Subject: Re: NetBSD/xen network problems (need help)
To: Mike M. Volokhov <mishka@intostroy.com>
From: Manuel Bouyer <bouyer@antioche.eu.org>
List: port-xen
Date: 01/26/2006 13:13:29
--X1bOJ3K7DJ5YkBrT
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline

On Thu, Jan 26, 2006 at 01:11:52PM +0200, Mike M. Volokhov wrote:
> > Always ip_forward.
> 
> Not sure now. Todays morning I've got yet another on dom1 (one xennet,
> but ip.forwarding is on):
> 
> login: panic: kernel diagnostic assertion "((pa ^ (pa + m->m_pkthdr.len)) & PG_FRAME) == 0" failed: file "../../../../arch/xen/xen/if_xennet.c", line 1037
> Stopped in pid 539.1 (imapd) at netbsd:cpu_Debugger+0x4:        leave
> cpu_Debugger(c08dc698,ffffffff,c09e4100,c08d4f38,c08d4f00) at netbsd:cpu_Debugger+0x4
> panic(c033c780,c03097e7,c0338e80,c0338c60,40d) at netbsd:panic+0x121
> __main(c03097e7,c0338c60,40d,c0338e80,cb4f7b74) at netbsd:__main
> xennet_start(c0988038,cb4f7bb0,c0988038,2,400) at netbsd:xennet_start+0x55a
> ether_output(c0988038,c08c0900,c08d746c,c08d839c,c08c0900) at netbsd:ether_output+0x38b
> ip_output(c08c0900,0,c08d7468,400,0) at netbsd:ip_output+0x547
> tcp_output(c08dfe10,c09e4b00,3,0,0) at netbsd:tcp_output+0xdbf
> tcp_usrreq(c08d6948,9,c09e4b00,0,0) at netbsd:tcp_usrreq+0x2f4
> sosend(c08d6948,0,cb4f7ec4,c09e4b00,0) at netbsd:sosend+0x37b
> soo_write(ca3d3888,ca3d38b0,cb4f7ec4,ca3c0000,1) at netbsd:soo_write+0x22
> dofilewrite(ca3c2b28,1,ca3d3888,806f800,86) at netbsd:dofilewrite+0x81
> sys_write(ca3cb8c4,cb4f7f64,cb4f7f5c,0,0) at netbsd:sys_write+0x59
> syscall_plain() at netbsd:syscall_plain+0x18f

OK, on this one the mbuf didn't come from the forward path.

> 
> Sure! (Because this patch is for if_xennet only, I've left xennetback
> with previous patch applied, in case if it matters.) Results (on

Yes, this is the right thing to do

> multihomed dom2):
> 
> pa 0x616cf38 len 200 m_flags 0x2
> panic: xennet_start1: buffer crosses page

OK, thanks. Here's a new patch which adds more debug printf.

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
     NetBSD: 26 ans d'experience feront toujours la difference
--

--X1bOJ3K7DJ5YkBrT
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment; filename=diff

Index: if_xennet.c
===================================================================
RCS file: /cvsroot/src/sys/arch/xen/xen/if_xennet.c,v
retrieving revision 1.13.2.19
diff -u -r1.13.2.19 if_xennet.c
--- if_xennet.c	22 Jan 2006 10:41:59 -0000	1.13.2.19
+++ if_xennet.c	26 Jan 2006 12:10:50 -0000
@@ -638,6 +638,7 @@
 	mmu_update_t *mmu = rx_mmu;
 	multicall_entry_t *mcl = rx_mcl;
 	struct mbuf *m;
+	void *pktp;
 
 	network_tx_buf_gc(sc);
 	ifp->if_flags &= ~IFF_OACTIVE;
@@ -660,7 +661,7 @@
 		/* XXXcl check rx->status for error */
 
                 MGETHDR(m, M_DONTWAIT, MT_DATA);
-                if (m == NULL) {
+                if (__predict_false(m == NULL)) {
 			printf("xennet: rx no mbuf\n");
 			break;
 		}
@@ -687,14 +688,12 @@
 
 		/* Do all the remapping work, and M->P updates, in one
 		 * big hypercall. */
-		if ((mcl - rx_mcl) != 0) {
-			mcl->op = __HYPERVISOR_mmu_update;
-			mcl->args[0] = (unsigned long)rx_mmu;
-			mcl->args[1] = mmu - rx_mmu;
-			mcl->args[2] = 0;
-			mcl++;
-			(void)HYPERVISOR_multicall(rx_mcl, mcl - rx_mcl);
-		}
+		mcl->op = __HYPERVISOR_mmu_update;
+		mcl->args[0] = (unsigned long)rx_mmu;
+		mcl->args[1] = mmu - rx_mmu;
+		mcl->args[2] = 0;
+		mcl++;
+		(void)HYPERVISOR_multicall(rx_mcl, mcl - rx_mcl);
 		if (0)
 		printf("page mapped at va %08lx -> %08x/%08lx\n",
 		    sc->sc_rx_bufa[rx->id].xb_rx.xbrx_va,
@@ -711,11 +710,23 @@
 		    (void *)(PTE_BASE[x86_btop
 			(sc->sc_rx_bufa[rx->id].xb_rx.xbrx_va)] & PG_FRAME)));
 
+		pktp = (void *)(sc->sc_rx_bufa[rx->id].xb_rx.xbrx_va +
+		    (rx->addr & PAGE_MASK));
+		if ((ifp->if_flags & IFF_PROMISC) == 0) {
+			struct ether_header *eh = pktp;
+			if (ETHER_IS_MULTICAST(eh->ether_dhost) == 0 &&
+			    memcmp(LLADDR(ifp->if_sadl), eh->ether_dhost,
+			    ETHER_ADDR_LEN) != 0) {
+				m_freem(m);
+				xennet_rx_push_buffer(sc, rx->id);
+				continue; /* packet not for us */
+			}
+		}
 		m->m_pkthdr.rcvif = ifp;
-		if (sc->sc_rx->req_prod != sc->sc_rx->resp_prod) {
+		if (__predict_true(
+		    sc->sc_rx->req_prod != sc->sc_rx->resp_prod)) {
 			m->m_len = m->m_pkthdr.len = rx->status;
-			MEXTADD(m, (void *)(sc->sc_rx_bufa[rx->id].xb_rx.
-			    xbrx_va + (rx->addr & PAGE_MASK)), rx->status,
+			MEXTADD(m, pktp, rx->status,
 			    M_DEVBUF, xennet_rx_mbuf_free,
 			    &sc->sc_rx_bufa[rx->id]);
 		} else {
@@ -726,14 +737,13 @@
 			 */
 			m->m_len = MHLEN;
 			m->m_pkthdr.len = 0;
-			m_copyback(m, 0, rx->status, 
-			    (caddr_t)(sc->sc_rx_bufa[rx->id].xb_rx.xbrx_va +
-			    (rx->addr & PAGE_MASK)));
+			m_copyback(m, 0, rx->status, pktp);
 			xennet_rx_push_buffer(sc, rx->id);
 			if (m->m_pkthdr.len < rx->status) {
+				/* out of memory, just drop packets */
 				ifp->if_ierrors++;
 				m_freem(m);
-				break;
+				continue;
 			}
 		}
 
@@ -1009,22 +1019,43 @@
 			m_copydata(m, 0, m->m_pkthdr.len, mtod(new_m, caddr_t));
 			new_m->m_len = new_m->m_pkthdr.len = m->m_pkthdr.len;
 
-			m_freem(m);
-			m = new_m;
-			if ((m->m_flags & M_EXT) != 0) {
-				pa = m->m_ext.ext_paddr;
-				KASSERT(m->m_data == m->m_ext.ext_buf);
+			if ((new_m->m_flags & M_EXT) != 0) {
+				pa = new_m->m_ext.ext_paddr;
+				KASSERT(new_m->m_data == new_m->m_ext.ext_buf);
 				KASSERT(pa != M_PADDR_INVALID);
 			} else {
-				pa = m->m_paddr;
+				pa = new_m->m_paddr;
 				KASSERT(pa != M_PADDR_INVALID);
-				KASSERT(m->m_data == M_BUFADDR(m));
-				pa += M_BUFOFFSET(m);
+				KASSERT(new_m->m_data == M_BUFADDR(new_m));
+				pa += M_BUFOFFSET(new_m);
+			}
+			if (((pa ^ (pa + new_m->m_pkthdr.len)) & PG_FRAME) != 0) {
+				printf("pa %p len %d m_flags 0x%x\n",
+				    (void *)pa, new_m->m_pkthdr.len, new_m->m_flags);
+				printf("new_m %p m_paddr %lx M_BUFOFFSET %d\n",
+				   new_m,
+				   ((new_m->m_flags & M_EXT) != 0) ?
+				       new_m->m_ext.ext_paddr : new_m->m_paddr,
+				   ((new_m->m_flags & M_EXT) != 0) ?
+				   0 : M_BUFOFFSET(new_m));
+				panic("xennet_start1: buffer crosses page");
 			}
+			m_freem(m);
+			m = new_m;
 		} else
 			IFQ_DEQUEUE(&ifp->if_snd, m);
 
-		KASSERT(((pa ^ (pa + m->m_pkthdr.len)) & PG_FRAME) == 0);
+		if (((pa ^ (pa + m->m_pkthdr.len)) & PG_FRAME) != 0) {
+			printf("pa %p len %d m_flags 0x%x\n",
+			    (void *)pa, m->m_pkthdr.len, m->m_flags);
+			printf("m %p m_paddr %lx M_BUFOFFSET %d\n",
+			   m,
+			   ((m->m_flags & M_EXT) != 0) ?
+			       m->m_ext.ext_paddr : m->m_paddr,
+			   ((m->m_flags & M_EXT) != 0) ?
+			   0 : M_BUFOFFSET(m));
+			panic("xennet_start2: buffer crosses page");
+		}
 
 		bufid = get_bufarray_entry(sc->sc_tx_bufa);
 		KASSERT(bufid < NETIF_TX_RING_SIZE);

--X1bOJ3K7DJ5YkBrT--