The inode can be different in a container, for example, a docker and host
both open the same uacce parent device, which uses the same uacce struct
but different inode, so uacce->inode is not enough.
What's worse, when docker stops, the inode will be destroyed as well,
causing use-after-free in uacce_remove.
So use q->filep->f_mapping to replace uacce->inode->i_mapping.
Signed-off-by: Weili Qian <[email protected]>
Signed-off-by: Zhangfei Gao <[email protected]>
---
drivers/misc/uacce/uacce.c | 16 +++++++++-------
include/linux/uacce.h | 4 ++--
2 files changed, 11 insertions(+), 9 deletions(-)
diff --git a/drivers/misc/uacce/uacce.c b/drivers/misc/uacce/uacce.c
index 346bd7cf2e94..740ace422baa 100644
--- a/drivers/misc/uacce/uacce.c
+++ b/drivers/misc/uacce/uacce.c
@@ -166,8 +166,8 @@ static int uacce_fops_open(struct inode *inode, struct file *filep)
init_waitqueue_head(&q->wait);
filep->private_data = q;
- uacce->inode = inode;
q->state = UACCE_Q_INIT;
+ q->private_data = filep;
mutex_init(&q->mutex);
list_add(&q->list, &uacce->queues);
mutex_unlock(&uacce->mutex);
@@ -574,12 +574,6 @@ void uacce_remove(struct uacce_device *uacce)
if (!uacce)
return;
- /*
- * unmap remaining mapping from user space, preventing user still
- * access the mmaped area while parent device is already removed
- */
- if (uacce->inode)
- unmap_mapping_range(uacce->inode->i_mapping, 0, 0, 1);
/*
* uacce_fops_open() may be running concurrently, even after we remove
@@ -589,6 +583,8 @@ void uacce_remove(struct uacce_device *uacce)
mutex_lock(&uacce->mutex);
/* ensure no open queue remains */
list_for_each_entry_safe(q, next_q, &uacce->queues, list) {
+ struct file *filep = q->private_data;
+
/*
* Taking q->mutex ensures that fops do not use the defunct
* uacce->ops after the queue is disabled.
@@ -597,6 +593,12 @@ void uacce_remove(struct uacce_device *uacce)
uacce_put_queue(q);
mutex_unlock(&q->mutex);
uacce_unbind_queue(q);
+
+ /*
+ * unmap remaining mapping from user space, preventing user still
+ * access the mmaped area while parent device is already removed
+ */
+ unmap_mapping_range(filep->f_mapping, 0, 0, 1);
}
/* disable sva now since no opened queues */
diff --git a/include/linux/uacce.h b/include/linux/uacce.h
index 0a81c3dfd26c..64b800b74436 100644
--- a/include/linux/uacce.h
+++ b/include/linux/uacce.h
@@ -86,6 +86,7 @@ enum uacce_q_state {
* @state: queue state machine
* @pasid: pasid associated to the mm
* @handle: iommu_sva handle returned by iommu_sva_bind_device()
+ * @private_data: private data for saving filep
*/
struct uacce_queue {
struct uacce_device *uacce;
@@ -97,6 +98,7 @@ struct uacce_queue {
enum uacce_q_state state;
u32 pasid;
struct iommu_sva *handle;
+ void *private_data;
};
/**
@@ -114,7 +116,6 @@ struct uacce_queue {
* @mutex: protects uacce operation
* @priv: private pointer of the uacce
* @queues: list of queues
- * @inode: core vfs
*/
struct uacce_device {
const char *algs;
@@ -130,7 +131,6 @@ struct uacce_device {
struct mutex mutex;
void *priv;
struct list_head queues;
- struct inode *inode;
};
#if IS_ENABLED(CONFIG_UACCE)
--
2.39.2 (Apple Git-143)
On Thu, May 11, 2023 at 10:15:53AM +0800, Zhangfei Gao wrote:
> The inode can be different in a container, for example, a docker and host
> both open the same uacce parent device, which uses the same uacce struct
> but different inode, so uacce->inode is not enough.
>
> What's worse, when docker stops, the inode will be destroyed as well,
> causing use-after-free in uacce_remove.
>
> So use q->filep->f_mapping to replace uacce->inode->i_mapping.
>
> Signed-off-by: Weili Qian <[email protected]>
> Signed-off-by: Zhangfei Gao <[email protected]>
> ---
> drivers/misc/uacce/uacce.c | 16 +++++++++-------
> include/linux/uacce.h | 4 ++--
> 2 files changed, 11 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/misc/uacce/uacce.c b/drivers/misc/uacce/uacce.c
> index 346bd7cf2e94..740ace422baa 100644
> --- a/drivers/misc/uacce/uacce.c
> +++ b/drivers/misc/uacce/uacce.c
> @@ -166,8 +166,8 @@ static int uacce_fops_open(struct inode *inode, struct file *filep)
>
> init_waitqueue_head(&q->wait);
> filep->private_data = q;
> - uacce->inode = inode;
> q->state = UACCE_Q_INIT;
> + q->private_data = filep;
> mutex_init(&q->mutex);
> list_add(&q->list, &uacce->queues);
> mutex_unlock(&uacce->mutex);
> @@ -574,12 +574,6 @@ void uacce_remove(struct uacce_device *uacce)
>
> if (!uacce)
> return;
> - /*
> - * unmap remaining mapping from user space, preventing user still
> - * access the mmaped area while parent device is already removed
> - */
> - if (uacce->inode)
> - unmap_mapping_range(uacce->inode->i_mapping, 0, 0, 1);
>
> /*
> * uacce_fops_open() may be running concurrently, even after we remove
> @@ -589,6 +583,8 @@ void uacce_remove(struct uacce_device *uacce)
> mutex_lock(&uacce->mutex);
> /* ensure no open queue remains */
> list_for_each_entry_safe(q, next_q, &uacce->queues, list) {
> + struct file *filep = q->private_data;
> +
> /*
> * Taking q->mutex ensures that fops do not use the defunct
> * uacce->ops after the queue is disabled.
> @@ -597,6 +593,12 @@ void uacce_remove(struct uacce_device *uacce)
> uacce_put_queue(q);
> mutex_unlock(&q->mutex);
> uacce_unbind_queue(q);
> +
> + /*
> + * unmap remaining mapping from user space, preventing user still
> + * access the mmaped area while parent device is already removed
> + */
> + unmap_mapping_range(filep->f_mapping, 0, 0, 1);
> }
>
> /* disable sva now since no opened queues */
> diff --git a/include/linux/uacce.h b/include/linux/uacce.h
> index 0a81c3dfd26c..64b800b74436 100644
> --- a/include/linux/uacce.h
> +++ b/include/linux/uacce.h
> @@ -86,6 +86,7 @@ enum uacce_q_state {
> * @state: queue state machine
> * @pasid: pasid associated to the mm
> * @handle: iommu_sva handle returned by iommu_sva_bind_device()
> + * @private_data: private data for saving filep
> */
> struct uacce_queue {
> struct uacce_device *uacce;
> @@ -97,6 +98,7 @@ struct uacce_queue {
> enum uacce_q_state state;
> u32 pasid;
> struct iommu_sva *handle;
> + void *private_data;
Make this a real pointer to the inode, no need to make this "void *",
right?
thanks,
greg k-h
On Thu, May 11, 2023 at 10:15:53AM +0800, Zhangfei Gao wrote:
> The inode can be different in a container, for example, a docker and host
> both open the same uacce parent device, which uses the same uacce struct
> but different inode, so uacce->inode is not enough.
>
> What's worse, when docker stops, the inode will be destroyed as well,
> causing use-after-free in uacce_remove.
>
> So use q->filep->f_mapping to replace uacce->inode->i_mapping.
> @@ -574,12 +574,6 @@ void uacce_remove(struct uacce_device *uacce)
>
> if (!uacce)
> return;
> - /*
> - * unmap remaining mapping from user space, preventing user still
> - * access the mmaped area while parent device is already removed
> - */
> - if (uacce->inode)
> - unmap_mapping_range(uacce->inode->i_mapping, 0, 0, 1);
>
> /*
> * uacce_fops_open() may be running concurrently, even after we remove
> @@ -589,6 +583,8 @@ void uacce_remove(struct uacce_device *uacce)
> mutex_lock(&uacce->mutex);
> /* ensure no open queue remains */
> list_for_each_entry_safe(q, next_q, &uacce->queues, list) {
> + struct file *filep = q->private_data;
> +
> /*
> * Taking q->mutex ensures that fops do not use the defunct
> * uacce->ops after the queue is disabled.
> @@ -597,6 +593,12 @@ void uacce_remove(struct uacce_device *uacce)
> uacce_put_queue(q);
> mutex_unlock(&q->mutex);
> uacce_unbind_queue(q);
> +
> + /*
> + * unmap remaining mapping from user space, preventing user still
> + * access the mmaped area while parent device is already removed
> + */
> + unmap_mapping_range(filep->f_mapping, 0, 0, 1);
IDGI. Going through uacce_queue instead of uacce_device is fine, but why
bother with file *or* inode? Just store a reference to struct address_space in
your uacce_queue and be done with that...
Another problem in that driver is uacce_vma_close(); this
if (vma->vm_pgoff < UACCE_MAX_REGION)
qfr = q->qfrs[vma->vm_pgoff];
kfree(qfr);
can't be right - you have q->qfrs left pointing to freed object. If nothing
else, subsequent mmap() will fail with -EEXIST, won't it?
v2:
using q->mapping instead of inode or file
vma_close clears q->qfrs as well
Zhangfei Gao (2):
uacce: use q->mapping to replace inode->i_mapping
uacce: vma_close clears q->qfrs when freeing qfrs
drivers/misc/uacce/uacce.c | 25 ++++++++++++++-----------
include/linux/uacce.h | 4 ++--
2 files changed, 16 insertions(+), 13 deletions(-)
--
2.39.2 (Apple Git-143)
On Thu, 11 May 2023 at 12:05, Al Viro <[email protected]> wrote:
>
> On Thu, May 11, 2023 at 10:15:53AM +0800, Zhangfei Gao wrote:
> > The inode can be different in a container, for example, a docker and host
> > both open the same uacce parent device, which uses the same uacce struct
> > but different inode, so uacce->inode is not enough.
> >
> > What's worse, when docker stops, the inode will be destroyed as well,
> > causing use-after-free in uacce_remove.
> >
> > So use q->filep->f_mapping to replace uacce->inode->i_mapping.
>
> > @@ -574,12 +574,6 @@ void uacce_remove(struct uacce_device *uacce)
> >
> > if (!uacce)
> > return;
> > - /*
> > - * unmap remaining mapping from user space, preventing user still
> > - * access the mmaped area while parent device is already removed
> > - */
> > - if (uacce->inode)
> > - unmap_mapping_range(uacce->inode->i_mapping, 0, 0, 1);
> >
> > /*
> > * uacce_fops_open() may be running concurrently, even after we remove
> > @@ -589,6 +583,8 @@ void uacce_remove(struct uacce_device *uacce)
> > mutex_lock(&uacce->mutex);
> > /* ensure no open queue remains */
> > list_for_each_entry_safe(q, next_q, &uacce->queues, list) {
> > + struct file *filep = q->private_data;
> > +
> > /*
> > * Taking q->mutex ensures that fops do not use the defunct
> > * uacce->ops after the queue is disabled.
> > @@ -597,6 +593,12 @@ void uacce_remove(struct uacce_device *uacce)
> > uacce_put_queue(q);
> > mutex_unlock(&q->mutex);
> > uacce_unbind_queue(q);
> > +
> > + /*
> > + * unmap remaining mapping from user space, preventing user still
> > + * access the mmaped area while parent device is already removed
> > + */
> > + unmap_mapping_range(filep->f_mapping, 0, 0, 1);
>
> IDGI. Going through uacce_queue instead of uacce_device is fine, but why
> bother with file *or* inode? Just store a reference to struct address_space in
> your uacce_queue and be done with that...
Yes, a struct address_space is enough.
>
> Another problem in that driver is uacce_vma_close(); this
> if (vma->vm_pgoff < UACCE_MAX_REGION)
> qfr = q->qfrs[vma->vm_pgoff];
>
> kfree(qfr);
> can't be right - you have q->qfrs left pointing to freed object. If nothing
> else, subsequent mmap() will fail with -EEXIST, won't it?
Good catch, will fix it.
Thanks
The inode can be different in a container, for example, a docker and host
both open the same uacce parent device, which uses the same uacce struct
but different inode, so uacce->inode is not enough.
What's worse, when docker stops, the inode will be destroyed as well,
causing use-after-free in uacce_remove.
So use q->mapping to replace uacce->inode->i_mapping.
Signed-off-by: Weili Qian <[email protected]>
Signed-off-by: Zhangfei Gao <[email protected]>
---
drivers/misc/uacce/uacce.c | 14 +++++++-------
include/linux/uacce.h | 4 ++--
2 files changed, 9 insertions(+), 9 deletions(-)
diff --git a/drivers/misc/uacce/uacce.c b/drivers/misc/uacce/uacce.c
index 346bd7cf2e94..338b59ef5493 100644
--- a/drivers/misc/uacce/uacce.c
+++ b/drivers/misc/uacce/uacce.c
@@ -166,8 +166,8 @@ static int uacce_fops_open(struct inode *inode, struct file *filep)
init_waitqueue_head(&q->wait);
filep->private_data = q;
- uacce->inode = inode;
q->state = UACCE_Q_INIT;
+ q->mapping = filep->f_mapping;
mutex_init(&q->mutex);
list_add(&q->list, &uacce->queues);
mutex_unlock(&uacce->mutex);
@@ -574,12 +574,6 @@ void uacce_remove(struct uacce_device *uacce)
if (!uacce)
return;
- /*
- * unmap remaining mapping from user space, preventing user still
- * access the mmaped area while parent device is already removed
- */
- if (uacce->inode)
- unmap_mapping_range(uacce->inode->i_mapping, 0, 0, 1);
/*
* uacce_fops_open() may be running concurrently, even after we remove
@@ -597,6 +591,12 @@ void uacce_remove(struct uacce_device *uacce)
uacce_put_queue(q);
mutex_unlock(&q->mutex);
uacce_unbind_queue(q);
+
+ /*
+ * unmap remaining mapping from user space, preventing user still
+ * access the mmaped area while parent device is already removed
+ */
+ unmap_mapping_range(q->mapping, 0, 0, 1);
}
/* disable sva now since no opened queues */
diff --git a/include/linux/uacce.h b/include/linux/uacce.h
index 0a81c3dfd26c..e290c0269944 100644
--- a/include/linux/uacce.h
+++ b/include/linux/uacce.h
@@ -86,6 +86,7 @@ enum uacce_q_state {
* @state: queue state machine
* @pasid: pasid associated to the mm
* @handle: iommu_sva handle returned by iommu_sva_bind_device()
+ * @mapping: user space mapping of the queue
*/
struct uacce_queue {
struct uacce_device *uacce;
@@ -97,6 +98,7 @@ struct uacce_queue {
enum uacce_q_state state;
u32 pasid;
struct iommu_sva *handle;
+ struct address_space *mapping;
};
/**
@@ -114,7 +116,6 @@ struct uacce_queue {
* @mutex: protects uacce operation
* @priv: private pointer of the uacce
* @queues: list of queues
- * @inode: core vfs
*/
struct uacce_device {
const char *algs;
@@ -130,7 +131,6 @@ struct uacce_device {
struct mutex mutex;
void *priv;
struct list_head queues;
- struct inode *inode;
};
#if IS_ENABLED(CONFIG_UACCE)
--
2.39.2 (Apple Git-143)
vma_close frees qfrs but not clears q->qfrs, which still points
to the freed object, leading to subsequent mmap fail.
So vma_close clears q->qfrs as well.
Suggested-by: Al Viro <[email protected]>
Signed-off-by: Zhangfei Gao <[email protected]>
---
drivers/misc/uacce/uacce.c | 11 +++++++----
1 file changed, 7 insertions(+), 4 deletions(-)
diff --git a/drivers/misc/uacce/uacce.c b/drivers/misc/uacce/uacce.c
index 338b59ef5493..930c252753a0 100644
--- a/drivers/misc/uacce/uacce.c
+++ b/drivers/misc/uacce/uacce.c
@@ -200,12 +200,15 @@ static int uacce_fops_release(struct inode *inode, struct file *filep)
static void uacce_vma_close(struct vm_area_struct *vma)
{
struct uacce_queue *q = vma->vm_private_data;
- struct uacce_qfile_region *qfr = NULL;
- if (vma->vm_pgoff < UACCE_MAX_REGION)
- qfr = q->qfrs[vma->vm_pgoff];
+ if (vma->vm_pgoff < UACCE_MAX_REGION) {
+ struct uacce_qfile_region *qfr = q->qfrs[vma->vm_pgoff];
- kfree(qfr);
+ mutex_lock(&q->mutex);
+ q->qfrs[vma->vm_pgoff] = NULL;
+ mutex_unlock(&q->mutex);
+ kfree(qfr);
+ }
}
static const struct vm_operations_struct uacce_vm_ops = {
--
2.39.2 (Apple Git-143)