xref: /freebsd/sys/kern/subr_pcpu.c (revision cc698b490001646c7174d14af6500400be9bd4ff)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 2001 Wind River Systems, Inc.
5  * All rights reserved.
6  * Written by: John Baldwin <jhb@FreeBSD.org>
7  *
8  * Copyright (c) 2009 Jeffrey Roberson <jeff@freebsd.org>
9  * All rights reserved.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the author nor the names of any co-contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35 
36 /*
37  * This module provides MI support for per-cpu data.
38  *
39  * Each architecture determines the mapping of logical CPU IDs to physical
40  * CPUs.  The requirements of this mapping are as follows:
41  *  - Logical CPU IDs must reside in the range 0 ... MAXCPU - 1.
42  *  - The mapping is not required to be dense.  That is, there may be
43  *    gaps in the mappings.
44  *  - The platform sets the value of MAXCPU in <machine/param.h>.
45  *  - It is suggested, but not required, that in the non-SMP case, the
46  *    platform define MAXCPU to be 1 and define the logical ID of the
47  *    sole CPU as 0.
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 #include "opt_ddb.h"
54 
55 #include <sys/param.h>
56 #include <sys/systm.h>
57 #include <sys/sysctl.h>
58 #include <sys/lock.h>
59 #include <sys/malloc.h>
60 #include <sys/pcpu.h>
61 #include <sys/proc.h>
62 #include <sys/smp.h>
63 #include <sys/sx.h>
64 #include <vm/uma.h>
65 #include <ddb/ddb.h>
66 
67 static MALLOC_DEFINE(M_PCPU, "Per-cpu", "Per-cpu resource accouting.");
68 
69 struct dpcpu_free {
70 	uintptr_t	df_start;
71 	int		df_len;
72 	TAILQ_ENTRY(dpcpu_free) df_link;
73 };
74 
75 DPCPU_DEFINE_STATIC(char, modspace[DPCPU_MODMIN] __aligned(__alignof(void *)));
76 static TAILQ_HEAD(, dpcpu_free) dpcpu_head = TAILQ_HEAD_INITIALIZER(dpcpu_head);
77 static struct sx dpcpu_lock;
78 uintptr_t dpcpu_off[MAXCPU];
79 struct pcpu *cpuid_to_pcpu[MAXCPU];
80 struct cpuhead cpuhead = STAILQ_HEAD_INITIALIZER(cpuhead);
81 
82 /*
83  * Initialize the MI portions of a struct pcpu.
84  */
85 void
86 pcpu_init(struct pcpu *pcpu, int cpuid, size_t size)
87 {
88 
89 	bzero(pcpu, size);
90 	KASSERT(cpuid >= 0 && cpuid < MAXCPU,
91 	    ("pcpu_init: invalid cpuid %d", cpuid));
92 	pcpu->pc_cpuid = cpuid;
93 	cpuid_to_pcpu[cpuid] = pcpu;
94 	STAILQ_INSERT_TAIL(&cpuhead, pcpu, pc_allcpu);
95 	cpu_pcpu_init(pcpu, cpuid, size);
96 	pcpu->pc_rm_queue.rmq_next = &pcpu->pc_rm_queue;
97 	pcpu->pc_rm_queue.rmq_prev = &pcpu->pc_rm_queue;
98 }
99 
100 void
101 dpcpu_init(void *dpcpu, int cpuid)
102 {
103 	struct pcpu *pcpu;
104 
105 	pcpu = pcpu_find(cpuid);
106 	pcpu->pc_dynamic = (uintptr_t)dpcpu - DPCPU_START;
107 
108 	/*
109 	 * Initialize defaults from our linker section.
110 	 */
111 	memcpy(dpcpu, (void *)DPCPU_START, DPCPU_BYTES);
112 
113 	/*
114 	 * Place it in the global pcpu offset array.
115 	 */
116 	dpcpu_off[cpuid] = pcpu->pc_dynamic;
117 }
118 
119 static void
120 dpcpu_startup(void *dummy __unused)
121 {
122 	struct dpcpu_free *df;
123 
124 	df = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
125 	df->df_start = (uintptr_t)&DPCPU_NAME(modspace);
126 	df->df_len = DPCPU_MODMIN;
127 	TAILQ_INSERT_HEAD(&dpcpu_head, df, df_link);
128 	sx_init(&dpcpu_lock, "dpcpu alloc lock");
129 }
130 SYSINIT(dpcpu, SI_SUB_KLD, SI_ORDER_FIRST, dpcpu_startup, NULL);
131 
132 /*
133  * UMA_PCPU_ZONE zones, that are available for all kernel
134  * consumers. Right now 64 bit zone is used for counter(9)
135  * and int zone is used for mount point counters.
136  */
137 
138 uma_zone_t pcpu_zone_int;
139 uma_zone_t pcpu_zone_64;
140 
141 static void
142 pcpu_zones_startup(void)
143 {
144 
145 	pcpu_zone_int = uma_zcreate("int pcpu", sizeof(int),
146 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_PCPU);
147 	pcpu_zone_64 = uma_zcreate("64 pcpu", sizeof(uint64_t),
148 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_PCPU);
149 }
150 SYSINIT(pcpu_zones, SI_SUB_VM, SI_ORDER_ANY, pcpu_zones_startup, NULL);
151 
152 /*
153  * First-fit extent based allocator for allocating space in the per-cpu
154  * region reserved for modules.  This is only intended for use by the
155  * kernel linkers to place module linker sets.
156  */
157 void *
158 dpcpu_alloc(int size)
159 {
160 	struct dpcpu_free *df;
161 	void *s;
162 
163 	s = NULL;
164 	size = roundup2(size, sizeof(void *));
165 	sx_xlock(&dpcpu_lock);
166 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
167 		if (df->df_len < size)
168 			continue;
169 		if (df->df_len == size) {
170 			s = (void *)df->df_start;
171 			TAILQ_REMOVE(&dpcpu_head, df, df_link);
172 			free(df, M_PCPU);
173 			break;
174 		}
175 		s = (void *)df->df_start;
176 		df->df_len -= size;
177 		df->df_start = df->df_start + size;
178 		break;
179 	}
180 	sx_xunlock(&dpcpu_lock);
181 
182 	return (s);
183 }
184 
185 /*
186  * Free dynamic per-cpu space at module unload time.
187  */
188 void
189 dpcpu_free(void *s, int size)
190 {
191 	struct dpcpu_free *df;
192 	struct dpcpu_free *dn;
193 	uintptr_t start;
194 	uintptr_t end;
195 
196 	size = roundup2(size, sizeof(void *));
197 	start = (uintptr_t)s;
198 	end = start + size;
199 	/*
200 	 * Free a region of space and merge it with as many neighbors as
201 	 * possible.  Keeping the list sorted simplifies this operation.
202 	 */
203 	sx_xlock(&dpcpu_lock);
204 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
205 		if (df->df_start > end)
206 			break;
207 		/*
208 		 * If we expand at the end of an entry we may have to
209 		 * merge it with the one following it as well.
210 		 */
211 		if (df->df_start + df->df_len == start) {
212 			df->df_len += size;
213 			dn = TAILQ_NEXT(df, df_link);
214 			if (df->df_start + df->df_len == dn->df_start) {
215 				df->df_len += dn->df_len;
216 				TAILQ_REMOVE(&dpcpu_head, dn, df_link);
217 				free(dn, M_PCPU);
218 			}
219 			sx_xunlock(&dpcpu_lock);
220 			return;
221 		}
222 		if (df->df_start == end) {
223 			df->df_start = start;
224 			df->df_len += size;
225 			sx_xunlock(&dpcpu_lock);
226 			return;
227 		}
228 	}
229 	dn = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
230 	dn->df_start = start;
231 	dn->df_len = size;
232 	if (df)
233 		TAILQ_INSERT_BEFORE(df, dn, df_link);
234 	else
235 		TAILQ_INSERT_TAIL(&dpcpu_head, dn, df_link);
236 	sx_xunlock(&dpcpu_lock);
237 }
238 
239 /*
240  * Initialize the per-cpu storage from an updated linker-set region.
241  */
242 void
243 dpcpu_copy(void *s, int size)
244 {
245 #ifdef SMP
246 	uintptr_t dpcpu;
247 	int i;
248 
249 	CPU_FOREACH(i) {
250 		dpcpu = dpcpu_off[i];
251 		if (dpcpu == 0)
252 			continue;
253 		memcpy((void *)(dpcpu + (uintptr_t)s), s, size);
254 	}
255 #else
256 	memcpy((void *)(dpcpu_off[0] + (uintptr_t)s), s, size);
257 #endif
258 }
259 
260 /*
261  * Destroy a struct pcpu.
262  */
263 void
264 pcpu_destroy(struct pcpu *pcpu)
265 {
266 
267 	STAILQ_REMOVE(&cpuhead, pcpu, pcpu, pc_allcpu);
268 	cpuid_to_pcpu[pcpu->pc_cpuid] = NULL;
269 	dpcpu_off[pcpu->pc_cpuid] = 0;
270 }
271 
272 /*
273  * Locate a struct pcpu by cpu id.
274  */
275 struct pcpu *
276 pcpu_find(u_int cpuid)
277 {
278 
279 	return (cpuid_to_pcpu[cpuid]);
280 }
281 
282 int
283 sysctl_dpcpu_quad(SYSCTL_HANDLER_ARGS)
284 {
285 	uintptr_t dpcpu;
286 	int64_t count;
287 	int i;
288 
289 	count = 0;
290 	CPU_FOREACH(i) {
291 		dpcpu = dpcpu_off[i];
292 		if (dpcpu == 0)
293 			continue;
294 		count += *(int64_t *)(dpcpu + (uintptr_t)arg1);
295 	}
296 	return (SYSCTL_OUT(req, &count, sizeof(count)));
297 }
298 
299 int
300 sysctl_dpcpu_long(SYSCTL_HANDLER_ARGS)
301 {
302 	uintptr_t dpcpu;
303 	long count;
304 	int i;
305 
306 	count = 0;
307 	CPU_FOREACH(i) {
308 		dpcpu = dpcpu_off[i];
309 		if (dpcpu == 0)
310 			continue;
311 		count += *(long *)(dpcpu + (uintptr_t)arg1);
312 	}
313 	return (SYSCTL_OUT(req, &count, sizeof(count)));
314 }
315 
316 int
317 sysctl_dpcpu_int(SYSCTL_HANDLER_ARGS)
318 {
319 	uintptr_t dpcpu;
320 	int count;
321 	int i;
322 
323 	count = 0;
324 	CPU_FOREACH(i) {
325 		dpcpu = dpcpu_off[i];
326 		if (dpcpu == 0)
327 			continue;
328 		count += *(int *)(dpcpu + (uintptr_t)arg1);
329 	}
330 	return (SYSCTL_OUT(req, &count, sizeof(count)));
331 }
332 
333 #ifdef DDB
334 DB_SHOW_COMMAND(dpcpu_off, db_show_dpcpu_off)
335 {
336 	int id;
337 
338 	CPU_FOREACH(id) {
339 		db_printf("dpcpu_off[%2d] = 0x%jx (+ DPCPU_START = %p)\n",
340 		    id, (uintmax_t)dpcpu_off[id],
341 		    (void *)(uintptr_t)(dpcpu_off[id] + DPCPU_START));
342 	}
343 }
344 
345 static void
346 show_pcpu(struct pcpu *pc)
347 {
348 	struct thread *td;
349 
350 	db_printf("cpuid        = %d\n", pc->pc_cpuid);
351 	db_printf("dynamic pcpu = %p\n", (void *)pc->pc_dynamic);
352 	db_printf("curthread    = ");
353 	td = pc->pc_curthread;
354 	if (td != NULL)
355 		db_printf("%p: pid %d tid %d \"%s\"\n", td, td->td_proc->p_pid,
356 		    td->td_tid, td->td_name);
357 	else
358 		db_printf("none\n");
359 	db_printf("curpcb       = %p\n", pc->pc_curpcb);
360 	db_printf("fpcurthread  = ");
361 	td = pc->pc_fpcurthread;
362 	if (td != NULL)
363 		db_printf("%p: pid %d \"%s\"\n", td, td->td_proc->p_pid,
364 		    td->td_name);
365 	else
366 		db_printf("none\n");
367 	db_printf("idlethread   = ");
368 	td = pc->pc_idlethread;
369 	if (td != NULL)
370 		db_printf("%p: tid %d \"%s\"\n", td, td->td_tid, td->td_name);
371 	else
372 		db_printf("none\n");
373 	db_show_mdpcpu(pc);
374 
375 #ifdef VIMAGE
376 	db_printf("curvnet      = %p\n", pc->pc_curthread->td_vnet);
377 #endif
378 
379 #ifdef WITNESS
380 	db_printf("spin locks held:\n");
381 	witness_list_locks(&pc->pc_spinlocks, db_printf);
382 #endif
383 }
384 
385 DB_SHOW_COMMAND(pcpu, db_show_pcpu)
386 {
387 	struct pcpu *pc;
388 	int id;
389 
390 	if (have_addr)
391 		id = ((addr >> 4) % 16) * 10 + (addr % 16);
392 	else
393 		id = PCPU_GET(cpuid);
394 	pc = pcpu_find(id);
395 	if (pc == NULL) {
396 		db_printf("CPU %d not found\n", id);
397 		return;
398 	}
399 	show_pcpu(pc);
400 }
401 
402 DB_SHOW_ALL_COMMAND(pcpu, db_show_cpu_all)
403 {
404 	struct pcpu *pc;
405 	int id;
406 
407 	db_printf("Current CPU: %d\n\n", PCPU_GET(cpuid));
408 	CPU_FOREACH(id) {
409 		pc = pcpu_find(id);
410 		if (pc != NULL) {
411 			show_pcpu(pc);
412 			db_printf("\n");
413 		}
414 	}
415 }
416 DB_SHOW_ALIAS(allpcpu, db_show_cpu_all);
417 #endif
418