Re: [BUG] net: chelsio: Possible buffer overflow caused by DMA failures/attacks

On Tue, May 05, 2020 at 11:50:28PM +0800, Jia-Ju Bai wrote:

In alloc_rx_resources():
    sge->respQ.entries =
        pci_alloc_consistent(pdev, size, &sge->respQ.dma_addr);

Thus, "sge->respQ.entries" is a DMA value, and it is assigned to
"e" in process_pure_responses():
    struct sge *sge = adapter->sge;
    struct respQ *q = &sge->respQ;
    struct respQ_e *e = &q->entries[q->cidx];

When DMA failures or attacks occur, the data stored in "e" can be
changed at any time. In this case, the value of "e->FreelistQid"
can be a large number to cause buffer overflow when the
following code is executed:
    const struct freelQ *fl = &sge->freelQ[e->FreelistQid];

Similarly, "sge->respQ.entries" is also assigned to "e" in
process_responses():
    struct sge *sge = adapter->sge;
    struct respQ *q = &sge->respQ;
    struct respQ_e *e = &q->entries[q->cidx];

When DMA failures or attacks occur, the data stored in "e" can be
changed at any time. In this case, the value of "e->FreelistQid"
can be a large number to cause buffer overflow when the
following code is executed:
    struct freelQ *fl = &sge->freelQ[e->FreelistQid];

Considering that DMA can fail or be attacked, I think that it is dangerous
to
use a DMA value (or any value tainted by it) as an array index or a
control-flow
condition. However, I have found many such dangerous cases in Linux device
drivers
through my static-analysis tool and code review.
I am not sure whether my opinion is correct, so I want to listen to your
points of view.

Can you create a patch to show what you think needs to be fixed?  That's
the best way to get feedback, reports like this are usually very
infrequently replied to.

thanks,

greg k-h