On Tue, Dec 6, 2016 at 10:32 AM, Johannes Thumshirn [off-list ref] wrote:
On Mon, Dec 05, 2016 at 07:03:39PM +0000, Al Viro wrote:
quoted
On Mon, Dec 05, 2016 at 04:17:53PM +0100, Johannes Thumshirn wrote:
quoted
633 hp = &srp->header;
[...]
646 hp->dxferp = (char __user *)buf + cmd_size;
quoted
So the memory for hp->dxferp comes from:
633 hp = &srp->header;
????
quoted
quoted
From my debug instrumentation I see that the dxferp ends up in the
iovec_iter's kvec->iov_base and the faulting address is always dxferp + n *
4k with n in [1, 16] (and we're copying 16 4k pages from the iovec into the
bio).
_Address_ of hp->dxferp comes from that assignment; the value is 'buf'
argument of sg_write() + small offset. In this case, it should point
inside a pipe buffer, which is, indeed, at a kernel address. Who'd
allocated srp is irrelevant.
Yes I realized that as well when I had enough distance between me and the
code...
quoted
And if you end up dereferencing more than one page worth there, you do have
a problem - pipe buffers are not going to be that large. Could you slap
WARN_ON((size_t)input_size > count);
right after the calculation of input_size in sg_write() and see if it triggers
on your reproducer?
I did and it didn't trigger. What triggers is (as expected) a
WARN_ON((size_t)mxsize > count);
We have count at 80 and mxsize (which ends in hp->dxfer_len) at 65499. But the
65499 bytes are the len of the data we're suppost to be copying in via the
iov. I'm still rather confused what's happening here, sorry.
I think the critical piece here is some kind of race or timing
condition. Note that the test program executes all of
memfd_create/write/open/sendfile twice. Second time the calls race
with each other, but they also can race with the first execution of
the calls.