Skip to contentSkip to conversation

vfio: hisi_acc_vfio_pci: fix integer overflow check in hisi_acc_vf_resume_write()

PATCH In review Dan Carpenter 2 participants 5 messages Jul 5, 2022 – Jul 7, 2022

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 vfio: hisi_acc_vfio_pci: fix integer overflow check in hisi_acc_vf_resume_write()

The casting on this makes the integer overflow check slightly wrong. "len" is an unsigned long. "*pos" and "requested_length" are signed long longs. Imagine "len" is ULONG_MAX and "*pos" is 2. "ULONG_MAX + 2 = 1". That's an integer overflow. However, if we cast the ULONG_MAX to long long then "-1 + 2 = 1". That's not an integer overflow.

It's simpler if "requested_length" length is an unsigned value so we don't have to worry about negatives.

I believe that the checks in the VFS layer and the check for "*pos < 0" probably prevent this bug in real life, but it's safer to just be sure.

Fixes b0eed085903e ("hisi_acc_vfio_pci: Add support for VFIO live migration")

---
It is strange that we are doing:

	pos = &filp->f_pos;

instead of using the passed in value of pos. The VFS layer ensures that the passed in value of "*pos + len" cannot overflow in rw_verify_area() so normally this check could have been removed.

 drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/vfio/pci/hisilicon/hisi_acc_vfio_pci.c +3 -3
index ea762e28c1cc..dcc34488b0c0 100644
@@ -701,7 +701,7 @@ static ssize_t hisi_acc_vf_resume_write(struct file *filp, const char __user *bu
 					size_t len, loff_t *pos)
 {
 	struct hisi_acc_vf_migration_file *migf = filp->private_data;
-	loff_t requested_length;
+	unsigned long requested_length;
 	ssize_t done = 0;
 	int ret;
 
@@ -709,8 +709,8 @@ static ssize_t hisi_acc_vf_resume_write(struct file *filp, const char __user *bu
 		return -ESPIPE;
 	pos = &filp->f_pos;
 
-	if (*pos < 0 ||
-	    check_add_overflow((loff_t)len, *pos, &requested_length))
+	if (*pos < 0 || *pos > ULONG_MAX ||
+	    check_add_overflow(len, (unsigned long)*pos, &requested_length))
 		return -EINVAL;
 
 	if (requested_length > sizeof(struct acc_vf_data))
Signature · 1 line
2.35.1
#1 lore raw
Quoted 4 lines · Dan Carpenter

The casting on this makes the integer overflow check slightly wrong. "len" is an unsigned long. "*pos" and "requested_length" are signed long longs. Imagine "len" is ULONG_MAX and "*pos" is 2. "ULONG_MAX + 2 = 1".

I wonder if this can happen, len is a kernel controlled value bounded by a memory allocation..

Quoted 3 lines
Fixes b0eed085903e ("hisi_acc_vfio_pci: Add support for VFIO live migration")

---

This code was copy and pasted from drivers/vfio/pci/mlx5/main.c, so it should be fixed too

Quoted 5 lines

It is strange that we are doing:

	pos = &filp->f_pos;

instead of using the passed in value of pos.

IIRC the way we have the struct file configured the pos argument is NULL.

Jason

#2 lore raw
Quoted 9 lines · Jason Gunthorpe
Quoted 4 lines · Dan Carpenter

The casting on this makes the integer overflow check slightly wrong. "len" is an unsigned long. "*pos" and "requested_length" are signed long longs. Imagine "len" is ULONG_MAX and "*pos" is 2. "ULONG_MAX + 2 = 1".

I wonder if this can happen, len is a kernel controlled value bounded by a memory allocation..

Oh. Smatch uses a model which says that all read/writes come from vfs_write(). The problem with tracking kernel read/writes is that recursion is tricky. So Smatch just deletes those from the DB.

Quoted 6 lines
Quoted 3 lines
Fixes b0eed085903e ("hisi_acc_vfio_pci: Add support for VFIO live migration")

---

This code was copy and pasted from drivers/vfio/pci/mlx5/main.c, so it should be fixed too

Sure.

I created a static checker warning for this type of thing but it didn't catch the issue in drivers/vfio/pci/mlx5/main.c because Smatch says that the bug is impossible. Which is true.

Smatch doesn't really parse rw_verify_area() accurately. I just hard coded that function as accepting values 0-1000000000 for both *ppos and count.

regards,
dan carpenter

#3 lore raw
Quoted 7 lines · Dan Carpenter
Quoted 6 lines · Jason Gunthorpe
Quoted 4 lines · Dan Carpenter

The casting on this makes the integer overflow check slightly wrong. "len" is an unsigned long. "*pos" and "requested_length" are signed long longs. Imagine "len" is ULONG_MAX and "*pos" is 2. "ULONG_MAX + 2 = 1".

Quoted 7 lines
Quoted 3 lines

I wonder if this can happen, len is a kernel controlled value bounded by a memory allocation..

Oh. Smatch uses a model which says that all read/writes come from vfs_write(). The problem with tracking kernel read/writes is that recursion is tricky. So Smatch just deletes those from the DB.

Oh, maybe I got it wrong, len is the user input, so yes that does look bad

Quoted 8 lines
Quoted 2 lines

This code was copy and pasted from drivers/vfio/pci/mlx5/main.c, so it should be fixed too

Sure.

I created a static checker warning for this type of thing but it didn't catch the issue in drivers/vfio/pci/mlx5/main.c because Smatch says that the bug is impossible. Which is true.

How come it is different?

Jason

#4 lore raw
Quoted 5 lines · Jason Gunthorpe
Quoted 3 lines · Dan Carpenter
Quoted 2 lines

This code was copy and pasted from drivers/vfio/pci/mlx5/main.c, so it should be fixed too

Quoted 7 lines
Quoted 5 lines

Sure.

I created a static checker warning for this type of thing but it didn't catch the issue in drivers/vfio/pci/mlx5/main.c because Smatch says that the bug is impossible. Which is true.

How come it is different?

No, it doesn't find either one. I don't think it's a real bug because of the rw_verify_area() thing. But I wrote the check based on noticing it during review just to see if there were similar issues and didn't find anything.

regards,
dan carpenter