xref: /titanic_50/usr/src/lib/libumem/common/vmem_sbrk.c (revision 3c112a2b34403220c06c3e2fcac403358cfba168)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * The structure of the sbrk backend:
31  *
32  * +-----------+
33  * | sbrk_top  |
34  * +-----------+
35  *      | (vmem_sbrk_alloc(), vmem_free())
36  *      |
37  * +-----------+
38  * | sbrk_heap |
39  * +-----------+
40  *   | | ... |  (vmem_alloc(), vmem_free())
41  * <other arenas>
42  *
43  * The sbrk_top arena holds all controlled memory.  vmem_sbrk_alloc() handles
44  * allocations from it, including growing the heap when we run low.
45  *
46  * Growing the heap is complicated by the fact that we have to extend the
47  * sbrk_top arena (using _vmem_extend_alloc()), and that can fail.  Since
48  * other threads may be actively allocating, we can't return the memory.
49  *
50  * Instead, we put it on a doubly-linked list, sbrk_fails, which we search
51  * before calling sbrk().
52  */
53 
54 #include <errno.h>
55 #include <limits.h>
56 #include <sys/sysmacros.h>
57 #include <sys/mman.h>
58 #include <unistd.h>
59 
60 #include "vmem_base.h"
61 
62 #include "misc.h"
63 
64 size_t vmem_sbrk_pagesize = 0; /* the preferred page size of the heap */
65 
66 #define	VMEM_SBRK_MINALLOC	(64 * 1024)
67 size_t vmem_sbrk_minalloc = VMEM_SBRK_MINALLOC; /* minimum allocation */
68 
69 static size_t real_pagesize;
70 static vmem_t *sbrk_heap;
71 
72 typedef struct sbrk_fail {
73 	struct sbrk_fail *sf_next;
74 	struct sbrk_fail *sf_prev;
75 	void *sf_base;			/* == the sbrk_fail's address */
76 	size_t sf_size;			/* the size of this buffer */
77 } sbrk_fail_t;
78 
79 static sbrk_fail_t sbrk_fails = {
80 	&sbrk_fails,
81 	&sbrk_fails,
82 	NULL,
83 	0
84 };
85 
86 static mutex_t sbrk_faillock = DEFAULTMUTEX;
87 
88 /*
89  * Try to extend src with [pos, pos + size).
90  *
91  * If it fails, add the block to the sbrk_fails list.
92  */
93 static void *
94 vmem_sbrk_extend_alloc(vmem_t *src, void *pos, size_t size, size_t alloc,
95     int vmflags)
96 {
97 	sbrk_fail_t *fnext, *fprev, *fp;
98 	void *ret;
99 
100 	ret = _vmem_extend_alloc(src, pos, size, alloc, vmflags);
101 	if (ret != NULL)
102 		return (ret);
103 
104 	fp = (sbrk_fail_t *)pos;
105 
106 	ASSERT(sizeof (sbrk_fail_t) <= size);
107 
108 	fp->sf_base = pos;
109 	fp->sf_size = size;
110 
111 	(void) mutex_lock(&sbrk_faillock);
112 	fp->sf_next = fnext = &sbrk_fails;
113 	fp->sf_prev = fprev = sbrk_fails.sf_prev;
114 	fnext->sf_prev = fp;
115 	fprev->sf_next = fp;
116 	(void) mutex_unlock(&sbrk_faillock);
117 
118 	return (NULL);
119 }
120 
121 /*
122  * Try to add at least size bytes to src, using the sbrk_fails list
123  */
124 static void *
125 vmem_sbrk_tryfail(vmem_t *src, size_t size, int vmflags)
126 {
127 	sbrk_fail_t *fp;
128 
129 	(void) mutex_lock(&sbrk_faillock);
130 	for (fp = sbrk_fails.sf_next; fp != &sbrk_fails; fp = fp->sf_next) {
131 		if (fp->sf_size >= size) {
132 			fp->sf_next->sf_prev = fp->sf_prev;
133 			fp->sf_prev->sf_next = fp->sf_next;
134 			fp->sf_next = fp->sf_prev = NULL;
135 			break;
136 		}
137 	}
138 	(void) mutex_unlock(&sbrk_faillock);
139 
140 	if (fp != &sbrk_fails) {
141 		ASSERT(fp->sf_base == (void *)fp);
142 		return (vmem_sbrk_extend_alloc(src, fp, fp->sf_size, size,
143 		    vmflags));
144 	}
145 	/*
146 	 * nothing of the right size on the freelist
147 	 */
148 	return (NULL);
149 }
150 
151 static void *
152 vmem_sbrk_alloc(vmem_t *src, size_t size, int vmflags)
153 {
154 	extern void *_sbrk_grow_aligned(size_t min_size, size_t low_align,
155 	    size_t high_align, size_t *actual_size);
156 
157 	void *ret;
158 	void *buf;
159 	size_t buf_size;
160 
161 	int old_errno = errno;
162 
163 	ret = vmem_alloc(src, size, VM_NOSLEEP);
164 	if (ret != NULL) {
165 		errno = old_errno;
166 		return (ret);
167 	}
168 
169 	/*
170 	 * The allocation failed.  We need to grow the heap.
171 	 *
172 	 * First, try to use any buffers which failed earlier.
173 	 */
174 	if (sbrk_fails.sf_next != &sbrk_fails &&
175 	    (ret = vmem_sbrk_tryfail(src, size, vmflags)) != NULL)
176 		return (ret);
177 
178 	buf_size = MAX(size, vmem_sbrk_minalloc);
179 
180 	/*
181 	 * buf_size gets overwritten with the actual allocated size
182 	 */
183 	buf = _sbrk_grow_aligned(buf_size, real_pagesize, vmem_sbrk_pagesize,
184 	    &buf_size);
185 
186 	if (buf != MAP_FAILED) {
187 		ret = vmem_sbrk_extend_alloc(src, buf, buf_size, size, vmflags);
188 		if (ret != NULL) {
189 			errno = old_errno;
190 			return (ret);
191 		}
192 	}
193 
194 	/*
195 	 * Growing the heap failed. The vmem_alloc() above called umem_reap().
196 	 */
197 	ASSERT((vmflags & VM_NOSLEEP) == VM_NOSLEEP);
198 
199 	errno = old_errno;
200 	return (NULL);
201 }
202 
203 /*
204  * fork1() support
205  */
206 void
207 vmem_sbrk_lockup(void)
208 {
209 	(void) mutex_lock(&sbrk_faillock);
210 }
211 
212 void
213 vmem_sbrk_release(void)
214 {
215 	(void) mutex_unlock(&sbrk_faillock);
216 }
217 
218 vmem_t *
219 vmem_sbrk_arena(vmem_alloc_t **a_out, vmem_free_t **f_out)
220 {
221 	if (sbrk_heap == NULL) {
222 		size_t heap_size;
223 
224 		real_pagesize = sysconf(_SC_PAGESIZE);
225 
226 		heap_size = vmem_sbrk_pagesize;
227 
228 		if (issetugid()) {
229 			heap_size = 0;
230 		} else if (heap_size != 0 && !ISP2(heap_size)) {
231 			heap_size = 0;
232 			log_message("ignoring bad pagesize: 0x%p\n", heap_size);
233 		}
234 		if (heap_size <= real_pagesize) {
235 			heap_size = real_pagesize;
236 		} else {
237 			struct memcntl_mha mha;
238 			mha.mha_cmd = MHA_MAPSIZE_BSSBRK;
239 			mha.mha_flags = 0;
240 			mha.mha_pagesize = heap_size;
241 
242 			if (memcntl(NULL, 0, MC_HAT_ADVISE, (char *)&mha, 0, 0)
243 			    == -1) {
244 				log_message("unable to set MAPSIZE_BSSBRK to "
245 				    "0x%p\n", heap_size);
246 				heap_size = real_pagesize;
247 			}
248 		}
249 		vmem_sbrk_pagesize = heap_size;
250 
251 		/* validate vmem_sbrk_minalloc */
252 		if (vmem_sbrk_minalloc < VMEM_SBRK_MINALLOC)
253 			vmem_sbrk_minalloc = VMEM_SBRK_MINALLOC;
254 		vmem_sbrk_minalloc = P2ROUNDUP(vmem_sbrk_minalloc, heap_size);
255 
256 		sbrk_heap = vmem_init("sbrk_top", real_pagesize,
257 		    vmem_sbrk_alloc, vmem_free,
258 		    "sbrk_heap", NULL, 0, real_pagesize,
259 		    vmem_alloc, vmem_free);
260 	}
261 
262 	if (a_out != NULL)
263 		*a_out = vmem_alloc;
264 	if (f_out != NULL)
265 		*f_out = vmem_free;
266 
267 	return (sbrk_heap);
268 }
269