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/cpuset.h>
20#include <sys/queue.h>
21
22/*
23 * Types of nodes in the topological tree.
24 */
25typedef 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. */
46typedef	unsigned int hwid_t;
47/* Logical CPU idenitifier. */
48typedef	int cpuid_t;
49
50/* A node in the topology. */
51struct 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
77struct 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	int32_t		cg_first;	/* First cpu in this group. */
83	int32_t		cg_last;	/* Last cpu in this group. */
84	int16_t		cg_children;	/* Number of children groups. */
85	int8_t		cg_level;	/* Shared cache level. */
86	int8_t		cg_flags;	/* Traversal modifiers. */
87};
88
89typedef struct cpu_group *cpu_group_t;
90
91/*
92 * Defines common resources for CPUs in the group.  The highest level
93 * resource should be used when multiple are shared.
94 */
95#define	CG_SHARE_NONE	0
96#define	CG_SHARE_L1	1
97#define	CG_SHARE_L2	2
98#define	CG_SHARE_L3	3
99
100#define MAX_CACHE_LEVELS	CG_SHARE_L3
101
102/*
103 * Behavior modifiers for load balancing and affinity.
104 */
105#define	CG_FLAG_HTT	0x01		/* Schedule the alternate core last. */
106#define	CG_FLAG_SMT	0x02		/* New age htt, less crippled. */
107#define	CG_FLAG_THREAD	(CG_FLAG_HTT | CG_FLAG_SMT)	/* Any threading. */
108#define	CG_FLAG_NODE	0x04		/* NUMA node. */
109
110/*
111 * Convenience routines for building and traversing topologies.
112 */
113#ifdef SMP
114void topo_init_node(struct topo_node *node);
115void topo_init_root(struct topo_node *root);
116struct topo_node * topo_add_node_by_hwid(struct topo_node *parent, int hwid,
117    topo_node_type type, uintptr_t subtype);
118struct topo_node * topo_find_node_by_hwid(struct topo_node *parent, int hwid,
119    topo_node_type type, uintptr_t subtype);
120void topo_promote_child(struct topo_node *child);
121struct topo_node * topo_next_node(struct topo_node *top,
122    struct topo_node *node);
123struct topo_node * topo_next_nonchild_node(struct topo_node *top,
124    struct topo_node *node);
125void topo_set_pu_id(struct topo_node *node, cpuid_t id);
126
127enum topo_level {
128	TOPO_LEVEL_PKG = 0,
129	/*
130	 * Some systems have useful sub-package core organizations.  On these,
131	 * a package has one or more subgroups.  Each subgroup contains one or
132	 * more cache groups (cores that share a last level cache).
133	 */
134	TOPO_LEVEL_GROUP,
135	TOPO_LEVEL_CACHEGROUP,
136	TOPO_LEVEL_CORE,
137	TOPO_LEVEL_THREAD,
138	TOPO_LEVEL_COUNT	/* Must be last */
139};
140struct topo_analysis {
141	int entities[TOPO_LEVEL_COUNT];
142};
143int topo_analyze(struct topo_node *topo_root, int all,
144    struct topo_analysis *results);
145
146#define	TOPO_FOREACH(i, root)	\
147	for (i = root; i != NULL; i = topo_next_node(root, i))
148
149struct cpu_group *smp_topo(void);
150struct cpu_group *smp_topo_alloc(u_int count);
151struct cpu_group *smp_topo_none(void);
152struct cpu_group *smp_topo_1level(int l1share, int l1count, int l1flags);
153struct cpu_group *smp_topo_2level(int l2share, int l2count, int l1share,
154    int l1count, int l1flags);
155struct cpu_group *smp_topo_find(struct cpu_group *top, int cpu);
156
157extern void (*cpustop_restartfunc)(void);
158/* The suspend/resume cpusets are x86 only, but minimize ifdefs. */
159extern volatile cpuset_t resuming_cpus;	/* woken up cpus in suspend pen */
160extern volatile cpuset_t started_cpus;	/* cpus to let out of stop pen */
161extern volatile cpuset_t stopped_cpus;	/* cpus in stop pen */
162extern volatile cpuset_t suspended_cpus; /* cpus [near] sleeping in susp pen */
163extern volatile cpuset_t toresume_cpus;	/* cpus to let out of suspend pen */
164extern cpuset_t hlt_cpus_mask;		/* XXX 'mask' is detail in old impl */
165extern cpuset_t logical_cpus_mask;
166#endif /* SMP */
167
168extern u_int mp_maxid;
169extern int mp_maxcpus;
170extern int mp_ncores;
171extern int mp_ncpus;
172extern int smp_cpus;
173extern volatile int smp_started;
174extern int smp_threads_per_core;
175
176extern cpuset_t all_cpus;
177extern cpuset_t cpuset_domain[MAXMEMDOM]; 	/* CPUs in each NUMA domain. */
178
179struct pcb;
180extern struct pcb *stoppcbs;
181
182/*
183 * Macro allowing us to determine whether a CPU is absent at any given
184 * time, thus permitting us to configure sparse maps of cpuid-dependent
185 * (per-CPU) structures.
186 */
187#define	CPU_ABSENT(x_cpu)	(!CPU_ISSET(x_cpu, &all_cpus))
188
189/*
190 * Macros to iterate over non-absent CPUs.  CPU_FOREACH() takes an
191 * integer iterator and iterates over the available set of CPUs.
192 * CPU_FIRST() returns the id of the first non-absent CPU.  CPU_NEXT()
193 * returns the id of the next non-absent CPU.  It will wrap back to
194 * CPU_FIRST() once the end of the list is reached.  The iterators are
195 * currently implemented via inline functions.
196 */
197#define	CPU_FOREACH(i)							\
198	for ((i) = 0; (i) <= mp_maxid; (i)++)				\
199		if (!CPU_ABSENT((i)))
200
201static __inline int
202cpu_first(void)
203{
204	int i;
205
206	for (i = 0;; i++)
207		if (!CPU_ABSENT(i))
208			return (i);
209}
210
211static __inline int
212cpu_next(int i)
213{
214
215	for (;;) {
216		i++;
217		if ((u_int)i > mp_maxid)
218			i = 0;
219		if (!CPU_ABSENT(i))
220			return (i);
221	}
222}
223
224#define	CPU_FIRST()	cpu_first()
225#define	CPU_NEXT(i)	cpu_next((i))
226
227#ifdef SMP
228/*
229 * Machine dependent functions used to initialize MP support.
230 *
231 * The cpu_mp_probe() should check to see if MP support is present and return
232 * zero if it is not or non-zero if it is.  If MP support is present, then
233 * cpu_mp_start() will be called so that MP can be enabled.  This function
234 * should do things such as startup secondary processors.  It should also
235 * setup mp_ncpus, all_cpus, and smp_cpus.  It should also ensure that
236 * smp_started is initialized at the appropriate time.
237 * Once cpu_mp_start() returns, machine independent MP startup code will be
238 * executed and a simple message will be output to the console.  Finally,
239 * cpu_mp_announce() will be called so that machine dependent messages about
240 * the MP support may be output to the console if desired.
241 *
242 * The cpu_setmaxid() function is called very early during the boot process
243 * so that the MD code may set mp_maxid to provide an upper bound on CPU IDs
244 * that other subsystems may use.  If a platform is not able to determine
245 * the exact maximum ID that early, then it may set mp_maxid to MAXCPU - 1.
246 */
247struct thread;
248
249struct cpu_group *cpu_topo(void);
250void	cpu_mp_announce(void);
251int	cpu_mp_probe(void);
252void	cpu_mp_setmaxid(void);
253void	cpu_mp_start(void);
254
255void	forward_signal(struct thread *);
256int	restart_cpus(cpuset_t);
257int	stop_cpus(cpuset_t);
258int	stop_cpus_hard(cpuset_t);
259#if defined(__amd64__) || defined(__i386__)
260int	suspend_cpus(cpuset_t);
261int	resume_cpus(cpuset_t);
262#endif
263
264void	smp_rendezvous_action(void);
265extern	struct mtx smp_ipi_mtx;
266
267#endif /* SMP */
268
269int	quiesce_all_cpus(const char *, int);
270int	quiesce_cpus(cpuset_t, const char *, int);
271void	quiesce_all_critical(void);
272void	cpus_fence_seq_cst(void);
273void	smp_no_rendezvous_barrier(void *);
274void	smp_rendezvous(void (*)(void *),
275		       void (*)(void *),
276		       void (*)(void *),
277		       void *arg);
278void	smp_rendezvous_cpus(cpuset_t,
279		       void (*)(void *),
280		       void (*)(void *),
281		       void (*)(void *),
282		       void *arg);
283
284struct smp_rendezvous_cpus_retry_arg {
285	cpuset_t cpus;
286};
287void	smp_rendezvous_cpus_retry(cpuset_t,
288		       void (*)(void *),
289		       void (*)(void *),
290		       void (*)(void *),
291		       void (*)(void *, int),
292		       struct smp_rendezvous_cpus_retry_arg *);
293
294void	smp_rendezvous_cpus_done(struct smp_rendezvous_cpus_retry_arg *);
295
296#endif /* !LOCORE */
297#endif /* _KERNEL */
298#endif /* _SYS_SMP_H_ */
299