dev_set_name(&pi->dev, "pata_parport.%u", pi->dev.id);
>  	if (device_register(&pi->dev)) {
> +		module_put(pi->proto->owner);
>  		put_device(&pi->dev);
>  		/* pata_parport_dev_release will do ida_free(dev->id) and kfree(pi) */
>  		return NULL;
>  	}
>  
> -	pi->proto = pr;
> -
> -	if (!try_module_get(pi->proto->owner))
> -		goto out_unreg_dev;

[Severity: Low]
Does removing this goto leave the out_unreg_dev label unused at the bottom
of pi_init_one()?

If the label is no longer referenced, compiling the kernel will emit a
-Wunused-label warning. This can cause build failures in configurations
that enforce -Werror.

>  	if (pi->proto->init_proto && pi->proto->init_proto(pi) < 0)
>  		goto out_module_put;

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/cover.1788349317.git.xiaopei01@kylinos.cn?part=1

Re: [PATCH 2/2] ata: pata_parport: unregister devices on protocol unregister

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] AB-BA deadlock between `pi_mutex` and parport claim during device probing.
- [High] Missing NULL check on dev_get_drvdata() exposes a NULL pointer dereference if the driver is unbound via sysfs.
--

commit 5b5d1ccae8e55ae326690dca369a3675ed2bc9bf
Author: Pei Xiao <xiaopei01@kylinos.cn>
Subject: ata: pata_parport: unregister devices on protocol unregister

When a protocol module registers multiple protocols and a later
registration fails, the rollback path unregisters the protocol but
leaves the dynamically created pi_adapter devices untouched. This tear
down ensures adapters are removed while the module is still alive.

> diff --git a/drivers/ata/pata_parport/pata_parport.c b/drivers/ata/pata_parport/pata_parport.c
> index 7462f9b1acc57..4ee5188070319 100644
> --- a/drivers/ata/pata_parport/pata_parport.c
> +++ b/drivers/ata/pata_parport/pata_parport.c
[ ... ]
> @@ -612,6 +612,18 @@ int pata_parport_register_driver(struct pi_protocol *pr)
>  }
>  EXPORT_SYMBOL_GPL(pata_parport_register_driver);
>  
> +static int pi_remove_by_proto(struct device *dev, void *data)
> +{
> +	struct pi_protocol *pr = data;
> +	struct ata_host *host = dev_get_drvdata(dev);
> +	struct pi_adapter *pi = host->private_data;

[Severity: High]
This isn't a regression introduced by this patch, but does this missing null
check expose a null pointer dereference if the driver is unbound via sysfs?

If a user unbinds the driver via sysfs, dev_set_drvdata(dev, NULL) is called.
Later, when pi_remove_by_proto() or pi_remove_one() is triggered, wouldn't
dev_get_drvdata() return NULL and cause a crash when dereferencing
host->private_data?

> +
> +	if (pi->proto == pr)
> +		pi_remove_one(dev);
> +
> +	return 0;
> +}
> +
[ ... ]

[Severity: High]
This is a pre-existing issue, but is there an AB-BA deadlock between pi_mutex
and the parport claim during device probing?

During module initialization, we loop over all available protocols for a given
parport while holding pi_mutex:

drivers/ata/pata_parport/pata_parport.c:pata_parport_attach() {
    ...
	mutex_lock(&pi_mutex);
    ...
	if (probe) {
		/* probe this port using all protocols */
		idr_for_each_entry(&protocols, pr, pr_num)
			pi_init_one(port, pr, -1, -1, -1);
	}
	mutex_unlock(&pi_mutex);
}

If a device successfully initializes with the first protocol, it permanently
claims the parport:

drivers/ata/pata_parport/pata_parport.c:pi_connect() {
    ...
	parport_claim_or_block(pi->pardev);
    ...
}

The loop then proceeds to the next protocol and attempts to initialize it for
the same port. Does the second initialization block forever while holding
pi_mutex?

drivers/ata/pata_parport/pata_parport.c:pi_probe_unit() {
    ...
	if (pi->proto->test_port) {
		parport_claim_or_block(pi->pardev);
		max = pi->proto->test_port(pi);
		parport_release(pi->pardev);
	}
    ...
}

Because the port is permanently held by the first adapter, it appears the
thread will block forever while holding pi_mutex, causing a system deadlock
blocking all future parport ata operations.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/cover.1788349317.git.xiaopei01@kylinos.cn?part=2