1 /*-
2 * ----------------------------------------------------------------------------
3 * "THE BEER-WARE LICENSE" (Revision 42):
4 * <phk@FreeBSD.org> wrote this file. As long as you retain this notice you
5 * can do whatever you want with this stuff. If we meet some day, and you think
6 * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp
7 * ----------------------------------------------------------------------------
8 *
9 * $FreeBSD$
10 */
11
12 #ifndef _SYS_SMP_H_
13 #define _SYS_SMP_H_
14
15 #ifdef _KERNEL
16
17 #ifndef LOCORE
18
19 #include <sys/cpuset.h>
20 #include <sys/queue.h>
21
22 /*
23 * Types of nodes in the topological tree.
24 */
25 typedef enum {
26 /* No node has this type; can be used in topo API calls. */
27 TOPO_TYPE_DUMMY,
28 /* Processing unit aka computing unit aka logical CPU. */
29 TOPO_TYPE_PU,
30 /* Physical subdivision of a package. */
31 TOPO_TYPE_CORE,
32 /* CPU L1/L2/L3 cache. */
33 TOPO_TYPE_CACHE,
34 /* Package aka chip, equivalent to socket. */
35 TOPO_TYPE_PKG,
36 /* NUMA node. */
37 TOPO_TYPE_NODE,
38 /* Other logical or physical grouping of PUs. */
39 /* E.g. PUs on the same dye, or PUs sharing an FPU. */
40 TOPO_TYPE_GROUP,
41 /* The whole system. */
42 TOPO_TYPE_SYSTEM
43 } topo_node_type;
44
45 /* Hardware indenitifier of a topology component. */
46 typedef unsigned int hwid_t;
47 /* Logical CPU idenitifier. */
48 typedef int cpuid_t;
49
50 /* A node in the topology. */
51 struct topo_node {
52 struct topo_node *parent;
53 TAILQ_HEAD(topo_children, topo_node) children;
54 TAILQ_ENTRY(topo_node) siblings;
55 cpuset_t cpuset;
56 topo_node_type type;
57 uintptr_t subtype;
58 hwid_t hwid;
59 cpuid_t id;
60 int nchildren;
61 int cpu_count;
62 };
63
64 /*
65 * Scheduling topology of a NUMA or SMP system.
66 *
67 * The top level topology is an array of pointers to groups. Each group
68 * contains a bitmask of cpus in its group or subgroups. It may also
69 * contain a pointer to an array of child groups.
70 *
71 * The bitmasks at non leaf groups may be used by consumers who support
72 * a smaller depth than the hardware provides.
73 *
74 * The topology may be omitted by systems where all CPUs are equal.
75 */
76
77 struct cpu_group {
78 struct cpu_group *cg_parent; /* Our parent group. */
79 struct cpu_group *cg_child; /* Optional children groups. */
80 cpuset_t cg_mask; /* Mask of cpus in this group. */
81 int32_t cg_count; /* Count of cpus in this group. */
82 int16_t cg_children; /* Number of children groups. */
83 int8_t cg_level; /* Shared cache level. */
84 int8_t cg_flags; /* Traversal modifiers. */
85 };
86
87 typedef struct cpu_group *cpu_group_t;
88
89 /*
90 * Defines common resources for CPUs in the group. The highest level
91 * resource should be used when multiple are shared.
92 */
93 #define CG_SHARE_NONE 0
94 #define CG_SHARE_L1 1
95 #define CG_SHARE_L2 2
96 #define CG_SHARE_L3 3
97
98 #define MAX_CACHE_LEVELS CG_SHARE_L3
99
100 /*
101 * Behavior modifiers for load balancing and affinity.
102 */
103 #define CG_FLAG_HTT 0x01 /* Schedule the alternate core last. */
104 #define CG_FLAG_SMT 0x02 /* New age htt, less crippled. */
105 #define CG_FLAG_THREAD (CG_FLAG_HTT | CG_FLAG_SMT) /* Any threading. */
106
107 /*
108 * Convenience routines for building and traversing topologies.
109 */
110 #ifdef SMP
111 void topo_init_node(struct topo_node *node);
112 void topo_init_root(struct topo_node *root);
113 struct topo_node * topo_add_node_by_hwid(struct topo_node *parent, int hwid,
114 topo_node_type type, uintptr_t subtype);
115 struct topo_node * topo_find_node_by_hwid(struct topo_node *parent, int hwid,
116 topo_node_type type, uintptr_t subtype);
117 void topo_promote_child(struct topo_node *child);
118 struct topo_node * topo_next_node(struct topo_node *top,
119 struct topo_node *node);
120 struct topo_node * topo_next_nonchild_node(struct topo_node *top,
121 struct topo_node *node);
122 void topo_set_pu_id(struct topo_node *node, cpuid_t id);
123 int topo_analyze(struct topo_node *topo_root, int all, int *pkg_count,
124 int *cores_per_pkg, int *thrs_per_core);
125
126 #define TOPO_FOREACH(i, root) \
127 for (i = root; i != NULL; i = topo_next_node(root, i))
128
129 struct cpu_group *smp_topo(void);
130 struct cpu_group *smp_topo_alloc(u_int count);
131 struct cpu_group *smp_topo_none(void);
132 struct cpu_group *smp_topo_1level(int l1share, int l1count, int l1flags);
133 struct cpu_group *smp_topo_2level(int l2share, int l2count, int l1share,
134 int l1count, int l1flags);
135 struct cpu_group *smp_topo_find(struct cpu_group *top, int cpu);
136
137 extern void (*cpustop_restartfunc)(void);
138 extern int smp_cpus;
139 /* The suspend/resume cpusets are x86 only, but minimize ifdefs. */
140 extern volatile cpuset_t resuming_cpus; /* woken up cpus in suspend pen */
141 extern volatile cpuset_t started_cpus; /* cpus to let out of stop pen */
142 extern volatile cpuset_t stopped_cpus; /* cpus in stop pen */
143 extern volatile cpuset_t suspended_cpus; /* cpus [near] sleeping in susp pen */
144 extern volatile cpuset_t toresume_cpus; /* cpus to let out of suspend pen */
145 extern cpuset_t hlt_cpus_mask; /* XXX 'mask' is detail in old impl */
146 extern cpuset_t logical_cpus_mask;
147 #endif /* SMP */
148
149 extern u_int mp_maxid;
150 extern int mp_maxcpus;
151 extern int mp_ncpus;
152 extern volatile int smp_started;
153
154 extern cpuset_t all_cpus;
155 extern cpuset_t cpuset_domain[MAXMEMDOM]; /* CPUs in each NUMA domain. */
156
157 /*
158 * Macro allowing us to determine whether a CPU is absent at any given
159 * time, thus permitting us to configure sparse maps of cpuid-dependent
160 * (per-CPU) structures.
161 */
162 #define CPU_ABSENT(x_cpu) (!CPU_ISSET(x_cpu, &all_cpus))
163
164 /*
165 * Macros to iterate over non-absent CPUs. CPU_FOREACH() takes an
166 * integer iterator and iterates over the available set of CPUs.
167 * CPU_FIRST() returns the id of the first non-absent CPU. CPU_NEXT()
168 * returns the id of the next non-absent CPU. It will wrap back to
169 * CPU_FIRST() once the end of the list is reached. The iterators are
170 * currently implemented via inline functions.
171 */
172 #define CPU_FOREACH(i) \
173 for ((i) = 0; (i) <= mp_maxid; (i)++) \
174 if (!CPU_ABSENT((i)))
175
176 static __inline int
cpu_first(void)177 cpu_first(void)
178 {
179 int i;
180
181 for (i = 0;; i++)
182 if (!CPU_ABSENT(i))
183 return (i);
184 }
185
186 static __inline int
cpu_next(int i)187 cpu_next(int i)
188 {
189
190 for (;;) {
191 i++;
192 if (i > mp_maxid)
193 i = 0;
194 if (!CPU_ABSENT(i))
195 return (i);
196 }
197 }
198
199 #define CPU_FIRST() cpu_first()
200 #define CPU_NEXT(i) cpu_next((i))
201
202 #ifdef SMP
203 /*
204 * Machine dependent functions used to initialize MP support.
205 *
206 * The cpu_mp_probe() should check to see if MP support is present and return
207 * zero if it is not or non-zero if it is. If MP support is present, then
208 * cpu_mp_start() will be called so that MP can be enabled. This function
209 * should do things such as startup secondary processors. It should also
210 * setup mp_ncpus, all_cpus, and smp_cpus. It should also ensure that
211 * smp_started is initialized at the appropriate time.
212 * Once cpu_mp_start() returns, machine independent MP startup code will be
213 * executed and a simple message will be output to the console. Finally,
214 * cpu_mp_announce() will be called so that machine dependent messages about
215 * the MP support may be output to the console if desired.
216 *
217 * The cpu_setmaxid() function is called very early during the boot process
218 * so that the MD code may set mp_maxid to provide an upper bound on CPU IDs
219 * that other subsystems may use. If a platform is not able to determine
220 * the exact maximum ID that early, then it may set mp_maxid to MAXCPU - 1.
221 */
222 struct thread;
223
224 struct cpu_group *cpu_topo(void);
225 void cpu_mp_announce(void);
226 int cpu_mp_probe(void);
227 void cpu_mp_setmaxid(void);
228 void cpu_mp_start(void);
229
230 void forward_signal(struct thread *);
231 int restart_cpus(cpuset_t);
232 int stop_cpus(cpuset_t);
233 int stop_cpus_hard(cpuset_t);
234 #if defined(__amd64__) || defined(__i386__)
235 int suspend_cpus(cpuset_t);
236 int resume_cpus(cpuset_t);
237 #endif
238
239 void smp_rendezvous_action(void);
240 extern struct mtx smp_ipi_mtx;
241
242 #endif /* SMP */
243
244 int quiesce_all_cpus(const char *, int);
245 int quiesce_cpus(cpuset_t, const char *, int);
246 /*
247 * smp_no_rendevous_barrier was renamed to smp_no_rendezvous_barrier
248 * in __FreeBSD_version 1101508, with the old name remaining in 11.x
249 * as an alias for compatibility. The old name will be gone in 12.0
250 * (__FreeBSD_version >= 1200028).
251 */
252 void smp_no_rendevous_barrier(void *);
253 void smp_no_rendezvous_barrier(void *);
254 void smp_rendezvous(void (*)(void *),
255 void (*)(void *),
256 void (*)(void *),
257 void *arg);
258 void smp_rendezvous_cpus(cpuset_t,
259 void (*)(void *),
260 void (*)(void *),
261 void (*)(void *),
262 void *arg);
263 #endif /* !LOCORE */
264 #endif /* _KERNEL */
265 #endif /* _SYS_SMP_H_ */
266