On Mon, Jan 11, 2021 at 9:31 PM Bui Quang Minh [off-list ref] wrote:

On Mon, Jan 11, 2021 at 01:00:31PM +0100, Oliver Neukum wrote:

quoted

Am Montag, den 11.01.2021, 10:49 +0000 schrieb Bui Quang Minh:

quoted

In mcba_usb_read_bulk_callback(), when we don't resubmit or fails to
resubmit the urb, we need to deallocate the transfer buffer that is
allocated in mcba_usb_start().

Reported-by: syzbot+57281c762a3922e14dfe@syzkaller.appspotmail.com
Signed-off-by: Bui Quang Minh <redacted>
---
v1: add memory leak fix when not resubmitting urb
v2: add memory leak fix when failing to resubmit urb

 drivers/net/can/usb/mcba_usb.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/drivers/net/can/usb/mcba_usb.c b/drivers/net/can/usb/mcba_usb.c
index df54eb7d4b36..30236e640116 100644
--- a/drivers/net/can/usb/mcba_usb.c
+++ b/drivers/net/can/usb/mcba_usb.c

@@ -584,6 +584,8 @@ static void mcba_usb_read_bulk_callback(struct urb *urb)
    case -EPIPE:
    case -EPROTO:
    case -ESHUTDOWN:
+           usb_free_coherent(urb->dev, urb->transfer_buffer_length,
+                             urb->transfer_buffer, urb->transfer_dma);
            return;

Can you call usb_free_coherent() in what can be hard IRQ context?

You are right, I digged in the code and saw some comments that on some
architectures, usb_free_coherent() cannot be called in hard IRQ context.
I see the usb_free_coherent() is called in write_bulk_callback too. I will
send a patch that uses usb_anchor to keep track of these urbs and cleanup
the transfer buffer later in disconnect().

Hi, I have sent a version 3 patch. However, I found out that usb_free_coherent()
is ok in this situation. In usb_free_coherent(), if the buffer is allocated via
dma_alloc_coherent() in usb_alloc_coherent(), dma_free_coherent() is called.
In dma_free_coherent(), ops->free() may be called in some cases which may
contains calls to vunmap() that is not permitted in interrupt context. However,
in usb_alloc_coherent(), buffer can be allocated from dma pool if the
size is less
than 2048 and the buffer size in mcba_usb is obviously less than 2048.
As a result,
usb_free_coherent() will at most fall in the path that calls
dma_pool_free(), which is
safe. Am I right?

Thanks,
Quang Minh.