xref: /freebsd/sys/kern/subr_pcpu.c (revision 9bd497b8354567454e075076d40c996e21bd6095)
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/linker_set.h>
57 #include <sys/lock.h>
58 #include <sys/malloc.h>
59 #include <sys/pcpu.h>
60 #include <sys/proc.h>
61 #include <sys/smp.h>
62 #include <sys/sx.h>
63 #include <ddb/ddb.h>
64 
65 MALLOC_DEFINE(M_PCPU, "Per-cpu", "Per-cpu resource accouting.");
66 
67 struct dpcpu_free {
68 	uintptr_t	df_start;
69 	int		df_len;
70 	TAILQ_ENTRY(dpcpu_free) df_link;
71 };
72 
73 static DPCPU_DEFINE(char, modspace[DPCPU_MODMIN]);
74 static TAILQ_HEAD(, dpcpu_free) dpcpu_head = TAILQ_HEAD_INITIALIZER(dpcpu_head);
75 static struct sx dpcpu_lock;
76 uintptr_t dpcpu_off[MAXCPU];
77 struct pcpu *cpuid_to_pcpu[MAXCPU];
78 struct cpuhead cpuhead = SLIST_HEAD_INITIALIZER(cpuhead);
79 
80 /*
81  * Initialize the MI portions of a struct pcpu.
82  */
83 void
84 pcpu_init(struct pcpu *pcpu, int cpuid, size_t size)
85 {
86 
87 	bzero(pcpu, size);
88 	KASSERT(cpuid >= 0 && cpuid < MAXCPU,
89 	    ("pcpu_init: invalid cpuid %d", cpuid));
90 	pcpu->pc_cpuid = cpuid;
91 	pcpu->pc_cpumask = 1 << cpuid;
92 	cpuid_to_pcpu[cpuid] = pcpu;
93 	SLIST_INSERT_HEAD(&cpuhead, pcpu, pc_allcpu);
94 	cpu_pcpu_init(pcpu, cpuid, size);
95 	pcpu->pc_rm_queue.rmq_next = &pcpu->pc_rm_queue;
96 	pcpu->pc_rm_queue.rmq_prev = &pcpu->pc_rm_queue;
97 #ifdef KTR
98 	snprintf(pcpu->pc_name, sizeof(pcpu->pc_name), "CPU %d", cpuid);
99 #endif
100 }
101 
102 void
103 dpcpu_init(void *dpcpu, int cpuid)
104 {
105 	struct pcpu *pcpu;
106 
107 	pcpu = pcpu_find(cpuid);
108 	pcpu->pc_dynamic = (uintptr_t)dpcpu - DPCPU_START;
109 
110 	/*
111 	 * Initialize defaults from our linker section.
112 	 */
113 	memcpy(dpcpu, (void *)DPCPU_START, DPCPU_BYTES);
114 
115 	/*
116 	 * Place it in the global pcpu offset array.
117 	 */
118 	dpcpu_off[cpuid] = pcpu->pc_dynamic;
119 }
120 
121 static void
122 dpcpu_startup(void *dummy __unused)
123 {
124 	struct dpcpu_free *df;
125 
126 	df = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
127 	df->df_start = (uintptr_t)&DPCPU_NAME(modspace);
128 	df->df_len = DPCPU_MODSIZE;
129 	TAILQ_INSERT_HEAD(&dpcpu_head, df, df_link);
130 	sx_init(&dpcpu_lock, "dpcpu alloc lock");
131 }
132 SYSINIT(dpcpu, SI_SUB_KLD, SI_ORDER_FIRST, dpcpu_startup, 0);
133 
134 /*
135  * First-fit extent based allocator for allocating space in the per-cpu
136  * region reserved for modules.  This is only intended for use by the
137  * kernel linkers to place module linker sets.
138  */
139 void *
140 dpcpu_alloc(int size)
141 {
142 	struct dpcpu_free *df;
143 	void *s;
144 
145 	s = NULL;
146 	size = roundup2(size, sizeof(void *));
147 	sx_xlock(&dpcpu_lock);
148 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
149 		if (df->df_len < size)
150 			continue;
151 		if (df->df_len == size) {
152 			s = (void *)df->df_start;
153 			TAILQ_REMOVE(&dpcpu_head, df, df_link);
154 			free(df, M_PCPU);
155 			break;
156 		}
157 		s = (void *)df->df_start;
158 		df->df_len -= size;
159 		df->df_start = df->df_start + size;
160 		break;
161 	}
162 	sx_xunlock(&dpcpu_lock);
163 
164 	return (s);
165 }
166 
167 /*
168  * Free dynamic per-cpu space at module unload time.
169  */
170 void
171 dpcpu_free(void *s, int size)
172 {
173 	struct dpcpu_free *df;
174 	struct dpcpu_free *dn;
175 	uintptr_t start;
176 	uintptr_t end;
177 
178 	size = roundup2(size, sizeof(void *));
179 	start = (uintptr_t)s;
180 	end = start + size;
181 	/*
182 	 * Free a region of space and merge it with as many neighbors as
183 	 * possible.  Keeping the list sorted simplifies this operation.
184 	 */
185 	sx_xlock(&dpcpu_lock);
186 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
187 		if (df->df_start > end)
188 			break;
189 		/*
190 		 * If we expand at the end of an entry we may have to
191 		 * merge it with the one following it as well.
192 		 */
193 		if (df->df_start + df->df_len == start) {
194 			df->df_len += size;
195 			dn = TAILQ_NEXT(df, df_link);
196 			if (df->df_start + df->df_len == dn->df_start) {
197 				df->df_len += dn->df_len;
198 				TAILQ_REMOVE(&dpcpu_head, dn, df_link);
199 				free(dn, M_PCPU);
200 			}
201 			sx_xunlock(&dpcpu_lock);
202 			return;
203 		}
204 		if (df->df_start == end) {
205 			df->df_start = start;
206 			df->df_len += size;
207 			sx_xunlock(&dpcpu_lock);
208 			return;
209 		}
210 	}
211 	dn = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
212 	dn->df_start = start;
213 	dn->df_len = size;
214 	if (df)
215 		TAILQ_INSERT_BEFORE(df, dn, df_link);
216 	else
217 		TAILQ_INSERT_TAIL(&dpcpu_head, dn, df_link);
218 	sx_xunlock(&dpcpu_lock);
219 }
220 
221 /*
222  * Initialize the per-cpu storage from an updated linker-set region.
223  */
224 void
225 dpcpu_copy(void *s, int size)
226 {
227 #ifdef SMP
228 	uintptr_t dpcpu;
229 	int i;
230 
231 	for (i = 0; i < mp_ncpus; ++i) {
232 		dpcpu = dpcpu_off[i];
233 		if (dpcpu == 0)
234 			continue;
235 		memcpy((void *)(dpcpu + (uintptr_t)s), s, size);
236 	}
237 #else
238 	memcpy((void *)(dpcpu_off[0] + (uintptr_t)s), s, size);
239 #endif
240 }
241 
242 /*
243  * Destroy a struct pcpu.
244  */
245 void
246 pcpu_destroy(struct pcpu *pcpu)
247 {
248 
249 	SLIST_REMOVE(&cpuhead, pcpu, pcpu, pc_allcpu);
250 	cpuid_to_pcpu[pcpu->pc_cpuid] = NULL;
251 	dpcpu_off[pcpu->pc_cpuid] = 0;
252 }
253 
254 /*
255  * Locate a struct pcpu by cpu id.
256  */
257 struct pcpu *
258 pcpu_find(u_int cpuid)
259 {
260 
261 	return (cpuid_to_pcpu[cpuid]);
262 }
263 
264 int
265 sysctl_dpcpu_quad(SYSCTL_HANDLER_ARGS)
266 {
267 	uintptr_t dpcpu;
268 	int64_t count;
269 	int i;
270 
271 	count = 0;
272 	for (i = 0; i < mp_ncpus; ++i) {
273 		dpcpu = dpcpu_off[i];
274 		if (dpcpu == 0)
275 			continue;
276 		count += *(int64_t *)(dpcpu + (uintptr_t)arg1);
277 	}
278 	return (SYSCTL_OUT(req, &count, sizeof(count)));
279 }
280 
281 int
282 sysctl_dpcpu_long(SYSCTL_HANDLER_ARGS)
283 {
284 	uintptr_t dpcpu;
285 	long count;
286 	int i;
287 
288 	count = 0;
289 	for (i = 0; i < mp_ncpus; ++i) {
290 		dpcpu = dpcpu_off[i];
291 		if (dpcpu == 0)
292 			continue;
293 		count += *(long *)(dpcpu + (uintptr_t)arg1);
294 	}
295 	return (SYSCTL_OUT(req, &count, sizeof(count)));
296 }
297 
298 int
299 sysctl_dpcpu_int(SYSCTL_HANDLER_ARGS)
300 {
301 	uintptr_t dpcpu;
302 	int count;
303 	int i;
304 
305 	count = 0;
306 	for (i = 0; i < mp_ncpus; ++i) {
307 		dpcpu = dpcpu_off[i];
308 		if (dpcpu == 0)
309 			continue;
310 		count += *(int *)(dpcpu + (uintptr_t)arg1);
311 	}
312 	return (SYSCTL_OUT(req, &count, sizeof(count)));
313 }
314 
315 #ifdef DDB
316 DB_SHOW_COMMAND(dpcpu_off, db_show_dpcpu_off)
317 {
318 	int id;
319 
320 	for (id = 0; id <= mp_maxid; id++) {
321 		if (CPU_ABSENT(id))
322 			continue;
323 		db_printf("dpcpu_off[%2d] = 0x%jx (+ DPCPU_START = %p)\n",
324 		    id, (uintmax_t)dpcpu_off[id],
325 		    (void *)(uintptr_t)(dpcpu_off[id] + DPCPU_START));
326 	}
327 }
328 
329 static void
330 show_pcpu(struct pcpu *pc)
331 {
332 	struct thread *td;
333 
334 	db_printf("cpuid        = %d\n", pc->pc_cpuid);
335 	db_printf("dynamic pcpu	= %p\n", (void *)pc->pc_dynamic);
336 	db_printf("curthread    = ");
337 	td = pc->pc_curthread;
338 	if (td != NULL)
339 		db_printf("%p: pid %d \"%s\"\n", td, td->td_proc->p_pid,
340 		    td->td_name);
341 	else
342 		db_printf("none\n");
343 	db_printf("curpcb       = %p\n", pc->pc_curpcb);
344 	db_printf("fpcurthread  = ");
345 	td = pc->pc_fpcurthread;
346 	if (td != NULL)
347 		db_printf("%p: pid %d \"%s\"\n", td, td->td_proc->p_pid,
348 		    td->td_name);
349 	else
350 		db_printf("none\n");
351 	db_printf("idlethread   = ");
352 	td = pc->pc_idlethread;
353 	if (td != NULL)
354 		db_printf("%p: pid %d \"%s\"\n", td, td->td_proc->p_pid,
355 		    td->td_name);
356 	else
357 		db_printf("none\n");
358 	db_show_mdpcpu(pc);
359 
360 #ifdef VIMAGE
361 	db_printf("curvnet      = %p\n", pc->pc_curthread->td_vnet);
362 #endif
363 
364 #ifdef WITNESS
365 	db_printf("spin locks held:\n");
366 	witness_list_locks(&pc->pc_spinlocks);
367 #endif
368 }
369 
370 DB_SHOW_COMMAND(pcpu, db_show_pcpu)
371 {
372 	struct pcpu *pc;
373 	int id;
374 
375 	if (have_addr)
376 		id = ((addr >> 4) % 16) * 10 + (addr % 16);
377 	else
378 		id = PCPU_GET(cpuid);
379 	pc = pcpu_find(id);
380 	if (pc == NULL) {
381 		db_printf("CPU %d not found\n", id);
382 		return;
383 	}
384 	show_pcpu(pc);
385 }
386 
387 DB_SHOW_ALL_COMMAND(pcpu, db_show_cpu_all)
388 {
389 	struct pcpu *pc;
390 	int id;
391 
392 	db_printf("Current CPU: %d\n\n", PCPU_GET(cpuid));
393 	for (id = 0; id <= mp_maxid; id++) {
394 		pc = pcpu_find(id);
395 		if (pc != NULL) {
396 			show_pcpu(pc);
397 			db_printf("\n");
398 		}
399 	}
400 }
401 DB_SHOW_ALIAS(allpcpu, db_show_cpu_all);
402 #endif
403