Nouveau - Nov 2020 - [PATCH 3/3] drm/nouveau: clean up all clients on device removal

This series fixes a number of use-after-frees in nouveau's postclose()
handler. It was discovered by pointing IGT's core_hotunplug tests at a
nouveau device, but the steps to reproduce it are simple:

1. Open the device file
2. Unbind the driver or remove the device
3. Close the file opened in step 1.

During the device removal, the nouveau_drm structure is de-allocated,
but is dereferenced in the postclose() handler.

One obvious solution is to ensure all the operations in the postclose()
handler are valid by extending the lifetime of the nouveau_drm
structure. This is possible with the new devm_drm_dev_alloc() interface,
but the change is somewhat invasive so I thought it best to submit that
work separately.

Instead, we make use of the drm_dev_unplug() API, clean up all clients
in the device removal call, and check to make sure the device has not
been unplugged in the postclose() handler. While this does not enable
hot-unplug support for nouveau, it's enough to avoid crashing the kernel
and leads to all the core_hotunplug tests to pass.

Jeremy Cline (3):
  drm/nouveau: use drm_dev_unplug() during device removal
  drm/nouveau: Add a dedicated mutex for the clients list
  drm/nouveau: clean up all clients on device removal

 drivers/gpu/drm/nouveau/nouveau_drm.c | 39 +++++++++++++++++++++++----
 drivers/gpu/drm/nouveau/nouveau_drv.h |  5 ++++
 2 files changed, 39 insertions(+), 5 deletions(-)

-- 
2.28.0

Nouveau does not currently support hot-unplugging, but it still makes
sense to switch from drm_dev_unregister() to drm_dev_unplug().
drm_dev_unplug() calls drm_dev_unregister() after marking the device as
unplugged, but only after any device critical sections are finished.

Since nouveau isn't using drm_dev_enter() and drm_dev_exit(), there are
no critical sections so this is nearly functionally equivalent. However,
the DRM layer does check to see if the device is unplugged, and if it is
returns appropriate error codes.

In the future nouveau can add critical sections in order to truly
support hot-unplugging.

Signed-off-by: Jeremy Cline <jcline at redhat.com>
---
 drivers/gpu/drm/nouveau/nouveau_drm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c
b/drivers/gpu/drm/nouveau/nouveau_drm.c
index d141a5f004af..4fe4d664c5f2 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
@@ -792,7 +792,7 @@ nouveau_drm_device_remove(struct drm_device *dev)
 	struct nvkm_client *client;
 	struct nvkm_device *device;
 
-	drm_dev_unregister(dev);
+	drm_dev_unplug(dev);
 
 	dev->irq_enabled = false;
 	client = nvxx_client(&drm->client.base);
-- 
2.28.0

Rather than protecting the nouveau_drm clients list with the lock within
the "client" nouveau_cli, add a dedicated lock to serialize access to
the list. This is both clearer and necessary to avoid lockdep being
upset with us when we need to iterate through all the clients in the
list and potentially lock their mutex, which is the same class as the
lock protecting the entire list.

Signed-off-by: Jeremy Cline <jcline at redhat.com>
---
 drivers/gpu/drm/nouveau/nouveau_drm.c | 9 +++++----
 drivers/gpu/drm/nouveau/nouveau_drv.h | 5 +++++
 2 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c
b/drivers/gpu/drm/nouveau/nouveau_drm.c
index 4fe4d664c5f2..d182b877258a 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
@@ -557,6 +557,7 @@ nouveau_drm_device_init(struct drm_device *dev)
 		nvkm_dbgopt(nouveau_debug, "DRM");
 
 	INIT_LIST_HEAD(&drm->clients);
+	mutex_init(&drm->clients_lock);
 	spin_lock_init(&drm->tile.lock);
 
 	/* workaround an odd issue on nvc1 by disabling the device's
@@ -1089,9 +1090,9 @@ nouveau_drm_open(struct drm_device *dev, struct drm_file
*fpriv)
 
 	fpriv->driver_priv = cli;
 
-	mutex_lock(&drm->client.mutex);
+	mutex_lock(&drm->clients_lock);
 	list_add(&cli->head, &drm->clients);
-	mutex_unlock(&drm->client.mutex);
+	mutex_unlock(&drm->clients_lock);
 
 done:
 	if (ret && cli) {
@@ -1117,9 +1118,9 @@ nouveau_drm_postclose(struct drm_device *dev, struct
drm_file *fpriv)
 		nouveau_abi16_fini(cli->abi16);
 	mutex_unlock(&cli->mutex);
 
-	mutex_lock(&drm->client.mutex);
+	mutex_lock(&drm->clients_lock);
 	list_del(&cli->head);
-	mutex_unlock(&drm->client.mutex);
+	mutex_unlock(&drm->clients_lock);
 
 	nouveau_cli_fini(cli);
 	kfree(cli);
diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h
b/drivers/gpu/drm/nouveau/nouveau_drv.h
index 9d04d1b36434..550e5f335146 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drv.h
+++ b/drivers/gpu/drm/nouveau/nouveau_drv.h
@@ -141,6 +141,11 @@ struct nouveau_drm {
 
 	struct list_head clients;
 
+	/**
+	 * @clients_lock: Protects access to the @clients list of &struct
nouveau_cli.
+	 */
+	struct mutex clients_lock;
+
 	u8 old_pm_cap;
 
 	struct {
-- 
2.28.0

The postclose handler can run after the device has been removed (or the
driver has been unbound) since userspace clients are free to hold the
file open the file as long as they want. Because the device removal
callback frees the entire nouveau_drm structure, any reference to it in
the postclose handler will result in a use-after-free.

To reproduce this, one must simply open the device file, unbind the
driver (or physically remove the device), and then close the device
file. This was found and can be reproduced easily with the IGT
core_hotunplug tests.

To avoid this, all clients are cleaned up in the device finialization
rather than deferring it to the postclose handler, and the postclose
handler is protected by a critical section which ensures the
drm_dev_unplug() and the postclose handler won't race.

This is not an ideal fix, since as I understand the proposed plan for
the kernel<->userspace interface for hotplug support, destroying the
client before the file is closed will cause problems. However, I believe
to properly fix this issue, the lifetime of the nouveau_drm structure
needs to be extended to match the drm_device, and this proved to be a
rather invasive change. Thus, I've broken this out so the fix can be
easily backported.

Signed-off-by: Jeremy Cline <jcline at redhat.com>
---
 drivers/gpu/drm/nouveau/nouveau_drm.c | 30 +++++++++++++++++++++++++++
 1 file changed, 30 insertions(+)

diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c
b/drivers/gpu/drm/nouveau/nouveau_drm.c
index d182b877258a..74fab777f4d0 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
@@ -628,6 +628,7 @@ nouveau_drm_device_init(struct drm_device *dev)
 static void
 nouveau_drm_device_fini(struct drm_device *dev)
 {
+	struct nouveau_cli *cli, *temp_cli;
 	struct nouveau_drm *drm = nouveau_drm(dev);
 
 	if (nouveau_pmops_runtime()) {
@@ -652,6 +653,24 @@ nouveau_drm_device_fini(struct drm_device *dev)
 	nouveau_ttm_fini(drm);
 	nouveau_vga_fini(drm);
 
+	/*
+	 * There may be existing clients from as-yet unclosed files. For now,
+	 * clean them up here rather than deferring until the file is closed,
+	 * but this likely not correct if we want to support hot-unplugging
+	 * properly.
+	 */
+	mutex_lock(&drm->clients_lock);
+	list_for_each_entry_safe(cli, temp_cli, &drm->clients, head) {
+		list_del(&cli->head);
+		mutex_lock(&cli->mutex);
+		if (cli->abi16)
+			nouveau_abi16_fini(cli->abi16);
+		mutex_unlock(&cli->mutex);
+		nouveau_cli_fini(cli);
+		kfree(cli);
+	}
+	mutex_unlock(&drm->clients_lock);
+
 	nouveau_cli_fini(&drm->client);
 	nouveau_cli_fini(&drm->master);
 	nvif_parent_dtor(&drm->parent);
@@ -1110,6 +1129,16 @@ nouveau_drm_postclose(struct drm_device *dev, struct
drm_file *fpriv)
 {
 	struct nouveau_cli *cli = nouveau_cli(fpriv);
 	struct nouveau_drm *drm = nouveau_drm(dev);
+	int dev_index;
+
+	/*
+	 * The device is gone, and as it currently stands all clients are
+	 * cleaned up in the removal codepath. In the future this may change
+	 * so that we can support hot-unplugging, but for now we immediately
+	 * return to avoid a double-free situation.
+	 */
+	if (!drm_dev_enter(dev, &dev_index))
+		return;
 
 	pm_runtime_get_sync(dev->dev);
 
@@ -1126,6 +1155,7 @@ nouveau_drm_postclose(struct drm_device *dev, struct
drm_file *fpriv)
 	kfree(cli);
 	pm_runtime_mark_last_busy(dev->dev);
 	pm_runtime_put_autosuspend(dev->dev);
+	drm_dev_exit(dev_index);
 }
 
 static const struct drm_ioctl_desc
-- 
2.28.0

This series fixes a number of use-after-frees in nouveau's postclose()
handler. It was discovered by pointing IGT's core_hotunplug tests at a
nouveau device, but the steps to reproduce it are simple:

1. Open the device file
2. Unbind the driver or remove the device
3. Close the file opened in step 1.

During the device removal, the nouveau_drm structure is de-allocated,
but is dereferenced in the postclose() handler.

One obvious solution is to ensure all the operations in the postclose()
handler are valid by extending the lifetime of the nouveau_drm
structure. This is possible with the new devm_drm_dev_alloc() interface,
but the change is somewhat invasive so I thought it best to submit that
work separately.

Instead, we make use of the drm_dev_unplug() API, clean up all clients
in the device removal call, and check to make sure the device has not
been unplugged in the postclose() handler. While this does not enable
hot-unplug support for nouveau, it's enough to avoid crashing the kernel
and leads to all the core_hotunplug tests to pass.

This series reroll addresses a missing mutex_destroy() call and a typo
in a commit message.

Jeremy Cline (3):
  drm/nouveau: use drm_dev_unplug() during device removal
  drm/nouveau: Add a dedicated mutex for the clients list
  drm/nouveau: clean up all clients on device removal

 drivers/gpu/drm/nouveau/nouveau_drm.c | 42 +++++++++++++++++++++++----
 drivers/gpu/drm/nouveau/nouveau_drv.h |  5 ++++
 2 files changed, 42 insertions(+), 5 deletions(-)

-- 
2.28.0