1=================
2Scheduler Domains
3=================
4
5Each CPU has a "base" scheduling domain (struct sched_domain). The domain
6hierarchy is built from these base domains via the ->parent pointer. ->parent
7MUST be NULL terminated, and domain structures should be per-CPU as they are
8locklessly updated.
9
10Each scheduling domain spans a number of CPUs (stored in the ->span field).
11A domain's span MUST be a superset of it child's span (this restriction could
12be relaxed if the need arises), and a base domain for CPU i MUST span at least
13i. The top domain for each CPU will generally span all CPUs in the system
14although strictly it doesn't have to, but this could lead to a case where some
15CPUs will never be given tasks to run unless the CPUs allowed mask is
16explicitly set. A sched domain's span means "balance process load among these
17CPUs".
18
19Each scheduling domain must have one or more CPU groups (struct sched_group)
20which are organised as a circular one way linked list from the ->groups
21pointer. The union of cpumasks of these groups MUST be the same as the
22domain's span. The intersection of cpumasks from any two of these groups
23MUST be the empty set. The group pointed to by the ->groups pointer MUST
24contain the CPU to which the domain belongs. Groups may be shared among
25CPUs as they contain read only data after they have been set up.
26
27Balancing within a sched domain occurs between groups. That is, each group
28is treated as one entity. The load of a group is defined as the sum of the
29load of each of its member CPUs, and only when the load of a group becomes
30out of balance are tasks moved between groups.
31
32In kernel/sched/core.c, trigger_load_balance() is run periodically on each CPU
33through scheduler_tick(). It raises a softirq after the next regularly scheduled
34rebalancing event for the current runqueue has arrived. The actual load
35balancing workhorse, run_rebalance_domains()->rebalance_domains(), is then run
36in softirq context (SCHED_SOFTIRQ).
37
38The latter function takes two arguments: the current CPU and whether it was idle
39at the time the scheduler_tick() happened and iterates over all sched domains
40our CPU is on, starting from its base domain and going up the ->parent chain.
41While doing that, it checks to see if the current domain has exhausted its
42rebalance interval. If so, it runs load_balance() on that domain. It then checks
43the parent sched_domain (if it exists), and the parent of the parent and so
44forth.
45
46Initially, load_balance() finds the busiest group in the current sched domain.
47If it succeeds, it looks for the busiest runqueue of all the CPUs' runqueues in
48that group. If it manages to find such a runqueue, it locks both our initial
49CPU's runqueue and the newly found busiest one and starts moving tasks from it
50to our runqueue. The exact number of tasks amounts to an imbalance previously
51computed while iterating over this sched domain's groups.
52
53Implementing sched domains
54==========================
55
56The "base" domain will "span" the first level of the hierarchy. In the case
57of SMT, you'll span all siblings of the physical CPU, with each group being
58a single virtual CPU.
59
60In SMP, the parent of the base domain will span all physical CPUs in the
61node. Each group being a single physical CPU. Then with NUMA, the parent
62of the SMP domain will span the entire machine, with each group having the
63cpumask of a node. Or, you could do multi-level NUMA or Opteron, for example,
64might have just one domain covering its one NUMA level.
65
66The implementor should read comments in include/linux/sched.h:
67struct sched_domain fields, SD_FLAG_*, SD_*_INIT to get an idea of
68the specifics and what to tune.
69
70Architectures may retain the regular override the default SD_*_INIT flags
71while using the generic domain builder in kernel/sched/core.c if they wish to
72retain the traditional SMT->SMP->NUMA topology (or some subset of that). This
73can be done by #define'ing ARCH_HASH_SCHED_TUNE.
74
75Alternatively, the architecture may completely override the generic domain
76builder by #define'ing ARCH_HASH_SCHED_DOMAIN, and exporting your
77arch_init_sched_domains function. This function will attach domains to all
78CPUs using cpu_attach_domain.
79
80The sched-domains debugging infrastructure can be enabled by enabling
81CONFIG_SCHED_DEBUG. This enables an error checking parse of the sched domains
82which should catch most possible errors (described above). It also prints out
83the domain structure in a visual format.
84