xref: /freebsd/sys/kern/subr_pcpu.c (revision 0b3105a37d7adcadcb720112fed4dc4e8040be99)
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/lock.h>
57 #include <sys/malloc.h>
58 #include <sys/pcpu.h>
59 #include <sys/proc.h>
60 #include <sys/smp.h>
61 #include <sys/sx.h>
62 #include <vm/uma.h>
63 #include <ddb/ddb.h>
64 
65 static 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 = STAILQ_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 	cpuid_to_pcpu[cpuid] = pcpu;
92 	STAILQ_INSERT_TAIL(&cpuhead, pcpu, pc_allcpu);
93 	cpu_pcpu_init(pcpu, cpuid, size);
94 	pcpu->pc_rm_queue.rmq_next = &pcpu->pc_rm_queue;
95 	pcpu->pc_rm_queue.rmq_prev = &pcpu->pc_rm_queue;
96 }
97 
98 void
99 dpcpu_init(void *dpcpu, int cpuid)
100 {
101 	struct pcpu *pcpu;
102 
103 	pcpu = pcpu_find(cpuid);
104 	pcpu->pc_dynamic = (uintptr_t)dpcpu - DPCPU_START;
105 
106 	/*
107 	 * Initialize defaults from our linker section.
108 	 */
109 	memcpy(dpcpu, (void *)DPCPU_START, DPCPU_BYTES);
110 
111 	/*
112 	 * Place it in the global pcpu offset array.
113 	 */
114 	dpcpu_off[cpuid] = pcpu->pc_dynamic;
115 }
116 
117 static void
118 dpcpu_startup(void *dummy __unused)
119 {
120 	struct dpcpu_free *df;
121 
122 	df = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
123 	df->df_start = (uintptr_t)&DPCPU_NAME(modspace);
124 	df->df_len = DPCPU_MODMIN;
125 	TAILQ_INSERT_HEAD(&dpcpu_head, df, df_link);
126 	sx_init(&dpcpu_lock, "dpcpu alloc lock");
127 }
128 SYSINIT(dpcpu, SI_SUB_KLD, SI_ORDER_FIRST, dpcpu_startup, 0);
129 
130 /*
131  * UMA_PCPU_ZONE zones, that are available for all kernel
132  * consumers. Right now 64 bit zone is used for counter(9)
133  * and pointer zone is used by flowtable.
134  */
135 
136 uma_zone_t pcpu_zone_64;
137 uma_zone_t pcpu_zone_ptr;
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 	if (sizeof(uint64_t) == sizeof(void *))
147 		pcpu_zone_ptr = pcpu_zone_64;
148 	else
149 		pcpu_zone_ptr = uma_zcreate("ptr pcpu", sizeof(void *),
150 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_PCPU);
151 }
152 SYSINIT(pcpu_zones, SI_SUB_KMEM, SI_ORDER_ANY, pcpu_zones_startup, NULL);
153 
154 /*
155  * First-fit extent based allocator for allocating space in the per-cpu
156  * region reserved for modules.  This is only intended for use by the
157  * kernel linkers to place module linker sets.
158  */
159 void *
160 dpcpu_alloc(int size)
161 {
162 	struct dpcpu_free *df;
163 	void *s;
164 
165 	s = NULL;
166 	size = roundup2(size, sizeof(void *));
167 	sx_xlock(&dpcpu_lock);
168 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
169 		if (df->df_len < size)
170 			continue;
171 		if (df->df_len == size) {
172 			s = (void *)df->df_start;
173 			TAILQ_REMOVE(&dpcpu_head, df, df_link);
174 			free(df, M_PCPU);
175 			break;
176 		}
177 		s = (void *)df->df_start;
178 		df->df_len -= size;
179 		df->df_start = df->df_start + size;
180 		break;
181 	}
182 	sx_xunlock(&dpcpu_lock);
183 
184 	return (s);
185 }
186 
187 /*
188  * Free dynamic per-cpu space at module unload time.
189  */
190 void
191 dpcpu_free(void *s, int size)
192 {
193 	struct dpcpu_free *df;
194 	struct dpcpu_free *dn;
195 	uintptr_t start;
196 	uintptr_t end;
197 
198 	size = roundup2(size, sizeof(void *));
199 	start = (uintptr_t)s;
200 	end = start + size;
201 	/*
202 	 * Free a region of space and merge it with as many neighbors as
203 	 * possible.  Keeping the list sorted simplifies this operation.
204 	 */
205 	sx_xlock(&dpcpu_lock);
206 	TAILQ_FOREACH(df, &dpcpu_head, df_link) {
207 		if (df->df_start > end)
208 			break;
209 		/*
210 		 * If we expand at the end of an entry we may have to
211 		 * merge it with the one following it as well.
212 		 */
213 		if (df->df_start + df->df_len == start) {
214 			df->df_len += size;
215 			dn = TAILQ_NEXT(df, df_link);
216 			if (df->df_start + df->df_len == dn->df_start) {
217 				df->df_len += dn->df_len;
218 				TAILQ_REMOVE(&dpcpu_head, dn, df_link);
219 				free(dn, M_PCPU);
220 			}
221 			sx_xunlock(&dpcpu_lock);
222 			return;
223 		}
224 		if (df->df_start == end) {
225 			df->df_start = start;
226 			df->df_len += size;
227 			sx_xunlock(&dpcpu_lock);
228 			return;
229 		}
230 	}
231 	dn = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO);
232 	dn->df_start = start;
233 	dn->df_len = size;
234 	if (df)
235 		TAILQ_INSERT_BEFORE(df, dn, df_link);
236 	else
237 		TAILQ_INSERT_TAIL(&dpcpu_head, dn, df_link);
238 	sx_xunlock(&dpcpu_lock);
239 }
240 
241 /*
242  * Initialize the per-cpu storage from an updated linker-set region.
243  */
244 void
245 dpcpu_copy(void *s, int size)
246 {
247 #ifdef SMP
248 	uintptr_t dpcpu;
249 	int i;
250 
251 	for (i = 0; i < mp_ncpus; ++i) {
252 		dpcpu = dpcpu_off[i];
253 		if (dpcpu == 0)
254 			continue;
255 		memcpy((void *)(dpcpu + (uintptr_t)s), s, size);
256 	}
257 #else
258 	memcpy((void *)(dpcpu_off[0] + (uintptr_t)s), s, size);
259 #endif
260 }
261 
262 /*
263  * Destroy a struct pcpu.
264  */
265 void
266 pcpu_destroy(struct pcpu *pcpu)
267 {
268 
269 	STAILQ_REMOVE(&cpuhead, pcpu, pcpu, pc_allcpu);
270 	cpuid_to_pcpu[pcpu->pc_cpuid] = NULL;
271 	dpcpu_off[pcpu->pc_cpuid] = 0;
272 }
273 
274 /*
275  * Locate a struct pcpu by cpu id.
276  */
277 struct pcpu *
278 pcpu_find(u_int cpuid)
279 {
280 
281 	return (cpuid_to_pcpu[cpuid]);
282 }
283 
284 int
285 sysctl_dpcpu_quad(SYSCTL_HANDLER_ARGS)
286 {
287 	uintptr_t dpcpu;
288 	int64_t count;
289 	int i;
290 
291 	count = 0;
292 	for (i = 0; i < mp_ncpus; ++i) {
293 		dpcpu = dpcpu_off[i];
294 		if (dpcpu == 0)
295 			continue;
296 		count += *(int64_t *)(dpcpu + (uintptr_t)arg1);
297 	}
298 	return (SYSCTL_OUT(req, &count, sizeof(count)));
299 }
300 
301 int
302 sysctl_dpcpu_long(SYSCTL_HANDLER_ARGS)
303 {
304 	uintptr_t dpcpu;
305 	long count;
306 	int i;
307 
308 	count = 0;
309 	for (i = 0; i < mp_ncpus; ++i) {
310 		dpcpu = dpcpu_off[i];
311 		if (dpcpu == 0)
312 			continue;
313 		count += *(long *)(dpcpu + (uintptr_t)arg1);
314 	}
315 	return (SYSCTL_OUT(req, &count, sizeof(count)));
316 }
317 
318 int
319 sysctl_dpcpu_int(SYSCTL_HANDLER_ARGS)
320 {
321 	uintptr_t dpcpu;
322 	int count;
323 	int i;
324 
325 	count = 0;
326 	for (i = 0; i < mp_ncpus; ++i) {
327 		dpcpu = dpcpu_off[i];
328 		if (dpcpu == 0)
329 			continue;
330 		count += *(int *)(dpcpu + (uintptr_t)arg1);
331 	}
332 	return (SYSCTL_OUT(req, &count, sizeof(count)));
333 }
334 
335 #ifdef DDB
336 DB_SHOW_COMMAND(dpcpu_off, db_show_dpcpu_off)
337 {
338 	int id;
339 
340 	CPU_FOREACH(id) {
341 		db_printf("dpcpu_off[%2d] = 0x%jx (+ DPCPU_START = %p)\n",
342 		    id, (uintmax_t)dpcpu_off[id],
343 		    (void *)(uintptr_t)(dpcpu_off[id] + DPCPU_START));
344 	}
345 }
346 
347 static void
348 show_pcpu(struct pcpu *pc)
349 {
350 	struct thread *td;
351 
352 	db_printf("cpuid        = %d\n", pc->pc_cpuid);
353 	db_printf("dynamic pcpu = %p\n", (void *)pc->pc_dynamic);
354 	db_printf("curthread    = ");
355 	td = pc->pc_curthread;
356 	if (td != NULL)
357 		db_printf("%p: pid %d \"%s\"\n", td, td->td_proc->p_pid,
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 	for (id = 0; id <= mp_maxid; 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