xref: /freebsd/sys/contrib/openzfs/module/zfs/aggsum.c (revision 02e9120893770924227138ba49df1edb3896112a)
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 2^aggsum_borrow_shift times the current request, so we will
82  * have to get the as_lock approximately every 2^aggsum_borrow_shift calls to
83  * aggsum_add().
84  */
85 static uint_t aggsum_borrow_shift = 4;
86 
87 void
88 aggsum_init(aggsum_t *as, uint64_t value)
89 {
90 	memset(as, 0, sizeof (*as));
91 	as->as_lower_bound = as->as_upper_bound = value;
92 	mutex_init(&as->as_lock, NULL, MUTEX_DEFAULT, NULL);
93 	/*
94 	 * Too many buckets may hurt read performance without improving
95 	 * write.  From 12 CPUs use bucket per 2 CPUs, from 48 per 4, etc.
96 	 */
97 	as->as_bucketshift = highbit64(boot_ncpus / 6) / 2;
98 	as->as_numbuckets = ((boot_ncpus - 1) >> as->as_bucketshift) + 1;
99 	as->as_buckets = kmem_zalloc(as->as_numbuckets *
100 	    sizeof (aggsum_bucket_t), KM_SLEEP);
101 	for (int i = 0; i < as->as_numbuckets; i++) {
102 		mutex_init(&as->as_buckets[i].asc_lock,
103 		    NULL, MUTEX_DEFAULT, NULL);
104 	}
105 }
106 
107 void
108 aggsum_fini(aggsum_t *as)
109 {
110 	for (int i = 0; i < as->as_numbuckets; i++)
111 		mutex_destroy(&as->as_buckets[i].asc_lock);
112 	kmem_free(as->as_buckets, as->as_numbuckets * sizeof (aggsum_bucket_t));
113 	mutex_destroy(&as->as_lock);
114 }
115 
116 int64_t
117 aggsum_lower_bound(aggsum_t *as)
118 {
119 	return (atomic_load_64((volatile uint64_t *)&as->as_lower_bound));
120 }
121 
122 uint64_t
123 aggsum_upper_bound(aggsum_t *as)
124 {
125 	return (atomic_load_64(&as->as_upper_bound));
126 }
127 
128 uint64_t
129 aggsum_value(aggsum_t *as)
130 {
131 	int64_t lb;
132 	uint64_t ub;
133 
134 	mutex_enter(&as->as_lock);
135 	lb = as->as_lower_bound;
136 	ub = as->as_upper_bound;
137 	if (lb == ub) {
138 		for (int i = 0; i < as->as_numbuckets; i++) {
139 			ASSERT0(as->as_buckets[i].asc_delta);
140 			ASSERT0(as->as_buckets[i].asc_borrowed);
141 		}
142 		mutex_exit(&as->as_lock);
143 		return (lb);
144 	}
145 	for (int i = 0; i < as->as_numbuckets; i++) {
146 		struct aggsum_bucket *asb = &as->as_buckets[i];
147 		if (asb->asc_borrowed == 0)
148 			continue;
149 		mutex_enter(&asb->asc_lock);
150 		lb += asb->asc_delta + asb->asc_borrowed;
151 		ub += asb->asc_delta - asb->asc_borrowed;
152 		asb->asc_delta = 0;
153 		asb->asc_borrowed = 0;
154 		mutex_exit(&asb->asc_lock);
155 	}
156 	ASSERT3U(lb, ==, ub);
157 	atomic_store_64((volatile uint64_t *)&as->as_lower_bound, lb);
158 	atomic_store_64(&as->as_upper_bound, lb);
159 	mutex_exit(&as->as_lock);
160 
161 	return (lb);
162 }
163 
164 void
165 aggsum_add(aggsum_t *as, int64_t delta)
166 {
167 	struct aggsum_bucket *asb;
168 	int64_t borrow;
169 
170 	asb = &as->as_buckets[(CPU_SEQID_UNSTABLE >> as->as_bucketshift) %
171 	    as->as_numbuckets];
172 
173 	/* Try fast path if we already borrowed enough before. */
174 	mutex_enter(&asb->asc_lock);
175 	if (asb->asc_delta + delta <= (int64_t)asb->asc_borrowed &&
176 	    asb->asc_delta + delta >= -(int64_t)asb->asc_borrowed) {
177 		asb->asc_delta += delta;
178 		mutex_exit(&asb->asc_lock);
179 		return;
180 	}
181 	mutex_exit(&asb->asc_lock);
182 
183 	/*
184 	 * We haven't borrowed enough.  Take the global lock and borrow
185 	 * considering what is requested now and what we borrowed before.
186 	 */
187 	borrow = (delta < 0 ? -delta : delta);
188 	borrow <<= aggsum_borrow_shift + as->as_bucketshift;
189 	mutex_enter(&as->as_lock);
190 	if (borrow >= asb->asc_borrowed)
191 		borrow -= asb->asc_borrowed;
192 	else
193 		borrow = (borrow - (int64_t)asb->asc_borrowed) / 4;
194 	mutex_enter(&asb->asc_lock);
195 	delta += asb->asc_delta;
196 	asb->asc_delta = 0;
197 	asb->asc_borrowed += borrow;
198 	mutex_exit(&asb->asc_lock);
199 	atomic_store_64((volatile uint64_t *)&as->as_lower_bound,
200 	    as->as_lower_bound + delta - borrow);
201 	atomic_store_64(&as->as_upper_bound,
202 	    as->as_upper_bound + delta + borrow);
203 	mutex_exit(&as->as_lock);
204 }
205 
206 /*
207  * Compare the aggsum value to target efficiently. Returns -1 if the value
208  * represented by the aggsum is less than target, 1 if it's greater, and 0 if
209  * they are equal.
210  */
211 int
212 aggsum_compare(aggsum_t *as, uint64_t target)
213 {
214 	int64_t lb;
215 	uint64_t ub;
216 	int i;
217 
218 	if (atomic_load_64(&as->as_upper_bound) < target)
219 		return (-1);
220 	lb = atomic_load_64((volatile uint64_t *)&as->as_lower_bound);
221 	if (lb > 0 && (uint64_t)lb > target)
222 		return (1);
223 	mutex_enter(&as->as_lock);
224 	lb = as->as_lower_bound;
225 	ub = as->as_upper_bound;
226 	for (i = 0; i < as->as_numbuckets; i++) {
227 		struct aggsum_bucket *asb = &as->as_buckets[i];
228 		if (asb->asc_borrowed == 0)
229 			continue;
230 		mutex_enter(&asb->asc_lock);
231 		lb += asb->asc_delta + asb->asc_borrowed;
232 		ub += asb->asc_delta - asb->asc_borrowed;
233 		asb->asc_delta = 0;
234 		asb->asc_borrowed = 0;
235 		mutex_exit(&asb->asc_lock);
236 		if (ub < target || (lb > 0 && (uint64_t)lb > target))
237 			break;
238 	}
239 	if (i >= as->as_numbuckets)
240 		ASSERT3U(lb, ==, ub);
241 	atomic_store_64((volatile uint64_t *)&as->as_lower_bound, lb);
242 	atomic_store_64(&as->as_upper_bound, ub);
243 	mutex_exit(&as->as_lock);
244 	return (ub < target ? -1 : (uint64_t)lb > target ? 1 : 0);
245 }
246