summaryrefslogtreecommitdiff
path: root/src/chunk_dss.c
blob: ee3f83888e0fa2e2fd4f11c1610a3b62ddc7baaf (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
#define	JEMALLOC_CHUNK_DSS_C_
#include "jemalloc/internal/jemalloc_internal.h"
/******************************************************************************/
/* Data. */

const char	*dss_prec_names[] = {
	"disabled",
	"primary",
	"secondary",
	"N/A"
};

/*
 * Current dss precedence default, used when creating new arenas.  NB: This is
 * stored as unsigned rather than dss_prec_t because in principle there's no
 * guarantee that sizeof(dss_prec_t) is the same as sizeof(unsigned), and we use
 * atomic operations to synchronize the setting.
 */
static unsigned		dss_prec_default = (unsigned)DSS_PREC_DEFAULT;

/* Base address of the DSS. */
static void		*dss_base;
/* Atomic boolean indicating whether the DSS is exhausted. */
static unsigned		dss_exhausted;
/* Atomic current upper limit on DSS addresses. */
static void		*dss_max;

/******************************************************************************/

static void *
chunk_dss_sbrk(intptr_t increment)
{

#ifdef JEMALLOC_DSS
	return (sbrk(increment));
#else
	not_implemented();
	return (NULL);
#endif
}

dss_prec_t
chunk_dss_prec_get(void)
{
	dss_prec_t ret;

	if (!have_dss)
		return (dss_prec_disabled);
	ret = (dss_prec_t)atomic_read_u(&dss_prec_default);
	return (ret);
}

bool
chunk_dss_prec_set(dss_prec_t dss_prec)
{

	if (!have_dss)
		return (dss_prec != dss_prec_disabled);
	atomic_write_u(&dss_prec_default, (unsigned)dss_prec);
	return (false);
}

static void *
chunk_dss_max_update(void *new_addr)
{
	void *max_cur;
	spin_t spinner;

	/*
	 * Get the current end of the DSS as max_cur and assure that dss_max is
	 * up to date.
	 */
	spin_init(&spinner);
	while (true) {
		void *max_prev = atomic_read_p(&dss_max);

		max_cur = chunk_dss_sbrk(0);
		if ((uintptr_t)max_prev > (uintptr_t)max_cur) {
			/*
			 * Another thread optimistically updated dss_max.  Wait
			 * for it to finish.
			 */
			spin_adaptive(&spinner);
			continue;
		}
		if (!atomic_cas_p(&dss_max, max_prev, max_cur))
			break;
	}
	/* Fixed new_addr can only be supported if it is at the edge of DSS. */
	if (new_addr != NULL && max_cur != new_addr)
		return (NULL);

	return (max_cur);
}

void *
chunk_alloc_dss(tsdn_t *tsdn, arena_t *arena, void *new_addr, size_t size,
    size_t alignment, bool *zero, bool *commit)
{
	cassert(have_dss);
	assert(size > 0 && (size & chunksize_mask) == 0);
	assert(alignment > 0 && (alignment & chunksize_mask) == 0);

	/*
	 * sbrk() uses a signed increment argument, so take care not to
	 * interpret a huge allocation request as a negative increment.
	 */
	if ((intptr_t)size < 0)
		return (NULL);

	if (!atomic_read_u(&dss_exhausted)) {
		/*
		 * The loop is necessary to recover from races with other
		 * threads that are using the DSS for something other than
		 * malloc.
		 */
		while (true) {
			void *ret, *cpad, *max_cur, *dss_next, *dss_prev;
			size_t gap_size, cpad_size;
			intptr_t incr;

			max_cur = chunk_dss_max_update(new_addr);
			if (max_cur == NULL)
				goto label_oom;

			/*
			 * Calculate how much padding is necessary to
			 * chunk-align the end of the DSS.
			 */
			gap_size = (chunksize - CHUNK_ADDR2OFFSET(dss_max)) &
			    chunksize_mask;
			/*
			 * Compute how much chunk-aligned pad space (if any) is
			 * necessary to satisfy alignment.  This space can be
			 * recycled for later use.
			 */
			cpad = (void *)((uintptr_t)dss_max + gap_size);
			ret = (void *)ALIGNMENT_CEILING((uintptr_t)dss_max,
			    alignment);
			cpad_size = (uintptr_t)ret - (uintptr_t)cpad;
			dss_next = (void *)((uintptr_t)ret + size);
			if ((uintptr_t)ret < (uintptr_t)dss_max ||
			    (uintptr_t)dss_next < (uintptr_t)dss_max)
				goto label_oom; /* Wrap-around. */
			incr = gap_size + cpad_size + size;

			/*
			 * Optimistically update dss_max, and roll back below if
			 * sbrk() fails.  No other thread will try to extend the
			 * DSS while dss_max is greater than the current DSS
			 * max reported by sbrk(0).
			 */
			if (atomic_cas_p(&dss_max, max_cur, dss_next))
				continue;

			/* Try to allocate. */
			dss_prev = chunk_dss_sbrk(incr);
			if (dss_prev == max_cur) {
				/* Success. */
				if (cpad_size != 0) {
					chunk_hooks_t chunk_hooks =
					    CHUNK_HOOKS_INITIALIZER;
					chunk_dalloc_wrapper(tsdn, arena,
					    &chunk_hooks, cpad, cpad_size,
					    arena_extent_sn_next(arena), false,
					    true);
				}
				if (*zero) {
					JEMALLOC_VALGRIND_MAKE_MEM_UNDEFINED(
					    ret, size);
					memset(ret, 0, size);
				}
				if (!*commit)
					*commit = pages_decommit(ret, size);
				return (ret);
			}

			/*
			 * Failure, whether due to OOM or a race with a raw
			 * sbrk() call from outside the allocator.  Try to roll
			 * back optimistic dss_max update; if rollback fails,
			 * it's due to another caller of this function having
			 * succeeded since this invocation started, in which
			 * case rollback is not necessary.
			 */
			atomic_cas_p(&dss_max, dss_next, max_cur);
			if (dss_prev == (void *)-1) {
				/* OOM. */
				atomic_write_u(&dss_exhausted, (unsigned)true);
				goto label_oom;
			}
		}
	}
label_oom:
	return (NULL);
}

static bool
chunk_in_dss_helper(void *chunk, void *max)
{

	return ((uintptr_t)chunk >= (uintptr_t)dss_base && (uintptr_t)chunk <
	    (uintptr_t)max);
}

bool
chunk_in_dss(void *chunk)
{

	cassert(have_dss);

	return (chunk_in_dss_helper(chunk, atomic_read_p(&dss_max)));
}

bool
chunk_dss_mergeable(void *chunk_a, void *chunk_b)
{
	void *max;

	cassert(have_dss);

	max = atomic_read_p(&dss_max);
	return (chunk_in_dss_helper(chunk_a, max) ==
	    chunk_in_dss_helper(chunk_b, max));
}

void
chunk_dss_boot(void)
{

	cassert(have_dss);

	dss_base = chunk_dss_sbrk(0);
	dss_exhausted = (unsigned)(dss_base == (void *)-1);
	dss_max = dss_base;
}

/******************************************************************************/