1 /*-
2 * SPDX-License-Identifier: Beerware
3 *
4 * ----------------------------------------------------------------------------
5 * "THE BEER-WARE LICENSE" (Revision 42):
6 * <phk@FreeBSD.org> wrote this file. As long as you retain this notice you
7 * can do whatever you want with this stuff. If we meet some day, and you think
8 * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp
9 * ----------------------------------------------------------------------------
10 */
11
12 #ifndef _SYS_SMP_H_
13 #define _SYS_SMP_H_
14
15 #ifdef _KERNEL
16
17 #ifndef LOCORE
18
19 #include <sys/types.h>
20 #include <sys/cpuset.h>
21 #include <sys/queue.h>
22
23 /*
24 * Types of nodes in the topological tree.
25 */
26 typedef enum {
27 /* No node has this type; can be used in topo API calls. */
28 TOPO_TYPE_DUMMY,
29 /* Processing unit aka computing unit aka logical CPU. */
30 TOPO_TYPE_PU,
31 /* Physical subdivision of a package. */
32 TOPO_TYPE_CORE,
33 /* CPU L1/L2/L3 cache. */
34 TOPO_TYPE_CACHE,
35 /* Package aka chip, equivalent to socket. */
36 TOPO_TYPE_PKG,
37 /* NUMA node. */
38 TOPO_TYPE_NODE,
39 /* Other logical or physical grouping of PUs. */
40 /* E.g. PUs on the same dye, or PUs sharing an FPU. */
41 TOPO_TYPE_GROUP,
42 /* The whole system. */
43 TOPO_TYPE_SYSTEM
44 } topo_node_type;
45
46 /* Hardware indenitifier of a topology component. */
47 typedef unsigned int hwid_t;
48 /* Logical CPU idenitifier. */
49 typedef int cpuid_t;
50
51 /* A node in the topology. */
52 struct topo_node {
53 struct topo_node *parent;
54 TAILQ_HEAD(topo_children, topo_node) children;
55 TAILQ_ENTRY(topo_node) siblings;
56 cpuset_t cpuset;
57 topo_node_type type;
58 uintptr_t subtype;
59 hwid_t hwid;
60 cpuid_t id;
61 int nchildren;
62 int cpu_count;
63 };
64
65 /*
66 * Scheduling topology of a NUMA or SMP system.
67 *
68 * The top level topology is an array of pointers to groups. Each group
69 * contains a bitmask of cpus in its group or subgroups. It may also
70 * contain a pointer to an array of child groups.
71 *
72 * The bitmasks at non leaf groups may be used by consumers who support
73 * a smaller depth than the hardware provides.
74 *
75 * The topology may be omitted by systems where all CPUs are equal.
76 */
77
78 struct cpu_group {
79 struct cpu_group *cg_parent; /* Our parent group. */
80 struct cpu_group *cg_child; /* Optional children groups. */
81 cpuset_t cg_mask; /* Mask of cpus in this group. */
82 int32_t cg_count; /* Count of cpus in this group. */
83 int32_t cg_first; /* First cpu in this group. */
84 int32_t cg_last; /* Last cpu in this group. */
85 int16_t cg_children; /* Number of children groups. */
86 int8_t cg_level; /* Shared cache level. */
87 int8_t cg_flags; /* Traversal modifiers. */
88 };
89
90 typedef struct cpu_group *cpu_group_t;
91
92 extern cpu_group_t cpu_top;
93
94 /*
95 * Defines common resources for CPUs in the group. The highest level
96 * resource should be used when multiple are shared.
97 */
98 #define CG_SHARE_NONE 0
99 #define CG_SHARE_L1 1
100 #define CG_SHARE_L2 2
101 #define CG_SHARE_L3 3
102
103 #define MAX_CACHE_LEVELS CG_SHARE_L3
104
105 /*
106 * Behavior modifiers for load balancing and affinity.
107 */
108 #define CG_FLAG_HTT 0x01 /* Schedule the alternate core last. */
109 #define CG_FLAG_SMT 0x02 /* New age htt, less crippled. */
110 #define CG_FLAG_THREAD (CG_FLAG_HTT | CG_FLAG_SMT) /* Any threading. */
111 #define CG_FLAG_NODE 0x04 /* NUMA node. */
112
113 /*
114 * Convenience routines for building and traversing topologies.
115 */
116 #ifdef SMP
117 void topo_init_node(struct topo_node *node);
118 void topo_init_root(struct topo_node *root);
119 struct topo_node * topo_add_node_by_hwid(struct topo_node *parent, int hwid,
120 topo_node_type type, uintptr_t subtype);
121 struct topo_node * topo_find_node_by_hwid(struct topo_node *parent, int hwid,
122 topo_node_type type, uintptr_t subtype);
123 void topo_promote_child(struct topo_node *child);
124 struct topo_node * topo_next_node(struct topo_node *top,
125 struct topo_node *node);
126 struct topo_node * topo_next_nonchild_node(struct topo_node *top,
127 struct topo_node *node);
128 void topo_set_pu_id(struct topo_node *node, cpuid_t id);
129
130 enum topo_level {
131 TOPO_LEVEL_PKG = 0,
132 /*
133 * Some systems have useful sub-package core organizations. On these,
134 * a package has one or more subgroups. Each subgroup contains one or
135 * more cache groups (cores that share a last level cache).
136 */
137 TOPO_LEVEL_GROUP,
138 TOPO_LEVEL_CACHEGROUP,
139 TOPO_LEVEL_CORE,
140 TOPO_LEVEL_THREAD,
141 TOPO_LEVEL_COUNT /* Must be last */
142 };
143 struct topo_analysis {
144 int entities[TOPO_LEVEL_COUNT];
145 };
146 int topo_analyze(struct topo_node *topo_root, int all,
147 struct topo_analysis *results);
148
149 #define TOPO_FOREACH(i, root) \
150 for (i = root; i != NULL; i = topo_next_node(root, i))
151
152 struct cpu_group *smp_topo_1level(int l1share, int l1count, int l1flags);
153 struct cpu_group *smp_topo_2level(int l2share, int l2count, int l1share,
154 int l1count, int l1flags);
155 struct cpu_group *smp_topo_find(struct cpu_group *top, int cpu);
156
157 extern void (*cpustop_restartfunc)(void);
158 /* The suspend/resume cpusets are x86 only, but minimize ifdefs. */
159 extern volatile cpuset_t resuming_cpus; /* woken up cpus in suspend pen */
160 extern volatile cpuset_t started_cpus; /* cpus to let out of stop pen */
161 extern volatile cpuset_t stopped_cpus; /* cpus in stop pen */
162 extern volatile cpuset_t suspended_cpus; /* cpus [near] sleeping in susp pen */
163 extern volatile cpuset_t toresume_cpus; /* cpus to let out of suspend pen */
164 extern cpuset_t logical_cpus_mask;
165 #endif /* SMP */
166
167 struct cpu_group *smp_topo(void);
168 struct cpu_group *smp_topo_alloc(u_int count);
169 struct cpu_group *smp_topo_none(void);
170
171 extern u_int mp_maxid;
172 extern int mp_maxcpus;
173 extern int mp_ncores;
174 extern int mp_ncpus;
175 extern int smp_cpus;
176 extern volatile int smp_started;
177 extern int smp_threads_per_core;
178
179 extern cpuset_t all_cpus;
180 extern cpuset_t cpuset_domain[MAXMEMDOM]; /* CPUs in each NUMA domain. */
181
182 struct pcb;
183 extern struct pcb *stoppcbs;
184
185 /*
186 * Macro allowing us to determine whether a CPU is absent at any given
187 * time, thus permitting us to configure sparse maps of cpuid-dependent
188 * (per-CPU) structures.
189 */
190 #define CPU_ABSENT(x_cpu) (!CPU_ISSET(x_cpu, &all_cpus))
191
192 /*
193 * Macros to iterate over non-absent CPUs. CPU_FOREACH() takes an
194 * integer iterator and iterates over the available set of CPUs.
195 * CPU_FIRST() returns the id of the first non-absent CPU. CPU_NEXT()
196 * returns the id of the next non-absent CPU. It will wrap back to
197 * CPU_FIRST() once the end of the list is reached. The iterators are
198 * currently implemented via inline functions.
199 */
200 #define CPU_FOREACH(i) \
201 for ((i) = 0; (i) <= mp_maxid; (i)++) \
202 if (!CPU_ABSENT((i)))
203
204 static __inline int
cpu_first(void)205 cpu_first(void)
206 {
207 int i;
208
209 for (i = 0;; i++)
210 if (!CPU_ABSENT(i))
211 return (i);
212 }
213
214 static __inline int
cpu_next(int i)215 cpu_next(int i)
216 {
217
218 for (;;) {
219 i++;
220 if ((u_int)i > mp_maxid)
221 i = 0;
222 if (!CPU_ABSENT(i))
223 return (i);
224 }
225 }
226
227 #define CPU_FIRST() cpu_first()
228 #define CPU_NEXT(i) cpu_next((i))
229
230 #ifdef SMP
231 /*
232 * Machine dependent functions used to initialize MP support.
233 *
234 * The cpu_mp_probe() should check to see if MP support is present and return
235 * zero if it is not or non-zero if it is. If MP support is present, then
236 * cpu_mp_start() will be called so that MP can be enabled. This function
237 * should do things such as startup secondary processors. It should also
238 * setup mp_ncpus, all_cpus, and smp_cpus. It should also ensure that
239 * smp_started is initialized at the appropriate time.
240 * Once cpu_mp_start() returns, machine independent MP startup code will be
241 * executed and a simple message will be output to the console. Finally,
242 * cpu_mp_announce() will be called so that machine dependent messages about
243 * the MP support may be output to the console if desired.
244 *
245 * The cpu_setmaxid() function is called very early during the boot process
246 * so that the MD code may set mp_maxid to provide an upper bound on CPU IDs
247 * that other subsystems may use. If a platform is not able to determine
248 * the exact maximum ID that early, then it may set mp_maxid to MAXCPU - 1.
249 */
250 struct thread;
251
252 struct cpu_group *cpu_topo(void);
253 void cpu_mp_announce(void);
254 int cpu_mp_probe(void);
255 void cpu_mp_setmaxid(void);
256 void cpu_mp_start(void);
257 void cpu_mp_stop(void); /* Go back to single-CPU */
258
259 void forward_signal(struct thread *);
260 int restart_cpus(cpuset_t);
261 int stop_cpus(cpuset_t);
262 int stop_cpus_hard(cpuset_t);
263 #if defined(__amd64__) || defined(__i386__)
264 int suspend_cpus(cpuset_t);
265 int resume_cpus(cpuset_t);
266 int offline_cpus(cpuset_t);
267 #endif
268
269 void smp_rendezvous_action(void);
270 extern struct mtx smp_ipi_mtx;
271
272 #endif /* SMP */
273
274 int quiesce_all_cpus(const char *, int);
275 int quiesce_cpus(cpuset_t, const char *, int);
276 void quiesce_all_critical(void);
277 void cpus_fence_seq_cst(void);
278 void smp_no_rendezvous_barrier(void *);
279 void smp_rendezvous(void (*)(void *),
280 void (*)(void *),
281 void (*)(void *),
282 void *arg);
283 void smp_rendezvous_cpus(cpuset_t,
284 void (*)(void *),
285 void (*)(void *),
286 void (*)(void *),
287 void *arg);
288 void smp_rendezvous_cpu(u_int,
289 void (*)(void *),
290 void (*)(void *),
291 void (*)(void *),
292 void *arg);
293
294 struct smp_rendezvous_cpus_retry_arg {
295 cpuset_t cpus;
296 };
297 void smp_rendezvous_cpus_retry(cpuset_t,
298 void (*)(void *),
299 void (*)(void *),
300 void (*)(void *),
301 void (*)(void *, int),
302 struct smp_rendezvous_cpus_retry_arg *);
303
304 void smp_rendezvous_cpus_done(struct smp_rendezvous_cpus_retry_arg *);
305
306 #endif /* !LOCORE */
307 #endif /* _KERNEL */
308 #endif /* _SYS_SMP_H_ */
309