On 13/09/19 06:46, Greg Kroah-Hartman wrote:
> USB drivers expect kmalloc to return DMA-able memory. I don't know
> about specific alignment issues, that should only an issue for the host
> controller being used here, which you do not say in the above list.
I have no idea, this is just the analysis of a syzkaller report. From
the backtrace, it's one that ends up calling kmalloc; all of them should
have the same issue with KASAN.
The specific alignment requirement for this bug comes from this call in
usbdev_mmap:
if (remap_pfn_range(vma, vma->vm_start,
virt_to_phys(usbm->mem) >> PAGE_SHIFT,
size, vma->vm_page_prot) < 0) {
> We have had some reports that usbdev_mmap() does not do the "correct
> thing" for all host controllers, but a lot of the DMA work that is in
> linux-next for 5.4-rc1 should have helped resolve those issues. What
> tree are you seeing these bug reports happening from?
It's in master, but the relevant code is the same in linux-next; in fact
in this case there is no DMA involved at all. hcd_buffer_alloc hits
the case "some USB hosts just use PIO".
On those host controllers, it should be reproducible with just this:
diff --git a/drivers/usb/core/usb.c b/drivers/usb/core/usb.c
index 7fcb9f782931..cc0460730bce 100644
--- a/drivers/usb/core/usb.c
+++ b/drivers/usb/core/usb.c
@@ -905,9 +905,12 @@ EXPORT_SYMBOL_GPL(__usb_get_extra_descriptor);
void *usb_alloc_coherent(struct usb_device *dev, size_t size, gfp_t mem_flags,
dma_addr_t *dma)
{
+ void *buf;
if (!dev || !dev->bus)
return NULL;
- return hcd_buffer_alloc(dev->bus, size, mem_flags, dma);
+ buf = hcd_buffer_alloc(dev->bus, size, mem_flags, dma);
+ WARN_ON_ONCE(virt_to_phys(buf) & ~PAGE_MASK);
+ return buf;
}
EXPORT_SYMBOL_GPL(usb_alloc_coherent);
and CONFIG_KASAN=y or possibly just CONFIG_DEBUG_SLAB=y. mmap-ing /dev/usb
should warn if my analysis is correct.
If you think the above patch makes sense, I can test it and submit it formally.
Paolo
