xref: /freebsd/sys/kern/subr_pcpu.c (revision f6a3b357e9be4c6423c85eff9a847163a0d307c8)
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  */
136 
137 uma_zone_t pcpu_zone_64;
138 
139 static void
140 pcpu_zones_startup(void)
141 {
142 
143 	pcpu_zone_64 = uma_zcreate("64 pcpu", sizeof(uint64_t),
144 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_PCPU);
145 }
146 SYSINIT(pcpu_zones, SI_SUB_VM, SI_ORDER_ANY, pcpu_zones_startup, NULL);
147 
148 /*
149  * First-fit extent based allocator for allocating space in the per-cpu
150  * region reserved for modules.  This is only intended for use by the
151  * kernel linkers to place module linker sets.
152  */
153 void *
154 dpcpu_alloc(int size)
155 {
156 	struct dpcpu_free *df;
157 	void *s;
158 
159 	s = NULL;
160 	size = roundup2(size, sizeof(void *));
161 	sx_xlock(&dpcpu_lock);
162 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
163 		if (df->df_len < size)
164 			continue;
165 		if (df->df_len == size) {
166 			s = (void *)df->df_start;
167 			TAILQ_REMOVE(&dpcpu_head, df, df_link);
168 			free(df, M_PCPU);
169 			break;
170 		}
171 		s = (void *)df->df_start;
172 		df->df_len -= size;
173 		df->df_start = df->df_start + size;
174 		break;
175 	}
176 	sx_xunlock(&dpcpu_lock);
177 
178 	return (s);
179 }
180 
181 /*
182  * Free dynamic per-cpu space at module unload time.
183  */
184 void
185 dpcpu_free(void *s, int size)
186 {
187 	struct dpcpu_free *df;
188 	struct dpcpu_free *dn;
189 	uintptr_t start;
190 	uintptr_t end;
191 
192 	size = roundup2(size, sizeof(void *));
193 	start = (uintptr_t)s;
194 	end = start + size;
195 	/*
196 	 * Free a region of space and merge it with as many neighbors as
197 	 * possible.  Keeping the list sorted simplifies this operation.
198 	 */
199 	sx_xlock(&dpcpu_lock);
200 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
201 		if (df->df_start > end)
202 			break;
203 		/*
204 		 * If we expand at the end of an entry we may have to
205 		 * merge it with the one following it as well.
206 		 */
207 		if (df->df_start + df->df_len == start) {
208 			df->df_len += size;
209 			dn = TAILQ_NEXT(df, df_link);
210 			if (df->df_start + df->df_len == dn->df_start) {
211 				df->df_len += dn->df_len;
212 				TAILQ_REMOVE(&dpcpu_head, dn, df_link);
213 				free(dn, M_PCPU);
214 			}
215 			sx_xunlock(&dpcpu_lock);
216 			return;
217 		}
218 		if (df->df_start == end) {
219 			df->df_start = start;
220 			df->df_len += size;
221 			sx_xunlock(&dpcpu_lock);
222 			return;
223 		}
224 	}
225 	dn = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
226 	dn->df_start = start;
227 	dn->df_len = size;
228 	if (df)
229 		TAILQ_INSERT_BEFORE(df, dn, df_link);
230 	else
231 		TAILQ_INSERT_TAIL(&dpcpu_head, dn, df_link);
232 	sx_xunlock(&dpcpu_lock);
233 }
234 
235 /*
236  * Initialize the per-cpu storage from an updated linker-set region.
237  */
238 void
239 dpcpu_copy(void *s, int size)
240 {
241 #ifdef SMP
242 	uintptr_t dpcpu;
243 	int i;
244 
245 	CPU_FOREACH(i) {
246 		dpcpu = dpcpu_off[i];
247 		if (dpcpu == 0)
248 			continue;
249 		memcpy((void *)(dpcpu + (uintptr_t)s), s, size);
250 	}
251 #else
252 	memcpy((void *)(dpcpu_off[0] + (uintptr_t)s), s, size);
253 #endif
254 }
255 
256 /*
257  * Destroy a struct pcpu.
258  */
259 void
260 pcpu_destroy(struct pcpu *pcpu)
261 {
262 
263 	STAILQ_REMOVE(&cpuhead, pcpu, pcpu, pc_allcpu);
264 	cpuid_to_pcpu[pcpu->pc_cpuid] = NULL;
265 	dpcpu_off[pcpu->pc_cpuid] = 0;
266 }
267 
268 /*
269  * Locate a struct pcpu by cpu id.
270  */
271 struct pcpu *
272 pcpu_find(u_int cpuid)
273 {
274 
275 	return (cpuid_to_pcpu[cpuid]);
276 }
277 
278 int
279 sysctl_dpcpu_quad(SYSCTL_HANDLER_ARGS)
280 {
281 	uintptr_t dpcpu;
282 	int64_t count;
283 	int i;
284 
285 	count = 0;
286 	CPU_FOREACH(i) {
287 		dpcpu = dpcpu_off[i];
288 		if (dpcpu == 0)
289 			continue;
290 		count += *(int64_t *)(dpcpu + (uintptr_t)arg1);
291 	}
292 	return (SYSCTL_OUT(req, &count, sizeof(count)));
293 }
294 
295 int
296 sysctl_dpcpu_long(SYSCTL_HANDLER_ARGS)
297 {
298 	uintptr_t dpcpu;
299 	long count;
300 	int i;
301 
302 	count = 0;
303 	CPU_FOREACH(i) {
304 		dpcpu = dpcpu_off[i];
305 		if (dpcpu == 0)
306 			continue;
307 		count += *(long *)(dpcpu + (uintptr_t)arg1);
308 	}
309 	return (SYSCTL_OUT(req, &count, sizeof(count)));
310 }
311 
312 int
313 sysctl_dpcpu_int(SYSCTL_HANDLER_ARGS)
314 {
315 	uintptr_t dpcpu;
316 	int count;
317 	int i;
318 
319 	count = 0;
320 	CPU_FOREACH(i) {
321 		dpcpu = dpcpu_off[i];
322 		if (dpcpu == 0)
323 			continue;
324 		count += *(int *)(dpcpu + (uintptr_t)arg1);
325 	}
326 	return (SYSCTL_OUT(req, &count, sizeof(count)));
327 }
328 
329 #ifdef DDB
330 DB_SHOW_COMMAND(dpcpu_off, db_show_dpcpu_off)
331 {
332 	int id;
333 
334 	CPU_FOREACH(id) {
335 		db_printf("dpcpu_off[%2d] = 0x%jx (+ DPCPU_START = %p)\n",
336 		    id, (uintmax_t)dpcpu_off[id],
337 		    (void *)(uintptr_t)(dpcpu_off[id] + DPCPU_START));
338 	}
339 }
340 
341 static void
342 show_pcpu(struct pcpu *pc)
343 {
344 	struct thread *td;
345 
346 	db_printf("cpuid        = %d\n", pc->pc_cpuid);
347 	db_printf("dynamic pcpu = %p\n", (void *)pc->pc_dynamic);
348 	db_printf("curthread    = ");
349 	td = pc->pc_curthread;
350 	if (td != NULL)
351 		db_printf("%p: pid %d tid %d \"%s\"\n", td, td->td_proc->p_pid,
352 		    td->td_tid, td->td_name);
353 	else
354 		db_printf("none\n");
355 	db_printf("curpcb       = %p\n", pc->pc_curpcb);
356 	db_printf("fpcurthread  = ");
357 	td = pc->pc_fpcurthread;
358 	if (td != NULL)
359 		db_printf("%p: pid %d \"%s\"\n", td, td->td_proc->p_pid,
360 		    td->td_name);
361 	else
362 		db_printf("none\n");
363 	db_printf("idlethread   = ");
364 	td = pc->pc_idlethread;
365 	if (td != NULL)
366 		db_printf("%p: tid %d \"%s\"\n", td, td->td_tid, td->td_name);
367 	else
368 		db_printf("none\n");
369 	db_show_mdpcpu(pc);
370 
371 #ifdef VIMAGE
372 	db_printf("curvnet      = %p\n", pc->pc_curthread->td_vnet);
373 #endif
374 
375 #ifdef WITNESS
376 	db_printf("spin locks held:\n");
377 	witness_list_locks(&pc->pc_spinlocks, db_printf);
378 #endif
379 }
380 
381 DB_SHOW_COMMAND(pcpu, db_show_pcpu)
382 {
383 	struct pcpu *pc;
384 	int id;
385 
386 	if (have_addr)
387 		id = ((addr >> 4) % 16) * 10 + (addr % 16);
388 	else
389 		id = PCPU_GET(cpuid);
390 	pc = pcpu_find(id);
391 	if (pc == NULL) {
392 		db_printf("CPU %d not found\n", id);
393 		return;
394 	}
395 	show_pcpu(pc);
396 }
397 
398 DB_SHOW_ALL_COMMAND(pcpu, db_show_cpu_all)
399 {
400 	struct pcpu *pc;
401 	int id;
402 
403 	db_printf("Current CPU: %d\n\n", PCPU_GET(cpuid));
404 	CPU_FOREACH(id) {
405 		pc = pcpu_find(id);
406 		if (pc != NULL) {
407 			show_pcpu(pc);
408 			db_printf("\n");
409 		}
410 	}
411 }
412 DB_SHOW_ALIAS(allpcpu, db_show_cpu_all);
413 #endif
414