summaryrefslogtreecommitdiff
path: root/Documentation/virtual/kvm/halt-polling.txt
blob: 4a841831876978cbc5396f8fc88232815cfd786c (plain)
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
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
The KVM halt polling system
===========================

The KVM halt polling system provides a feature within KVM whereby the latency
of a guest can, under some circumstances, be reduced by polling in the host
for some time period after the guest has elected to no longer run by cedeing.
That is, when a guest vcpu has ceded, or in the case of powerpc when all of the
vcpus of a single vcore have ceded, the host kernel polls for wakeup conditions
before giving up the cpu to the scheduler in order to let something else run.

Polling provides a latency advantage in cases where the guest can be run again
very quickly by at least saving us a trip through the scheduler, normally on
the order of a few micro-seconds, although performance benefits are workload
dependant. In the event that no wakeup source arrives during the polling
interval or some other task on the runqueue is runnable the scheduler is
invoked. Thus halt polling is especially useful on workloads with very short
wakeup periods where the time spent halt polling is minimised and the time
savings of not invoking the scheduler are distinguishable.

The generic halt polling code is implemented in:

	virt/kvm/kvm_main.c: kvm_vcpu_block()

The powerpc kvm-hv specific case is implemented in:

	arch/powerpc/kvm/book3s_hv.c: kvmppc_vcore_blocked()

Halt Polling Interval
=====================

The maximum time for which to poll before invoking the scheduler, referred to
as the halt polling interval, is increased and decreased based on the perceived
effectiveness of the polling in an attempt to limit pointless polling.
This value is stored in either the vcpu struct:

	kvm_vcpu->halt_poll_ns

or in the case of powerpc kvm-hv, in the vcore struct:

	kvmppc_vcore->halt_poll_ns

Thus this is a per vcpu (or vcore) value.

During polling if a wakeup source is received within the halt polling interval,
the interval is left unchanged. In the event that a wakeup source isn't
received during the polling interval (and thus schedule is invoked) there are
two options, either the polling interval and total block time[0] were less than
the global max polling interval (see module params below), or the total block
time was greater than the global max polling interval.

In the event that both the polling interval and total block time were less than
the global max polling interval then the polling interval can be increased in
the hope that next time during the longer polling interval the wake up source
will be received while the host is polling and the latency benefits will be
received. The polling interval is grown in the function grow_halt_poll_ns() and
is multiplied by the module parameter halt_poll_ns_grow.

In the event that the total block time was greater than the global max polling
interval then the host will never poll for long enough (limited by the global
max) to wakeup during the polling interval so it may as well be shrunk in order
to avoid pointless polling. The polling interval is shrunk in the function
shrink_halt_poll_ns() and is divided by the module parameter
halt_poll_ns_shrink, or set to 0 iff halt_poll_ns_shrink == 0.

It is worth noting that this adjustment process attempts to hone in on some
steady state polling interval but will only really do a good job for wakeups
which come at an approximately constant rate, otherwise there will be constant
adjustment of the polling interval.

[0] total block time: the time between when the halt polling function is
		      invoked and a wakeup source received (irrespective of
		      whether the scheduler is invoked within that function).

Module Parameters
=================

The kvm module has 3 tuneable module parameters to adjust the global max
polling interval as well as the rate at which the polling interval is grown and
shrunk. These variables are defined in include/linux/kvm_host.h and as module
parameters in virt/kvm/kvm_main.c, or arch/powerpc/kvm/book3s_hv.c in the
powerpc kvm-hv case.

Module Parameter    |	     Description	      |	     Default Value
--------------------------------------------------------------------------------
halt_poll_ns	    | The global max polling interval | KVM_HALT_POLL_NS_DEFAULT
		    | which defines the ceiling value |
		    | of the polling interval for     | (per arch value)
		    | each vcpu. 		      |
--------------------------------------------------------------------------------
halt_poll_ns_grow   | The value by which the halt     |	2
		    | polling interval is multiplied  |
		    | in the grow_halt_poll_ns()      |
		    | function.			      |
--------------------------------------------------------------------------------
halt_poll_ns_shrink | The value by which the halt     |	0
		    | polling interval is divided in  |
		    | the shrink_halt_poll_ns()	      |
		    | function.			      |
--------------------------------------------------------------------------------

These module parameters can be set from the debugfs files in:

	/sys/module/kvm/parameters/

Note: that these module parameters are system wide values and are not able to
      be tuned on a per vm basis.

Further Notes
=============

- Care should be taken when setting the halt_poll_ns module parameter as a
large value has the potential to drive the cpu usage to 100% on a machine which
would be almost entirely idle otherwise. This is because even if a guest has
wakeups during which very little work is done and which are quite far apart, if
the period is shorter than the global max polling interval (halt_poll_ns) then
the host will always poll for the entire block time and thus cpu utilisation
will go to 100%.

- Halt polling essentially presents a trade off between power usage and latency
and the module parameters should be used to tune the affinity for this. Idle
cpu time is essentially converted to host kernel time with the aim of decreasing
latency when entering the guest.

- Halt polling will only be conducted by the host when no other tasks are
runnable on that cpu, otherwise the polling will cease immediately and
schedule will be invoked to allow that other task to run. Thus this doesn't
allow a guest to denial of service the cpu.