xref: /freebsd/sys/kern/subr_pcpu.c (revision ae41709ab46305df80f7f35bb478a3c8ebf22ebb)
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 	pcpu->pc_zpcpu_offset = zpcpu_offset_cpu(cpuid);
99 }
100 
101 void
102 dpcpu_init(void *dpcpu, int cpuid)
103 {
104 	struct pcpu *pcpu;
105 
106 	pcpu = pcpu_find(cpuid);
107 	pcpu->pc_dynamic = (uintptr_t)dpcpu - DPCPU_START;
108 
109 	/*
110 	 * Initialize defaults from our linker section.
111 	 */
112 	memcpy(dpcpu, (void *)DPCPU_START, DPCPU_BYTES);
113 
114 	/*
115 	 * Place it in the global pcpu offset array.
116 	 */
117 	dpcpu_off[cpuid] = pcpu->pc_dynamic;
118 }
119 
120 static void
121 dpcpu_startup(void *dummy __unused)
122 {
123 	struct dpcpu_free *df;
124 
125 	df = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
126 	df->df_start = (uintptr_t)&DPCPU_NAME(modspace);
127 	df->df_len = DPCPU_MODMIN;
128 	TAILQ_INSERT_HEAD(&dpcpu_head, df, df_link);
129 	sx_init(&dpcpu_lock, "dpcpu alloc lock");
130 }
131 SYSINIT(dpcpu, SI_SUB_KLD, SI_ORDER_FIRST, dpcpu_startup, NULL);
132 
133 /*
134  * UMA_ZONE_PCPU zones for general kernel use.
135  */
136 
137 uma_zone_t pcpu_zone_4;
138 uma_zone_t pcpu_zone_8;
139 
140 static void
141 pcpu_zones_startup(void)
142 {
143 
144 	pcpu_zone_4 = uma_zcreate("pcpu-4", sizeof(uint32_t),
145 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_PCPU);
146 	pcpu_zone_8 = uma_zcreate("pcpu-8", sizeof(uint64_t),
147 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_PCPU);
148 }
149 SYSINIT(pcpu_zones, SI_SUB_COUNTER, SI_ORDER_FIRST, pcpu_zones_startup, NULL);
150 
151 /*
152  * First-fit extent based allocator for allocating space in the per-cpu
153  * region reserved for modules.  This is only intended for use by the
154  * kernel linkers to place module linker sets.
155  */
156 void *
157 dpcpu_alloc(int size)
158 {
159 	struct dpcpu_free *df;
160 	void *s;
161 
162 	s = NULL;
163 	size = roundup2(size, sizeof(void *));
164 	sx_xlock(&dpcpu_lock);
165 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
166 		if (df->df_len < size)
167 			continue;
168 		if (df->df_len == size) {
169 			s = (void *)df->df_start;
170 			TAILQ_REMOVE(&dpcpu_head, df, df_link);
171 			free(df, M_PCPU);
172 			break;
173 		}
174 		s = (void *)df->df_start;
175 		df->df_len -= size;
176 		df->df_start = df->df_start + size;
177 		break;
178 	}
179 	sx_xunlock(&dpcpu_lock);
180 
181 	return (s);
182 }
183 
184 /*
185  * Free dynamic per-cpu space at module unload time.
186  */
187 void
188 dpcpu_free(void *s, int size)
189 {
190 	struct dpcpu_free *df;
191 	struct dpcpu_free *dn;
192 	uintptr_t start;
193 	uintptr_t end;
194 
195 	size = roundup2(size, sizeof(void *));
196 	start = (uintptr_t)s;
197 	end = start + size;
198 	/*
199 	 * Free a region of space and merge it with as many neighbors as
200 	 * possible.  Keeping the list sorted simplifies this operation.
201 	 */
202 	sx_xlock(&dpcpu_lock);
203 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
204 		if (df->df_start > end)
205 			break;
206 		/*
207 		 * If we expand at the end of an entry we may have to
208 		 * merge it with the one following it as well.
209 		 */
210 		if (df->df_start + df->df_len == start) {
211 			df->df_len += size;
212 			dn = TAILQ_NEXT(df, df_link);
213 			if (df->df_start + df->df_len == dn->df_start) {
214 				df->df_len += dn->df_len;
215 				TAILQ_REMOVE(&dpcpu_head, dn, df_link);
216 				free(dn, M_PCPU);
217 			}
218 			sx_xunlock(&dpcpu_lock);
219 			return;
220 		}
221 		if (df->df_start == end) {
222 			df->df_start = start;
223 			df->df_len += size;
224 			sx_xunlock(&dpcpu_lock);
225 			return;
226 		}
227 	}
228 	dn = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
229 	dn->df_start = start;
230 	dn->df_len = size;
231 	if (df)
232 		TAILQ_INSERT_BEFORE(df, dn, df_link);
233 	else
234 		TAILQ_INSERT_TAIL(&dpcpu_head, dn, df_link);
235 	sx_xunlock(&dpcpu_lock);
236 }
237 
238 /*
239  * Initialize the per-cpu storage from an updated linker-set region.
240  */
241 void
242 dpcpu_copy(void *s, int size)
243 {
244 #ifdef SMP
245 	uintptr_t dpcpu;
246 	int i;
247 
248 	CPU_FOREACH(i) {
249 		dpcpu = dpcpu_off[i];
250 		if (dpcpu == 0)
251 			continue;
252 		memcpy((void *)(dpcpu + (uintptr_t)s), s, size);
253 	}
254 #else
255 	memcpy((void *)(dpcpu_off[0] + (uintptr_t)s), s, size);
256 #endif
257 }
258 
259 /*
260  * Destroy a struct pcpu.
261  */
262 void
263 pcpu_destroy(struct pcpu *pcpu)
264 {
265 
266 	STAILQ_REMOVE(&cpuhead, pcpu, pcpu, pc_allcpu);
267 	cpuid_to_pcpu[pcpu->pc_cpuid] = NULL;
268 	dpcpu_off[pcpu->pc_cpuid] = 0;
269 }
270 
271 /*
272  * Locate a struct pcpu by cpu id.
273  */
274 struct pcpu *
275 pcpu_find(u_int cpuid)
276 {
277 
278 	return (cpuid_to_pcpu[cpuid]);
279 }
280 
281 int
282 sysctl_dpcpu_quad(SYSCTL_HANDLER_ARGS)
283 {
284 	uintptr_t dpcpu;
285 	int64_t count;
286 	int i;
287 
288 	count = 0;
289 	CPU_FOREACH(i) {
290 		dpcpu = dpcpu_off[i];
291 		if (dpcpu == 0)
292 			continue;
293 		count += *(int64_t *)(dpcpu + (uintptr_t)arg1);
294 	}
295 	return (SYSCTL_OUT(req, &count, sizeof(count)));
296 }
297 
298 int
299 sysctl_dpcpu_long(SYSCTL_HANDLER_ARGS)
300 {
301 	uintptr_t dpcpu;
302 	long count;
303 	int i;
304 
305 	count = 0;
306 	CPU_FOREACH(i) {
307 		dpcpu = dpcpu_off[i];
308 		if (dpcpu == 0)
309 			continue;
310 		count += *(long *)(dpcpu + (uintptr_t)arg1);
311 	}
312 	return (SYSCTL_OUT(req, &count, sizeof(count)));
313 }
314 
315 int
316 sysctl_dpcpu_int(SYSCTL_HANDLER_ARGS)
317 {
318 	uintptr_t dpcpu;
319 	int count;
320 	int i;
321 
322 	count = 0;
323 	CPU_FOREACH(i) {
324 		dpcpu = dpcpu_off[i];
325 		if (dpcpu == 0)
326 			continue;
327 		count += *(int *)(dpcpu + (uintptr_t)arg1);
328 	}
329 	return (SYSCTL_OUT(req, &count, sizeof(count)));
330 }
331 
332 #ifdef DDB
333 DB_SHOW_COMMAND(dpcpu_off, db_show_dpcpu_off)
334 {
335 	int id;
336 
337 	CPU_FOREACH(id) {
338 		db_printf("dpcpu_off[%2d] = 0x%jx (+ DPCPU_START = %p)\n",
339 		    id, (uintmax_t)dpcpu_off[id],
340 		    (void *)(uintptr_t)(dpcpu_off[id] + DPCPU_START));
341 	}
342 }
343 
344 static void
345 show_pcpu(struct pcpu *pc)
346 {
347 	struct thread *td;
348 
349 	db_printf("cpuid        = %d\n", pc->pc_cpuid);
350 	db_printf("dynamic pcpu = %p\n", (void *)pc->pc_dynamic);
351 	db_printf("curthread    = ");
352 	td = pc->pc_curthread;
353 	if (td != NULL)
354 		db_printf("%p: pid %d tid %d critnest %d \"%s\"\n", td,
355 		    td->td_proc->p_pid, td->td_tid, td->td_critnest,
356 		    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