1 /*
2 * CDDL HEADER START
3 *
4 * This file and its contents are supplied under the terms of the
5 * Common Development and Distribution License ("CDDL"), version 1.0.
6 * You may only use this file in accordance with the terms of version
7 * 1.0 of the CDDL.
8 *
9 * A full copy of the text of the CDDL should have accompanied this
10 * source. A copy of the CDDL is also available via the Internet at
11 * http://www.illumos.org/license/CDDL.
12 *
13 * CDDL HEADER END
14 */
15 /*
16 * Copyright (c) 2017, 2018 by Delphix. All rights reserved.
17 */
18
19 #include <sys/zfs_context.h>
20 #include <sys/aggsum.h>
21
22 /*
23 * Aggregate-sum counters are a form of fanned-out counter, used when atomic
24 * instructions on a single field cause enough CPU cache line contention to
25 * slow system performance. Due to their increased overhead and the expense
26 * involved with precisely reading from them, they should only be used in cases
27 * where the write rate (increment/decrement) is much higher than the read rate
28 * (get value).
29 *
30 * Aggregate sum counters are comprised of two basic parts, the core and the
31 * buckets. The core counter contains a lock for the entire counter, as well
32 * as the current upper and lower bounds on the value of the counter. The
33 * aggsum_bucket structure contains a per-bucket lock to protect the contents of
34 * the bucket, the current amount that this bucket has changed from the global
35 * counter (called the delta), and the amount of increment and decrement we have
36 * "borrowed" from the core counter.
37 *
38 * The basic operation of an aggsum is simple. Threads that wish to modify the
39 * counter will modify one bucket's counter (determined by their current CPU, to
40 * help minimize lock and cache contention). If the bucket already has
41 * sufficient capacity borrowed from the core structure to handle their request,
42 * they simply modify the delta and return. If the bucket does not, we clear
43 * the bucket's current state (to prevent the borrowed amounts from getting too
44 * large), and borrow more from the core counter. Borrowing is done by adding to
45 * the upper bound (or subtracting from the lower bound) of the core counter,
46 * and setting the borrow value for the bucket to the amount added (or
47 * subtracted). Clearing the bucket is the opposite; we add the current delta
48 * to both the lower and upper bounds of the core counter, subtract the borrowed
49 * incremental from the upper bound, and add the borrowed decrement from the
50 * lower bound. Note that only borrowing and clearing require access to the
51 * core counter; since all other operations access CPU-local resources,
52 * performance can be much higher than a traditional counter.
53 *
54 * Threads that wish to read from the counter have a slightly more challenging
55 * task. It is fast to determine the upper and lower bounds of the aggum; this
56 * does not require grabbing any locks. This suffices for cases where an
57 * approximation of the aggsum's value is acceptable. However, if one needs to
58 * know whether some specific value is above or below the current value in the
59 * aggsum, they invoke aggsum_compare(). This function operates by repeatedly
60 * comparing the target value to the upper and lower bounds of the aggsum, and
61 * then clearing a bucket. This proceeds until the target is outside of the
62 * upper and lower bounds and we return a response, or the last bucket has been
63 * cleared and we know that the target is equal to the aggsum's value. Finally,
64 * the most expensive operation is determining the precise value of the aggsum.
65 * To do this, we clear every bucket and then return the upper bound (which must
66 * be equal to the lower bound). What makes aggsum_compare() and aggsum_value()
67 * expensive is clearing buckets. This involves grabbing the global lock
68 * (serializing against themselves and borrow operations), grabbing a bucket's
69 * lock (preventing threads on those CPUs from modifying their delta), and
70 * zeroing out the borrowed value (forcing that thread to borrow on its next
71 * request, which will also be expensive). This is what makes aggsums well
72 * suited for write-many read-rarely operations.
73 *
74 * Note that the aggsums do not expand if more CPUs are hot-added. In that
75 * case, we will have less fanout than boot_ncpus, but we don't want to always
76 * reserve the RAM necessary to create the extra slots for additional CPUs up
77 * front, and dynamically adding them is a complex task.
78 */
79
80 /*
81 * We will borrow aggsum_borrow_multiplier times the current request, so we will
82 * have to get the as_lock approximately every aggsum_borrow_multiplier calls to
83 * aggsum_delta().
84 */
85 static uint_t aggsum_borrow_multiplier = 10;
86
87 void
aggsum_init(aggsum_t * as,uint64_t value)88 aggsum_init(aggsum_t *as, uint64_t value)
89 {
90 bzero(as, sizeof (*as));
91 as->as_lower_bound = as->as_upper_bound = value;
92 mutex_init(&as->as_lock, NULL, MUTEX_DEFAULT, NULL);
93 as->as_numbuckets = boot_ncpus;
94 as->as_buckets = kmem_zalloc(boot_ncpus * sizeof (aggsum_bucket_t),
95 KM_SLEEP);
96 for (int i = 0; i < as->as_numbuckets; i++) {
97 mutex_init(&as->as_buckets[i].asc_lock,
98 NULL, MUTEX_DEFAULT, NULL);
99 }
100 }
101
102 void
aggsum_fini(aggsum_t * as)103 aggsum_fini(aggsum_t *as)
104 {
105 for (int i = 0; i < as->as_numbuckets; i++)
106 mutex_destroy(&as->as_buckets[i].asc_lock);
107 kmem_free(as->as_buckets, as->as_numbuckets * sizeof (aggsum_bucket_t));
108 mutex_destroy(&as->as_lock);
109 }
110
111 int64_t
aggsum_lower_bound(aggsum_t * as)112 aggsum_lower_bound(aggsum_t *as)
113 {
114 return (as->as_lower_bound);
115 }
116
117 int64_t
aggsum_upper_bound(aggsum_t * as)118 aggsum_upper_bound(aggsum_t *as)
119 {
120 return (as->as_upper_bound);
121 }
122
123 static void
aggsum_flush_bucket(aggsum_t * as,struct aggsum_bucket * asb)124 aggsum_flush_bucket(aggsum_t *as, struct aggsum_bucket *asb)
125 {
126 ASSERT(MUTEX_HELD(&as->as_lock));
127 ASSERT(MUTEX_HELD(&asb->asc_lock));
128
129 /*
130 * We use atomic instructions for this because we read the upper and
131 * lower bounds without the lock, so we need stores to be atomic.
132 */
133 atomic_add_64((volatile uint64_t *)&as->as_lower_bound,
134 asb->asc_delta + asb->asc_borrowed);
135 atomic_add_64((volatile uint64_t *)&as->as_upper_bound,
136 asb->asc_delta - asb->asc_borrowed);
137 asb->asc_delta = 0;
138 asb->asc_borrowed = 0;
139 }
140
141 uint64_t
aggsum_value(aggsum_t * as)142 aggsum_value(aggsum_t *as)
143 {
144 int64_t rv;
145
146 mutex_enter(&as->as_lock);
147 if (as->as_lower_bound == as->as_upper_bound) {
148 rv = as->as_lower_bound;
149 for (int i = 0; i < as->as_numbuckets; i++) {
150 ASSERT0(as->as_buckets[i].asc_delta);
151 ASSERT0(as->as_buckets[i].asc_borrowed);
152 }
153 mutex_exit(&as->as_lock);
154 return (rv);
155 }
156 for (int i = 0; i < as->as_numbuckets; i++) {
157 struct aggsum_bucket *asb = &as->as_buckets[i];
158 mutex_enter(&asb->asc_lock);
159 aggsum_flush_bucket(as, asb);
160 mutex_exit(&asb->asc_lock);
161 }
162 VERIFY3U(as->as_lower_bound, ==, as->as_upper_bound);
163 rv = as->as_lower_bound;
164 mutex_exit(&as->as_lock);
165
166 return (rv);
167 }
168
169 void
aggsum_add(aggsum_t * as,int64_t delta)170 aggsum_add(aggsum_t *as, int64_t delta)
171 {
172 struct aggsum_bucket *asb;
173 int64_t borrow;
174
175 asb = &as->as_buckets[CPU_SEQID_UNSTABLE % as->as_numbuckets];
176
177 /* Try fast path if we already borrowed enough before. */
178 mutex_enter(&asb->asc_lock);
179 if (asb->asc_delta + delta <= (int64_t)asb->asc_borrowed &&
180 asb->asc_delta + delta >= -(int64_t)asb->asc_borrowed) {
181 asb->asc_delta += delta;
182 mutex_exit(&asb->asc_lock);
183 return;
184 }
185 mutex_exit(&asb->asc_lock);
186
187 /*
188 * We haven't borrowed enough. Take the global lock and borrow
189 * considering what is requested now and what we borrowed before.
190 */
191 borrow = (delta < 0 ? -delta : delta) * aggsum_borrow_multiplier;
192 mutex_enter(&as->as_lock);
193 mutex_enter(&asb->asc_lock);
194 delta += asb->asc_delta;
195 asb->asc_delta = 0;
196 if (borrow >= asb->asc_borrowed)
197 borrow -= asb->asc_borrowed;
198 else
199 borrow = (borrow - (int64_t)asb->asc_borrowed) / 4;
200 asb->asc_borrowed += borrow;
201 atomic_add_64((volatile uint64_t *)&as->as_lower_bound,
202 delta - borrow);
203 atomic_add_64((volatile uint64_t *)&as->as_upper_bound,
204 delta + borrow);
205 mutex_exit(&asb->asc_lock);
206 mutex_exit(&as->as_lock);
207 }
208
209 /*
210 * Compare the aggsum value to target efficiently. Returns -1 if the value
211 * represented by the aggsum is less than target, 1 if it's greater, and 0 if
212 * they are equal.
213 */
214 int
aggsum_compare(aggsum_t * as,uint64_t target)215 aggsum_compare(aggsum_t *as, uint64_t target)
216 {
217 if (as->as_upper_bound < target)
218 return (-1);
219 if (as->as_lower_bound > target)
220 return (1);
221 mutex_enter(&as->as_lock);
222 for (int i = 0; i < as->as_numbuckets; i++) {
223 struct aggsum_bucket *asb = &as->as_buckets[i];
224 mutex_enter(&asb->asc_lock);
225 aggsum_flush_bucket(as, asb);
226 mutex_exit(&asb->asc_lock);
227 if (as->as_upper_bound < target) {
228 mutex_exit(&as->as_lock);
229 return (-1);
230 }
231 if (as->as_lower_bound > target) {
232 mutex_exit(&as->as_lock);
233 return (1);
234 }
235 }
236 VERIFY3U(as->as_lower_bound, ==, as->as_upper_bound);
237 ASSERT3U(as->as_lower_bound, ==, target);
238 mutex_exit(&as->as_lock);
239 return (0);
240 }
241