mirror of
https://github.com/AuxXxilium/linux_dsm_epyc7002.git
synced 2024-11-24 15:31:14 +07:00
ef2820a735
Implementation of (a)rwnd calculation might lead to severe performance issues
and associations completely stalling. These problems are described and solution
is proposed which improves lksctp's robustness in congestion state.
1) Sudden drop of a_rwnd and incomplete window recovery afterwards
Data accounted in sctp_assoc_rwnd_decrease takes only payload size (sctp data),
but size of sk_buff, which is blamed against receiver buffer, is not accounted
in rwnd. Theoretically, this should not be the problem as actual size of buffer
is double the amount requested on the socket (SO_RECVBUF). Problem here is
that this will have bad scaling for data which is less then sizeof sk_buff.
E.g. in 4G (LTE) networks, link interfacing radio side will have a large portion
of traffic of this size (less then 100B).
An example of sudden drop and incomplete window recovery is given below. Node B
exhibits problematic behavior. Node A initiates association and B is configured
to advertise rwnd of 10000. A sends messages of size 43B (size of typical sctp
message in 4G (LTE) network). On B data is left in buffer by not reading socket
in userspace.
Lets examine when we will hit pressure state and declare rwnd to be 0 for
scenario with above stated parameters (rwnd == 10000, chunk size == 43, each
chunk is sent in separate sctp packet)
Logic is implemented in sctp_assoc_rwnd_decrease:
socket_buffer (see below) is maximum size which can be held in socket buffer
(sk_rcvbuf). current_alloced is amount of data currently allocated (rx_count)
A simple expression is given for which it will be examined after how many
packets for above stated parameters we enter pressure state:
We start by condition which has to be met in order to enter pressure state:
socket_buffer < currently_alloced;
currently_alloced is represented as size of sctp packets received so far and not
yet delivered to userspace. x is the number of chunks/packets (since there is no
bundling, and each chunk is delivered in separate packet, we can observe each
chunk also as sctp packet, and what is important here, having its own sk_buff):
socket_buffer < x*each_sctp_packet;
each_sctp_packet is sctp chunk size + sizeof(struct sk_buff). socket_buffer is
twice the amount of initially requested size of socket buffer, which is in case
of sctp, twice the a_rwnd requested:
2*rwnd < x*(payload+sizeof(struc sk_buff));
sizeof(struct sk_buff) is 190 (3.13.0-rc4+). Above is stated that rwnd is 10000
and each payload size is 43
20000 < x(43+190);
x > 20000/233;
x ~> 84;
After ~84 messages, pressure state is entered and 0 rwnd is advertised while
received 84*43B ~= 3612B sctp data. This is why external observer notices sudden
drop from 6474 to 0, as it will be now shown in example:
IP A.34340 > B.12345: sctp (1) [INIT] [init tag: 1875509148] [rwnd: 81920] [OS: 10] [MIS: 65535] [init TSN: 1096057017]
IP B.12345 > A.34340: sctp (1) [INIT ACK] [init tag: 3198966556] [rwnd: 10000] [OS: 10] [MIS: 10] [init TSN: 902132839]
IP A.34340 > B.12345: sctp (1) [COOKIE ECHO]
IP B.12345 > A.34340: sctp (1) [COOKIE ACK]
IP A.34340 > B.12345: sctp (1) [DATA] (B)(E) [TSN: 1096057017] [SID: 0] [SSEQ 0] [PPID 0x18]
IP B.12345 > A.34340: sctp (1) [SACK] [cum ack 1096057017] [a_rwnd 9957] [#gap acks 0] [#dup tsns 0]
IP A.34340 > B.12345: sctp (1) [DATA] (B)(E) [TSN: 1096057018] [SID: 0] [SSEQ 1] [PPID 0x18]
IP B.12345 > A.34340: sctp (1) [SACK] [cum ack 1096057018] [a_rwnd 9957] [#gap acks 0] [#dup tsns 0]
IP A.34340 > B.12345: sctp (1) [DATA] (B)(E) [TSN: 1096057019] [SID: 0] [SSEQ 2] [PPID 0x18]
IP B.12345 > A.34340: sctp (1) [SACK] [cum ack 1096057019] [a_rwnd 9914] [#gap acks 0] [#dup tsns 0]
<...>
IP A.34340 > B.12345: sctp (1) [DATA] (B)(E) [TSN: 1096057098] [SID: 0] [SSEQ 81] [PPID 0x18]
IP B.12345 > A.34340: sctp (1) [SACK] [cum ack 1096057098] [a_rwnd 6517] [#gap acks 0] [#dup tsns 0]
IP A.34340 > B.12345: sctp (1) [DATA] (B)(E) [TSN: 1096057099] [SID: 0] [SSEQ 82] [PPID 0x18]
IP B.12345 > A.34340: sctp (1) [SACK] [cum ack 1096057099] [a_rwnd 6474] [#gap acks 0] [#dup tsns 0]
IP A.34340 > B.12345: sctp (1) [DATA] (B)(E) [TSN: 1096057100] [SID: 0] [SSEQ 83] [PPID 0x18]
--> Sudden drop
IP B.12345 > A.34340: sctp (1) [SACK] [cum ack 1096057100] [a_rwnd 0] [#gap acks 0] [#dup tsns 0]
At this point, rwnd_press stores current rwnd value so it can be later restored
in sctp_assoc_rwnd_increase. This however doesn't happen as condition to start
slowly increasing rwnd until rwnd_press is returned to rwnd is never met. This
condition is not met since rwnd, after it hit 0, must first reach rwnd_press by
adding amount which is read from userspace. Let us observe values in above
example. Initial a_rwnd is 10000, pressure was hit when rwnd was ~6500 and the
amount of actual sctp data currently waiting to be delivered to userspace
is ~3500. When userspace starts to read, sctp_assoc_rwnd_increase will be blamed
only for sctp data, which is ~3500. Condition is never met, and when userspace
reads all data, rwnd stays on 3569.
IP B.12345 > A.34340: sctp (1) [SACK] [cum ack 1096057100] [a_rwnd 1505] [#gap acks 0] [#dup tsns 0]
IP B.12345 > A.34340: sctp (1) [SACK] [cum ack 1096057100] [a_rwnd 3010] [#gap acks 0] [#dup tsns 0]
IP A.34340 > B.12345: sctp (1) [DATA] (B)(E) [TSN: 1096057101] [SID: 0] [SSEQ 84] [PPID 0x18]
IP B.12345 > A.34340: sctp (1) [SACK] [cum ack 1096057101] [a_rwnd 3569] [#gap acks 0] [#dup tsns 0]
--> At this point userspace read everything, rwnd recovered only to 3569
IP A.34340 > B.12345: sctp (1) [DATA] (B)(E) [TSN: 1096057102] [SID: 0] [SSEQ 85] [PPID 0x18]
IP B.12345 > A.34340: sctp (1) [SACK] [cum ack 1096057102] [a_rwnd 3569] [#gap acks 0] [#dup tsns 0]
Reproduction is straight forward, it is enough for sender to send packets of
size less then sizeof(struct sk_buff) and receiver keeping them in its buffers.
2) Minute size window for associations sharing the same socket buffer
In case multiple associations share the same socket, and same socket buffer
(sctp.rcvbuf_policy == 0), different scenarios exist in which congestion on one
of the associations can permanently drop rwnd of other association(s).
Situation will be typically observed as one association suddenly having rwnd
dropped to size of last packet received and never recovering beyond that point.
Different scenarios will lead to it, but all have in common that one of the
associations (let it be association from 1)) nearly depleted socket buffer, and
the other association blames socket buffer just for the amount enough to start
the pressure. This association will enter pressure state, set rwnd_press and
announce 0 rwnd.
When data is read by userspace, similar situation as in 1) will occur, rwnd will
increase just for the size read by userspace but rwnd_press will be high enough
so that association doesn't have enough credit to reach rwnd_press and restore
to previous state. This case is special case of 1), being worse as there is, in
the worst case, only one packet in buffer for which size rwnd will be increased.
Consequence is association which has very low maximum rwnd ('minute size', in
our case down to 43B - size of packet which caused pressure) and as such
unusable.
Scenario happened in the field and labs frequently after congestion state (link
breaks, different probabilities of packet drop, packet reordering) and with
scenario 1) preceding. Here is given a deterministic scenario for reproduction:
>From node A establish two associations on the same socket, with rcvbuf_policy
being set to share one common buffer (sctp.rcvbuf_policy == 0). On association 1
repeat scenario from 1), that is, bring it down to 0 and restore up. Observe
scenario 1). Use small payload size (here we use 43). Once rwnd is 'recovered',
bring it down close to 0, as in just one more packet would close it. This has as
a consequence that association number 2 is able to receive (at least) one more
packet which will bring it in pressure state. E.g. if association 2 had rwnd of
10000, packet received was 43, and we enter at this point into pressure,
rwnd_press will have 9957. Once payload is delivered to userspace, rwnd will
increase for 43, but conditions to restore rwnd to original state, just as in
1), will never be satisfied.
--> Association 1, between A.y and B.12345
IP A.55915 > B.12345: sctp (1) [INIT] [init tag: 836880897] [rwnd: 10000] [OS: 10] [MIS: 65535] [init TSN: 4032536569]
IP B.12345 > A.55915: sctp (1) [INIT ACK] [init tag: 2873310749] [rwnd: 81920] [OS: 10] [MIS: 10] [init TSN: 3799315613]
IP A.55915 > B.12345: sctp (1) [COOKIE ECHO]
IP B.12345 > A.55915: sctp (1) [COOKIE ACK]
--> Association 2, between A.z and B.12346
IP A.55915 > B.12346: sctp (1) [INIT] [init tag: 534798321] [rwnd: 10000] [OS: 10] [MIS: 65535] [init TSN: 2099285173]
IP B.12346 > A.55915: sctp (1) [INIT ACK] [init tag: 516668823] [rwnd: 81920] [OS: 10] [MIS: 10] [init TSN: 3676403240]
IP A.55915 > B.12346: sctp (1) [COOKIE ECHO]
IP B.12346 > A.55915: sctp (1) [COOKIE ACK]
--> Deplete socket buffer by sending messages of size 43B over association 1
IP B.12345 > A.55915: sctp (1) [DATA] (B)(E) [TSN: 3799315613] [SID: 0] [SSEQ 0] [PPID 0x18]
IP A.55915 > B.12345: sctp (1) [SACK] [cum ack 3799315613] [a_rwnd 9957] [#gap acks 0] [#dup tsns 0]
<...>
IP A.55915 > B.12345: sctp (1) [SACK] [cum ack 3799315696] [a_rwnd 6388] [#gap acks 0] [#dup tsns 0]
IP B.12345 > A.55915: sctp (1) [DATA] (B)(E) [TSN: 3799315697] [SID: 0] [SSEQ 84] [PPID 0x18]
IP A.55915 > B.12345: sctp (1) [SACK] [cum ack 3799315697] [a_rwnd 6345] [#gap acks 0] [#dup tsns 0]
--> Sudden drop on 1
IP B.12345 > A.55915: sctp (1) [DATA] (B)(E) [TSN: 3799315698] [SID: 0] [SSEQ 85] [PPID 0x18]
IP A.55915 > B.12345: sctp (1) [SACK] [cum ack 3799315698] [a_rwnd 0] [#gap acks 0] [#dup tsns 0]
--> Here userspace read, rwnd 'recovered' to 3698, now deplete again using
association 1 so there is place in buffer for only one more packet
IP B.12345 > A.55915: sctp (1) [DATA] (B)(E) [TSN: 3799315799] [SID: 0] [SSEQ 186] [PPID 0x18]
IP A.55915 > B.12345: sctp (1) [SACK] [cum ack 3799315799] [a_rwnd 86] [#gap acks 0] [#dup tsns 0]
IP B.12345 > A.55915: sctp (1) [DATA] (B)(E) [TSN: 3799315800] [SID: 0] [SSEQ 187] [PPID 0x18]
IP A.55915 > B.12345: sctp (1) [SACK] [cum ack 3799315800] [a_rwnd 43] [#gap acks 0] [#dup tsns 0]
--> Socket buffer is almost depleted, but there is space for one more packet,
send them over association 2, size 43B
IP B.12346 > A.55915: sctp (1) [DATA] (B)(E) [TSN: 3676403240] [SID: 0] [SSEQ 0] [PPID 0x18]
IP A.55915 > B.12346: sctp (1) [SACK] [cum ack 3676403240] [a_rwnd 0] [#gap acks 0] [#dup tsns 0]
--> Immediate drop
IP A.60995 > B.12346: sctp (1) [SACK] [cum ack 387491510] [a_rwnd 0] [#gap acks 0] [#dup tsns 0]
--> Read everything from the socket, both association recover up to maximum rwnd
they are capable of reaching, note that association 1 recovered up to 3698,
and association 2 recovered only to 43
IP A.55915 > B.12345: sctp (1) [SACK] [cum ack 3799315800] [a_rwnd 1548] [#gap acks 0] [#dup tsns 0]
IP A.55915 > B.12345: sctp (1) [SACK] [cum ack 3799315800] [a_rwnd 3053] [#gap acks 0] [#dup tsns 0]
IP B.12345 > A.55915: sctp (1) [DATA] (B)(E) [TSN: 3799315801] [SID: 0] [SSEQ 188] [PPID 0x18]
IP A.55915 > B.12345: sctp (1) [SACK] [cum ack 3799315801] [a_rwnd 3698] [#gap acks 0] [#dup tsns 0]
IP B.12346 > A.55915: sctp (1) [DATA] (B)(E) [TSN: 3676403241] [SID: 0] [SSEQ 1] [PPID 0x18]
IP A.55915 > B.12346: sctp (1) [SACK] [cum ack 3676403241] [a_rwnd 43] [#gap acks 0] [#dup tsns 0]
A careful reader might wonder why it is necessary to reproduce 1) prior
reproduction of 2). It is simply easier to observe when to send packet over
association 2 which will push association into the pressure state.
Proposed solution:
Both problems share the same root cause, and that is improper scaling of socket
buffer with rwnd. Solution in which sizeof(sk_buff) is taken into concern while
calculating rwnd is not possible due to fact that there is no linear
relationship between amount of data blamed in increase/decrease with IP packet
in which payload arrived. Even in case such solution would be followed,
complexity of the code would increase. Due to nature of current rwnd handling,
slow increase (in sctp_assoc_rwnd_increase) of rwnd after pressure state is
entered is rationale, but it gives false representation to the sender of current
buffer space. Furthermore, it implements additional congestion control mechanism
which is defined on implementation, and not on standard basis.
Proposed solution simplifies whole algorithm having on mind definition from rfc:
o Receiver Window (rwnd): This gives the sender an indication of the space
available in the receiver's inbound buffer.
Core of the proposed solution is given with these lines:
sctp_assoc_rwnd_update:
if ((asoc->base.sk->sk_rcvbuf - rx_count) > 0)
asoc->rwnd = (asoc->base.sk->sk_rcvbuf - rx_count) >> 1;
else
asoc->rwnd = 0;
We advertise to sender (half of) actual space we have. Half is in the braces
depending whether you would like to observe size of socket buffer as SO_RECVBUF
or twice the amount, i.e. size is the one visible from userspace, that is,
from kernelspace.
In this way sender is given with good approximation of our buffer space,
regardless of the buffer policy - we always advertise what we have. Proposed
solution fixes described problems and removes necessity for rwnd restoration
algorithm. Finally, as proposed solution is simplification, some lines of code,
along with some bytes in struct sctp_association are saved.
Version 2 of the patch addressed comments from Vlad. Name of the function is set
to be more descriptive, and two parts of code are changed, in one removing the
superfluous call to sctp_assoc_rwnd_update since call would not result in update
of rwnd, and the other being reordering of the code in a way that call to
sctp_assoc_rwnd_update updates rwnd. Version 3 corrected change introduced in v2
in a way that existing function is not reordered/copied in line, but it is
correctly called. Thanks Vlad for suggesting.
Signed-off-by: Matija Glavinic Pecotic <matija.glavinic-pecotic.ext@nsn.com>
Reviewed-by: Alexander Sverdlin <alexander.sverdlin@nsn.com>
Acked-by: Vlad Yasevich <vyasevich@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
||
|---|---|---|
| .. | ||
| auth.h | ||
| checksum.h | ||
| command.h | ||
| constants.h | ||
| sctp.h | ||
| sm.h | ||
| structs.h | ||
| tsnmap.h | ||
| ulpevent.h | ||
| ulpqueue.h | ||