misc. questions
Colin Percival
cperciva at freebsd.org
Thu Mar 7 12:42:56 UTC 2013
On 03/07/13 04:35, Mark Felder wrote:
> On Wed, 06 Mar 2013 16:40:38 -0600, Colin Percival <cperciva at freebsd.org> wrote:
>> You'll want to turn off tso, since it produces long mbuf chains which most
>> xn netbacks choke on. (I have a very ugly workaround patch for this which I
>> use on EC2, but simply turning off tso is enough unless you need Gbps+ speeds).
>
> Can you link me to this patch? I have an environment that might warrant using it
> for now.
Attached.
And remember that I said it was a *very ugly* workaround... :-)
--
Colin Percival
Security Officer Emeritus, FreeBSD | The power to serve
Founder, Tarsnap | www.tarsnap.com | Online backups for the truly paranoid
-------------- next part --------------
--- sys/kern/uipc_mbuf.c (revision 223824)
+++ sys/kern/uipc_mbuf.c (working copy)
@@ -525,12 +525,14 @@
* only their reference counts are incremented.
*/
struct mbuf *
-m_copym(struct mbuf *m, int off0, int len, int wait)
+m_copy_nbufs(struct mbuf *m, int off0, int len, int wait, long * outlen,
+ int nbufmax)
{
struct mbuf *n, **np;
int off = off0;
struct mbuf *top;
int copyhdr = 0;
+ int len_orig = len;
KASSERT(off >= 0, ("m_copym, negative off %d", off));
KASSERT(len >= 0, ("m_copym, negative len %d", len));
@@ -546,7 +548,7 @@
}
np = ⊤
top = 0;
- while (len > 0) {
+ while (len > 0 && nbufmax-- > 0) {
if (m == NULL) {
KASSERT(len == M_COPYALL,
("m_copym, length > size of mbuf chain"));
@@ -584,6 +586,9 @@
if (top == NULL)
mbstat.m_mcfail++; /* XXX: No consistency. */
+ if (outlen)
+ *outlen = len_orig - len;
+
return (top);
nospace:
m_freem(top);
@@ -591,6 +596,13 @@
return (NULL);
}
+struct mbuf *
+m_copym(struct mbuf *m, int off0, int len, int wait)
+{
+
+ return (m_copy_nbufs(m, off0, len, wait, NULL, INT_MAX));
+}
+
/*
* Returns mbuf chain with new head for the prepending case.
* Copies from mbuf (chain) n from off for len to mbuf (chain) m
--- sys/netinet/tcp_output.c (revision 228872)
+++ sys/netinet/tcp_output.c (working copy)
@@ -183,6 +183,7 @@
int sack_rxmit, sack_bytes_rxmt;
struct sackhole *p;
int tso;
+ int max_mbuf_chain_len = 16; /* XXX Set this based on interface? */
struct tcpopt to;
#if 0
int maxburst = TCP_MAXBURST;
@@ -806,16 +807,6 @@
struct mbuf *mb;
u_int moff;
- if ((tp->t_flags & TF_FORCEDATA) && len == 1)
- TCPSTAT_INC(tcps_sndprobe);
- else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
- tp->t_sndrexmitpack++;
- TCPSTAT_INC(tcps_sndrexmitpack);
- TCPSTAT_ADD(tcps_sndrexmitbyte, len);
- } else {
- TCPSTAT_INC(tcps_sndpack);
- TCPSTAT_ADD(tcps_sndbyte, len);
- }
MGETHDR(m, M_DONTWAIT, MT_DATA);
if (m == NULL) {
SOCKBUF_UNLOCK(&so->so_snd);
@@ -847,7 +838,8 @@
mtod(m, caddr_t) + hdrlen);
m->m_len += len;
} else {
- m->m_next = m_copy(mb, moff, (int)len);
+ m->m_next = m_copy_nbufs(mb, moff, len, M_DONTWAIT,
+ &len, max_mbuf_chain_len);
if (m->m_next == NULL) {
SOCKBUF_UNLOCK(&so->so_snd);
(void) m_free(m);
@@ -856,6 +848,18 @@
}
}
+ /* Update stats here as m_copy_nbufs may have adjusted len. */
+ if ((tp->t_flags & TF_FORCEDATA) && len == 1)
+ TCPSTAT_INC(tcps_sndprobe);
+ else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
+ tp->t_sndrexmitpack++;
+ TCPSTAT_INC(tcps_sndrexmitpack);
+ TCPSTAT_ADD(tcps_sndrexmitbyte, len);
+ } else {
+ TCPSTAT_INC(tcps_sndpack);
+ TCPSTAT_ADD(tcps_sndbyte, len);
+ }
+
/*
* If we're sending everything we've got, set PUSH.
* (This will keep happy those implementations which only
--- sys/sys/mbuf.h (revision 223824)
+++ sys/sys/mbuf.h (working copy)
@@ -849,6 +849,7 @@
int, int, int, int);
struct mbuf *m_copypacket(struct mbuf *, int);
void m_copy_pkthdr(struct mbuf *, struct mbuf *);
+struct mbuf *m_copy_nbufs(struct mbuf *, int, int, int, long *, int);
struct mbuf *m_copyup(struct mbuf *n, int len, int dstoff);
struct mbuf *m_defrag(struct mbuf *, int);
void m_demote(struct mbuf *, int);
More information about the freebsd-xen
mailing list