xref: /freebsd/sys/kern/subr_pcpu.c (revision 1f474190)
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_PCPU_ZONE zones, that are available for all kernel
135  * consumers. Right now 64 bit zone is used for counter(9)
136  * and int zone is used for mount point counters.
137  */
138 
139 uma_zone_t pcpu_zone_int;
140 uma_zone_t pcpu_zone_64;
141 
142 static void
143 pcpu_zones_startup(void)
144 {
145 
146 	pcpu_zone_int = uma_zcreate("int pcpu", sizeof(int),
147 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_PCPU);
148 	pcpu_zone_64 = uma_zcreate("64 pcpu", sizeof(uint64_t),
149 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_PCPU);
150 }
151 SYSINIT(pcpu_zones, SI_SUB_COUNTER, SI_ORDER_FIRST, pcpu_zones_startup, NULL);
152 
153 /*
154  * First-fit extent based allocator for allocating space in the per-cpu
155  * region reserved for modules.  This is only intended for use by the
156  * kernel linkers to place module linker sets.
157  */
158 void *
159 dpcpu_alloc(int size)
160 {
161 	struct dpcpu_free *df;
162 	void *s;
163 
164 	s = NULL;
165 	size = roundup2(size, sizeof(void *));
166 	sx_xlock(&dpcpu_lock);
167 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
168 		if (df->df_len < size)
169 			continue;
170 		if (df->df_len == size) {
171 			s = (void *)df->df_start;
172 			TAILQ_REMOVE(&dpcpu_head, df, df_link);
173 			free(df, M_PCPU);
174 			break;
175 		}
176 		s = (void *)df->df_start;
177 		df->df_len -= size;
178 		df->df_start = df->df_start + size;
179 		break;
180 	}
181 	sx_xunlock(&dpcpu_lock);
182 
183 	return (s);
184 }
185 
186 /*
187  * Free dynamic per-cpu space at module unload time.
188  */
189 void
190 dpcpu_free(void *s, int size)
191 {
192 	struct dpcpu_free *df;
193 	struct dpcpu_free *dn;
194 	uintptr_t start;
195 	uintptr_t end;
196 
197 	size = roundup2(size, sizeof(void *));
198 	start = (uintptr_t)s;
199 	end = start + size;
200 	/*
201 	 * Free a region of space and merge it with as many neighbors as
202 	 * possible.  Keeping the list sorted simplifies this operation.
203 	 */
204 	sx_xlock(&dpcpu_lock);
205 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
206 		if (df->df_start > end)
207 			break;
208 		/*
209 		 * If we expand at the end of an entry we may have to
210 		 * merge it with the one following it as well.
211 		 */
212 		if (df->df_start + df->df_len == start) {
213 			df->df_len += size;
214 			dn = TAILQ_NEXT(df, df_link);
215 			if (df->df_start + df->df_len == dn->df_start) {
216 				df->df_len += dn->df_len;
217 				TAILQ_REMOVE(&dpcpu_head, dn, df_link);
218 				free(dn, M_PCPU);
219 			}
220 			sx_xunlock(&dpcpu_lock);
221 			return;
222 		}
223 		if (df->df_start == end) {
224 			df->df_start = start;
225 			df->df_len += size;
226 			sx_xunlock(&dpcpu_lock);
227 			return;
228 		}
229 	}
230 	dn = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
231 	dn->df_start = start;
232 	dn->df_len = size;
233 	if (df)
234 		TAILQ_INSERT_BEFORE(df, dn, df_link);
235 	else
236 		TAILQ_INSERT_TAIL(&dpcpu_head, dn, df_link);
237 	sx_xunlock(&dpcpu_lock);
238 }
239 
240 /*
241  * Initialize the per-cpu storage from an updated linker-set region.
242  */
243 void
244 dpcpu_copy(void *s, int size)
245 {
246 #ifdef SMP
247 	uintptr_t dpcpu;
248 	int i;
249 
250 	CPU_FOREACH(i) {
251 		dpcpu = dpcpu_off[i];
252 		if (dpcpu == 0)
253 			continue;
254 		memcpy((void *)(dpcpu + (uintptr_t)s), s, size);
255 	}
256 #else
257 	memcpy((void *)(dpcpu_off[0] + (uintptr_t)s), s, size);
258 #endif
259 }
260 
261 /*
262  * Destroy a struct pcpu.
263  */
264 void
265 pcpu_destroy(struct pcpu *pcpu)
266 {
267 
268 	STAILQ_REMOVE(&cpuhead, pcpu, pcpu, pc_allcpu);
269 	cpuid_to_pcpu[pcpu->pc_cpuid] = NULL;
270 	dpcpu_off[pcpu->pc_cpuid] = 0;
271 }
272 
273 /*
274  * Locate a struct pcpu by cpu id.
275  */
276 struct pcpu *
277 pcpu_find(u_int cpuid)
278 {
279 
280 	return (cpuid_to_pcpu[cpuid]);
281 }
282 
283 int
284 sysctl_dpcpu_quad(SYSCTL_HANDLER_ARGS)
285 {
286 	uintptr_t dpcpu;
287 	int64_t count;
288 	int i;
289 
290 	count = 0;
291 	CPU_FOREACH(i) {
292 		dpcpu = dpcpu_off[i];
293 		if (dpcpu == 0)
294 			continue;
295 		count += *(int64_t *)(dpcpu + (uintptr_t)arg1);
296 	}
297 	return (SYSCTL_OUT(req, &count, sizeof(count)));
298 }
299 
300 int
301 sysctl_dpcpu_long(SYSCTL_HANDLER_ARGS)
302 {
303 	uintptr_t dpcpu;
304 	long count;
305 	int i;
306 
307 	count = 0;
308 	CPU_FOREACH(i) {
309 		dpcpu = dpcpu_off[i];
310 		if (dpcpu == 0)
311 			continue;
312 		count += *(long *)(dpcpu + (uintptr_t)arg1);
313 	}
314 	return (SYSCTL_OUT(req, &count, sizeof(count)));
315 }
316 
317 int
318 sysctl_dpcpu_int(SYSCTL_HANDLER_ARGS)
319 {
320 	uintptr_t dpcpu;
321 	int count;
322 	int i;
323 
324 	count = 0;
325 	CPU_FOREACH(i) {
326 		dpcpu = dpcpu_off[i];
327 		if (dpcpu == 0)
328 			continue;
329 		count += *(int *)(dpcpu + (uintptr_t)arg1);
330 	}
331 	return (SYSCTL_OUT(req, &count, sizeof(count)));
332 }
333 
334 #ifdef DDB
335 DB_SHOW_COMMAND(dpcpu_off, db_show_dpcpu_off)
336 {
337 	int id;
338 
339 	CPU_FOREACH(id) {
340 		db_printf("dpcpu_off[%2d] = 0x%jx (+ DPCPU_START = %p)\n",
341 		    id, (uintmax_t)dpcpu_off[id],
342 		    (void *)(uintptr_t)(dpcpu_off[id] + DPCPU_START));
343 	}
344 }
345 
346 static void
347 show_pcpu(struct pcpu *pc)
348 {
349 	struct thread *td;
350 
351 	db_printf("cpuid        = %d\n", pc->pc_cpuid);
352 	db_printf("dynamic pcpu = %p\n", (void *)pc->pc_dynamic);
353 	db_printf("curthread    = ");
354 	td = pc->pc_curthread;
355 	if (td != NULL)
356 		db_printf("%p: pid %d tid %d critnest %d \"%s\"\n", td,
357 		    td->td_proc->p_pid, td->td_tid, td->td_critnest,
358 		    td->td_name);
359 	else
360 		db_printf("none\n");
361 	db_printf("curpcb       = %p\n", pc->pc_curpcb);
362 	db_printf("fpcurthread  = ");
363 	td = pc->pc_fpcurthread;
364 	if (td != NULL)
365 		db_printf("%p: pid %d \"%s\"\n", td, td->td_proc->p_pid,
366 		    td->td_name);
367 	else
368 		db_printf("none\n");
369 	db_printf("idlethread   = ");
370 	td = pc->pc_idlethread;
371 	if (td != NULL)
372 		db_printf("%p: tid %d \"%s\"\n", td, td->td_tid, td->td_name);
373 	else
374 		db_printf("none\n");
375 	db_show_mdpcpu(pc);
376 
377 #ifdef VIMAGE
378 	db_printf("curvnet      = %p\n", pc->pc_curthread->td_vnet);
379 #endif
380 
381 #ifdef WITNESS
382 	db_printf("spin locks held:\n");
383 	witness_list_locks(&pc->pc_spinlocks, db_printf);
384 #endif
385 }
386 
387 DB_SHOW_COMMAND(pcpu, db_show_pcpu)
388 {
389 	struct pcpu *pc;
390 	int id;
391 
392 	if (have_addr)
393 		id = ((addr >> 4) % 16) * 10 + (addr % 16);
394 	else
395 		id = PCPU_GET(cpuid);
396 	pc = pcpu_find(id);
397 	if (pc == NULL) {
398 		db_printf("CPU %d not found\n", id);
399 		return;
400 	}
401 	show_pcpu(pc);
402 }
403 
404 DB_SHOW_ALL_COMMAND(pcpu, db_show_cpu_all)
405 {
406 	struct pcpu *pc;
407 	int id;
408 
409 	db_printf("Current CPU: %d\n\n", PCPU_GET(cpuid));
410 	CPU_FOREACH(id) {
411 		pc = pcpu_find(id);
412 		if (pc != NULL) {
413 			show_pcpu(pc);
414 			db_printf("\n");
415 		}
416 	}
417 }
418 DB_SHOW_ALIAS(allpcpu, db_show_cpu_all);
419 #endif
420