Emilio G. Cota <c...@braap.org> writes: > This is a hash table with optional auto-resizing and MRU promotion for > reads and writes. Its implementation goal is to stay fast while > scaling for read-mostly workloads. > > A hash table with these features will be necessary for the scalability > of the ongoing MTTCG work; before those changes arrive we can already > benefit from the single-threaded speedup that qht also provides. > > Signed-off-by: Emilio G. Cota <c...@braap.org> > --- > include/qemu/qht.h | 150 ++++++++++++++++++ > util/Makefile.objs | 2 +- > util/qht.c | 458 > +++++++++++++++++++++++++++++++++++++++++++++++++++++ > 3 files changed, 609 insertions(+), 1 deletion(-) > create mode 100644 include/qemu/qht.h > create mode 100644 util/qht.c > > diff --git a/include/qemu/qht.h b/include/qemu/qht.h > new file mode 100644 > index 0000000..63244e4 > --- /dev/null > +++ b/include/qemu/qht.h > @@ -0,0 +1,150 @@ > +/* > + * Copyright (C) 2016, Emilio G. Cota <c...@braap.org> > + * > + * License: GNU GPL, version 2 or later. > + * See the COPYING file in the top-level directory. > + */ > +#ifndef QHT_H > +#define QHT_H > + > +#include <stdbool.h> > +#include <assert.h> > +#include <stdint.h> > +#include <stdio.h> > + > +#include "qemu/osdep.h" > +#include "qemu-common.h" > +#include "qemu/spinlock.h" > +#include "qemu/seqlock.h" > +#include "qemu/rcu.h" > + > +#if HOST_LONG_BITS == 32 > +#define QHT_BUCKET_ENTRIES 6 > +#else /* 64-bit */ > +#define QHT_BUCKET_ENTRIES 4 > +#endif
Are these purely a function of the total size of qht_bucket bellow and keeping it nicely cacheline aligned? > + > +struct qht_bucket { > + QemuSpinLock lock; > + QemuSeqLock sequence; > + uint32_t hashes[QHT_BUCKET_ENTRIES]; > + void *pointers[QHT_BUCKET_ENTRIES]; > + struct qht_bucket *next; > +} QEMU_CACHELINE_ALIGNED; For the TLB code we have compile time checks QEMU_BUILD_BUG_ON() to warn the developers if we go over a boundary. Might be worth adding a similar here. > +struct qht_map { > + struct qht_bucket *buckets; > + uint64_t n; > + uint64_t n_items; > + uint64_t n_items_threshold; > + struct rcu_head rcu; > +}; > + > +struct qht { > + struct qht_map *map; > + unsigned int mode; > +}; > + > +typedef bool (*qht_lookup_func_t)(const void *obj, const void *userp); > +typedef void (*qht_iter_func_t)(struct qht *ht, void *p, uint32_t h, void > *up); > + > +#define QHT_MODE_MRU_LOOKUP 0x1 /* move looked-up items to head */ > +#define QHT_MODE_MRU_INSERT 0x2 /* insert new elements at the head */ > +#define QHT_MODE_AUTO_RESIZE 0x4 /* auto-resize when heavily loaded */ > + > +void qht_init(struct qht *ht, uint64_t n_elems, unsigned int mode); > + > +/* call only when there are no readers left */ > +void qht_destroy(struct qht *ht); > + What's the rationale for making the caller responsible for locking here? Are we going to be protected by tb_lock in the MTTCG case rather than a finer grained locking inside the qht? > +/* call with an external lock held */ > +void qht_reset(struct qht *ht); > + > +/* call with an external lock held */ > +void qht_reset_size(struct qht *ht, uint64_t n_elems); > + > +/* call with an external lock held */ > +void qht_insert(struct qht *ht, void *p, uint32_t hash); > + > +/* call with an external lock held */ > +bool qht_remove(struct qht *ht, const void *p, uint32_t hash); > + > +/* call with an external lock held */ > +void qht_iter(struct qht *ht, qht_iter_func_t func, void *userp); > + > +/* call with an external lock held */ > +void qht_grow(struct qht *ht); > + > +void qht_bucket_mru(struct qht_bucket *b, struct qht_bucket *orig, > + const void *p, int pos); > + > +static inline > +struct qht_bucket *qht_map_to_bucket(struct qht_map *map, uint32_t hash) > +{ > + return &map->buckets[hash & (map->n - 1)]; > +} > + > +static inline > +void *__qht_lookup(struct qht_bucket *b, struct qht_bucket **far_bucket, > + int *pos, qht_lookup_func_t func, const void *userp, > + uint32_t hash) Aside the inline is there any reason to keep such an implementation detail in the header. I certainly couldn't measure a difference after I moved them into qht.c. > +{ > + unsigned int count = 0; > + int i; > + > + do { > + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { > + if (b->hashes[i] == hash) { > + void *p = atomic_read(&b->pointers[i]); > + > + /* > + * p could be NULL if we're racing with a writer. We could > use > + * barriers here but for performance we only issue the ones > + * in the seqlock. > + */ > + if (likely(p && func(p, userp))) { A comment about saving the position for MRU might be worthwhile here. > + if (unlikely(count)) { > + *far_bucket = b; > + *pos = i; > + } > + return p; > + } > + } > + } > + count++; > + b = atomic_read(&b->next); > + /* > + * This barrier guarantees that we will read a properly initialized > + * b->next; it is paired with an smp_wmb() before setting b->next. > + */ > + smp_rmb(); > + } while (b); > + return NULL; > +} > + > +static inline void *qht_lookup(struct qht *ht, qht_lookup_func_t func, > + const void *userp, uint32_t hash) > +{ > + struct qht_bucket *far_bucket = NULL; > + struct qht_bucket *b; > + struct qht_map *map; > + uint32_t version; > + int pos = 0; > + void *ret; > + > + map = atomic_read(&ht->map); > + /* paired with smp_wmb() before setting ht->map */ > + smp_rmb(); > + b = qht_map_to_bucket(map, hash); > + > + do { > + version = seqlock_read_begin(&b->sequence); > + ret = __qht_lookup(b, &far_bucket, &pos, func, userp, hash); > + } while (seqlock_read_retry(&b->sequence, version)); OK I was slightly confused by this until I got further along. Maybe some comments in the qht_bucket structure to point out the seqlock is important when a bucket is at the head of the list. > + if ((ht->mode & QHT_MODE_MRU_LOOKUP) && unlikely(far_bucket)) { > + qht_bucket_mru(b, far_bucket, ret, pos); > + } > + return ret; > +} > + > +#endif /* QHT_H */ > diff --git a/util/Makefile.objs b/util/Makefile.objs > index a8a777e..ae893dc 100644 > --- a/util/Makefile.objs > +++ b/util/Makefile.objs > @@ -1,4 +1,4 @@ > -util-obj-y = osdep.o cutils.o unicode.o qemu-timer-common.o > +util-obj-y = osdep.o cutils.o unicode.o qemu-timer-common.o qht.o > util-obj-$(CONFIG_POSIX) += compatfd.o > util-obj-$(CONFIG_POSIX) += event_notifier-posix.o > util-obj-$(CONFIG_POSIX) += mmap-alloc.o > diff --git a/util/qht.c b/util/qht.c > new file mode 100644 > index 0000000..590cd8a > --- /dev/null > +++ b/util/qht.c > @@ -0,0 +1,458 @@ > +/* > + * qht.c - QEMU Hash Table, designed to scale for read-mostly workloads. > + * > + * Copyright (C) 2016, Emilio G. Cota <c...@braap.org> > + * > + * License: GNU GPL, version 2 or later. > + * See the COPYING file in the top-level directory. > + * > + * Assumptions: > + * - Writers and iterators must take an external lock. > + * - A hash value of 0 is invalid. > + * - NULL cannot be inserted as a pointer value. > + * > + * Features: > + * - Optional auto-resizing: the hash table resizes up if the load surpasses > + * a certain threshold. Resizing is done concurrently with readers. > + * - Optional bucket MRU promotion policy. > + * > + * The key structure is the bucket, which is cacheline-sized. Buckets > + * contain a few hash values and pointers; the u32 hash values are stored in > + * full so that resizing is fast. Having this structure instead of directly > + * chaining items has three advantages: > + * - Failed lookups fail fast, and touch a minimum number of cache lines. > + * - Resizing the hash table with concurrent lookups is easy. > + * - We can have a few Most-Recently-Used (MRU) hash-pointer pairs in the > same > + * head bucket. This helps scalability, since MRU promotions (i.e. writes > to > + * the bucket) become less common. > + * > + * For concurrent lookups we use a per-bucket seqlock; per-bucket spinlocks > + * allow readers (lookups) to upgrade to writers and thus implement an MRU > + * promotion policy; these MRU-induced writes do not touch the cache lines of > + * other head buckets. > + * > + * Note that there are two types of buckets: > + * 1. "head" buckets are the ones allocated in the array of buckets in > qht_map. > + * 2. all "non-head" buckets (i.e. all others) are members of a chain that > + * starts from a head bucket. > + * Note that the seqlock and spinlock of a head bucket applies to all buckets > + * chained to it; these two fields are unused in non-head buckets. > + * > + * Resizing is done by taking all spinlocks (so that no > readers-turned-writers > + * can race with us) and then placing all elements into a new hash table. > Last, > + * the ht->map pointer is set, and the old map is freed once no RCU readers > can > + * see it anymore. > + * > + * Related Work: > + * - Idea of cacheline-sized buckets with full hashes taken from: > + * David, Guerraoui & Trigonakis, "Asynchronized Concurrency: > + * The Secret to Scaling Concurrent Search Data Structures", ASPLOS'15. > + * - Why not RCU-based hash tables? They would allow us to get rid of the > + * seqlock, but resizing would take forever since RCU read critical > + * sections in QEMU take quite a long time. > + * More info on relativistic hash tables: > + * + Triplett, McKenney & Walpole, "Resizable, Scalable, Concurrent Hash > + * Tables via Relativistic Programming", USENIX ATC'11. > + * + Corbet, "Relativistic hash tables, part 1: Algorithms", @ lwn.net, > 2014. > + * https://lwn.net/Articles/612021/ Very helpful links, thanks ;-) > + */ > +#include "qemu/qht.h" > +#include "qemu/atomic.h" > + > +static inline uint64_t qht_elems_to_buckets(uint64_t n_elems) > +{ > + return pow2ceil(n_elems / QHT_BUCKET_ENTRIES); > +} > + > +static inline void qht_head_init(struct qht_bucket *b) > +{ > + memset(b, 0, sizeof(*b)); > + qemu_spinlock_init(&b->lock); > + seqlock_init(&b->sequence); > +} > + > +static inline void qht_chain_destroy(struct qht_bucket *head) > +{ > + struct qht_bucket *curr = head->next; > + struct qht_bucket *prev; > + > + while (curr) { > + prev = curr; > + curr = curr->next; > + qemu_vfree(prev); > + } > +} > + > +/* pass only an orphan map */ > +static void qht_map_destroy(struct qht_map *map) > +{ > + uint64_t i; > + > + for (i = 0; i < map->n; i++) { > + qht_chain_destroy(&map->buckets[i]); > + } > + qemu_vfree(map->buckets); > + g_free(map); > +} > + > +static void qht_map_reclaim(struct rcu_head *rcu) > +{ > + struct qht_map *map = container_of(rcu, struct qht_map, rcu); > + > + qht_map_destroy(map); > +} > + > +static struct qht_map *qht_map_create(uint64_t n) > +{ > + struct qht_map *map; > + uint64_t i; > + > + assert(n <= (1ULL << 32)); /* we're using a 32-bit hash func */ > + map = g_malloc(sizeof(*map)); > + map->n = n; > + map->n_items = 0; > + map->n_items_threshold = n * QHT_BUCKET_ENTRIES / 2; > + map->buckets = qemu_memalign(QEMU_CACHELINE, sizeof(*map->buckets) * n); > + for (i = 0; i < n; i++) { > + qht_head_init(&map->buckets[i]); > + } > + return map; > +} > + > +static inline void qht_publish(struct qht *ht, struct qht_map *new) > +{ > + /* Readers should see a properly initialized map; pair with smp_rmb() */ > + smp_wmb(); > + atomic_set(&ht->map, new); > +} > + > +void qht_init(struct qht *ht, uint64_t n_elems, unsigned int mode) > +{ > + struct qht_map *map; > + uint64_t n = qht_elems_to_buckets(n_elems); > + > + map = qht_map_create(n); > + ht->mode = mode; > + qht_publish(ht, map); > +} > + > +/* call only when there are no readers left */ > +void qht_destroy(struct qht *ht) > +{ > + qht_map_destroy(ht->map); > + memset(ht, 0, sizeof(*ht)); > +} > + > +static void __qht_bucket_reset(struct qht_bucket *b) I think Paolo has already mentioned __ names. > +{ > + int i; > + > + do { > + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { > + atomic_set(&b->hashes[i], 0); > + atomic_set(&b->pointers[i], NULL); > + } > + b = b->next; > + } while (b); > +} > + > +static void qht_bucket_reset(struct qht_bucket *b) > +{ > + qemu_spinlock_lock(&b->lock); > + seqlock_write_begin(&b->sequence); > + __qht_bucket_reset(b); Personally I'd just import the __qht_bucket_reset into here to make it clear the seqlock protects us from using values halfway from being reset. Otherwise a quick comment to repeat the head text that effect on the above function. > + seqlock_write_end(&b->sequence); > + qemu_spinlock_unlock(&b->lock); > +} > + > +/* call with an external lock held */ > +void qht_reset(struct qht *ht) > +{ > + struct qht_map *map = ht->map; > + uint64_t i; > + > + for (i = 0; i < map->n; i++) { > + qht_bucket_reset(&map->buckets[i]); > + } > +} > + > +/* call with an external lock held */ > +void qht_reset_size(struct qht *ht, uint64_t n_elems) > +{ > + struct qht_map *old = ht->map; > + > + qht_reset(ht); > + if (old->n == qht_elems_to_buckets(n_elems)) { > + return; > + } > + qht_init(ht, n_elems, ht->mode); > + call_rcu1(&old->rcu, qht_map_reclaim); > +} > + > +/* Can only be called when at least orig is the 3rd link i.e. > head->2nd->orig */ > +static void > +__qht_move_bucket_to_front(struct qht_bucket *head, struct qht_bucket *orig) > +{ > + struct qht_bucket *b = head->next; > + > + for (;;) { > + if (b->next == orig) { > + atomic_set(&b->next, orig->next); > + atomic_set(&orig->next, head->next); > + atomic_set(&head->next, orig); > + return; > + } > + b = b->next; > + } > +} > + > +static inline void __qht_bucket_mru_head(struct qht_bucket *b, int pos) > +{ > + uint32_t orig_hash = b->hashes[pos]; > + void *orig_p = b->pointers[pos]; > + int i; > + > + for (i = 0; i < pos; i++) { > + atomic_set(&b->hashes[i + 1], b->hashes[i]); > + atomic_set(&b->pointers[i + 1], b->pointers[i]); > + } > + atomic_set(&b->hashes[0], orig_hash); > + atomic_set(&b->pointers[0], orig_p); > +} > + > + > +/* call with head->lock held */ > +static inline void > +__qht_bucket_mru(struct qht_bucket *head, struct qht_bucket *orig, int pos) > +{ > + uint32_t *dest_hash; > + void **dest_p; > + void *p; > + uint32_t hash; > + int i; > + > + if (head == orig) { > + return __qht_bucket_mru_head(head, pos); > + } OK this is calling out for a comment because I'm having trouble following the logic. I think we are pushing the last entry in the head bucket out the bucket the MRU entry is coming from while shuffling the rest up for the new entry. > + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { > + if (i == QHT_BUCKET_ENTRIES - 1) { > + dest_hash = &orig->hashes[pos]; > + dest_p = &o rig->pointers[pos]; > + } else { > + dest_hash = &head->hashes[i + 1]; > + dest_p = &head->pointers[i + 1]; > + } > + hash = *dest_hash; > + p = *dest_p; > + > + atomic_set(dest_hash, head->hashes[i]); > + atomic_set(dest_p, head->pointers[i]); > + > + atomic_set(&head->hashes[i], hash); > + atomic_set(&head->pointers[i], p); > + } So the bucket that has just swapped the MRU entry for an evicted entry gets placed after head. Is there a chance we end up just bouncing the head->next bucket each hit? > + if (head->next != orig) { > + __qht_move_bucket_to_front(head, orig); > + } > +} > + > +void qht_bucket_mru(struct qht_bucket *head, struct qht_bucket *orig, > + const void *p, int pos) > +{ > + qemu_spinlock_lock(&head->lock); > + if (unlikely(orig->pointers[pos] != p)) { > + /* while we acquired the lock, the bucket was updated, so bail out */ > + goto out; > + } > + seqlock_write_begin(&head->sequence); > + __qht_bucket_mru(head, orig, pos); > + seqlock_write_end(&head->sequence); > + out: > + qemu_spinlock_unlock(&head->lock); > +} > + > +/* call with b->lock held */ > +static void __qht_insert(struct qht *ht, struct qht_map *map, > + struct qht_bucket *b, void *p, uint32_t hash) > +{ > + struct qht_bucket *head = b; > + struct qht_bucket *prev = NULL; > + struct qht_bucket *new = NULL; > + unsigned int count = 0; > + int i; > + > + for (;;) { > + if (b == NULL) { > + b = qemu_memalign(QEMU_CACHELINE, sizeof(*b)); > + memset(b, 0, sizeof(*b)); > + new = b; > + } > + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { > + if (b->hashes[i]) { > + continue; > + } > + /* found an empty key: acquire the seqlock and write */ > + seqlock_write_begin(&head->sequence); > + if (new) { > + /* > + * This barrier is paired with smp_rmb() after reading > + * b->next when not holding b->lock. > + */ > + smp_wmb(); > + atomic_set(&prev->next, b); > + } > + atomic_set(&b->hashes[i], hash); > + atomic_set(&b->pointers[i], p); > + if ((ht->mode & QHT_MODE_MRU_INSERT) && count) { > + __qht_bucket_mru(head, b, i); > + } > + seqlock_write_end(&head->sequence); > + map->n_items++; > + return; > + } > + prev = b; > + b = b->next; > + count++; > + } > +} > + > +/* call with an external lock held */ > +void qht_insert(struct qht *ht, void *p, uint32_t hash) > +{ > + struct qht_map *map = ht->map; > + struct qht_bucket *b = qht_map_to_bucket(map, hash); > + > + qemu_spinlock_lock(&b->lock); > + __qht_insert(ht, map, b, p, hash); > + qemu_spinlock_unlock(&b->lock); > + > + if ((ht->mode & QHT_MODE_AUTO_RESIZE) && > + unlikely(map->n_items > map->n_items_threshold)) { > + qht_grow(ht); > + } > +} > + > +/* call with b->lock held */ > +static inline bool __qht_remove(struct qht_map *map, struct qht_bucket *b, > + const void *p, uint32_t hash) > +{ > + int i; > + > + do { > + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { > + if (b->hashes[i] == hash && b->pointers[i] == p) { > + atomic_set(&b->hashes[i], 0); > + atomic_set(&b->pointers[i], NULL); > + map->n_items--; > + return true; > + } > + } > + b = b->next; > + } while (b); > + return false; > +} > + > +/* call with an external lock held */ > +bool qht_remove(struct qht *ht, const void *p, uint32_t hash) > +{ > + struct qht_map *map = ht->map; > + struct qht_bucket *b = qht_map_to_bucket(map, hash); > + bool ret; > + > + qemu_spinlock_lock(&b->lock); > + seqlock_write_begin(&b->sequence); > + ret = __qht_remove(map, b, p, hash); > + seqlock_write_end(&b->sequence); > + qemu_spinlock_unlock(&b->lock); > + return ret; > +} > + > +/* > + * acquire all spinlocks from a map, so that writers cannot race with > + * readers-turned-writers. > + */ > +static void qht_lock(struct qht *ht) > +{ > + struct qht_map *map = ht->map; > + uint64_t i; > + > + /* if readers cannot upgrade, do nothing */ > + if (!(ht->mode & QHT_MODE_MRU_LOOKUP)) { > + return; > + } > + for (i = 0; i < map->n; i++) { > + struct qht_bucket *b = &map->buckets[i]; > + > + qemu_spinlock_lock(&b->lock); > + } > +} > + > +static void qht_unlock(struct qht *ht) > +{ > + struct qht_map *map = ht->map; > + uint64_t i; > + > + if (!(ht->mode & QHT_MODE_MRU_LOOKUP)) { > + return; > + } > + for (i = 0; i < map->n; i++) { > + struct qht_bucket *b = &map->buckets[i]; > + > + qemu_spinlock_unlock(&b->lock); > + } > +} > + > +/* external lock + all of the map's locks held (if !MRU_LOOKUP) */ > +static inline void __qht_map_iter(struct qht *ht, struct qht_map *map, > + qht_iter_func_t func, void *userp) > +{ > + uint64_t i; > + > + for (i = 0; i < map->n; i++) { > + struct qht_bucket *b = &map->buckets[i]; > + int j; > + > + do { > + for (j = 0; j < QHT_BUCKET_ENTRIES; j++) { > + if (b->hashes[j]) { > + func(ht, b->pointers[j], b->hashes[j], userp); > + } > + } > + b = b->next; > + } while (b); > + } > +} > + > +/* call with an external lock held */ > +void qht_iter(struct qht *ht, qht_iter_func_t func, void *userp) > +{ > + qht_lock(ht); > + __qht_map_iter(ht, ht->map, func, userp); > + qht_unlock(ht); > +} > + > +static void qht_map_copy(struct qht *ht, void *p, uint32_t hash, void *userp) > +{ > + struct qht_map *new = userp; > + struct qht_bucket *b = qht_map_to_bucket(new, hash); > + > + /* no need to acquire b->lock because no thread has seen this map yet */ > + __qht_insert(ht, new, b, p, hash); > +} > + > +/* call with an external lock held */ > +void qht_grow(struct qht *ht) > +{ > + struct qht_map *old = ht->map; > + struct qht_map *new; > + uint64_t n = old->n * 2; > + > + if (unlikely(n > (1ULL << 32))) { > + return; > + } > + new = qht_map_create(n); > + qht_iter(ht, qht_map_copy, new); > + > + qht_publish(ht, new); > + call_rcu1(&old->rcu, qht_map_reclaim); > +} This looks good and I can see this being a useful utility function across QEMU. My main problem was following exactly what happens when entries are moved around for the MRU case. As users don't have to have a deep knowledge of the implementation details this isn't a major issue although a few more expansive comments or ASCII diagrams may help if you are feeling up to it ;-) -- Alex Bennée