2 * SPDX-License-Identifier: Beerware
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 * ----------------------------------------------------------------------------
21 #include <sys/cpuset.h>
22 #include <sys/queue.h>
25 * Types of nodes in the topological tree.
28 /* No node has this type; can be used in topo API calls. */
30 /* Processing unit aka computing unit aka logical CPU. */
32 /* Physical subdivision of a package. */
34 /* CPU L1/L2/L3 cache. */
36 /* Package aka chip, equivalent to socket. */
40 /* Other logical or physical grouping of PUs. */
41 /* E.g. PUs on the same dye, or PUs sharing an FPU. */
43 /* The whole system. */
47 /* Hardware indenitifier of a topology component. */
48 typedef unsigned int hwid_t;
49 /* Logical CPU idenitifier. */
52 /* A node in the topology. */
54 struct topo_node *parent;
55 TAILQ_HEAD(topo_children, topo_node) children;
56 TAILQ_ENTRY(topo_node) siblings;
67 * Scheduling topology of a NUMA or SMP system.
69 * The top level topology is an array of pointers to groups. Each group
70 * contains a bitmask of cpus in its group or subgroups. It may also
71 * contain a pointer to an array of child groups.
73 * The bitmasks at non leaf groups may be used by consumers who support
74 * a smaller depth than the hardware provides.
76 * The topology may be omitted by systems where all CPUs are equal.
80 struct cpu_group *cg_parent; /* Our parent group. */
81 struct cpu_group *cg_child; /* Optional children groups. */
82 cpuset_t cg_mask; /* Mask of cpus in this group. */
83 int32_t cg_count; /* Count of cpus in this group. */
84 int32_t cg_first; /* First cpu in this group. */
85 int32_t cg_last; /* Last cpu in this group. */
86 int16_t cg_children; /* Number of children groups. */
87 int8_t cg_level; /* Shared cache level. */
88 int8_t cg_flags; /* Traversal modifiers. */
91 typedef struct cpu_group *cpu_group_t;
94 * Defines common resources for CPUs in the group. The highest level
95 * resource should be used when multiple are shared.
97 #define CG_SHARE_NONE 0
100 #define CG_SHARE_L3 3
102 #define MAX_CACHE_LEVELS CG_SHARE_L3
105 * Behavior modifiers for load balancing and affinity.
107 #define CG_FLAG_HTT 0x01 /* Schedule the alternate core last. */
108 #define CG_FLAG_SMT 0x02 /* New age htt, less crippled. */
109 #define CG_FLAG_THREAD (CG_FLAG_HTT | CG_FLAG_SMT) /* Any threading. */
110 #define CG_FLAG_NODE 0x04 /* NUMA node. */
113 * Convenience routines for building and traversing topologies.
116 void topo_init_node(struct topo_node *node);
117 void topo_init_root(struct topo_node *root);
118 struct topo_node * topo_add_node_by_hwid(struct topo_node *parent, int hwid,
119 topo_node_type type, uintptr_t subtype);
120 struct topo_node * topo_find_node_by_hwid(struct topo_node *parent, int hwid,
121 topo_node_type type, uintptr_t subtype);
122 void topo_promote_child(struct topo_node *child);
123 struct topo_node * topo_next_node(struct topo_node *top,
124 struct topo_node *node);
125 struct topo_node * topo_next_nonchild_node(struct topo_node *top,
126 struct topo_node *node);
127 void topo_set_pu_id(struct topo_node *node, cpuid_t id);
132 * Some systems have useful sub-package core organizations. On these,
133 * a package has one or more subgroups. Each subgroup contains one or
134 * more cache groups (cores that share a last level cache).
137 TOPO_LEVEL_CACHEGROUP,
140 TOPO_LEVEL_COUNT /* Must be last */
142 struct topo_analysis {
143 int entities[TOPO_LEVEL_COUNT];
145 int topo_analyze(struct topo_node *topo_root, int all,
146 struct topo_analysis *results);
148 #define TOPO_FOREACH(i, root) \
149 for (i = root; i != NULL; i = topo_next_node(root, i))
151 struct cpu_group *smp_topo(void);
152 struct cpu_group *smp_topo_alloc(u_int count);
153 struct cpu_group *smp_topo_none(void);
154 struct cpu_group *smp_topo_1level(int l1share, int l1count, int l1flags);
155 struct cpu_group *smp_topo_2level(int l2share, int l2count, int l1share,
156 int l1count, int l1flags);
157 struct cpu_group *smp_topo_find(struct cpu_group *top, int cpu);
159 extern void (*cpustop_restartfunc)(void);
160 /* The suspend/resume cpusets are x86 only, but minimize ifdefs. */
161 extern volatile cpuset_t resuming_cpus; /* woken up cpus in suspend pen */
162 extern volatile cpuset_t started_cpus; /* cpus to let out of stop pen */
163 extern volatile cpuset_t stopped_cpus; /* cpus in stop pen */
164 extern volatile cpuset_t suspended_cpus; /* cpus [near] sleeping in susp pen */
165 extern volatile cpuset_t toresume_cpus; /* cpus to let out of suspend pen */
166 extern cpuset_t hlt_cpus_mask; /* XXX 'mask' is detail in old impl */
167 extern cpuset_t logical_cpus_mask;
170 extern u_int mp_maxid;
171 extern int mp_maxcpus;
172 extern int mp_ncores;
175 extern volatile int smp_started;
176 extern int smp_threads_per_core;
178 extern cpuset_t all_cpus;
179 extern cpuset_t cpuset_domain[MAXMEMDOM]; /* CPUs in each NUMA domain. */
182 * Macro allowing us to determine whether a CPU is absent at any given
183 * time, thus permitting us to configure sparse maps of cpuid-dependent
184 * (per-CPU) structures.
186 #define CPU_ABSENT(x_cpu) (!CPU_ISSET(x_cpu, &all_cpus))
189 * Macros to iterate over non-absent CPUs. CPU_FOREACH() takes an
190 * integer iterator and iterates over the available set of CPUs.
191 * CPU_FIRST() returns the id of the first non-absent CPU. CPU_NEXT()
192 * returns the id of the next non-absent CPU. It will wrap back to
193 * CPU_FIRST() once the end of the list is reached. The iterators are
194 * currently implemented via inline functions.
196 #define CPU_FOREACH(i) \
197 for ((i) = 0; (i) <= mp_maxid; (i)++) \
198 if (!CPU_ABSENT((i)))
223 #define CPU_FIRST() cpu_first()
224 #define CPU_NEXT(i) cpu_next((i))
228 * Machine dependent functions used to initialize MP support.
230 * The cpu_mp_probe() should check to see if MP support is present and return
231 * zero if it is not or non-zero if it is. If MP support is present, then
232 * cpu_mp_start() will be called so that MP can be enabled. This function
233 * should do things such as startup secondary processors. It should also
234 * setup mp_ncpus, all_cpus, and smp_cpus. It should also ensure that
235 * smp_started is initialized at the appropriate time.
236 * Once cpu_mp_start() returns, machine independent MP startup code will be
237 * executed and a simple message will be output to the console. Finally,
238 * cpu_mp_announce() will be called so that machine dependent messages about
239 * the MP support may be output to the console if desired.
241 * The cpu_setmaxid() function is called very early during the boot process
242 * so that the MD code may set mp_maxid to provide an upper bound on CPU IDs
243 * that other subsystems may use. If a platform is not able to determine
244 * the exact maximum ID that early, then it may set mp_maxid to MAXCPU - 1.
248 struct cpu_group *cpu_topo(void);
249 void cpu_mp_announce(void);
250 int cpu_mp_probe(void);
251 void cpu_mp_setmaxid(void);
252 void cpu_mp_start(void);
254 void forward_signal(struct thread *);
255 int restart_cpus(cpuset_t);
256 int stop_cpus(cpuset_t);
257 int stop_cpus_hard(cpuset_t);
258 #if defined(__amd64__) || defined(__i386__)
259 int suspend_cpus(cpuset_t);
260 int resume_cpus(cpuset_t);
263 void smp_rendezvous_action(void);
264 extern struct mtx smp_ipi_mtx;
268 int quiesce_all_cpus(const char *, int);
269 int quiesce_cpus(cpuset_t, const char *, int);
270 void quiesce_all_critical(void);
271 void cpus_fence_seq_cst(void);
272 void smp_no_rendezvous_barrier(void *);
273 void smp_rendezvous(void (*)(void *),
277 void smp_rendezvous_cpus(cpuset_t,
283 struct smp_rendezvous_cpus_retry_arg {
286 void smp_rendezvous_cpus_retry(cpuset_t,
290 void (*)(void *, int),
291 struct smp_rendezvous_cpus_retry_arg *);
293 void smp_rendezvous_cpus_done(struct smp_rendezvous_cpus_retry_arg *);
297 #endif /* _SYS_SMP_H_ */