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
75/*
76 * We will borrow aggsum_borrow_multiplier times the current request, so we will
77 * have to get the as_lock approximately every aggsum_borrow_multiplier calls to
78 * aggsum_delta().
79 */
80static uint_t aggsum_borrow_multiplier = 10;
81
82void
83aggsum_init(aggsum_t *as, uint64_t value)
84{
85	bzero(as, sizeof (*as));
86	as->as_lower_bound = as->as_upper_bound = value;
87	mutex_init(&as->as_lock, NULL, MUTEX_DEFAULT, NULL);
88	as->as_numbuckets = boot_ncpus;
89	as->as_buckets = kmem_zalloc(boot_ncpus * sizeof (aggsum_bucket_t),
90	    KM_SLEEP);
91	for (int i = 0; i < as->as_numbuckets; i++) {
92		mutex_init(&as->as_buckets[i].asc_lock,
93		    NULL, MUTEX_DEFAULT, NULL);
94	}
95}
96
97void
98aggsum_fini(aggsum_t *as)
99{
100	for (int i = 0; i < as->as_numbuckets; i++)
101		mutex_destroy(&as->as_buckets[i].asc_lock);
102	kmem_free(as->as_buckets, as->as_numbuckets * sizeof (aggsum_bucket_t));
103	mutex_destroy(&as->as_lock);
104}
105
106int64_t
107aggsum_lower_bound(aggsum_t *as)
108{
109	return (as->as_lower_bound);
110}
111
112int64_t
113aggsum_upper_bound(aggsum_t *as)
114{
115	return (as->as_upper_bound);
116}
117
118static void
119aggsum_flush_bucket(aggsum_t *as, struct aggsum_bucket *asb)
120{
121	ASSERT(MUTEX_HELD(&as->as_lock));
122	ASSERT(MUTEX_HELD(&asb->asc_lock));
123
124	/*
125	 * We use atomic instructions for this because we read the upper and
126	 * lower bounds without the lock, so we need stores to be atomic.
127	 */
128	atomic_add_64((volatile uint64_t *)&as->as_lower_bound,
129	    asb->asc_delta + asb->asc_borrowed);
130	atomic_add_64((volatile uint64_t *)&as->as_upper_bound,
131	    asb->asc_delta - asb->asc_borrowed);
132	asb->asc_delta = 0;
133	asb->asc_borrowed = 0;
134}
135
136uint64_t
137aggsum_value(aggsum_t *as)
138{
139	int64_t rv;
140
141	mutex_enter(&as->as_lock);
142	if (as->as_lower_bound == as->as_upper_bound) {
143		rv = as->as_lower_bound;
144		for (int i = 0; i < as->as_numbuckets; i++) {
145			ASSERT0(as->as_buckets[i].asc_delta);
146			ASSERT0(as->as_buckets[i].asc_borrowed);
147		}
148		mutex_exit(&as->as_lock);
149		return (rv);
150	}
151	for (int i = 0; i < as->as_numbuckets; i++) {
152		struct aggsum_bucket *asb = &as->as_buckets[i];
153		mutex_enter(&asb->asc_lock);
154		aggsum_flush_bucket(as, asb);
155		mutex_exit(&asb->asc_lock);
156	}
157	VERIFY3U(as->as_lower_bound, ==, as->as_upper_bound);
158	rv = as->as_lower_bound;
159	mutex_exit(&as->as_lock);
160
161	return (rv);
162}
163
164void
165aggsum_add(aggsum_t *as, int64_t delta)
166{
167	struct aggsum_bucket *asb =
168	    &as->as_buckets[CPU_SEQID % as->as_numbuckets];
169	int64_t borrow;
170
171	/* Try fast path if we already borrowed enough before. */
172	mutex_enter(&asb->asc_lock);
173	if (asb->asc_delta + delta <= (int64_t)asb->asc_borrowed &&
174	    asb->asc_delta + delta >= -(int64_t)asb->asc_borrowed) {
175		asb->asc_delta += delta;
176		mutex_exit(&asb->asc_lock);
177		return;
178	}
179	mutex_exit(&asb->asc_lock);
180
181	/*
182	 * We haven't borrowed enough.  Take the global lock and borrow
183	 * considering what is requested now and what we borrowed before.
184	 */
185	borrow = (delta < 0 ? -delta : delta) * aggsum_borrow_multiplier;
186	mutex_enter(&as->as_lock);
187	mutex_enter(&asb->asc_lock);
188	delta += asb->asc_delta;
189	asb->asc_delta = 0;
190	if (borrow >= asb->asc_borrowed)
191		borrow -= asb->asc_borrowed;
192	else
193		borrow = (borrow - (int64_t)asb->asc_borrowed) / 4;
194	asb->asc_borrowed += borrow;
195	atomic_add_64((volatile uint64_t *)&as->as_lower_bound,
196	    delta - borrow);
197	atomic_add_64((volatile uint64_t *)&as->as_upper_bound,
198	    delta + borrow);
199	mutex_exit(&asb->asc_lock);
200	mutex_exit(&as->as_lock);
201}
202
203/*
204 * Compare the aggsum value to target efficiently. Returns -1 if the value
205 * represented by the aggsum is less than target, 1 if it's greater, and 0 if
206 * they are equal.
207 */
208int
209aggsum_compare(aggsum_t *as, uint64_t target)
210{
211	if (as->as_upper_bound < target)
212		return (-1);
213	if (as->as_lower_bound > target)
214		return (1);
215	mutex_enter(&as->as_lock);
216	for (int i = 0; i < as->as_numbuckets; i++) {
217		struct aggsum_bucket *asb = &as->as_buckets[i];
218		mutex_enter(&asb->asc_lock);
219		aggsum_flush_bucket(as, asb);
220		mutex_exit(&asb->asc_lock);
221		if (as->as_upper_bound < target) {
222			mutex_exit(&as->as_lock);
223			return (-1);
224		}
225		if (as->as_lower_bound > target) {
226			mutex_exit(&as->as_lock);
227			return (1);
228		}
229	}
230	VERIFY3U(as->as_lower_bound, ==, as->as_upper_bound);
231	ASSERT3U(as->as_lower_bound, ==, target);
232	mutex_exit(&as->as_lock);
233	return (0);
234}
235