aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorPaul Mackerras <paulus@samba.org>2005-05-12 19:47:12 -0400
committerJeff Garzik <jgarzik@pobox.com>2005-05-12 19:47:12 -0400
commit516cd15f1c0dd6eada3619915b113b4e5baccc7a (patch)
tree0ca69fcdeff9aa93b3b76722f97b512cbb5e403b
parent88d7bd8cb9eb8d64bf7997600b0d64f7834047c5 (diff)
[PATCH] PPP multilink fragmentation improvements
Here's a patch for -mm for now. Not sure whose territory this falls in, so I'm sending it to everyone I can think of. :) Some time ago I did some experiments with using PPP multilink over largish numbers of channels (up to 32). The TCP performance was woeful due to wildly fluctuating packet latencies, which turned out to be because we would sometimes split a packet across all 32 channels, and sometimes we would send a whole packet down a single channel. This patch fixes those problems by being a bit cleverer about how the packets are split across the available channels, and in particular, it waits until at least half of the channels can take another fragment before starting to split up the next packet. The patch also fixes a buglet in the multilink reconstruction code where it would discard incoming packets that had just the multilink header and no data. Such packets are valid and shouldn't be discarded. Signed-off-by: Paul Mackerras <paulus@samba.org> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
-rw-r--r--drivers/net/ppp_generic.c177
1 files changed, 110 insertions, 67 deletions
diff --git a/drivers/net/ppp_generic.c b/drivers/net/ppp_generic.c
index 3b377f6cd4a0..ad4b58af6b76 100644
--- a/drivers/net/ppp_generic.c
+++ b/drivers/net/ppp_generic.c
@@ -1217,36 +1217,43 @@ ppp_push(struct ppp *ppp)
1217 */ 1217 */
1218static int ppp_mp_explode(struct ppp *ppp, struct sk_buff *skb) 1218static int ppp_mp_explode(struct ppp *ppp, struct sk_buff *skb)
1219{ 1219{
1220 int nch, len, fragsize; 1220 int len, fragsize;
1221 int i, bits, hdrlen, mtu; 1221 int i, bits, hdrlen, mtu;
1222 int flen, fnb; 1222 int flen;
1223 int navail, nfree;
1224 int nbigger;
1223 unsigned char *p, *q; 1225 unsigned char *p, *q;
1224 struct list_head *list; 1226 struct list_head *list;
1225 struct channel *pch; 1227 struct channel *pch;
1226 struct sk_buff *frag; 1228 struct sk_buff *frag;
1227 struct ppp_channel *chan; 1229 struct ppp_channel *chan;
1228 1230
1229 nch = 0; 1231 nfree = 0; /* # channels which have no packet already queued */
1232 navail = 0; /* total # of usable channels (not deregistered) */
1230 hdrlen = (ppp->flags & SC_MP_XSHORTSEQ)? MPHDRLEN_SSN: MPHDRLEN; 1233 hdrlen = (ppp->flags & SC_MP_XSHORTSEQ)? MPHDRLEN_SSN: MPHDRLEN;
1234 i = 0;
1231 list = &ppp->channels; 1235 list = &ppp->channels;
1232 while ((list = list->next) != &ppp->channels) { 1236 while ((list = list->next) != &ppp->channels) {
1233 pch = list_entry(list, struct channel, clist); 1237 pch = list_entry(list, struct channel, clist);
1234 nch += pch->avail = (skb_queue_len(&pch->file.xq) == 0); 1238 navail += pch->avail = (pch->chan != NULL);
1235 /* 1239 if (pch->avail) {
1236 * If a channel hasn't had a fragment yet, it has to get 1240 if (skb_queue_len(&pch->file.xq) == 0
1237 * one before we send any fragments on later channels. 1241 || !pch->had_frag) {
1238 * If it can't take a fragment now, don't give any 1242 pch->avail = 2;
1239 * to subsequent channels. 1243 ++nfree;
1240 */
1241 if (!pch->had_frag && !pch->avail) {
1242 while ((list = list->next) != &ppp->channels) {
1243 pch = list_entry(list, struct channel, clist);
1244 pch->avail = 0;
1245 } 1244 }
1246 break; 1245 if (!pch->had_frag && i < ppp->nxchan)
1246 ppp->nxchan = i;
1247 } 1247 }
1248 ++i;
1248 } 1249 }
1249 if (nch == 0) 1250
1251 /*
1252 * Don't start sending this packet unless at least half of
1253 * the channels are free. This gives much better TCP
1254 * performance if we have a lot of channels.
1255 */
1256 if (nfree == 0 || nfree < navail / 2)
1250 return 0; /* can't take now, leave it in xmit_pending */ 1257 return 0; /* can't take now, leave it in xmit_pending */
1251 1258
1252 /* Do protocol field compression (XXX this should be optional) */ 1259 /* Do protocol field compression (XXX this should be optional) */
@@ -1257,14 +1264,19 @@ static int ppp_mp_explode(struct ppp *ppp, struct sk_buff *skb)
1257 --len; 1264 --len;
1258 } 1265 }
1259 1266
1260 /* decide on fragment size */ 1267 /*
1268 * Decide on fragment size.
1269 * We create a fragment for each free channel regardless of
1270 * how small they are (i.e. even 0 length) in order to minimize
1271 * the time that it will take to detect when a channel drops
1272 * a fragment.
1273 */
1261 fragsize = len; 1274 fragsize = len;
1262 if (nch > 1) { 1275 if (nfree > 1)
1263 int maxch = ROUNDUP(len, MIN_FRAG_SIZE); 1276 fragsize = ROUNDUP(fragsize, nfree);
1264 if (nch > maxch) 1277 /* nbigger channels get fragsize bytes, the rest get fragsize-1,
1265 nch = maxch; 1278 except if nbigger==0, then they all get fragsize. */
1266 fragsize = ROUNDUP(fragsize, nch); 1279 nbigger = len % nfree;
1267 }
1268 1280
1269 /* skip to the channel after the one we last used 1281 /* skip to the channel after the one we last used
1270 and start at that one */ 1282 and start at that one */
@@ -1278,7 +1290,7 @@ static int ppp_mp_explode(struct ppp *ppp, struct sk_buff *skb)
1278 1290
1279 /* create a fragment for each channel */ 1291 /* create a fragment for each channel */
1280 bits = B; 1292 bits = B;
1281 do { 1293 while (nfree > 0 || len > 0) {
1282 list = list->next; 1294 list = list->next;
1283 if (list == &ppp->channels) { 1295 if (list == &ppp->channels) {
1284 i = 0; 1296 i = 0;
@@ -1289,61 +1301,92 @@ static int ppp_mp_explode(struct ppp *ppp, struct sk_buff *skb)
1289 if (!pch->avail) 1301 if (!pch->avail)
1290 continue; 1302 continue;
1291 1303
1304 /*
1305 * Skip this channel if it has a fragment pending already and
1306 * we haven't given a fragment to all of the free channels.
1307 */
1308 if (pch->avail == 1) {
1309 if (nfree > 0)
1310 continue;
1311 } else {
1312 --nfree;
1313 pch->avail = 1;
1314 }
1315
1292 /* check the channel's mtu and whether it is still attached. */ 1316 /* check the channel's mtu and whether it is still attached. */
1293 spin_lock_bh(&pch->downl); 1317 spin_lock_bh(&pch->downl);
1294 if (pch->chan == 0 || (mtu = pch->chan->mtu) < hdrlen) { 1318 if (pch->chan == NULL) {
1295 /* can't use this channel */ 1319 /* can't use this channel, it's being deregistered */
1296 spin_unlock_bh(&pch->downl); 1320 spin_unlock_bh(&pch->downl);
1297 pch->avail = 0; 1321 pch->avail = 0;
1298 if (--nch == 0) 1322 if (--navail == 0)
1299 break; 1323 break;
1300 continue; 1324 continue;
1301 } 1325 }
1302 1326
1303 /* 1327 /*
1304 * We have to create multiple fragments for this channel 1328 * Create a fragment for this channel of
1305 * if fragsize is greater than the channel's mtu. 1329 * min(max(mtu+2-hdrlen, 4), fragsize, len) bytes.
1330 * If mtu+2-hdrlen < 4, that is a ridiculously small
1331 * MTU, so we use mtu = 2 + hdrlen.
1306 */ 1332 */
1307 if (fragsize > len) 1333 if (fragsize > len)
1308 fragsize = len; 1334 fragsize = len;
1309 for (flen = fragsize; flen > 0; flen -= fnb) { 1335 flen = fragsize;
1310 fnb = flen; 1336 mtu = pch->chan->mtu + 2 - hdrlen;
1311 if (fnb > mtu + 2 - hdrlen) 1337 if (mtu < 4)
1312 fnb = mtu + 2 - hdrlen; 1338 mtu = 4;
1313 if (fnb >= len) 1339 if (flen > mtu)
1314 bits |= E; 1340 flen = mtu;
1315 frag = alloc_skb(fnb + hdrlen, GFP_ATOMIC); 1341 if (flen == len && nfree == 0)
1316 if (frag == 0) 1342 bits |= E;
1317 goto noskb; 1343 frag = alloc_skb(flen + hdrlen + (flen == 0), GFP_ATOMIC);
1318 q = skb_put(frag, fnb + hdrlen); 1344 if (frag == 0)
1319 /* make the MP header */ 1345 goto noskb;
1320 q[0] = PPP_MP >> 8; 1346 q = skb_put(frag, flen + hdrlen);
1321 q[1] = PPP_MP; 1347
1322 if (ppp->flags & SC_MP_XSHORTSEQ) { 1348 /* make the MP header */
1323 q[2] = bits + ((ppp->nxseq >> 8) & 0xf); 1349 q[0] = PPP_MP >> 8;
1324 q[3] = ppp->nxseq; 1350 q[1] = PPP_MP;
1325 } else { 1351 if (ppp->flags & SC_MP_XSHORTSEQ) {
1326 q[2] = bits; 1352 q[2] = bits + ((ppp->nxseq >> 8) & 0xf);
1327 q[3] = ppp->nxseq >> 16; 1353 q[3] = ppp->nxseq;
1328 q[4] = ppp->nxseq >> 8; 1354 } else {
1329 q[5] = ppp->nxseq; 1355 q[2] = bits;
1330 } 1356 q[3] = ppp->nxseq >> 16;
1331 1357 q[4] = ppp->nxseq >> 8;
1332 /* copy the data in */ 1358 q[5] = ppp->nxseq;
1333 memcpy(q + hdrlen, p, fnb);
1334
1335 /* try to send it down the channel */
1336 chan = pch->chan;
1337 if (!chan->ops->start_xmit(chan, frag))
1338 skb_queue_tail(&pch->file.xq, frag);
1339 pch->had_frag = 1;
1340 p += fnb;
1341 len -= fnb;
1342 ++ppp->nxseq;
1343 bits = 0;
1344 } 1359 }
1360
1361 /*
1362 * Copy the data in.
1363 * Unfortunately there is a bug in older versions of
1364 * the Linux PPP multilink reconstruction code where it
1365 * drops 0-length fragments. Therefore we make sure the
1366 * fragment has at least one byte of data. Any bytes
1367 * we add in this situation will end up as padding on the
1368 * end of the reconstructed packet.
1369 */
1370 if (flen == 0)
1371 *skb_put(frag, 1) = 0;
1372 else
1373 memcpy(q + hdrlen, p, flen);
1374
1375 /* try to send it down the channel */
1376 chan = pch->chan;
1377 if (skb_queue_len(&pch->file.xq)
1378 || !chan->ops->start_xmit(chan, frag))
1379 skb_queue_tail(&pch->file.xq, frag);
1380 pch->had_frag = 1;
1381 p += flen;
1382 len -= flen;
1383 ++ppp->nxseq;
1384 bits = 0;
1345 spin_unlock_bh(&pch->downl); 1385 spin_unlock_bh(&pch->downl);
1346 } while (len > 0); 1386
1387 if (--nbigger == 0 && fragsize > 0)
1388 --fragsize;
1389 }
1347 ppp->nxchan = i; 1390 ppp->nxchan = i;
1348 1391
1349 return 1; 1392 return 1;
@@ -1422,7 +1465,7 @@ ppp_input(struct ppp_channel *chan, struct sk_buff *skb)
1422 kfree_skb(skb); 1465 kfree_skb(skb);
1423 return; 1466 return;
1424 } 1467 }
1425 1468
1426 proto = PPP_PROTO(skb); 1469 proto = PPP_PROTO(skb);
1427 read_lock_bh(&pch->upl); 1470 read_lock_bh(&pch->upl);
1428 if (pch->ppp == 0 || proto >= 0xc000 || proto == PPP_CCPFRAG) { 1471 if (pch->ppp == 0 || proto >= 0xc000 || proto == PPP_CCPFRAG) {
@@ -1691,7 +1734,7 @@ ppp_receive_mp_frame(struct ppp *ppp, struct sk_buff *skb, struct channel *pch)
1691 struct list_head *l; 1734 struct list_head *l;
1692 int mphdrlen = (ppp->flags & SC_MP_SHORTSEQ)? MPHDRLEN_SSN: MPHDRLEN; 1735 int mphdrlen = (ppp->flags & SC_MP_SHORTSEQ)? MPHDRLEN_SSN: MPHDRLEN;
1693 1736
1694 if (!pskb_may_pull(skb, mphdrlen + 1) || ppp->mrru == 0) 1737 if (!pskb_may_pull(skb, mphdrlen) || ppp->mrru == 0)
1695 goto err; /* no good, throw it away */ 1738 goto err; /* no good, throw it away */
1696 1739
1697 /* Decode sequence number and begin/end bits */ 1740 /* Decode sequence number and begin/end bits */