bpf, test_run: fix crashes due to XDP frame overwriting/corruption

syzbot and Ilya faced the splats when %XDP_PASS happens for bpf_test_run
after skb PP recycling was enabled for {__,}xdp_build_skb_from_frame():

BUG: kernel NULL pointer dereference, address: 0000000000000d28
RIP: 0010:memset_erms+0xd/0x20 arch/x86/lib/memset_64.S:66
[...]
Call Trace:
 <TASK>
 __finalize_skb_around net/core/skbuff.c:321 [inline]
 __build_skb_around+0x232/0x3a0 net/core/skbuff.c:379
 build_skb_around+0x32/0x290 net/core/skbuff.c:444
 __xdp_build_skb_from_frame+0x121/0x760 net/core/xdp.c:622
 xdp_recv_frames net/bpf/test_run.c:248 [inline]
 xdp_test_run_batch net/bpf/test_run.c:334 [inline]
 bpf_test_run_xdp_live+0x1289/0x1930 net/bpf/test_run.c:362
 bpf_prog_test_run_xdp+0xa05/0x14e0 net/bpf/test_run.c:1418
[...]

This happens due to that it calls xdp_scrub_frame(), which nullifies
xdpf->data. bpf_test_run code doesn't reinit the frame when the XDP
program doesn't adjust head or tail. Previously, %XDP_PASS meant the
page will be released from the pool and returned to the MM layer, but
now it does return to the Pool with the nullified xdpf->data, which
doesn't get reinitialized then.
So, in addition to checking whether the head and/or tail have been
adjusted, check also for a potential XDP frame corruption. xdpf->data
is 100% affected and also xdpf->flags is the field closest to the
metadata / frame start. Checking for these two should be enough for
non-extreme cases.

Fixes: 9c94bbf ("xdp: recycle Page Pool backed skbs built from XDP frames")
Reported-by: [email protected]
Link: https://lore.kernel.org/bpf/[email protected]
Reported-by: Ilya Leoshkevich <[email protected]>
Link: https://lore.kernel.org/bpf/[email protected]
Signed-off-by: Alexander Lobakin <[email protected]>

Author: alobakin

net/bpf/test_run.c

@@ -208,6 +208,16 @@ static void xdp_test_run_teardown(struct xdp_test_data *xdp)
 	kfree(xdp->skbs);
 }
 
+static bool frame_was_changed(const struct xdp_page_head *head)
+{
+	/* xdp_scrub_frame() zeroes the data pointer, flags is the last field,
+	 * i.e. has the highest chances to be overwritten. If those two are
+	 * untouched, it's most likely safe to skip the context reset.
+	 */
+	return head->frm.data != head->orig_ctx.data ||
+	       head->frm.flags != head->orig_ctx.flags;
+}
+
 static bool ctx_was_changed(struct xdp_page_head *head)
 {
 	return head->orig_ctx.data != head->ctx.data ||
@@ -217,7 +227,7 @@ static bool ctx_was_changed(struct xdp_page_head *head)
 
 static void reset_ctx(struct xdp_page_head *head)
 {
-	if (likely(!ctx_was_changed(head)))
+	if (likely(!frame_was_changed(head) && !ctx_was_changed(head)))
 		return;
 
 	head->ctx.data = head->orig_ctx.data;