Skip to contentSkip to conversation

staging: vme_user: bound slave windows to DMA buffers

PATCH In review Yousef Alhouseen 3 participants 5 messages Jun 25

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 staging: vme_user: bound slave windows to DMA buffers

vme_user allocates a fixed PCI_BUF_SIZE coherent DMA buffer for each slave image, but VME_SET_SLAVE passes the user-supplied size to vme_slave_set() unchanged. An enabled window larger than PCI_BUF_SIZE lets the bridge expose DMA addresses beyond the allocation.

The character device read/write paths also derive their bounds from vme_get_size(), so the oversized programmed window can make buffer_to_user() and buffer_from_user() access past image[minor].kern_buf.

Reject enabled slave windows that do not fit in the backing buffer and use the same capped size for slave read/write/llseek bounds. Also convert the read/write limit checks to subtraction-based form so offset + count cannot wrap around the image size check.

---
 drivers/staging/vme_user/vme_user.c | 38 ++++++++++++++++++++++-------
 1 file changed, 29 insertions(+), 9 deletions(-)
drivers/staging/vme_user/vme_user.c +29 -9
index 11e25c2f6..24d3c6ec3 100644
@@ -175,11 +175,22 @@ static ssize_t buffer_from_user(unsigned int minor, const char __user *buf,
 	return count;
 }
 
+static size_t vme_user_get_image_size(unsigned int minor)
+{
+	size_t image_size = vme_get_size(image[minor].resource);
+
+	if (type[minor] == SLAVE_MINOR)
+		image_size = min_t(size_t, image_size, image[minor].size_buf);
+
+	return image_size;
+}
+
 static ssize_t vme_user_read(struct file *file, char __user *buf, size_t count,
 			     loff_t *ppos)
 {
 	unsigned int minor = iminor(file_inode(file));
 	ssize_t retval;
+	size_t offset;
 	size_t image_size;
 
 	if (minor == CONTROL_MINOR)
@@ -188,17 +199,19 @@ static ssize_t vme_user_read(struct file *file, char __user *buf, size_t count,
 	mutex_lock(&image[minor].mutex);
 
 	/* XXX Do we *really* want this helper - we can use vme_*_get ? */
-	image_size = vme_get_size(image[minor].resource);
+	image_size = vme_user_get_image_size(minor);
 
 	/* Ensure we are starting at a valid location */
-	if ((*ppos < 0) || (*ppos > (image_size - 1))) {
+	if ((*ppos < 0) || ((u64)*ppos >= image_size)) {
 		mutex_unlock(&image[minor].mutex);
 		return 0;
 	}
 
+	offset = *ppos;
+
 	/* Ensure not reading past end of the image */
-	if (*ppos + count > image_size)
-		count = image_size - *ppos;
+	if (count > image_size - offset)
+		count = image_size - offset;
 
 	switch (type[minor]) {
 	case MASTER_MINOR:
@@ -223,6 +236,7 @@ static ssize_t vme_user_write(struct file *file, const char __user *buf,
 {
 	unsigned int minor = iminor(file_inode(file));
 	ssize_t retval;
+	size_t offset;
 	size_t image_size;
 
 	if (minor == CONTROL_MINOR)
@@ -230,17 +244,19 @@ static ssize_t vme_user_write(struct file *file, const char __user *buf,
 
 	mutex_lock(&image[minor].mutex);
 
-	image_size = vme_get_size(image[minor].resource);
+	image_size = vme_user_get_image_size(minor);
 
 	/* Ensure we are starting at a valid location */
-	if ((*ppos < 0) || (*ppos > (image_size - 1))) {
+	if ((*ppos < 0) || ((u64)*ppos >= image_size)) {
 		mutex_unlock(&image[minor].mutex);
 		return 0;
 	}
 
+	offset = *ppos;
+
 	/* Ensure not reading past end of the image */
-	if (*ppos + count > image_size)
-		count = image_size - *ppos;
+	if (count > image_size - offset)
+		count = image_size - offset;
 
 	switch (type[minor]) {
 	case MASTER_MINOR:
@@ -271,7 +287,7 @@ static loff_t vme_user_llseek(struct file *file, loff_t off, int whence)
 	case MASTER_MINOR:
 	case SLAVE_MINOR:
 		mutex_lock(&image[minor].mutex);
-		image_size = vme_get_size(image[minor].resource);
+		image_size = vme_user_get_image_size(minor);
 		res = fixed_size_llseek(file, off, whence, image_size);
 		mutex_unlock(&image[minor].mutex);
 		return res;
@@ -394,6 +410,10 @@ static int vme_user_ioctl(struct inode *inode, struct file *file,
 				return -EFAULT;
 			}
 
+			if (slave.enable &&
+			    (!slave.size || slave.size > image[minor].size_buf))
+				return -EINVAL;
+
 			/* XXX	We do not want to push aspace, cycle and width
 			 *	to userspace as they are
 			 */
Signature · 1 line
2.54.0
#1 lore raw
Quoted 15 lines · Yousef Alhouseen

vme_user allocates a fixed PCI_BUF_SIZE coherent DMA buffer for each slave image, but VME_SET_SLAVE passes the user-supplied size to vme_slave_set() unchanged. An enabled window larger than PCI_BUF_SIZE lets the bridge expose DMA addresses beyond the allocation.

The character device read/write paths also derive their bounds from vme_get_size(), so the oversized programmed window can make buffer_to_user() and buffer_from_user() access past image[minor].kern_buf.

Reject enabled slave windows that do not fit in the backing buffer and use the same capped size for slave read/write/llseek bounds. Also convert the read/write limit checks to subtraction-based form so offset + count cannot wrap around the image size check.

Quoted 4 lines
---
 drivers/staging/vme_user/vme_user.c | 38 ++++++++++++++++++++++-------
 1 file changed, 29 insertions(+), 9 deletions(-)

What tool did you use to find/make this change, and how was it tested?

thanks,

greg k-h

#2 lore raw
In reply to Yousef Alhouseen

There are two bug fixes in this patch:

1) The driver allows you to set any value for size.
2) The vme_get_size() function returns zero on error but nothing checks for error.

There is a third bug as well which is not addressed.

3) The slave.vme_addr is not checked and it allows you to read/write to any memory.

It sort of looks like bugs 1 and 2 only affect drivers/staging/vme_user/vme_fake.c. I feel like people hacked together some debug code and weren't too worried about the security since it wasn't intended to be used on real systems... The temptation is to fix this by deleting at least vme_fake.c.

Quoted 3 lines · Yousef Alhouseen

vme_user allocates a fixed PCI_BUF_SIZE coherent DMA buffer for each slave image, but VME_SET_SLAVE passes the user-supplied size to vme_slave_set() unchanged.

Obviously you meant checked instead of changed... There is some checking in vme_check_window() but it's not suffient.

Quoted 8 lines

An enabled window larger than PCI_BUF_SIZE lets the bridge expose DMA addresses beyond the allocation.

The character device read/write paths also derive their bounds from vme_get_size(), so the oversized programmed window can make buffer_to_user() and buffer_from_user() access past image[minor].kern_buf.

Quoted 4 lines

Reject enabled slave windows that do not fit in the backing buffer and use the same capped size for slave read/write/llseek bounds. Also convert the read/write limit checks to subtraction-based form so offset + count cannot wrap around the image size check.

An "also" in the commit message, means split it out into its own patch.

Quoted 3 lines

---

Fixes tag. Please say when you are using AI.

Quoted 10 lines
 drivers/staging/vme_user/vme_user.c | 38 ++++++++++++++++++++++-------
 1 file changed, 29 insertions(+), 9 deletions(-)
drivers/staging/vme_user/vme_user.c +0 -0
index 11e25c2f6..24d3c6ec3 100644
@@ -175,11 +175,22 @@ static ssize_t buffer_from_user(unsigned int minor, const char __user *buf,
 	return count;
 }

drivers/staging/vme_user/vme_user.c static ssize_t buffer_from_user(unsigned int minor, const char __user *buf, · line 186

Quoted 10 lines
drivers/staging/vme_user/vme_user.c +9 -0
 
+static size_t vme_user_get_image_size(unsigned int minor)
+{
+	size_t image_size = vme_get_size(image[minor].resource);
+
+	if (type[minor] == SLAVE_MINOR)
+		image_size = min_t(size_t, image_size, image[minor].size_buf);
+
+	return image_size;
+}

No, vme_get_size() should return the correct size. If it's really a bug then fix that instead of adding a work-around here.

drivers/staging/vme_user/vme_user.c static ssize_t buffer_from_user(unsigned int minor, const char __user *buf, · line 193

Quoted 7 lines
drivers/staging/vme_user/vme_user.c +2 -0
+
 static ssize_t vme_user_read(struct file *file, char __user *buf, size_t count,
 			     loff_t *ppos)
 {
 	unsigned int minor = iminor(file_inode(file));
 	ssize_t retval;
+	size_t offset;

No need for this variable.

drivers/staging/vme_user/vme_user.c static ssize_t vme_user_read(struct file *file, char __user *buf, size_t count, · line 202

Quoted 9 lines
drivers/staging/vme_user/vme_user.c +1 -1
 	size_t image_size;
 
 	if (minor == CONTROL_MINOR)
@@ -188,17 +199,19 @@ static ssize_t vme_user_read(struct file *file, char __user *buf, size_t count,
 	mutex_lock(&image[minor].mutex);
 
 	/* XXX Do we *really* want this helper - we can use vme_*_get ? */
-	image_size = vme_get_size(image[minor].resource);
+	image_size = vme_user_get_image_size(minor);

vme_get_size() returns zero on error. Just add an check for errors:

	if (image_size == 0) {
		mutex_unlock(&image[minor].mutex);
		return 0;
	}

drivers/staging/vme_user/vme_user.c static ssize_t vme_user_read(struct file *file, char __user *buf, size_t count, · line 205

Quoted 4 lines
drivers/staging/vme_user/vme_user.c +1 -1
 	/* Ensure we are starting at a valid location */
-	if ((*ppos < 0) || (*ppos > (image_size - 1))) {
+	if ((*ppos < 0) || ((u64)*ppos >= image_size)) {

No need for this cast.

	if ((*ppos < 0) || (*ppos >= image_size)) {

But actually if we check for zero then this is unnecesssary.

drivers/staging/vme_user/vme_user.c static ssize_t vme_user_read(struct file *file, char __user *buf, size_t count, · line 214

Quoted 11 lines
drivers/staging/vme_user/vme_user.c +4 -2
 		mutex_unlock(&image[minor].mutex);
 		return 0;
 	}
 
+	offset = *ppos;
+
 	/* Ensure not reading past end of the image */
-	if (*ppos + count > image_size)
-		count = image_size - *ppos;
+	if (count > image_size - offset)
+		count = image_size - offset;

The kernel ensures that "*ppos + count" can't overflow in rw_verify_area() so this change is unnecessary.

Quoted 45 lines
drivers/staging/vme_user/vme_user.c +8 -5
 	switch (type[minor]) {
 	case MASTER_MINOR:
@@ -223,6 +236,7 @@ static ssize_t vme_user_write(struct file *file, const char __user *buf,
 {
 	unsigned int minor = iminor(file_inode(file));
 	ssize_t retval;
+	size_t offset;
 	size_t image_size;
 
 	if (minor == CONTROL_MINOR)
@@ -230,17 +244,19 @@ static ssize_t vme_user_write(struct file *file, const char __user *buf,
 
 	mutex_lock(&image[minor].mutex);
 
-	image_size = vme_get_size(image[minor].resource);
+	image_size = vme_user_get_image_size(minor);
 
 	/* Ensure we are starting at a valid location */
-	if ((*ppos < 0) || (*ppos > (image_size - 1))) {
+	if ((*ppos < 0) || ((u64)*ppos >= image_size)) {
 		mutex_unlock(&image[minor].mutex);
 		return 0;
 	}
 
+	offset = *ppos;
+
 	/* Ensure not reading past end of the image */
-	if (*ppos + count > image_size)
-		count = image_size - *ppos;
+	if (count > image_size - offset)
+		count = image_size - offset;
 
 	switch (type[minor]) {
 	case MASTER_MINOR:
@@ -271,7 +287,7 @@ static loff_t vme_user_llseek(struct file *file, loff_t off, int whence)
 	case MASTER_MINOR:
 	case SLAVE_MINOR:
 		mutex_lock(&image[minor].mutex);
-		image_size = vme_get_size(image[minor].resource);
+		image_size = vme_user_get_image_size(minor);
 		res = fixed_size_llseek(file, off, whence, image_size);
 		mutex_unlock(&image[minor].mutex);
 		return res;

drivers/staging/vme_user/vme_user.c static int vme_user_ioctl(struct inode *inode, struct file *file, · line 415

Quoted 6 lines
drivers/staging/vme_user/vme_user.c +3 -0
@@ -394,6 +410,10 @@ static int vme_user_ioctl(struct inode *inode, struct file *file,
 				return -EFAULT;
 			}
 
+			if (slave.enable &&
+			    (!slave.size || slave.size > image[minor].size_buf))
+				return -EINVAL;

This isn't the right place to check... It should probably be done in vme_check_window()?

This code is quite tricky and I don't really understand it all. Why does resource_to/from_user() check for buffer overflow but buffer_to/from_user() does not? The checking in resource_to/from_user() is also insufficient because it doesn't take *ppos into consideration but fortunately, vme_master_read/write() does. This code would be safer if we added bounds checking to buffer_to/from_user().

regards,
dan carpenter

Quoted 7 lines
drivers/staging/vme_user/vme_user.c +1 -0
+
 			/* XXX	We do not want to push aspace, cycle and width
 			 *	to userspace as they are
 			 */
Signature · 1 line
2.54.0
#3 lore raw
Quoted 6 lines · Dan Carpenter

There are two bug fixes in this patch:

1) The driver allows you to set any value for size.
2) The vme_get_size() function returns zero on error but nothing checks for error.

Quoted 7 lines

There is a third bug as well which is not addressed.

3) The slave.vme_addr is not checked and it allows you to read/write to any memory.

It sort of looks like bugs 1 and 2 only affect
drivers/staging/vme_user/vme_fake.c.

Actually, no. The bugs seem to affect drivers/staging/vme_user/vme_tsi148.c but the variables are stored to registers instead of variables so it was a little more convoluted.

regards,
dan carpenter

#4 lore raw

Hi Dan, Greg,

Thanks for going through it. Agreed, this version mixes separate fixes and the helper is the wrong shape. Please drop it.

I found it while reading the vme_user ioctl and read/write paths. Testing was only git diff --check and checkpatch; I don't have VME hardware.

I'll rework from the core/bridge side and split out anything that still holds up.

Thanks,
Yousef

Quoted 21 lines · Dan Carpenter
Quoted 6 lines · Dan Carpenter

There are two bug fixes in this patch:

1) The driver allows you to set any value for size.
2) The vme_get_size() function returns zero on error but nothing checks for error.

Quoted 7 lines

There is a third bug as well which is not addressed.

3) The slave.vme_addr is not checked and it allows you to read/write to any memory.

It sort of looks like bugs 1 and 2 only affect
drivers/staging/vme_user/vme_fake.c.

Actually, no. The bugs seem to affect drivers/staging/vme_user/vme_tsi148.c but the variables are stored to registers instead of variables so it was a little more convoluted.

regards,
dan carpenter