2 * SPDX-License-Identifier: BSD-2-Clause
4 * Copyright (c) 2019,2020 Jeffrey Roberson <jeff@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice unmodified, this list of conditions, and the following
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28 #include <sys/param.h>
29 #include <sys/systm.h>
30 #include <sys/counter.h>
31 #include <sys/kernel.h>
32 #include <sys/limits.h>
36 #include <sys/sysctl.h>
41 * Global Unbounded Sequences (GUS)
43 * This is a novel safe memory reclamation technique inspired by
44 * epoch based reclamation from Samy Al Bahra's concurrency kit which
45 * in turn was based on work described in:
46 * Fraser, K. 2004. Practical Lock-Freedom. PhD Thesis, University
47 * of Cambridge Computing Laboratory.
48 * And shares some similarities with:
49 * Wang, Stamler, Parmer. 2016 Parallel Sections: Scaling System-Level
52 * This is not an implementation of hazard pointers or related
53 * techniques. The term safe memory reclamation is used as a
54 * generic descriptor for algorithms that defer frees to avoid
55 * use-after-free errors with lockless datastructures or as
56 * a mechanism to detect quiescence for writer synchronization.
58 * The basic approach is to maintain a monotonic write sequence
59 * number that is updated on some application defined granularity.
60 * Readers record the most recent write sequence number they have
61 * observed. A shared read sequence number records the lowest
62 * sequence number observed by any reader as of the last poll. Any
63 * write older than this value has been observed by all readers
64 * and memory can be reclaimed. Like Epoch we also detect idle
65 * readers by storing an invalid sequence number in the per-cpu
66 * state when the read section exits. Like Parsec we establish
67 * a global write clock that is used to mark memory on free.
69 * The write and read sequence numbers can be thought of as a two
70 * handed clock with readers always advancing towards writers. GUS
71 * maintains the invariant that all readers can safely access memory
72 * that was visible at the time they loaded their copy of the sequence
73 * number. Periodically the read sequence or hand is polled and
74 * advanced as far towards the write sequence as active readers allow.
75 * Memory which was freed between the old and new global read sequence
76 * number can now be reclaimed. When the system is idle the two hands
77 * meet and no deferred memory is outstanding. Readers never advance
78 * any sequence number, they only observe them. The shared read
79 * sequence number is consequently never higher than the write sequence.
80 * A stored sequence number that falls outside of this range has expired
81 * and needs no scan to reclaim.
83 * A notable distinction between GUS and Epoch, qsbr, rcu, etc. is
84 * that advancing the sequence number is decoupled from detecting its
85 * observation. That is to say, the delta between read and write
86 * sequence numbers is not bound. This can be thought of as a more
87 * generalized form of epoch which requires them at most one step
88 * apart. This results in a more granular assignment of sequence
89 * numbers even as read latencies prohibit all or some expiration.
90 * It also allows writers to advance the sequence number and save the
91 * poll for expiration until a later time when it is likely to
92 * complete without waiting. The batch granularity and free-to-use
93 * latency is dynamic and can be significantly smaller than in more
96 * This mechanism is primarily intended to be used in coordination with
97 * UMA. By integrating with the allocator we avoid all of the callout
98 * queue machinery and are provided with an efficient way to batch
99 * sequence advancement and waiting. The allocator accumulates a full
100 * per-cpu cache of memory before advancing the sequence. It then
101 * delays waiting for this sequence to expire until the memory is
102 * selected for reuse. In this way we only increment the sequence
103 * value once for n=cache-size frees and the waits are done long
104 * after the sequence has been expired so they need only be verified
105 * to account for pathological conditions and to advance the read
106 * sequence. Tying the sequence number to the bucket size has the
107 * nice property that as the zone gets busier the buckets get larger
108 * and the sequence writes become fewer. If the coherency of advancing
109 * the write sequence number becomes too costly we can advance
110 * it for every N buckets in exchange for higher free-to-use
111 * latency and consequently higher memory consumption.
113 * If the read overhead of accessing the shared cacheline becomes
114 * especially burdensome an invariant TSC could be used in place of the
115 * sequence. The algorithm would then only need to maintain the minimum
116 * observed tsc. This would trade potential cache synchronization
117 * overhead for local serialization and cpu timestamp overhead.
121 * A simplified diagram:
124 * | -------------------- sequence number space -------------------- |
126 * | ----- valid sequence numbers ---- |
128 * | -- free -- | --------- deferred frees -------- | ---- free ---- |
131 * In this example cpuA has the lowest sequence number and poll can
132 * advance rd seq. cpuB is not running and is considered to observe
135 * Freed memory that is tagged with a sequence number between rd seq and
136 * wr seq can not be safely reclaimed because cpuA may hold a reference to
137 * it. Any other memory is guaranteed to be unreferenced.
139 * Any writer is free to advance wr seq at any time however it may busy
140 * poll in pathological cases.
143 static uma_zone_t smr_shared_zone;
144 static uma_zone_t smr_zone;
147 #define SMR_SEQ_INIT 1 /* All valid sequence numbers are odd. */
148 #define SMR_SEQ_INCR 2
151 * SMR_SEQ_MAX_DELTA is the maximum distance allowed between rd_seq and
152 * wr_seq. For the modular arithmetic to work a value of UNIT_MAX / 2
153 * would be possible but it is checked after we increment the wr_seq so
154 * a safety margin is left to prevent overflow.
156 * We will block until SMR_SEQ_MAX_ADVANCE sequence numbers have progressed
157 * to prevent integer wrapping. See smr_advance() for more details.
159 #define SMR_SEQ_MAX_DELTA (UINT_MAX / 4)
160 #define SMR_SEQ_MAX_ADVANCE (SMR_SEQ_MAX_DELTA - 1024)
162 /* We want to test the wrapping feature in invariants kernels. */
163 #define SMR_SEQ_INCR (UINT_MAX / 10000)
164 #define SMR_SEQ_INIT (UINT_MAX - 100000)
165 /* Force extra polls to test the integer overflow detection. */
166 #define SMR_SEQ_MAX_DELTA (SMR_SEQ_INCR * 32)
167 #define SMR_SEQ_MAX_ADVANCE SMR_SEQ_MAX_DELTA / 2
171 * The grace period for lazy (tick based) SMR.
173 * Hardclock is responsible for advancing ticks on a single CPU while every
174 * CPU receives a regular clock interrupt. The clock interrupts are flushing
175 * the store buffers and any speculative loads that may violate our invariants.
176 * Because these interrupts are not synchronized we must wait one additional
177 * tick in the future to be certain that all processors have had their state
178 * synchronized by an interrupt.
180 * This assumes that the clock interrupt will only be delayed by other causes
181 * that will flush the store buffer or prevent access to the section protected
182 * data. For example, an idle processor, or an system management interrupt,
185 #define SMR_LAZY_GRACE 2
186 #define SMR_LAZY_INCR (SMR_LAZY_GRACE * SMR_SEQ_INCR)
189 * The maximum sequence number ahead of wr_seq that may still be valid. The
190 * sequence may not be advanced on write for lazy or deferred SMRs. In this
191 * case poll needs to attempt to forward the sequence number if the goal is
192 * within wr_seq + SMR_SEQ_ADVANCE.
194 #define SMR_SEQ_ADVANCE SMR_LAZY_INCR
196 static SYSCTL_NODE(_debug, OID_AUTO, smr, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
198 static COUNTER_U64_DEFINE_EARLY(advance);
199 SYSCTL_COUNTER_U64(_debug_smr, OID_AUTO, advance, CTLFLAG_RW, &advance, "");
200 static COUNTER_U64_DEFINE_EARLY(advance_wait);
201 SYSCTL_COUNTER_U64(_debug_smr, OID_AUTO, advance_wait, CTLFLAG_RW, &advance_wait, "");
202 static COUNTER_U64_DEFINE_EARLY(poll);
203 SYSCTL_COUNTER_U64(_debug_smr, OID_AUTO, poll, CTLFLAG_RW, &poll, "");
204 static COUNTER_U64_DEFINE_EARLY(poll_scan);
205 SYSCTL_COUNTER_U64(_debug_smr, OID_AUTO, poll_scan, CTLFLAG_RW, &poll_scan, "");
206 static COUNTER_U64_DEFINE_EARLY(poll_fail);
207 SYSCTL_COUNTER_U64(_debug_smr, OID_AUTO, poll_fail, CTLFLAG_RW, &poll_fail, "");
210 * Advance a lazy write sequence number. These move forward at the rate of
211 * ticks. Grace is SMR_LAZY_INCR (2 ticks) in the future.
213 * This returns the goal write sequence number.
216 smr_lazy_advance(smr_t smr, smr_shared_t s)
218 union s_wr s_wr, old;
221 CRITICAL_ASSERT(curthread);
224 * Load the stored ticks value before the current one. This way the
225 * current value can only be the same or larger.
227 old._pair = s_wr._pair = atomic_load_acq_64(&s->s_wr._pair);
231 * The most probable condition that the update already took place.
234 if (__predict_true(d == 0))
236 /* Cap the rate of advancement and handle long idle periods. */
237 if (d > SMR_LAZY_GRACE || d < 0)
240 s_wr.seq += d * SMR_SEQ_INCR;
243 * This can only fail if another thread races to call advance().
244 * Strong cmpset semantics mean we are guaranteed that the update
247 atomic_cmpset_64(&s->s_wr._pair, old._pair, s_wr._pair);
249 return (s_wr.seq + SMR_LAZY_INCR);
253 * Increment the shared write sequence by 2. Since it is initialized
254 * to 1 this means the only valid values are odd and an observed value
255 * of 0 in a particular CPU means it is not currently in a read section.
258 smr_shared_advance(smr_shared_t s)
261 return (atomic_fetchadd_int(&s->s_wr.seq, SMR_SEQ_INCR) + SMR_SEQ_INCR);
265 * Advance the write sequence number for a normal smr section. If the
266 * write sequence is too far behind the read sequence we have to poll
267 * to advance rd_seq and prevent undetectable wraps.
270 smr_default_advance(smr_t smr, smr_shared_t s)
272 smr_seq_t goal, s_rd_seq;
274 CRITICAL_ASSERT(curthread);
275 KASSERT((zpcpu_get(smr)->c_flags & SMR_LAZY) == 0,
276 ("smr_default_advance: called with lazy smr."));
279 * Load the current read seq before incrementing the goal so
280 * we are guaranteed it is always < goal.
282 s_rd_seq = atomic_load_acq_int(&s->s_rd_seq);
283 goal = smr_shared_advance(s);
286 * Force a synchronization here if the goal is getting too
287 * far ahead of the read sequence number. This keeps the
288 * wrap detecting arithmetic working in pathological cases.
290 if (SMR_SEQ_DELTA(goal, s_rd_seq) >= SMR_SEQ_MAX_DELTA) {
291 counter_u64_add(advance_wait, 1);
292 smr_wait(smr, goal - SMR_SEQ_MAX_ADVANCE);
294 counter_u64_add(advance, 1);
300 * Deferred SMRs conditionally update s_wr_seq based on an
301 * cpu local interval count.
304 smr_deferred_advance(smr_t smr, smr_shared_t s, smr_t self)
307 if (++self->c_deferred < self->c_limit)
308 return (smr_shared_current(s) + SMR_SEQ_INCR);
309 self->c_deferred = 0;
310 return (smr_default_advance(smr, s));
314 * Advance the write sequence and return the value for use as the
315 * wait goal. This guarantees that any changes made by the calling
316 * thread prior to this call will be visible to all threads after
317 * rd_seq meets or exceeds the return value.
319 * This function may busy loop if the readers are roughly 1 billion
320 * sequence numbers behind the writers.
322 * Lazy SMRs will not busy loop and the wrap happens every 25 days
323 * at 1khz and 60 hours at 10khz. Readers can block for no longer
324 * than half of this for SMR_SEQ_ macros to continue working.
327 smr_advance(smr_t smr)
335 * It is illegal to enter while in an smr section.
337 SMR_ASSERT_NOT_ENTERED(smr);
340 * Modifications not done in a smr section need to be visible
341 * before advancing the seq.
343 atomic_thread_fence_rel();
346 /* Try to touch the line once. */
347 self = zpcpu_get(smr);
349 flags = self->c_flags;
350 goal = SMR_SEQ_INVALID;
351 if ((flags & (SMR_LAZY | SMR_DEFERRED)) == 0)
352 goal = smr_default_advance(smr, s);
353 else if ((flags & SMR_LAZY) != 0)
354 goal = smr_lazy_advance(smr, s);
355 else if ((flags & SMR_DEFERRED) != 0)
356 goal = smr_deferred_advance(smr, s, self);
363 * Poll to determine the currently observed sequence number on a cpu
364 * and spinwait if the 'wait' argument is true.
367 smr_poll_cpu(smr_t c, smr_seq_t s_rd_seq, smr_seq_t goal, bool wait)
371 c_seq = SMR_SEQ_INVALID;
373 c_seq = atomic_load_int(&c->c_seq);
374 if (c_seq == SMR_SEQ_INVALID)
378 * There is a race described in smr.h:smr_enter that
379 * can lead to a stale seq value but not stale data
380 * access. If we find a value out of range here we
381 * pin it to the current min to prevent it from
382 * advancing until that stale section has expired.
384 * The race is created when a cpu loads the s_wr_seq
385 * value in a local register and then another thread
386 * advances s_wr_seq and calls smr_poll() which will
387 * oberve no value yet in c_seq and advance s_rd_seq
388 * up to s_wr_seq which is beyond the register
389 * cached value. This is only likely to happen on
390 * hypervisor or with a system management interrupt.
392 if (SMR_SEQ_LT(c_seq, s_rd_seq))
396 * If the sequence number meets the goal we are done
399 if (SMR_SEQ_LEQ(goal, c_seq))
411 * Loop until all cores have observed the goal sequence or have
412 * gone inactive. Returns the oldest sequence currently active;
414 * This function assumes a snapshot of sequence values has
415 * been obtained and validated by smr_poll().
418 smr_poll_scan(smr_t smr, smr_shared_t s, smr_seq_t s_rd_seq,
419 smr_seq_t s_wr_seq, smr_seq_t goal, bool wait)
421 smr_seq_t rd_seq, c_seq;
424 CRITICAL_ASSERT(curthread);
425 counter_u64_add_protected(poll_scan, 1);
428 * The read sequence can be no larger than the write sequence at
429 * the start of the poll.
434 * Query the active sequence on this cpu. If we're not
435 * waiting and we don't meet the goal we will still scan
436 * the rest of the cpus to update s_rd_seq before returning
439 c_seq = smr_poll_cpu(zpcpu_get_cpu(smr, i), s_rd_seq, goal,
443 * Limit the minimum observed rd_seq whether we met the goal
446 if (c_seq != SMR_SEQ_INVALID)
447 rd_seq = SMR_SEQ_MIN(rd_seq, c_seq);
451 * Advance the rd_seq as long as we observed a more recent value.
453 s_rd_seq = atomic_load_int(&s->s_rd_seq);
454 if (SMR_SEQ_GT(rd_seq, s_rd_seq)) {
455 atomic_cmpset_int(&s->s_rd_seq, s_rd_seq, rd_seq);
463 * Poll to determine whether all readers have observed the 'goal' write
466 * If wait is true this will spin until the goal is met.
468 * This routine will updated the minimum observed read sequence number in
469 * s_rd_seq if it does a scan. It may not do a scan if another call has
470 * advanced s_rd_seq beyond the callers goal already.
472 * Returns true if the goal is met and false if not.
475 smr_poll(smr_t smr, smr_seq_t goal, bool wait)
479 smr_seq_t s_wr_seq, s_rd_seq;
485 * It is illegal to enter while in an smr section.
487 KASSERT(!wait || !SMR_ENTERED(smr),
488 ("smr_poll: Blocking not allowed in a SMR section."));
489 KASSERT(!wait || (zpcpu_get(smr)->c_flags & SMR_LAZY) == 0,
490 ("smr_poll: Blocking not allowed on lazy smrs."));
493 * Use a critical section so that we can avoid ABA races
494 * caused by long preemption sleeps.
498 /* Attempt to load from self only once. */
499 self = zpcpu_get(smr);
501 flags = self->c_flags;
502 counter_u64_add_protected(poll, 1);
505 * Conditionally advance the lazy write clock on any writer
508 if ((flags & SMR_LAZY) != 0)
509 smr_lazy_advance(smr, s);
512 * Acquire barrier loads s_wr_seq after s_rd_seq so that we can not
513 * observe an updated read sequence that is larger than write.
515 s_rd_seq = atomic_load_acq_int(&s->s_rd_seq);
518 * If we have already observed the sequence number we can immediately
519 * return success. Most polls should meet this criterion.
521 if (SMR_SEQ_LEQ(goal, s_rd_seq))
525 * wr_seq must be loaded prior to any c_seq value so that a
526 * stale c_seq can only reference time after this wr_seq.
528 s_wr_seq = atomic_load_acq_int(&s->s_wr.seq);
531 * This is the distance from s_wr_seq to goal. Positive values
534 delta = SMR_SEQ_DELTA(goal, s_wr_seq);
537 * Detect a stale wr_seq.
539 * This goal may have come from a deferred advance or a lazy
540 * smr. If we are not blocking we can not succeed but the
541 * sequence number is valid.
543 if (delta > 0 && delta <= SMR_SEQ_ADVANCE &&
544 (flags & (SMR_LAZY | SMR_DEFERRED)) != 0) {
549 /* LAZY is always !wait. */
550 s_wr_seq = smr_shared_advance(s);
555 * Detect an invalid goal.
557 * The goal must be in the range of s_wr_seq >= goal >= s_rd_seq for
558 * it to be valid. If it is not then the caller held on to it and
559 * the integer wrapped. If we wrapped back within range the caller
560 * will harmlessly scan.
565 /* Determine the lowest visible sequence number. */
566 s_rd_seq = smr_poll_scan(smr, s, s_rd_seq, s_wr_seq, goal, wait);
567 success = SMR_SEQ_LEQ(goal, s_rd_seq);
570 counter_u64_add_protected(poll_fail, 1);
574 * Serialize with smr_advance()/smr_exit(). The caller is now free
575 * to modify memory as expected.
577 atomic_thread_fence_acq();
579 KASSERT(success || !wait, ("%s: blocking poll failed", __func__));
584 smr_create(const char *name, int limit, int flags)
590 s = uma_zalloc(smr_shared_zone, M_WAITOK);
591 smr = uma_zalloc_pcpu(smr_zone, M_WAITOK);
594 s->s_rd_seq = s->s_wr.seq = SMR_SEQ_INIT;
595 s->s_wr.ticks = ticks;
597 /* Initialize all CPUS, not just those running. */
598 for (i = 0; i <= mp_maxid; i++) {
599 c = zpcpu_get_cpu(smr, i);
600 c->c_seq = SMR_SEQ_INVALID;
606 atomic_thread_fence_seq_cst();
612 smr_destroy(smr_t smr)
615 smr_synchronize(smr);
616 uma_zfree(smr_shared_zone, smr->c_shared);
617 uma_zfree_pcpu(smr_zone, smr);
621 * Initialize the UMA slab zone.
627 smr_shared_zone = uma_zcreate("SMR SHARED", sizeof(struct smr_shared),
628 NULL, NULL, NULL, NULL, (CACHE_LINE_SIZE * 2) - 1, 0);
629 smr_zone = uma_zcreate("SMR CPU", sizeof(struct smr),
630 NULL, NULL, NULL, NULL, (CACHE_LINE_SIZE * 2) - 1, UMA_ZONE_PCPU);