2020-03-07 13:10:54

by Måns Rullgård

Subject: [PATCH] usb: musb: fix crash with highmen PIO and usbmon

When handling a PIO bulk transfer with highmem buffer, a temporary
mapping is assigned to urb->transfer_buffer. After the transfer is
complete, an invalid address is left behind in this pointer. This is
not ordinarily a problem since nothing touches that buffer before the
urb is released. However, when usbmon is active, usbmon_urb_complete()
calls (indirectly) mon_bin_get_data() which does access the transfer
buffer if it is set. To prevent an invalid memory access here, reset
urb->tranfer_buffer to NULL when finished.

Fixes: 8e8a55165469 ("usb: musb: host: Handle highmem in PIO mode")
Signed-off-by: Mans Rullgard <[email protected]>
---
drivers/usb/musb/musb_host.c | 8 ++++++--
1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/drivers/usb/musb/musb_host.c b/drivers/usb/musb/musb_host.c
index 1c813c37462a..b67b40de1947 100644
--- a/drivers/usb/musb/musb_host.c
+++ b/drivers/usb/musb/musb_host.c
@@ -1459,8 +1459,10 @@ void musb_host_tx(struct musb *musb, u8 epnum)
qh->segsize = length;

if (qh->use_sg) {
- if (offset + length >= urb->transfer_buffer_length)
+ if (offset + length >= urb->transfer_buffer_length) {
qh->use_sg = false;
+ urb->transfer_buffer = NULL;
+ }
}

musb_ep_select(mbase, epnum);
@@ -1977,8 +1979,10 @@ void musb_host_rx(struct musb *musb, u8 epnum)
urb->actual_length += xfer_len;
qh->offset += xfer_len;
if (done) {
- if (qh->use_sg)
+ if (qh->use_sg) {
qh->use_sg = false;
+ urb->transfer_buffer = NULL;
+ }

if (urb->status == -EINPROGRESS)
urb->status = status;
--
2.25.1

2020-03-09 13:56:55

by Bin Liu

Subject: Re: [PATCH] usb: musb: fix crash with highmen PIO and usbmon

Hi Mans,

On Sat, Mar 07, 2020 at 01:07:20PM +0000, Mans Rullgard wrote:
> When handling a PIO bulk transfer with highmem buffer, a temporary
> mapping is assigned to urb->transfer_buffer. After the transfer is
> complete, an invalid address is left behind in this pointer. This is
> not ordinarily a problem since nothing touches that buffer before the
> urb is released. However, when usbmon is active, usbmon_urb_complete()
> calls (indirectly) mon_bin_get_data() which does access the transfer
> buffer if it is set. To prevent an invalid memory access here, reset
> urb->tranfer_buffer to NULL when finished.
>
> Fixes: 8e8a55165469 ("usb: musb: host: Handle highmem in PIO mode")
> Signed-off-by: Mans Rullgard <[email protected]>

Thanks for fixing the bug.

> ---
> drivers/usb/musb/musb_host.c | 8 ++++++--
> 1 file changed, 6 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/usb/musb/musb_host.c b/drivers/usb/musb/musb_host.c
> index 1c813c37462a..b67b40de1947 100644
> --- a/drivers/usb/musb/musb_host.c
> +++ b/drivers/usb/musb/musb_host.c
> @@ -1459,8 +1459,10 @@ void musb_host_tx(struct musb *musb, u8 epnum)
> qh->segsize = length;
>
> if (qh->use_sg) {
> - if (offset + length >= urb->transfer_buffer_length)
> + if (offset + length >= urb->transfer_buffer_length) {
> qh->use_sg = false;
> + urb->transfer_buffer = NULL;
> + }

In this tx case, can you directly pass qh->sg_miter.addr to
musb_write_fifo() so that urb->transfer_buffer is not touched at all?

-Bin.

2020-03-09 14:51:10

by Måns Rullgård

Subject: Re: [PATCH] usb: musb: fix crash with highmen PIO and usbmon

Bin Liu <[email protected]> writes:

> Hi Mans,
>
> On Sat, Mar 07, 2020 at 01:07:20PM +0000, Mans Rullgard wrote:
>> When handling a PIO bulk transfer with highmem buffer, a temporary
>> mapping is assigned to urb->transfer_buffer. After the transfer is
>> complete, an invalid address is left behind in this pointer. This is
>> not ordinarily a problem since nothing touches that buffer before the
>> urb is released. However, when usbmon is active, usbmon_urb_complete()
>> calls (indirectly) mon_bin_get_data() which does access the transfer
>> buffer if it is set. To prevent an invalid memory access here, reset
>> urb->tranfer_buffer to NULL when finished.
>>
>> Fixes: 8e8a55165469 ("usb: musb: host: Handle highmem in PIO mode")
>> Signed-off-by: Mans Rullgard <[email protected]>
>
> Thanks for fixing the bug.
>
>> ---
>> drivers/usb/musb/musb_host.c | 8 ++++++--
>> 1 file changed, 6 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/usb/musb/musb_host.c b/drivers/usb/musb/musb_host.c
>> index 1c813c37462a..b67b40de1947 100644
>> --- a/drivers/usb/musb/musb_host.c
>> +++ b/drivers/usb/musb/musb_host.c
>> @@ -1459,8 +1459,10 @@ void musb_host_tx(struct musb *musb, u8 epnum)
>> qh->segsize = length;
>>
>> if (qh->use_sg) {
>> - if (offset + length >= urb->transfer_buffer_length)
>> + if (offset + length >= urb->transfer_buffer_length) {
>> qh->use_sg = false;
>> + urb->transfer_buffer = NULL;
>> + }
>
> In this tx case, can you directly pass qh->sg_miter.addr to
> musb_write_fifo() so that urb->transfer_buffer is not touched at all?

Yes, that seems to work. I'll prepare a new patch.

--
M?ns Rullg?rd