xref: /freebsd/contrib/jemalloc/src/eset.c (revision c43cad87172039ccf38172129c79755ea79e6102)
1 #include "jemalloc/internal/jemalloc_preamble.h"
2 #include "jemalloc/internal/jemalloc_internal_includes.h"
3 
4 #include "jemalloc/internal/eset.h"
5 
6 #define ESET_NPSIZES (SC_NPSIZES + 1)
7 
8 static void
9 eset_bin_init(eset_bin_t *bin) {
10 	edata_heap_new(&bin->heap);
11 	/*
12 	 * heap_min doesn't need initialization; it gets filled in when the bin
13 	 * goes from non-empty to empty.
14 	 */
15 }
16 
17 static void
18 eset_bin_stats_init(eset_bin_stats_t *bin_stats) {
19 	atomic_store_zu(&bin_stats->nextents, 0, ATOMIC_RELAXED);
20 	atomic_store_zu(&bin_stats->nbytes, 0, ATOMIC_RELAXED);
21 }
22 
23 void
24 eset_init(eset_t *eset, extent_state_t state) {
25 	for (unsigned i = 0; i < ESET_NPSIZES; i++) {
26 		eset_bin_init(&eset->bins[i]);
27 		eset_bin_stats_init(&eset->bin_stats[i]);
28 	}
29 	fb_init(eset->bitmap, ESET_NPSIZES);
30 	edata_list_inactive_init(&eset->lru);
31 	eset->state = state;
32 }
33 
34 size_t
35 eset_npages_get(eset_t *eset) {
36 	return atomic_load_zu(&eset->npages, ATOMIC_RELAXED);
37 }
38 
39 size_t
40 eset_nextents_get(eset_t *eset, pszind_t pind) {
41 	return atomic_load_zu(&eset->bin_stats[pind].nextents, ATOMIC_RELAXED);
42 }
43 
44 size_t
45 eset_nbytes_get(eset_t *eset, pszind_t pind) {
46 	return atomic_load_zu(&eset->bin_stats[pind].nbytes, ATOMIC_RELAXED);
47 }
48 
49 static void
50 eset_stats_add(eset_t *eset, pszind_t pind, size_t sz) {
51 	size_t cur = atomic_load_zu(&eset->bin_stats[pind].nextents,
52 	    ATOMIC_RELAXED);
53 	atomic_store_zu(&eset->bin_stats[pind].nextents, cur + 1,
54 	    ATOMIC_RELAXED);
55 	cur = atomic_load_zu(&eset->bin_stats[pind].nbytes, ATOMIC_RELAXED);
56 	atomic_store_zu(&eset->bin_stats[pind].nbytes, cur + sz,
57 	    ATOMIC_RELAXED);
58 }
59 
60 static void
61 eset_stats_sub(eset_t *eset, pszind_t pind, size_t sz) {
62 	size_t cur = atomic_load_zu(&eset->bin_stats[pind].nextents,
63 	    ATOMIC_RELAXED);
64 	atomic_store_zu(&eset->bin_stats[pind].nextents, cur - 1,
65 	    ATOMIC_RELAXED);
66 	cur = atomic_load_zu(&eset->bin_stats[pind].nbytes, ATOMIC_RELAXED);
67 	atomic_store_zu(&eset->bin_stats[pind].nbytes, cur - sz,
68 	    ATOMIC_RELAXED);
69 }
70 
71 void
72 eset_insert(eset_t *eset, edata_t *edata) {
73 	assert(edata_state_get(edata) == eset->state);
74 
75 	size_t size = edata_size_get(edata);
76 	size_t psz = sz_psz_quantize_floor(size);
77 	pszind_t pind = sz_psz2ind(psz);
78 
79 	edata_cmp_summary_t edata_cmp_summary = edata_cmp_summary_get(edata);
80 	if (edata_heap_empty(&eset->bins[pind].heap)) {
81 		fb_set(eset->bitmap, ESET_NPSIZES, (size_t)pind);
82 		/* Only element is automatically the min element. */
83 		eset->bins[pind].heap_min = edata_cmp_summary;
84 	} else {
85 		/*
86 		 * There's already a min element; update the summary if we're
87 		 * about to insert a lower one.
88 		 */
89 		if (edata_cmp_summary_comp(edata_cmp_summary,
90 		    eset->bins[pind].heap_min) < 0) {
91 			eset->bins[pind].heap_min = edata_cmp_summary;
92 		}
93 	}
94 	edata_heap_insert(&eset->bins[pind].heap, edata);
95 
96 	if (config_stats) {
97 		eset_stats_add(eset, pind, size);
98 	}
99 
100 	edata_list_inactive_append(&eset->lru, edata);
101 	size_t npages = size >> LG_PAGE;
102 	/*
103 	 * All modifications to npages hold the mutex (as asserted above), so we
104 	 * don't need an atomic fetch-add; we can get by with a load followed by
105 	 * a store.
106 	 */
107 	size_t cur_eset_npages =
108 	    atomic_load_zu(&eset->npages, ATOMIC_RELAXED);
109 	atomic_store_zu(&eset->npages, cur_eset_npages + npages,
110 	    ATOMIC_RELAXED);
111 }
112 
113 void
114 eset_remove(eset_t *eset, edata_t *edata) {
115 	assert(edata_state_get(edata) == eset->state ||
116 	    edata_state_in_transition(edata_state_get(edata)));
117 
118 	size_t size = edata_size_get(edata);
119 	size_t psz = sz_psz_quantize_floor(size);
120 	pszind_t pind = sz_psz2ind(psz);
121 	if (config_stats) {
122 		eset_stats_sub(eset, pind, size);
123 	}
124 
125 	edata_cmp_summary_t edata_cmp_summary = edata_cmp_summary_get(edata);
126 	edata_heap_remove(&eset->bins[pind].heap, edata);
127 	if (edata_heap_empty(&eset->bins[pind].heap)) {
128 		fb_unset(eset->bitmap, ESET_NPSIZES, (size_t)pind);
129 	} else {
130 		/*
131 		 * This is a little weird; we compare if the summaries are
132 		 * equal, rather than if the edata we removed was the heap
133 		 * minimum.  The reason why is that getting the heap minimum
134 		 * can cause a pairing heap merge operation.  We can avoid this
135 		 * if we only update the min if it's changed, in which case the
136 		 * summaries of the removed element and the min element should
137 		 * compare equal.
138 		 */
139 		if (edata_cmp_summary_comp(edata_cmp_summary,
140 		    eset->bins[pind].heap_min) == 0) {
141 			eset->bins[pind].heap_min = edata_cmp_summary_get(
142 			    edata_heap_first(&eset->bins[pind].heap));
143 		}
144 	}
145 	edata_list_inactive_remove(&eset->lru, edata);
146 	size_t npages = size >> LG_PAGE;
147 	/*
148 	 * As in eset_insert, we hold eset->mtx and so don't need atomic
149 	 * operations for updating eset->npages.
150 	 */
151 	size_t cur_extents_npages =
152 	    atomic_load_zu(&eset->npages, ATOMIC_RELAXED);
153 	assert(cur_extents_npages >= npages);
154 	atomic_store_zu(&eset->npages,
155 	    cur_extents_npages - (size >> LG_PAGE), ATOMIC_RELAXED);
156 }
157 
158 /*
159  * Find an extent with size [min_size, max_size) to satisfy the alignment
160  * requirement.  For each size, try only the first extent in the heap.
161  */
162 static edata_t *
163 eset_fit_alignment(eset_t *eset, size_t min_size, size_t max_size,
164     size_t alignment) {
165         pszind_t pind = sz_psz2ind(sz_psz_quantize_ceil(min_size));
166         pszind_t pind_max = sz_psz2ind(sz_psz_quantize_ceil(max_size));
167 
168 	for (pszind_t i =
169 	    (pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)pind);
170 	    i < pind_max;
171 	    i = (pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)i + 1)) {
172 		assert(i < SC_NPSIZES);
173 		assert(!edata_heap_empty(&eset->bins[i].heap));
174 		edata_t *edata = edata_heap_first(&eset->bins[i].heap);
175 		uintptr_t base = (uintptr_t)edata_base_get(edata);
176 		size_t candidate_size = edata_size_get(edata);
177 		assert(candidate_size >= min_size);
178 
179 		uintptr_t next_align = ALIGNMENT_CEILING((uintptr_t)base,
180 		    PAGE_CEILING(alignment));
181 		if (base > next_align || base + candidate_size <= next_align) {
182 			/* Overflow or not crossing the next alignment. */
183 			continue;
184 		}
185 
186 		size_t leadsize = next_align - base;
187 		if (candidate_size - leadsize >= min_size) {
188 			return edata;
189 		}
190 	}
191 
192 	return NULL;
193 }
194 
195 /*
196  * Do first-fit extent selection, i.e. select the oldest/lowest extent that is
197  * large enough.
198  *
199  * lg_max_fit is the (log of the) maximum ratio between the requested size and
200  * the returned size that we'll allow.  This can reduce fragmentation by
201  * avoiding reusing and splitting large extents for smaller sizes.  In practice,
202  * it's set to opt_lg_extent_max_active_fit for the dirty eset and SC_PTR_BITS
203  * for others.
204  */
205 static edata_t *
206 eset_first_fit(eset_t *eset, size_t size, bool exact_only,
207     unsigned lg_max_fit) {
208 	edata_t *ret = NULL;
209 	edata_cmp_summary_t ret_summ JEMALLOC_CC_SILENCE_INIT({0});
210 
211 	pszind_t pind = sz_psz2ind(sz_psz_quantize_ceil(size));
212 
213 	if (exact_only) {
214 		return edata_heap_empty(&eset->bins[pind].heap) ? NULL :
215 		    edata_heap_first(&eset->bins[pind].heap);
216 	}
217 
218 	for (pszind_t i =
219 	    (pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)pind);
220 	    i < ESET_NPSIZES;
221 	    i = (pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)i + 1)) {
222 		assert(!edata_heap_empty(&eset->bins[i].heap));
223 		if (lg_max_fit == SC_PTR_BITS) {
224 			/*
225 			 * We'll shift by this below, and shifting out all the
226 			 * bits is undefined.  Decreasing is safe, since the
227 			 * page size is larger than 1 byte.
228 			 */
229 			lg_max_fit = SC_PTR_BITS - 1;
230 		}
231 		if ((sz_pind2sz(i) >> lg_max_fit) > size) {
232 			break;
233 		}
234 		if (ret == NULL || edata_cmp_summary_comp(
235 		    eset->bins[i].heap_min, ret_summ) < 0) {
236 			/*
237 			 * We grab the edata as early as possible, even though
238 			 * we might change it later.  Practically, a large
239 			 * portion of eset_fit calls succeed at the first valid
240 			 * index, so this doesn't cost much, and we get the
241 			 * effect of prefetching the edata as early as possible.
242 			 */
243 			edata_t *edata = edata_heap_first(&eset->bins[i].heap);
244 			assert(edata_size_get(edata) >= size);
245 			assert(ret == NULL || edata_snad_comp(edata, ret) < 0);
246 			assert(ret == NULL || edata_cmp_summary_comp(
247 			    eset->bins[i].heap_min,
248 			    edata_cmp_summary_get(edata)) == 0);
249 			ret = edata;
250 			ret_summ = eset->bins[i].heap_min;
251 		}
252 		if (i == SC_NPSIZES) {
253 			break;
254 		}
255 		assert(i < SC_NPSIZES);
256 	}
257 
258 	return ret;
259 }
260 
261 edata_t *
262 eset_fit(eset_t *eset, size_t esize, size_t alignment, bool exact_only,
263     unsigned lg_max_fit) {
264 	size_t max_size = esize + PAGE_CEILING(alignment) - PAGE;
265 	/* Beware size_t wrap-around. */
266 	if (max_size < esize) {
267 		return NULL;
268 	}
269 
270 	edata_t *edata = eset_first_fit(eset, max_size, exact_only, lg_max_fit);
271 
272 	if (alignment > PAGE && edata == NULL) {
273 		/*
274 		 * max_size guarantees the alignment requirement but is rather
275 		 * pessimistic.  Next we try to satisfy the aligned allocation
276 		 * with sizes in [esize, max_size).
277 		 */
278 		edata = eset_fit_alignment(eset, esize, max_size, alignment);
279 	}
280 
281 	return edata;
282 }
283