xref: /freebsd/sys/kern/subr_pcpu.c (revision 8d20be1e22095c27faf8fe8b2f0d089739cc742e)
1 /*-
2  * Copyright (c) 2001 Wind River Systems, Inc.
3  * All rights reserved.
4  * Written by: John Baldwin <jhb@FreeBSD.org>
5  *
6  * Copyright (c) 2009 Jeffrey Roberson <jeff@freebsd.org>
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 4. Neither the name of the author nor the names of any co-contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * This module provides MI support for per-cpu data.
36  *
37  * Each architecture determines the mapping of logical CPU IDs to physical
38  * CPUs.  The requirements of this mapping are as follows:
39  *  - Logical CPU IDs must reside in the range 0 ... MAXCPU - 1.
40  *  - The mapping is not required to be dense.  That is, there may be
41  *    gaps in the mappings.
42  *  - The platform sets the value of MAXCPU in <machine/param.h>.
43  *  - It is suggested, but not required, that in the non-SMP case, the
44  *    platform define MAXCPU to be 1 and define the logical ID of the
45  *    sole CPU as 0.
46  */
47 
48 #include <sys/cdefs.h>
49 __FBSDID("$FreeBSD$");
50 
51 #include "opt_ddb.h"
52 
53 #include <sys/param.h>
54 #include <sys/systm.h>
55 #include <sys/sysctl.h>
56 #include <sys/lock.h>
57 #include <sys/malloc.h>
58 #include <sys/pcpu.h>
59 #include <sys/proc.h>
60 #include <sys/smp.h>
61 #include <sys/sx.h>
62 #include <ddb/ddb.h>
63 
64 static MALLOC_DEFINE(M_PCPU, "Per-cpu", "Per-cpu resource accouting.");
65 
66 struct dpcpu_free {
67 	uintptr_t	df_start;
68 	int		df_len;
69 	TAILQ_ENTRY(dpcpu_free) df_link;
70 };
71 
72 static DPCPU_DEFINE(char, modspace[DPCPU_MODMIN]);
73 static TAILQ_HEAD(, dpcpu_free) dpcpu_head = TAILQ_HEAD_INITIALIZER(dpcpu_head);
74 static struct sx dpcpu_lock;
75 uintptr_t dpcpu_off[MAXCPU];
76 struct pcpu *cpuid_to_pcpu[MAXCPU];
77 struct cpuhead cpuhead = STAILQ_HEAD_INITIALIZER(cpuhead);
78 
79 /*
80  * Initialize the MI portions of a struct pcpu.
81  */
82 void
83 pcpu_init(struct pcpu *pcpu, int cpuid, size_t size)
84 {
85 
86 	bzero(pcpu, size);
87 	KASSERT(cpuid >= 0 && cpuid < MAXCPU,
88 	    ("pcpu_init: invalid cpuid %d", cpuid));
89 	pcpu->pc_cpuid = cpuid;
90 	cpuid_to_pcpu[cpuid] = pcpu;
91 	STAILQ_INSERT_TAIL(&cpuhead, pcpu, pc_allcpu);
92 	cpu_pcpu_init(pcpu, cpuid, size);
93 	pcpu->pc_rm_queue.rmq_next = &pcpu->pc_rm_queue;
94 	pcpu->pc_rm_queue.rmq_prev = &pcpu->pc_rm_queue;
95 }
96 
97 void
98 dpcpu_init(void *dpcpu, int cpuid)
99 {
100 	struct pcpu *pcpu;
101 
102 	pcpu = pcpu_find(cpuid);
103 	pcpu->pc_dynamic = (uintptr_t)dpcpu - DPCPU_START;
104 
105 	/*
106 	 * Initialize defaults from our linker section.
107 	 */
108 	memcpy(dpcpu, (void *)DPCPU_START, DPCPU_BYTES);
109 
110 	/*
111 	 * Place it in the global pcpu offset array.
112 	 */
113 	dpcpu_off[cpuid] = pcpu->pc_dynamic;
114 }
115 
116 static void
117 dpcpu_startup(void *dummy __unused)
118 {
119 	struct dpcpu_free *df;
120 
121 	df = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
122 	df->df_start = (uintptr_t)&DPCPU_NAME(modspace);
123 	df->df_len = DPCPU_MODMIN;
124 	TAILQ_INSERT_HEAD(&dpcpu_head, df, df_link);
125 	sx_init(&dpcpu_lock, "dpcpu alloc lock");
126 }
127 SYSINIT(dpcpu, SI_SUB_KLD, SI_ORDER_FIRST, dpcpu_startup, 0);
128 
129 /*
130  * First-fit extent based allocator for allocating space in the per-cpu
131  * region reserved for modules.  This is only intended for use by the
132  * kernel linkers to place module linker sets.
133  */
134 void *
135 dpcpu_alloc(int size)
136 {
137 	struct dpcpu_free *df;
138 	void *s;
139 
140 	s = NULL;
141 	size = roundup2(size, sizeof(void *));
142 	sx_xlock(&dpcpu_lock);
143 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
144 		if (df->df_len < size)
145 			continue;
146 		if (df->df_len == size) {
147 			s = (void *)df->df_start;
148 			TAILQ_REMOVE(&dpcpu_head, df, df_link);
149 			free(df, M_PCPU);
150 			break;
151 		}
152 		s = (void *)df->df_start;
153 		df->df_len -= size;
154 		df->df_start = df->df_start + size;
155 		break;
156 	}
157 	sx_xunlock(&dpcpu_lock);
158 
159 	return (s);
160 }
161 
162 /*
163  * Free dynamic per-cpu space at module unload time.
164  */
165 void
166 dpcpu_free(void *s, int size)
167 {
168 	struct dpcpu_free *df;
169 	struct dpcpu_free *dn;
170 	uintptr_t start;
171 	uintptr_t end;
172 
173 	size = roundup2(size, sizeof(void *));
174 	start = (uintptr_t)s;
175 	end = start + size;
176 	/*
177 	 * Free a region of space and merge it with as many neighbors as
178 	 * possible.  Keeping the list sorted simplifies this operation.
179 	 */
180 	sx_xlock(&dpcpu_lock);
181 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
182 		if (df->df_start > end)
183 			break;
184 		/*
185 		 * If we expand at the end of an entry we may have to
186 		 * merge it with the one following it as well.
187 		 */
188 		if (df->df_start + df->df_len == start) {
189 			df->df_len += size;
190 			dn = TAILQ_NEXT(df, df_link);
191 			if (df->df_start + df->df_len == dn->df_start) {
192 				df->df_len += dn->df_len;
193 				TAILQ_REMOVE(&dpcpu_head, dn, df_link);
194 				free(dn, M_PCPU);
195 			}
196 			sx_xunlock(&dpcpu_lock);
197 			return;
198 		}
199 		if (df->df_start == end) {
200 			df->df_start = start;
201 			df->df_len += size;
202 			sx_xunlock(&dpcpu_lock);
203 			return;
204 		}
205 	}
206 	dn = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
207 	dn->df_start = start;
208 	dn->df_len = size;
209 	if (df)
210 		TAILQ_INSERT_BEFORE(df, dn, df_link);
211 	else
212 		TAILQ_INSERT_TAIL(&dpcpu_head, dn, df_link);
213 	sx_xunlock(&dpcpu_lock);
214 }
215 
216 /*
217  * Initialize the per-cpu storage from an updated linker-set region.
218  */
219 void
220 dpcpu_copy(void *s, int size)
221 {
222 #ifdef SMP
223 	uintptr_t dpcpu;
224 	int i;
225 
226 	for (i = 0; i < mp_ncpus; ++i) {
227 		dpcpu = dpcpu_off[i];
228 		if (dpcpu == 0)
229 			continue;
230 		memcpy((void *)(dpcpu + (uintptr_t)s), s, size);
231 	}
232 #else
233 	memcpy((void *)(dpcpu_off[0] + (uintptr_t)s), s, size);
234 #endif
235 }
236 
237 /*
238  * Destroy a struct pcpu.
239  */
240 void
241 pcpu_destroy(struct pcpu *pcpu)
242 {
243 
244 	STAILQ_REMOVE(&cpuhead, pcpu, pcpu, pc_allcpu);
245 	cpuid_to_pcpu[pcpu->pc_cpuid] = NULL;
246 	dpcpu_off[pcpu->pc_cpuid] = 0;
247 }
248 
249 /*
250  * Locate a struct pcpu by cpu id.
251  */
252 struct pcpu *
253 pcpu_find(u_int cpuid)
254 {
255 
256 	return (cpuid_to_pcpu[cpuid]);
257 }
258 
259 int
260 sysctl_dpcpu_quad(SYSCTL_HANDLER_ARGS)
261 {
262 	uintptr_t dpcpu;
263 	int64_t count;
264 	int i;
265 
266 	count = 0;
267 	for (i = 0; i < mp_ncpus; ++i) {
268 		dpcpu = dpcpu_off[i];
269 		if (dpcpu == 0)
270 			continue;
271 		count += *(int64_t *)(dpcpu + (uintptr_t)arg1);
272 	}
273 	return (SYSCTL_OUT(req, &count, sizeof(count)));
274 }
275 
276 int
277 sysctl_dpcpu_long(SYSCTL_HANDLER_ARGS)
278 {
279 	uintptr_t dpcpu;
280 	long count;
281 	int i;
282 
283 	count = 0;
284 	for (i = 0; i < mp_ncpus; ++i) {
285 		dpcpu = dpcpu_off[i];
286 		if (dpcpu == 0)
287 			continue;
288 		count += *(long *)(dpcpu + (uintptr_t)arg1);
289 	}
290 	return (SYSCTL_OUT(req, &count, sizeof(count)));
291 }
292 
293 int
294 sysctl_dpcpu_int(SYSCTL_HANDLER_ARGS)
295 {
296 	uintptr_t dpcpu;
297 	int count;
298 	int i;
299 
300 	count = 0;
301 	for (i = 0; i < mp_ncpus; ++i) {
302 		dpcpu = dpcpu_off[i];
303 		if (dpcpu == 0)
304 			continue;
305 		count += *(int *)(dpcpu + (uintptr_t)arg1);
306 	}
307 	return (SYSCTL_OUT(req, &count, sizeof(count)));
308 }
309 
310 #ifdef DDB
311 DB_SHOW_COMMAND(dpcpu_off, db_show_dpcpu_off)
312 {
313 	int id;
314 
315 	CPU_FOREACH(id) {
316 		db_printf("dpcpu_off[%2d] = 0x%jx (+ DPCPU_START = %p)\n",
317 		    id, (uintmax_t)dpcpu_off[id],
318 		    (void *)(uintptr_t)(dpcpu_off[id] + DPCPU_START));
319 	}
320 }
321 
322 static void
323 show_pcpu(struct pcpu *pc)
324 {
325 	struct thread *td;
326 
327 	db_printf("cpuid        = %d\n", pc->pc_cpuid);
328 	db_printf("dynamic pcpu = %p\n", (void *)pc->pc_dynamic);
329 	db_printf("curthread    = ");
330 	td = pc->pc_curthread;
331 	if (td != NULL)
332 		db_printf("%p: pid %d \"%s\"\n", td, td->td_proc->p_pid,
333 		    td->td_name);
334 	else
335 		db_printf("none\n");
336 	db_printf("curpcb       = %p\n", pc->pc_curpcb);
337 	db_printf("fpcurthread  = ");
338 	td = pc->pc_fpcurthread;
339 	if (td != NULL)
340 		db_printf("%p: pid %d \"%s\"\n", td, td->td_proc->p_pid,
341 		    td->td_name);
342 	else
343 		db_printf("none\n");
344 	db_printf("idlethread   = ");
345 	td = pc->pc_idlethread;
346 	if (td != NULL)
347 		db_printf("%p: tid %d \"%s\"\n", td, td->td_tid, td->td_name);
348 	else
349 		db_printf("none\n");
350 	db_show_mdpcpu(pc);
351 
352 #ifdef VIMAGE
353 	db_printf("curvnet      = %p\n", pc->pc_curthread->td_vnet);
354 #endif
355 
356 #ifdef WITNESS
357 	db_printf("spin locks held:\n");
358 	witness_list_locks(&pc->pc_spinlocks, db_printf);
359 #endif
360 }
361 
362 DB_SHOW_COMMAND(pcpu, db_show_pcpu)
363 {
364 	struct pcpu *pc;
365 	int id;
366 
367 	if (have_addr)
368 		id = ((addr >> 4) % 16) * 10 + (addr % 16);
369 	else
370 		id = PCPU_GET(cpuid);
371 	pc = pcpu_find(id);
372 	if (pc == NULL) {
373 		db_printf("CPU %d not found\n", id);
374 		return;
375 	}
376 	show_pcpu(pc);
377 }
378 
379 DB_SHOW_ALL_COMMAND(pcpu, db_show_cpu_all)
380 {
381 	struct pcpu *pc;
382 	int id;
383 
384 	db_printf("Current CPU: %d\n\n", PCPU_GET(cpuid));
385 	for (id = 0; id <= mp_maxid; id++) {
386 		pc = pcpu_find(id);
387 		if (pc != NULL) {
388 			show_pcpu(pc);
389 			db_printf("\n");
390 		}
391 	}
392 }
393 DB_SHOW_ALIAS(allpcpu, db_show_cpu_all);
394 #endif
395