aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--include/linux/tcp.h3
-rw-r--r--include/net/tcp.h1
-rw-r--r--net/ipv4/tcp_input.c271
-rw-r--r--net/ipv4/tcp_output.c14
4 files changed, 172 insertions, 117 deletions
diff --git a/include/linux/tcp.h b/include/linux/tcp.h
index 794497c7d755..08027f1d7f31 100644
--- a/include/linux/tcp.h
+++ b/include/linux/tcp.h
@@ -343,10 +343,7 @@ struct tcp_sock {
343 struct sk_buff *scoreboard_skb_hint; 343 struct sk_buff *scoreboard_skb_hint;
344 struct sk_buff *retransmit_skb_hint; 344 struct sk_buff *retransmit_skb_hint;
345 struct sk_buff *forward_skb_hint; 345 struct sk_buff *forward_skb_hint;
346 struct sk_buff *fastpath_skb_hint;
347 346
348 int fastpath_cnt_hint; /* Lags behind by current skb's pcount
349 * compared to respective fackets_out */
350 int lost_cnt_hint; 347 int lost_cnt_hint;
351 int retransmit_cnt_hint; 348 int retransmit_cnt_hint;
352 349
diff --git a/include/net/tcp.h b/include/net/tcp.h
index 0ede804b16d4..f0c5e7a29400 100644
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -1081,7 +1081,6 @@ static inline void tcp_clear_retrans_hints_partial(struct tcp_sock *tp)
1081static inline void tcp_clear_all_retrans_hints(struct tcp_sock *tp) 1081static inline void tcp_clear_all_retrans_hints(struct tcp_sock *tp)
1082{ 1082{
1083 tcp_clear_retrans_hints_partial(tp); 1083 tcp_clear_retrans_hints_partial(tp);
1084 tp->fastpath_skb_hint = NULL;
1085} 1084}
1086 1085
1087/* MD5 Signature */ 1086/* MD5 Signature */
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index a287747e9dd6..3ad6a19ad30f 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -1333,6 +1333,88 @@ static int tcp_sacktag_one(struct sk_buff *skb, struct tcp_sock *tp,
1333 return flag; 1333 return flag;
1334} 1334}
1335 1335
1336static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1337 struct tcp_sack_block *next_dup,
1338 u32 start_seq, u32 end_seq,
1339 int dup_sack_in, int *fack_count,
1340 int *reord, int *flag)
1341{
1342 struct tcp_sock *tp = tcp_sk(sk);
1343
1344 tcp_for_write_queue_from(skb, sk) {
1345 int in_sack = 0;
1346 int dup_sack = dup_sack_in;
1347
1348 if (skb == tcp_send_head(sk))
1349 break;
1350
1351 /* queue is in-order => we can short-circuit the walk early */
1352 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1353 break;
1354
1355 if ((next_dup != NULL) &&
1356 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1357 in_sack = tcp_match_skb_to_sack(sk, skb,
1358 next_dup->start_seq,
1359 next_dup->end_seq);
1360 if (in_sack > 0)
1361 dup_sack = 1;
1362 }
1363
1364 if (in_sack <= 0)
1365 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, end_seq);
1366 if (unlikely(in_sack < 0))
1367 break;
1368
1369 if (in_sack)
1370 *flag |= tcp_sacktag_one(skb, tp, reord, dup_sack, *fack_count);
1371
1372 *fack_count += tcp_skb_pcount(skb);
1373 }
1374 return skb;
1375}
1376
1377/* Avoid all extra work that is being done by sacktag while walking in
1378 * a normal way
1379 */
1380static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1381 u32 skip_to_seq)
1382{
1383 tcp_for_write_queue_from(skb, sk) {
1384 if (skb == tcp_send_head(sk))
1385 break;
1386
1387 if (before(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1388 break;
1389 }
1390 return skb;
1391}
1392
1393static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1394 struct sock *sk,
1395 struct tcp_sack_block *next_dup,
1396 u32 skip_to_seq,
1397 int *fack_count, int *reord,
1398 int *flag)
1399{
1400 if (next_dup == NULL)
1401 return skb;
1402
1403 if (before(next_dup->start_seq, skip_to_seq)) {
1404 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1405 tcp_sacktag_walk(skb, sk, NULL,
1406 next_dup->start_seq, next_dup->end_seq,
1407 1, fack_count, reord, flag);
1408 }
1409
1410 return skb;
1411}
1412
1413static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1414{
1415 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1416}
1417
1336static int 1418static int
1337tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) 1419tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
1338{ 1420{
@@ -1342,16 +1424,16 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_
1342 TCP_SKB_CB(ack_skb)->sacked); 1424 TCP_SKB_CB(ack_skb)->sacked);
1343 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1425 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1344 struct tcp_sack_block sp[4]; 1426 struct tcp_sack_block sp[4];
1345 struct sk_buff *cached_skb; 1427 struct tcp_sack_block *cache;
1428 struct sk_buff *skb;
1346 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; 1429 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
1347 int used_sacks; 1430 int used_sacks;
1348 int reord = tp->packets_out; 1431 int reord = tp->packets_out;
1349 int flag = 0; 1432 int flag = 0;
1350 int found_dup_sack = 0; 1433 int found_dup_sack = 0;
1351 int cached_fack_count; 1434 int fack_count;
1352 int i; 1435 int i, j;
1353 int first_sack_index; 1436 int first_sack_index;
1354 int force_one_sack;
1355 1437
1356 if (!tp->sacked_out) { 1438 if (!tp->sacked_out) {
1357 if (WARN_ON(tp->fackets_out)) 1439 if (WARN_ON(tp->fackets_out))
@@ -1409,132 +1491,123 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_
1409 used_sacks++; 1491 used_sacks++;
1410 } 1492 }
1411 1493
1412 /* SACK fastpath: 1494 /* order SACK blocks to allow in order walk of the retrans queue */
1413 * if the only SACK change is the increase of the end_seq of 1495 for (i = used_sacks - 1; i > 0; i--) {
1414 * the first block then only apply that SACK block 1496 for (j = 0; j < i; j++){
1415 * and use retrans queue hinting otherwise slowpath */ 1497 if (after(sp[j].start_seq, sp[j+1].start_seq)) {
1416 force_one_sack = 1; 1498 struct tcp_sack_block tmp;
1417 for (i = 0; i < used_sacks; i++) {
1418 u32 start_seq = sp[i].start_seq;
1419 u32 end_seq = sp[i].end_seq;
1420
1421 if (i == 0) {
1422 if (tp->recv_sack_cache[i].start_seq != start_seq)
1423 force_one_sack = 0;
1424 } else {
1425 if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1426 (tp->recv_sack_cache[i].end_seq != end_seq))
1427 force_one_sack = 0;
1428 }
1429 tp->recv_sack_cache[i].start_seq = start_seq;
1430 tp->recv_sack_cache[i].end_seq = end_seq;
1431 }
1432 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1433 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1434 tp->recv_sack_cache[i].start_seq = 0;
1435 tp->recv_sack_cache[i].end_seq = 0;
1436 }
1437 1499
1438 if (force_one_sack) 1500 tmp = sp[j];
1439 used_sacks = 1; 1501 sp[j] = sp[j+1];
1440 else { 1502 sp[j+1] = tmp;
1441 int j;
1442 tp->fastpath_skb_hint = NULL;
1443
1444 /* order SACK blocks to allow in order walk of the retrans queue */
1445 for (i = used_sacks - 1; i > 0; i--) {
1446 for (j = 0; j < i; j++){
1447 if (after(sp[j].start_seq, sp[j+1].start_seq)) {
1448 struct tcp_sack_block tmp;
1449
1450 tmp = sp[j];
1451 sp[j] = sp[j+1];
1452 sp[j+1] = tmp;
1453
1454 /* Track where the first SACK block goes to */
1455 if (j == first_sack_index)
1456 first_sack_index = j+1;
1457 }
1458 1503
1504 /* Track where the first SACK block goes to */
1505 if (j == first_sack_index)
1506 first_sack_index = j+1;
1459 } 1507 }
1460 } 1508 }
1461 } 1509 }
1462 1510
1463 /* Use SACK fastpath hint if valid */ 1511 skb = tcp_write_queue_head(sk);
1464 cached_skb = tp->fastpath_skb_hint; 1512 fack_count = 0;
1465 cached_fack_count = tp->fastpath_cnt_hint; 1513 i = 0;
1466 if (!cached_skb) { 1514
1467 cached_skb = tcp_write_queue_head(sk); 1515 if (!tp->sacked_out) {
1468 cached_fack_count = 0; 1516 /* It's already past, so skip checking against it */
1517 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1518 } else {
1519 cache = tp->recv_sack_cache;
1520 /* Skip empty blocks in at head of the cache */
1521 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1522 !cache->end_seq)
1523 cache++;
1469 } 1524 }
1470 1525
1471 for (i = 0; i < used_sacks; i++) { 1526 while (i < used_sacks) {
1472 struct sk_buff *skb;
1473 u32 start_seq = sp[i].start_seq; 1527 u32 start_seq = sp[i].start_seq;
1474 u32 end_seq = sp[i].end_seq; 1528 u32 end_seq = sp[i].end_seq;
1475 int fack_count;
1476 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1529 int dup_sack = (found_dup_sack && (i == first_sack_index));
1477 int next_dup = (found_dup_sack && (i+1 == first_sack_index)); 1530 struct tcp_sack_block *next_dup = NULL;
1478 1531
1479 skb = cached_skb; 1532 if (found_dup_sack && ((i + 1) == first_sack_index))
1480 fack_count = cached_fack_count; 1533 next_dup = &sp[i + 1];
1481 1534
1482 /* Event "B" in the comment above. */ 1535 /* Event "B" in the comment above. */
1483 if (after(end_seq, tp->high_seq)) 1536 if (after(end_seq, tp->high_seq))
1484 flag |= FLAG_DATA_LOST; 1537 flag |= FLAG_DATA_LOST;
1485 1538
1486 tcp_for_write_queue_from(skb, sk) { 1539 /* Skip too early cached blocks */
1487 int in_sack = 0; 1540 while (tcp_sack_cache_ok(tp, cache) &&
1488 1541 !before(start_seq, cache->end_seq))
1489 if (skb == tcp_send_head(sk)) 1542 cache++;
1490 break; 1543
1491 1544 /* Can skip some work by looking recv_sack_cache? */
1492 cached_skb = skb; 1545 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1493 cached_fack_count = fack_count; 1546 after(end_seq, cache->start_seq)) {
1494 if (i == first_sack_index) { 1547
1495 tp->fastpath_skb_hint = skb; 1548 /* Head todo? */
1496 tp->fastpath_cnt_hint = fack_count; 1549 if (before(start_seq, cache->start_seq)) {
1550 skb = tcp_sacktag_skip(skb, sk, start_seq);
1551 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq,
1552 cache->start_seq, dup_sack,
1553 &fack_count, &reord, &flag);
1497 } 1554 }
1498 1555
1499 /* The retransmission queue is always in order, so 1556 /* Rest of the block already fully processed? */
1500 * we can short-circuit the walk early. 1557 if (!after(end_seq, cache->end_seq)) {
1501 */ 1558 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, cache->end_seq,
1502 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1559 &fack_count, &reord, &flag);
1503 break; 1560 goto advance_sp;
1504 1561 }
1505 dup_sack = (found_dup_sack && (i == first_sack_index));
1506 1562
1507 /* Due to sorting DSACK may reside within this SACK block! */ 1563 /* ...tail remains todo... */
1508 if (next_dup) { 1564 if (TCP_SKB_CB(tp->highest_sack)->end_seq == cache->end_seq) {
1509 u32 dup_start = sp[i+1].start_seq; 1565 /* ...but better entrypoint exists! Check that DSACKs are
1510 u32 dup_end = sp[i+1].end_seq; 1566 * properly accounted while skipping here
1567 */
1568 tcp_maybe_skipping_dsack(skb, sk, next_dup, cache->end_seq,
1569 &fack_count, &reord, &flag);
1511 1570
1512 if (before(TCP_SKB_CB(skb)->seq, dup_end)) { 1571 skb = tcp_write_queue_next(sk, tp->highest_sack);
1513 in_sack = tcp_match_skb_to_sack(sk, skb, dup_start, dup_end); 1572 fack_count = tp->fackets_out;
1514 if (in_sack > 0) 1573 cache++;
1515 dup_sack = 1; 1574 goto walk;
1516 }
1517 } 1575 }
1518 1576
1519 /* DSACK info lost if out-of-mem, try SACK still */ 1577 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1520 if (in_sack <= 0) 1578 /* Check overlap against next cached too (past this one already) */
1521 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, end_seq); 1579 cache++;
1522 if (unlikely(in_sack < 0)) 1580 continue;
1523 break; 1581 }
1524
1525 if (in_sack)
1526 flag |= tcp_sacktag_one(skb, tp, &reord, dup_sack, fack_count);
1527 1582
1528 fack_count += tcp_skb_pcount(skb); 1583 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1584 skb = tcp_write_queue_next(sk, tp->highest_sack);
1585 fack_count = tp->fackets_out;
1529 } 1586 }
1587 skb = tcp_sacktag_skip(skb, sk, start_seq);
1588
1589walk:
1590 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq,
1591 dup_sack, &fack_count, &reord, &flag);
1530 1592
1593advance_sp:
1531 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct 1594 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1532 * due to in-order walk 1595 * due to in-order walk
1533 */ 1596 */
1534 if (after(end_seq, tp->frto_highmark)) 1597 if (after(end_seq, tp->frto_highmark))
1535 flag &= ~FLAG_ONLY_ORIG_SACKED; 1598 flag &= ~FLAG_ONLY_ORIG_SACKED;
1599
1600 i++;
1536 } 1601 }
1537 1602
1603 /* Clear the head of the cache sack blocks so we can skip it next time */
1604 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1605 tp->recv_sack_cache[i].start_seq = 0;
1606 tp->recv_sack_cache[i].end_seq = 0;
1607 }
1608 for (j = 0; j < used_sacks; j++)
1609 tp->recv_sack_cache[i++] = sp[j];
1610
1538 flag |= tcp_mark_lost_retrans(sk); 1611 flag |= tcp_mark_lost_retrans(sk);
1539 1612
1540 tcp_verify_left_out(tp); 1613 tcp_verify_left_out(tp);
@@ -2821,9 +2894,7 @@ static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p,
2821 } 2894 }
2822 2895
2823 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 2896 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2824 /* hint's skb might be NULL but we don't need to care */ 2897
2825 tp->fastpath_cnt_hint -= min_t(u32, pkts_acked,
2826 tp->fastpath_cnt_hint);
2827 if (ca_ops->pkts_acked) { 2898 if (ca_ops->pkts_acked) {
2828 s32 rtt_us = -1; 2899 s32 rtt_us = -1;
2829 2900
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index ce506af5ce07..030fc69ea217 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -653,9 +653,7 @@ static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned
653} 653}
654 654
655/* When a modification to fackets out becomes necessary, we need to check 655/* When a modification to fackets out becomes necessary, we need to check
656 * skb is counted to fackets_out or not. Another important thing is to 656 * skb is counted to fackets_out or not.
657 * tweak SACK fastpath hint too as it would overwrite all changes unless
658 * hint is also changed.
659 */ 657 */
660static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, 658static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
661 int decr) 659 int decr)
@@ -667,11 +665,6 @@ static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
667 665
668 if (!before(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 666 if (!before(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
669 tp->fackets_out -= decr; 667 tp->fackets_out -= decr;
670
671 /* cnt_hint is "off-by-one" compared with fackets_out (see sacktag) */
672 if (tp->fastpath_skb_hint != NULL &&
673 after(TCP_SKB_CB(tp->fastpath_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
674 tp->fastpath_cnt_hint -= decr;
675} 668}
676 669
677/* Function to create two new TCP segments. Shrinks the given segment 670/* Function to create two new TCP segments. Shrinks the given segment
@@ -1753,11 +1746,6 @@ static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int m
1753 1746
1754 /* changed transmit queue under us so clear hints */ 1747 /* changed transmit queue under us so clear hints */
1755 tcp_clear_retrans_hints_partial(tp); 1748 tcp_clear_retrans_hints_partial(tp);
1756 /* manually tune sacktag skb hint */
1757 if (tp->fastpath_skb_hint == next_skb) {
1758 tp->fastpath_skb_hint = skb;
1759 tp->fastpath_cnt_hint -= tcp_skb_pcount(skb);
1760 }
1761 1749
1762 sk_stream_free_skb(sk, next_skb); 1750 sk_stream_free_skb(sk, next_skb);
1763 } 1751 }