diff options
author | Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> | 2008-11-25 05:20:15 (GMT) |
---|---|---|
committer | David S. Miller <davem@davemloft.net> | 2008-11-25 05:20:15 (GMT) |
commit | 832d11c5cd076abc0aa1eaf7be96c81d1a59ce41 (patch) | |
tree | 95b22ad16d1ff414cab39578ed8c927c2ce08723 /net/ipv4 | |
parent | f58b22fd3c16444edc393a217a74208f1894b601 (diff) | |
download | linux-fsl-qoriq-832d11c5cd076abc0aa1eaf7be96c81d1a59ce41.tar.xz |
tcp: Try to restore large SKBs while SACK processing
During SACK processing, most of the benefits of TSO are eaten by
the SACK blocks that one-by-one fragment SKBs to MSS sized chunks.
Then we're in problems when cleanup work for them has to be done
when a large cumulative ACK comes. Try to return back to pre-split
state already while more and more SACK info gets discovered by
combining newly discovered SACK areas with the previous skb if
that's SACKed as well.
This approach has a number of benefits:
1) The processing overhead is spread more equally over the RTT
2) Write queue has less skbs to process (affect everything
which has to walk in the queue past the sacked areas)
3) Write queue is consistent whole the time, so no other parts
of TCP has to be aware of this (this was not the case with
some other approach that was, well, quite intrusive all
around).
4) Clean_rtx_queue can release most of the pages using single
put_page instead of previous PAGE_SIZE/mss+1 calls
In case a hole is fully filled by the new SACK block, we attempt
to combine the next skb too which allows construction of skbs
that are even larger than what tso split them to and it handles
hole per on every nth patterns that often occur during slow start
overshoot pretty nicely. Though this to be really useful also
a retransmission would have to get lost since cumulative ACKs
advance one hole at a time in the most typical case.
TODO: handle upwards only merging. That should be rather easy
when segment is fully sacked but I'm leaving that as future
work item (it won't make very large difference anyway since
this current approach already covers quite a lot of normal
cases).
I was earlier thinking of some sophisticated way of tracking
timestamps of the first and the last segment but later on
realized that it won't be that necessary at all to store the
timestamp of the last segment. The cases that can occur are
basically either:
1) ambiguous => no sensible measurement can be taken anyway
2) non-ambiguous is due to reordering => having the timestamp
of the last segment there is just skewing things more off
than does some good since the ack got triggered by one of
the holes (besides some substle issues that would make
determining right hole/skb even harder problem). Anyway,
it has nothing to do with this change then.
I choose to route some abnormal looking cases with goto noop,
some could be handled differently (eg., by stopping the
walking at that skb but again). In general, they either
shouldn't happen at all or are rare enough to make no difference
in practice.
In theory this change (as whole) could cause some macroscale
regression (global) because of cache misses that are taken over
the round-trip time but it gets very likely better because of much
less (local) cache misses per other write queue walkers and the
big recovery clearing cumulative ack.
Worth to note that these benefits would be very easy to get also
without TSO/GSO being on as long as the data is in pages so that
we can merge them. Currently I won't let that happen because
DSACK splitting at fragment that would mess up pcounts due to
sk_can_gso in tcp_set_skb_tso_segs. Once DSACKs fragments gets
avoided, we have some conditions that can be made less strict.
TODO: I will probably have to convert the excessive pointer
passing to struct sacktag_state... :-)
My testing revealed that considerable amount of skbs couldn't
be shifted because they were cloned (most likely still awaiting
tx reclaim)...
[The rest is considering future work instead since I got
repeatably EFAULT to tcpdump's recvfrom when I added
pskb_expand_head to deal with clones, so I separated that
into another, later patch]
...To counter that, I gave up on the fifth advantage:
5) When growing previous SACK block, less allocs for new skbs
are done, basically a new alloc is needed only when new hole
is detected and when the previous skb runs out of frags space
...which now only happens of if reclaim is fast enough to dispose
the clone before the SACK block comes in (the window is RTT long),
otherwise we'll have to alloc some.
With clones being handled I got these numbers (will be somewhat
worse without that), taken with fine-grained mibs:
TCPSackShifted 398
TCPSackMerged 877
TCPSackShiftFallback 320
TCPSACKCOLLAPSEFALLBACKGSO 0
TCPSACKCOLLAPSEFALLBACKSKBBITS 0
TCPSACKCOLLAPSEFALLBACKSKBDATA 0
TCPSACKCOLLAPSEFALLBACKBELOW 0
TCPSACKCOLLAPSEFALLBACKFIRST 1
TCPSACKCOLLAPSEFALLBACKPREVBITS 318
TCPSACKCOLLAPSEFALLBACKMSS 1
TCPSACKCOLLAPSEFALLBACKNOHEAD 0
TCPSACKCOLLAPSEFALLBACKSHIFT 0
TCPSACKCOLLAPSENOOPSEQ 0
TCPSACKCOLLAPSENOOPSMALLPCOUNT 0
TCPSACKCOLLAPSENOOPSMALLLEN 0
TCPSACKCOLLAPSEHOLE 12
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'net/ipv4')
-rw-r--r-- | net/ipv4/tcp_input.c | 256 |
1 files changed, 249 insertions, 7 deletions
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 3c8e297..97d5767 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -1242,6 +1242,8 @@ static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb, * aligned portion of it that matches. Therefore we might need to fragment * which may fail and creates some hassle (caller must handle error case * returns). + * + * FIXME: this could be merged to shift decision code */ static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, u32 start_seq, u32 end_seq) @@ -1353,9 +1355,6 @@ static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk, if (fack_count > tp->fackets_out) tp->fackets_out = fack_count; - - if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) - tcp_advance_highest_sack(sk, skb); } /* D-SACK. We can detect redundant retransmission in S|R and plain R @@ -1370,12 +1369,231 @@ static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk, return flag; } +static int tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, + struct sk_buff *skb, unsigned int pcount, + int shifted, int fack_count, int *reord, + int *flag, int mss) +{ + struct tcp_sock *tp = tcp_sk(sk); + u8 dummy_sacked = TCP_SKB_CB(skb)->sacked; /* We discard results */ + + BUG_ON(!pcount); + + TCP_SKB_CB(prev)->end_seq += shifted; + TCP_SKB_CB(skb)->seq += shifted; + + skb_shinfo(prev)->gso_segs += pcount; + BUG_ON(skb_shinfo(skb)->gso_segs < pcount); + skb_shinfo(skb)->gso_segs -= pcount; + + /* When we're adding to gso_segs == 1, gso_size will be zero, + * in theory this shouldn't be necessary but as long as DSACK + * code can come after this skb later on it's better to keep + * setting gso_size to something. + */ + if (!skb_shinfo(prev)->gso_size) { + skb_shinfo(prev)->gso_size = mss; + skb_shinfo(prev)->gso_type = sk->sk_gso_type; + } + + /* CHECKME: To clear or not to clear? Mimics normal skb currently */ + if (skb_shinfo(skb)->gso_segs <= 1) { + skb_shinfo(skb)->gso_size = 0; + skb_shinfo(skb)->gso_type = 0; + } + + *flag |= tcp_sacktag_one(skb, sk, reord, 0, fack_count, &dummy_sacked, + pcount); + + /* Difference in this won't matter, both ACKed by the same cumul. ACK */ + TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); + + tcp_clear_all_retrans_hints(tp); + + if (skb->len > 0) { + BUG_ON(!tcp_skb_pcount(skb)); + return 0; + } + + /* Whole SKB was eaten :-) */ + + TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags; + if (skb == tcp_highest_sack(sk)) + tcp_advance_highest_sack(sk, skb); + + tcp_unlink_write_queue(skb, sk); + sk_wmem_free_skb(sk, skb); + + return 1; +} + +/* I wish gso_size would have a bit more sane initialization than + * something-or-zero which complicates things + */ +static int tcp_shift_mss(struct sk_buff *skb) +{ + int mss = tcp_skb_mss(skb); + + if (!mss) + mss = skb->len; + + return mss; +} + +/* Shifting pages past head area doesn't work */ +static int skb_can_shift(struct sk_buff *skb) +{ + return !skb_headlen(skb) && skb_is_nonlinear(skb); +} + +/* Try collapsing SACK blocks spanning across multiple skbs to a single + * skb. + */ +static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, + u32 start_seq, u32 end_seq, + int dup_sack, int *fack_count, + int *reord, int *flag) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct sk_buff *prev; + int mss; + int pcount = 0; + int len; + int in_sack; + + if (!sk_can_gso(sk)) + goto fallback; + + /* Normally R but no L won't result in plain S */ + if (!dup_sack && + (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) == TCPCB_SACKED_RETRANS) + goto fallback; + if (!skb_can_shift(skb)) + goto fallback; + /* This frame is about to be dropped (was ACKed). */ + if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) + goto fallback; + + /* Can only happen with delayed DSACK + discard craziness */ + if (unlikely(skb == tcp_write_queue_head(sk))) + goto fallback; + prev = tcp_write_queue_prev(sk, skb); + + if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) + goto fallback; + + in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && + !before(end_seq, TCP_SKB_CB(skb)->end_seq); + + if (in_sack) { + len = skb->len; + pcount = tcp_skb_pcount(skb); + mss = tcp_shift_mss(skb); + + /* TODO: Fix DSACKs to not fragment already SACKed and we can + * drop this restriction as unnecessary + */ + if (mss != tcp_shift_mss(prev)) + goto fallback; + } else { + if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) + goto noop; + /* CHECKME: This is non-MSS split case only?, this will + * cause skipped skbs due to advancing loop btw, original + * has that feature too + */ + if (tcp_skb_pcount(skb) <= 1) + goto noop; + + in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); + if (!in_sack) { + /* TODO: head merge to next could be attempted here + * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), + * though it might not be worth of the additional hassle + * + * ...we can probably just fallback to what was done + * previously. We could try merging non-SACKed ones + * as well but it probably isn't going to buy off + * because later SACKs might again split them, and + * it would make skb timestamp tracking considerably + * harder problem. + */ + goto fallback; + } + + len = end_seq - TCP_SKB_CB(skb)->seq; + BUG_ON(len < 0); + BUG_ON(len > skb->len); + + /* MSS boundaries should be honoured or else pcount will + * severely break even though it makes things bit trickier. + * Optimize common case to avoid most of the divides + */ + mss = tcp_skb_mss(skb); + + /* TODO: Fix DSACKs to not fragment already SACKed and we can + * drop this restriction as unnecessary + */ + if (mss != tcp_shift_mss(prev)) + goto fallback; + + if (len == mss) { + pcount = 1; + } else if (len < mss) { + goto noop; + } else { + pcount = len / mss; + len = pcount * mss; + } + } + + if (!skb_shift(prev, skb, len)) + goto fallback; + if (!tcp_shifted_skb(sk, prev, skb, pcount, len, *fack_count, reord, + flag, mss)) + goto out; + + /* Hole filled allows collapsing with the next as well, this is very + * useful when hole on every nth skb pattern happens + */ + if (prev == tcp_write_queue_tail(sk)) + goto out; + skb = tcp_write_queue_next(sk, prev); + + if (!skb_can_shift(skb)) + goto out; + if (skb == tcp_send_head(sk)) + goto out; + if ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) + goto out; + + len = skb->len; + if (skb_shift(prev, skb, len)) { + pcount += tcp_skb_pcount(skb); + tcp_shifted_skb(sk, prev, skb, tcp_skb_pcount(skb), len, + *fack_count, reord, flag, mss); + } + +out: + *fack_count += pcount; + return prev; + +noop: + return skb; + +fallback: + return NULL; +} + static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, struct tcp_sack_block *next_dup, u32 start_seq, u32 end_seq, int dup_sack_in, int *fack_count, int *reord, int *flag) { + struct tcp_sock *tp = tcp_sk(sk); + struct sk_buff *tmp; + tcp_for_write_queue_from(skb, sk) { int in_sack = 0; int dup_sack = dup_sack_in; @@ -1396,18 +1614,42 @@ static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, dup_sack = 1; } - if (in_sack <= 0) - in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, - end_seq); + /* skb reference here is a bit tricky to get right, since + * shifting can eat and free both this skb and the next, + * so not even _safe variant of the loop is enough. + */ + if (in_sack <= 0) { + tmp = tcp_shift_skb_data(sk, skb, start_seq, + end_seq, dup_sack, + fack_count, reord, flag); + if (tmp != NULL) { + if (tmp != skb) { + skb = tmp; + continue; + } + + in_sack = 0; + } else { + in_sack = tcp_match_skb_to_sack(sk, skb, + start_seq, + end_seq); + } + } + if (unlikely(in_sack < 0)) break; - if (in_sack) + if (in_sack) { *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack, *fack_count, &(TCP_SKB_CB(skb)->sacked), tcp_skb_pcount(skb)); + if (!before(TCP_SKB_CB(skb)->seq, + tcp_highest_sack_seq(tp))) + tcp_advance_highest_sack(sk, skb); + } + *fack_count += tcp_skb_pcount(skb); } return skb; |