[ECOS] How can we fix MBUF problem on old 1.3.1 code base

[ECOS] How can we fix MBUF problem on old 1.3.1 code base

[AshCan]

We are avoiding rebuilding the new anonymous CVS code base.  Therefore, we want to fix the existing network stack and driver code to eliminate "Out of MBUF" panics.  

We have MBUF problems both when:

1) when sending thousands of packets using multiple, concurrent ping commands.

2) When staying connected to a single socket and moving long audio data streams through the socket.

Are there are few simple fixes to the 1.3.1 code base to fix the out of MBUFs panic?  What are they?


thanks,

Dan Ash
SonicBox

Re: [ECOS] How can we fix MBUF problem on old 1.3.1 code base

[Jonathan Larmour]

AshCan@aol.com wrote:
> 
> Are there are few simple fixes to the 1.3.1 code base to fix the out of MBUFs panic?  What are they?

As luck would have it, I just happened to remember seeing the relevant
patch go by. Try the attached patch, but you'll probably have to
effectively apply it by hand, or tweak the patch - I doubt it will apply
straight off. I give no guarantees whether what you end up with will work
well though - it may rely on other patches being present, so take a backup
of your existing sources.

Jifl
-- 
Red Hat, 35 Cambridge Place, Cambridge, UK. CB2 1NS  Tel: +44 (1223) 728762
"Plan to be spontaneous tomorrow."  ||  These opinions are all my own fault
Index: net/drivers/eth/common/current/ChangeLog
===================================================================
RCS file: /local/cvsfiles/ecc/ecc/net/drivers/eth/common/current/ChangeLog,v
retrieving revision 1.9
diff -u -5 -p -r1.9 ChangeLog
--- net/drivers/eth/common/current/ChangeLog    2000/03/08 19:11:19     1.9
+++ net/drivers/eth/common/current/ChangeLog    2000/03/29 13:28:42
@@ -1,5 +1,10 @@
+2000-03-28  Gary Thomas  <gthomas@redhat.com>
+ 
+        * src/eth_drv.c (eth_drv_recv): Tolerate running out of MBUFs
+        instead of "panic"ing.
+ 
 2000-03-08  Gary Thomas  <gthomas@redhat.com>
 
        * src/eth_drv.c: Add some function [block] comments.
        (eth_drv_send): Use eCos scheduler lock instead of interrupt lock.
 
Index: net/drivers/eth/common/current/src/eth_drv.c
===================================================================
RCS file: /local/cvsfiles/ecc/ecc/net/drivers/eth/common/current/src/eth_drv.c,v
retrieving revision 1.7
diff -u -5 -p -r1.7 eth_drv.c
--- net/drivers/eth/common/current/src/eth_drv.c        2000/03/08 19:11:20     1.7
+++ net/drivers/eth/common/current/src/eth_drv.c        2000/03/29 13:25:07
@@ -336,11 +336,11 @@ eth_drv_recv(struct eth_drv_sc *sc, int 
     int sg_len;
 
     /* Pull packet off interface. */
     MGETHDR(m, M_DONTWAIT, MT_DATA);
     if (m == 0) {
-        panic("out of MBUFs");
+        diag_printf("warning: eth_recv out of MBUFs\n");
     }
 
     // Set up buffers
     // Unload ethernet header separately so IP/UDP/TCP headers are aligned
     sg_list[0].buf = (CYG_ADDRESS)eh;
@@ -348,30 +348,43 @@ eth_drv_recv(struct eth_drv_sc *sc, int 
     sg_len = 1;
 
     // Compute total length (minus ethernet header)
     total_len -= sizeof(*eh);
 
-    m->m_pkthdr.rcvif = ifp;
-    m->m_pkthdr.len = total_len;
-    mlen = MHLEN;
     top = 0;
+    mlen = MHLEN;
     mp = &top;
 
+    if (m) {
+        m->m_pkthdr.rcvif = ifp;
+        m->m_pkthdr.len = total_len;
+    } else {
+        sg_list[sg_len].buf = (CYG_ADDRESS)0;
+        sg_list[sg_len].len = min(total_len, MCLBYTES);
+        sg_len++;
+        total_len = 0;
+    }
+
     while (total_len > 0) {
         if (top) {
             MGET(m, M_DONTWAIT, MT_DATA);
             if (m == 0) {
                 m_freem(top);
-                panic("out of MBUFs");
+                panic("out of MBUFs [2]");
             }
             mlen = MLEN;
         }
         if (total_len >= MINCLSIZE) {
             MCLGET(m, M_DONTWAIT);
             if ((m->m_flags & M_EXT) == 0) {
                 m_freem(top);
-                panic("out of MBUFs");
+                diag_printf("warning: eth_recv out of MBUFs\n");
+                sg_list[sg_len].buf = (CYG_ADDRESS)0;
+                sg_list[sg_len].len = min(total_len, MCLBYTES);
+                sg_len++;
+                top = 0;
+                break;
             }
             mlen = MCLBYTES;
         }
         m->m_len = mlen = min(total_len, mlen);
         total_len -= mlen;
@@ -386,11 +399,13 @@ eth_drv_recv(struct eth_drv_sc *sc, int 
     // Ask hardware to unload buffers
     (sc->funs->recv)(sc, sg_list, sg_len);
 
     if (net_debug) {
         for (i = 0;  i < sg_len;  i++) {
-            diag_dump_buf((void *)sg_list[i].buf, sg_list[i].len);
+            if (sg_list[i].buf) {
+                diag_dump_buf((void *)sg_list[i].buf, sg_list[i].len);
+            }
         }
     }
 
     m = top;
     if (m == 0) {
Index: net/drivers/eth/edb7xxx/current/ChangeLog
===================================================================
RCS file: /local/cvsfiles/ecc/ecc/net/drivers/eth/edb7xxx/current/ChangeLog,v
retrieving revision 1.8
diff -u -5 -p -r1.8 ChangeLog
--- net/drivers/eth/edb7xxx/current/ChangeLog   2000/03/06 15:52:28     1.8
+++ net/drivers/eth/edb7xxx/current/ChangeLog   2000/03/29 13:28:17
@@ -1,5 +1,10 @@
+2000-03-28  Gary Thomas  <gthomas@redhat.com>
+
+        * src/if_edb7xxx.c (cs8900_recv): Handle case where there were
+        no buffers (and thus the sg_list[] has NULL pointers).
+
 2000-03-06  Gary Thomas  <gthomas@redhat.com>
 
        * src/if_edb7xxx.c: Use new driver API.
 
 2000-02-29  Gary Thomas  <gthomas@cygnus.co.uk>
Index: net/drivers/eth/edb7xxx/current/src/if_edb7xxx.c
===================================================================
RCS file: /local/cvsfiles/ecc/ecc/net/drivers/eth/edb7xxx/current/src/if_edb7xxx.c,v
retrieving revision 1.6
diff -u -5 -p -r1.6 if_edb7xxx.c
--- net/drivers/eth/edb7xxx/current/src/if_edb7xxx.c    2000/03/06 15:52:29     1.6
+++ net/drivers/eth/edb7xxx/current/src/if_edb7xxx.c    2000/03/29 13:25:07
@@ -330,24 +330,29 @@ cs8900_RxEvent(struct eth_drv_sc *sc)
 //
 static void
 cs8900_recv(struct eth_drv_sc *sc, struct eth_drv_sg *sg_list, int sg_len)
 {
     int i, mlen;
-    unsigned short *data;
-    unsigned char *cp;
+    unsigned short *data, val;
+    unsigned char *cp, cval;
 
     for (i = 0;  i < sg_len;  i++) {
         data = (unsigned short *)sg_list[i].buf;
         mlen = sg_list[i].len;
         while (mlen >= sizeof(*data)) {
-            *data++ = CS8900_RTDATA;
+            val = CS8900_RTDATA;
+            if (data) {
+                *data++ = val;
+            }
             mlen -= sizeof(*data);
         }
         if (mlen) {
             // Fetch last odd byte
-            cp = (unsigned char *)data;
-            *cp = CS8900_RTDATA & 0xFF;
+            cval = CS8900_RTDATA & 0xFF;
+            if (cp = (unsigned char *)data) {
+                *cp = cval;
+            }
         }
     }
 }
 
 static void
Index: net/drivers/eth/quicc/current/ChangeLog
===================================================================
RCS file: /local/cvsfiles/ecc/ecc/net/drivers/eth/quicc/current/ChangeLog,v
retrieving revision 1.8
diff -u -5 -p -r1.8 ChangeLog
--- net/drivers/eth/quicc/current/ChangeLog     2000/03/06 15:52:29     1.8
+++ net/drivers/eth/quicc/current/ChangeLog     2000/03/29 13:27:47
@@ -1,5 +1,10 @@
+2000-03-28  Gary Thomas  <gthomas@redhat.com>
+
+        * src/if_quicc.c (quicc_eth_recv): Handle case where there were
+        no buffers (and thus the sg_list[] contains NULL pointers).
+
 2000-03-06  Gary Thomas  <gthomas@redhat.com>
 
        * src/if_quicc.c: New driver API.
 
 2000-03-05  Gary Thomas  <gthomas@redhat.com>
Index: net/drivers/eth/quicc/current/src/if_quicc.c
===================================================================
RCS file: /local/cvsfiles/ecc/ecc/net/drivers/eth/quicc/current/src/if_quicc.c,v
retrieving revision 1.9
diff -u -5 -p -r1.9 if_quicc.c
--- net/drivers/eth/quicc/current/src/if_quicc.c        2000/03/06 15:52:30     1.9
+++ net/drivers/eth/quicc/current/src/if_quicc.c        2000/03/29 13:25:07
@@ -463,12 +463,14 @@ quicc_eth_recv(struct eth_drv_sc *sc, st
     HAL_DCACHE_IS_ENABLED(cache_state);
     if (cache_state) {
         HAL_DCACHE_INVALIDATE(qi->rxbd->buffer, qi->rxbd->length);  // Make sure no stale data
     }
     for (i = 0;  i < sg_len;  i++) {
-        bcopy(bp, (void *)sg_list[i].buf, sg_list[i].len);
-        bp += sg_list[i].len;
+        if (sg_list[i].buf != 0) {
+            bcopy(bp, (void *)sg_list[i].buf, sg_list[i].len);
+            bp += sg_list[i].len;
+        }
     }
 }
 
 static void
 quicc_eth_TxEvent(struct eth_drv_sc *sc, int stat)

Re: [ECOS] How can we fix MBUF problem on old 1.3.1 code base

[AshCan]

Jonathon,

thanks for those fixes.  The out of MBUF crashes are 
gone, but ofcourse I start gagging (playing network 
audio) instead.  I guess I need to pause playback to 
let the network catch up.  My code lets me push a button to close and reopen the socket and it is happy 
again.

What should the proper way to handle the condition.  I 
can set a semaphore that my network thread tests for, 
but at that point, what do I do when the semaphore is 
set?

Would increasing the number of MBUFs help?  It looks 
like I only have 16 total.  

thanks,

Dan Ash