Patchwork [v2] acpi: fix a potential inconsistency caused by double-fetch

login
register
mail settings
Submitter Kangjie Lu
Date Jan. 9, 2019, 8:14 a.m.
Message ID <20190109081423.10781-1-kjlu@umn.edu>
Download mbox | patch
Permalink /patch/695509/
State New
Headers show

Comments

Kangjie Lu - Jan. 9, 2019, 8:14 a.m.
"user_buf->length" is in user space, and copied in twice. The second
copy is after it passes the security check. If a user program races to
change user_buf->length in user space, the data fetched in the second
copy may invalidate the security check. The fix avoids the double-fetch
issue by using the value passing the security check.

Signed-off-by: Kangjie Lu <kjlu@umn.edu>
---
 drivers/acpi/custom_method.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
Rafael J. Wysocki - Jan. 14, 2019, 11:15 a.m.
On Wed, Jan 9, 2019 at 9:14 AM Kangjie Lu <kjlu@umn.edu> wrote:
>
> "user_buf->length" is in user space, and copied in twice. The second
> copy is after it passes the security check. If a user program races to
> change user_buf->length in user space, the data fetched in the second
> copy may invalidate the security check. The fix avoids the double-fetch
> issue by using the value passing the security check.

AFAICS the patch really does two things: it avoids the issue described
above and avoids using the (redundant) 'table' local variable on the
stack.  Arguably, you could fix the issue without getting rid of the
redundant variable.

>
> Signed-off-by: Kangjie Lu <kjlu@umn.edu>
> ---
>  drivers/acpi/custom_method.c | 10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/acpi/custom_method.c b/drivers/acpi/custom_method.c
> index 4451877f83b6..f10ee0519033 100644
> --- a/drivers/acpi/custom_method.c
> +++ b/drivers/acpi/custom_method.c
> @@ -26,17 +26,16 @@ static ssize_t cm_write(struct file *file, const char __user * user_buf,
>         static u32 max_size;
>         static u32 uncopied_bytes;
>
> -       struct acpi_table_header table;
>         acpi_status status;
>
>         if (!(*ppos)) {
>                 /* parse the table header to get the table length */
>                 if (count <= sizeof(struct acpi_table_header))
>                         return -EINVAL;
> -               if (copy_from_user(&table, user_buf,
> -                                  sizeof(struct acpi_table_header)))
> +               if (get_user(max_size,
> +                                       &((struct acpi_table_header *)user_buf)->length))
>                         return -EFAULT;
> -               uncopied_bytes = max_size = table.length;
> +               uncopied_bytes = max_size;
>                 buf = kzalloc(max_size, GFP_KERNEL);
>                 if (!buf)
>                         return -ENOMEM;
> @@ -57,6 +56,9 @@ static ssize_t cm_write(struct file *file, const char __user * user_buf,
>                 return -EFAULT;
>         }
>
> +       /* Ensure table length is not changed in the second copy */
> +       ((struct acpi_table_header *)(buf + (*ppos)))->length = max_size;

Why don't you return -EFAULT if max_size is different from ->length?
Surely, the table should not be used at all in that case.

Moreover, wouldn't it be even better to compare the entire header with
the one read previously and return -EFAULT if they don't match?

> +
>         uncopied_bytes -= count;
>         *ppos += count;
>
> --
Rafael J. Wysocki - Jan. 15, 2019, 7:05 p.m.
On Tue, Jan 15, 2019 at 7:05 AM Kangjie Lu <kjlu@umn.edu> wrote:
>
>
>
> On Mon, Jan 14, 2019 at 5:15 AM Rafael J. Wysocki <rafael@kernel.org> wrote:
>>
>> On Wed, Jan 9, 2019 at 9:14 AM Kangjie Lu <kjlu@umn.edu> wrote:
>> >
>> > "user_buf->length" is in user space, and copied in twice. The second
>> > copy is after it passes the security check. If a user program races to
>> > change user_buf->length in user space, the data fetched in the second
>> > copy may invalidate the security check. The fix avoids the double-fetch
>> > issue by using the value passing the security check.
>>
>> AFAICS the patch really does two things: it avoids the issue described
>> above and avoids using the (redundant) 'table' local variable on the
>> stack.  Arguably, you could fix the issue without getting rid of the
>> redundant variable.
>>
>> >
>> > Signed-off-by: Kangjie Lu <kjlu@umn.edu>
>> > ---
>> >  drivers/acpi/custom_method.c | 10 ++++++----
>> >  1 file changed, 6 insertions(+), 4 deletions(-)
>> >
>> > diff --git a/drivers/acpi/custom_method.c b/drivers/acpi/custom_method.c
>> > index 4451877f83b6..f10ee0519033 100644
>> > --- a/drivers/acpi/custom_method.c
>> > +++ b/drivers/acpi/custom_method.c
>> > @@ -26,17 +26,16 @@ static ssize_t cm_write(struct file *file, const char __user * user_buf,
>> >         static u32 max_size;
>> >         static u32 uncopied_bytes;
>> >
>> > -       struct acpi_table_header table;
>> >         acpi_status status;
>> >
>> >         if (!(*ppos)) {
>> >                 /* parse the table header to get the table length */
>> >                 if (count <= sizeof(struct acpi_table_header))
>> >                         return -EINVAL;
>> > -               if (copy_from_user(&table, user_buf,
>> > -                                  sizeof(struct acpi_table_header)))
>> > +               if (get_user(max_size,
>> > +                                       &((struct acpi_table_header *)user_buf)->length))
>> >                         return -EFAULT;
>> > -               uncopied_bytes = max_size = table.length;
>> > +               uncopied_bytes = max_size;
>> >                 buf = kzalloc(max_size, GFP_KERNEL);
>> >                 if (!buf)
>> >                         return -ENOMEM;
>> > @@ -57,6 +56,9 @@ static ssize_t cm_write(struct file *file, const char __user * user_buf,
>> >                 return -EFAULT;
>> >         }
>> >
>> > +       /* Ensure table length is not changed in the second copy */
>> > +       ((struct acpi_table_header *)(buf + (*ppos)))->length = max_size;
>>
>> Why don't you return -EFAULT if max_size is different from ->length?
>> Surely, the table should not be used at all in that case.
>
>
> We could do either, but I didn't see one is clearly better than the other.

As I said, why would you use any inconsistent data instead of
returning an error?

>>
>> Moreover, wouldn't it be even better to compare the entire header with
>> the one read previously and return -EFAULT if they don't match?
>
>
> If other fields are not critical and thus not checked, we don't have to compare
> the entire header for better performance reasons.

If you really care about consistency, performance doesn't matter that much.

Thanks,
Rafael

Patch

diff --git a/drivers/acpi/custom_method.c b/drivers/acpi/custom_method.c
index 4451877f83b6..f10ee0519033 100644
--- a/drivers/acpi/custom_method.c
+++ b/drivers/acpi/custom_method.c
@@ -26,17 +26,16 @@  static ssize_t cm_write(struct file *file, const char __user * user_buf,
 	static u32 max_size;
 	static u32 uncopied_bytes;
 
-	struct acpi_table_header table;
 	acpi_status status;
 
 	if (!(*ppos)) {
 		/* parse the table header to get the table length */
 		if (count <= sizeof(struct acpi_table_header))
 			return -EINVAL;
-		if (copy_from_user(&table, user_buf,
-				   sizeof(struct acpi_table_header)))
+		if (get_user(max_size,
+					&((struct acpi_table_header *)user_buf)->length))
 			return -EFAULT;
-		uncopied_bytes = max_size = table.length;
+		uncopied_bytes = max_size;
 		buf = kzalloc(max_size, GFP_KERNEL);
 		if (!buf)
 			return -ENOMEM;
@@ -57,6 +56,9 @@  static ssize_t cm_write(struct file *file, const char __user * user_buf,
 		return -EFAULT;
 	}
 
+	/* Ensure table length is not changed in the second copy */
+	((struct acpi_table_header *)(buf + (*ppos)))->length = max_size;
+
 	uncopied_bytes -= count;
 	*ppos += count;