blob: c7a1810c1cff29e5b50aef5ebdaa59003dec986b (
plain) (
blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
|
performance: 0.984
virtual: 0.922
debug: 0.904
device: 0.788
kernel: 0.773
files: 0.742
network: 0.691
ppc: 0.690
permissions: 0.690
vnc: 0.690
PID: 0.677
graphic: 0.672
TCG: 0.627
socket: 0.571
hypervisor: 0.525
semantic: 0.508
risc-v: 0.497
peripherals: 0.468
arm: 0.464
architecture: 0.455
VMM: 0.417
register: 0.414
x86: 0.379
boot: 0.310
i386: 0.309
mistranslation: 0.282
user-level: 0.238
assembly: 0.154
KVM: 0.115
TCP performance problems - GSO/TSO, MSS, 8139 related (Ignores lower MTU from PMTUD/MSS)
Description of problem:
MTU handling on guests using an RTL8139 virtualized NIC is broken; net/hw/8139.c works with a static MTU of 1500b for TCP offloading, leading to low throughput when clients connect from sub 1500MTU networks. PMTUD is ignored, and locking to a lower MTU in the OS mitigates the issue.
Steps to reproduce:
1. Create a guest with an RTL8139 nic
2. Try to retrieve a file from a client behind a sub 1500 MTU link
3. Observe low bandwidth due to retransmits
Additional information:
I just debugged this issue for an NGO which, for whatever reason, had an RTL8139 NIC in their guest. After i finally traced this to the RTL8139, i found this qemu-devel/netdev thread from six years ago, which apparently already debugged this issue and proposed a patch: https://lore.kernel.org/all/20161114162505.GD26664@stefanha-x1.localdomain/
I did not test the patch proposed there, but note that `net/hw/8139.c` still looks as discussed in that qemu-devel/netdev thread. As i haven't found a bug report in the archives, i figured you might want to know.
|