1 /* Copyright 2017 Facebook.
2 *
3 * Use and distribution licensed under the BSD license. See
4 * the LICENSE file for full text.
5 */
6
7 /* -*- Mode: C; tab-width: 4; c-basic-offset: 4; indent-tabs-mode: nil -*- */
8 #include "memcached.h"
9 #include "slab_automove_extstore.h"
10 #include <stdlib.h>
11 #include <string.h>
12
13 #define MIN_PAGES_FOR_SOURCE 2
14 #define MIN_PAGES_FOR_RECLAIM 2.5
15 #define MIN_PAGES_FREE 1.5
16
17 struct window_data {
18 uint64_t age;
19 uint64_t dirty;
20 uint64_t evicted;
21 unsigned int excess_free;
22 unsigned int relaxed;
23 };
24
25 typedef struct {
26 struct window_data *window_data;
27 struct settings *settings;
28 uint32_t window_size;
29 uint32_t window_cur;
30 uint32_t item_size;
31 double max_age_ratio;
32 double free_ratio;
33 bool pool_filled_once;
34 unsigned int global_pool_watermark;
35 item_stats_automove iam_before[MAX_NUMBER_OF_SLAB_CLASSES];
36 item_stats_automove iam_after[MAX_NUMBER_OF_SLAB_CLASSES];
37 slab_stats_automove sam_before[MAX_NUMBER_OF_SLAB_CLASSES];
38 slab_stats_automove sam_after[MAX_NUMBER_OF_SLAB_CLASSES];
39 } slab_automove;
40
slab_automove_extstore_init(struct settings * settings)41 void *slab_automove_extstore_init(struct settings *settings) {
42 uint32_t window_size = settings->slab_automove_window;
43 double max_age_ratio = settings->slab_automove_ratio;
44 slab_automove *a = calloc(1, sizeof(slab_automove));
45 if (a == NULL)
46 return NULL;
47 a->window_data = calloc(window_size * MAX_NUMBER_OF_SLAB_CLASSES, sizeof(struct window_data));
48 a->window_size = window_size;
49 a->max_age_ratio = max_age_ratio;
50 a->free_ratio = settings->slab_automove_freeratio;
51 a->item_size = settings->ext_item_size;
52 a->settings = settings;
53 a->pool_filled_once = false;
54 if (a->window_data == NULL) {
55 if (a->window_data)
56 free(a->window_data);
57 free(a);
58 return NULL;
59 }
60
61 // do a dry run to fill the before structs
62 fill_item_stats_automove(a->iam_before);
63 fill_slab_stats_automove(a->sam_before);
64
65 return (void *)a;
66 }
67
slab_automove_extstore_free(void * arg)68 void slab_automove_extstore_free(void *arg) {
69 slab_automove *a = (slab_automove *)arg;
70 free(a->window_data);
71 free(a);
72 }
73
window_sum(struct window_data * wd,struct window_data * w,uint32_t size)74 static void window_sum(struct window_data *wd, struct window_data *w,
75 uint32_t size) {
76 for (int x = 0; x < size; x++) {
77 struct window_data *d = &wd[x];
78 w->age += d->age;
79 w->dirty += d->dirty;
80 w->evicted += d->evicted;
81 w->excess_free += d->excess_free;
82 w->relaxed += d->relaxed;
83 }
84 }
85
global_pool_check(slab_automove * a)86 static int global_pool_check(slab_automove *a) {
87 bool mem_limit_reached;
88 unsigned int free = a->global_pool_watermark;
89 unsigned int count = global_page_pool_size(&mem_limit_reached);
90 if (!mem_limit_reached)
91 return 0;
92 if (count < free) {
93 a->pool_filled_once = true;
94 return 1;
95 } else {
96 a->pool_filled_once = true;
97 }
98 return 0;
99 }
100
101 /* A percentage of memory is configured to be held "free" as buffers for the
102 * external storage system.
103 * % of global memory is desired in the global page pool
104 * each slab class has a % of free chunks desired based on how much memory is
105 * currently in the class. This allows time for extstore to flush data when
106 * spikes or waves of set data arrive.
107 * The global page pool reserve acts as a secondary buffer for any slab class,
108 * which helps absorb shifts in which class is active.
109 */
memcheck(slab_automove * a)110 static void memcheck(slab_automove *a) {
111 unsigned int total_pages = 0;
112
113 // FIXME: is there a cached counter for total pages alloced?
114 // technically we only really need to do this once as the pages are
115 // prefilled and ratio isn't a runtime change.
116 for (int n = 1; n < MAX_NUMBER_OF_SLAB_CLASSES; n++) {
117 slab_stats_automove *sam = &a->sam_after[n];
118 total_pages += sam->total_pages;
119 }
120 // always update what remains in the global page pool
121 total_pages += a->sam_after[0].total_pages;
122 a->global_pool_watermark = total_pages * a->free_ratio;
123 if (a->global_pool_watermark < 2)
124 a->global_pool_watermark = 2;
125 settings.ext_global_pool_min = a->global_pool_watermark;
126 }
127
get_window_data(slab_automove * a,int class)128 static struct window_data *get_window_data(slab_automove *a, int class) {
129 int w_offset = class * a->window_size;
130 return &a->window_data[w_offset + (a->window_cur % a->window_size)];
131 }
132
slab_automove_extstore_run(void * arg,int * src,int * dst)133 void slab_automove_extstore_run(void *arg, int *src, int *dst) {
134 slab_automove *a = (slab_automove *)arg;
135 int n;
136 struct window_data w_sum;
137 int oldest = -1;
138 uint64_t oldest_age = 0;
139 bool too_free = false;
140 *src = -1;
141 *dst = -1;
142
143 int global_low = global_pool_check(a);
144 // fill after structs
145 fill_item_stats_automove(a->iam_after);
146 fill_slab_stats_automove(a->sam_after);
147 a->window_cur++;
148
149 memcheck(a);
150
151 // iterate slabs
152 for (n = POWER_SMALLEST; n < MAX_NUMBER_OF_SLAB_CLASSES; n++) {
153 bool small_slab = a->sam_before[n].chunk_size < a->item_size
154 ? true : false;
155 struct window_data *wd = get_window_data(a, n);
156 int w_offset = n * a->window_size;
157 memset(wd, 0, sizeof(struct window_data));
158 unsigned int free_target = a->sam_after[n].chunks_per_page * MIN_PAGES_FREE;
159
160 // if page delta, oom, or evicted delta, mark window dirty
161 // classes marked dirty cannot donate memory back to global pool.
162 if (a->iam_after[n].evicted - a->iam_before[n].evicted > 0 ||
163 a->iam_after[n].outofmemory - a->iam_before[n].outofmemory > 0) {
164 wd->evicted = 1;
165 wd->dirty = 1;
166 }
167 if (a->sam_after[n].total_pages - a->sam_before[n].total_pages > 0) {
168 wd->dirty = 1;
169 }
170 // double the free requirements means we may have memory we can
171 // reclaim to global, if it stays this way for the whole window.
172 if (a->sam_after[n].free_chunks > (free_target * 2)) {
173 wd->excess_free = 1;
174 }
175
176 // set age into window
177 wd->age = a->iam_after[n].age;
178
179 // summarize the window-up-to-now.
180 memset(&w_sum, 0, sizeof(struct window_data));
181 window_sum(&a->window_data[w_offset], &w_sum, a->window_size);
182
183 // grab age as average of window total
184 uint64_t age = w_sum.age / a->window_size;
185
186 // if > N free chunks and not dirty, reclaim memory
187 // small slab classes aren't age balanced and rely more on global
188 // pool. reclaim them more aggressively.
189 if (a->sam_after[n].free_chunks > a->sam_after[n].chunks_per_page * MIN_PAGES_FOR_RECLAIM
190 && w_sum.dirty == 0) {
191 if (small_slab) {
192 *src = n;
193 *dst = 0;
194 too_free = true;
195 } else if (!small_slab && w_sum.excess_free >= a->window_size) {
196 // If large slab and free chunks haven't decreased for a full
197 // window, reclaim pages.
198 *src = n;
199 *dst = 0;
200 too_free = true;
201 }
202 }
203
204 if (!small_slab) {
205 // if oldest and have enough pages, is oldest
206 if (age > oldest_age
207 && a->sam_after[n].total_pages > MIN_PAGES_FOR_SOURCE) {
208 oldest = n;
209 oldest_age = age;
210 }
211
212 }
213 }
214
215 memcpy(a->iam_before, a->iam_after,
216 sizeof(item_stats_automove) * MAX_NUMBER_OF_SLAB_CLASSES);
217 memcpy(a->sam_before, a->sam_after,
218 sizeof(slab_stats_automove) * MAX_NUMBER_OF_SLAB_CLASSES);
219 // only make decisions if window has filled once.
220 if (a->window_cur < a->window_size)
221 return;
222
223 if (!too_free && global_low && oldest != -1) {
224 *src = oldest;
225 *dst = 0;
226 }
227 return;
228 }
229