From patchwork Wed May 25 01:13:17 2016 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Emilio Cota X-Patchwork-Id: 9134575 Return-Path: Received: from mail.wl.linuxfoundation.org (pdx-wl-mail.web.codeaurora.org [172.30.200.125]) by pdx-korg-patchwork.web.codeaurora.org (Postfix) with ESMTP id 470B060221 for ; Wed, 25 May 2016 01:23:52 +0000 (UTC) Received: from mail.wl.linuxfoundation.org (localhost [127.0.0.1]) by mail.wl.linuxfoundation.org (Postfix) with ESMTP id 36E1627BCA for ; Wed, 25 May 2016 01:23:52 +0000 (UTC) Received: by mail.wl.linuxfoundation.org (Postfix, from userid 486) id 2AE2A28164; Wed, 25 May 2016 01:23:52 +0000 (UTC) X-Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on pdx-wl-mail.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-6.8 required=2.0 tests=BAYES_00,DKIM_SIGNED, RCVD_IN_DNSWL_HI,T_DKIM_INVALID autolearn=ham version=3.3.1 Received: from lists.gnu.org (lists.gnu.org [208.118.235.17]) (using TLSv1 with cipher AES256-SHA (256/256 bits)) (No client certificate requested) by mail.wl.linuxfoundation.org (Postfix) with ESMTPS id 5283927BCA for ; Wed, 25 May 2016 01:23:50 +0000 (UTC) Received: from localhost ([::1]:56530 helo=lists.gnu.org) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1b5NY5-0001Jf-DI for patchwork-qemu-devel@patchwork.kernel.org; Tue, 24 May 2016 21:23:49 -0400 Received: from eggs.gnu.org ([2001:4830:134:3::10]:49855) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1b5NOo-0001m8-0p for qemu-devel@nongnu.org; Tue, 24 May 2016 21:14:28 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1b5NOH-0007g9-UZ for qemu-devel@nongnu.org; Tue, 24 May 2016 21:14:13 -0400 Received: from out2-smtp.messagingengine.com ([66.111.4.26]:53131) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1b5NOF-0007cX-Ft for qemu-devel@nongnu.org; Tue, 24 May 2016 21:13:41 -0400 Received: from compute1.internal (compute1.nyi.internal [10.202.2.41]) by mailout.nyi.internal (Postfix) with ESMTP id 8A95F20CA8; Tue, 24 May 2016 21:13:26 -0400 (EDT) Received: from frontend2 ([10.202.2.161]) by compute1.internal (MEProxy); Tue, 24 May 2016 21:13:26 -0400 DKIM-Signature: v=1; a=rsa-sha1; c=relaxed/relaxed; d=braap.org; h=cc :date:from:in-reply-to:message-id:references:subject:to :x-sasl-enc:x-sasl-enc; s=mesmtp; bh=pdME4VRdSu+ekyXvrGavnQ3cpaY =; b=xfqLGY5AibSLrOia535EgMToGfvbZNjAzljl1rzPx3Ex8wNuIfoY+ncYihg gkjwuzuzUUDWhp3XPOj5Cmb6kQ3+tAkfFXBmVtmkbSKhv00Ra/bHn0jF6RpZw7mP 9IHzpsiYokfSPyvPbPNWrTUIA0mduea0Yu+jr5ZuHe4Hh1OM= DKIM-Signature: v=1; a=rsa-sha1; c=relaxed/relaxed; d= messagingengine.com; h=cc:date:from:in-reply-to:message-id :references:subject:to:x-sasl-enc:x-sasl-enc; s=smtpout; bh=pdME 4VRdSu+ekyXvrGavnQ3cpaY=; b=gWb5fSaC1N8cy2berryQ5csO+BK9g875j3KJ Rnss8mQqDxn0tqiSzDNkG96wNuV5YDut3KsBfyvrL8cVtM163Np/IhlbXtywOq4q dSpegX06bCIT0jYOTBX1gaJQ4jbWZ41HQWcV3f21qOf+I2pO/zlyKUp969OV69mX ZgIdsg8= X-Sasl-enc: dio164kZEDVUnxKENbkYAW+KxRWvZPCdXQJSF1Eh2RTK 1464138806 Received: from localhost (flamenco.cs.columbia.edu [128.59.20.216]) by mail.messagingengine.com (Postfix) with ESMTPA id 3CD6FCCD2E; Tue, 24 May 2016 21:13:26 -0400 (EDT) From: "Emilio G. Cota" To: QEMU Developers , MTTCG Devel Date: Tue, 24 May 2016 21:13:17 -0400 Message-Id: <1464138802-23503-11-git-send-email-cota@braap.org> X-Mailer: git-send-email 2.5.0 In-Reply-To: <1464138802-23503-1-git-send-email-cota@braap.org> References: <1464138802-23503-1-git-send-email-cota@braap.org> X-detected-operating-system: by eggs.gnu.org: GNU/Linux 2.2.x-3.x [generic] X-Received-From: 66.111.4.26 Subject: [Qemu-devel] [PATCH v6 10/15] qht: QEMU's fast, resizable and scalable Hash Table X-BeenThere: qemu-devel@nongnu.org X-Mailman-Version: 2.1.21 Precedence: list List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: Paolo Bonzini , Sergey Fedorov , =?UTF-8?q?Alex=20Benn=C3=A9e?= , Richard Henderson Errors-To: qemu-devel-bounces+patchwork-qemu-devel=patchwork.kernel.org@nongnu.org Sender: "Qemu-devel" X-Virus-Scanned: ClamAV using ClamSMTP This is a fast, scalable chained hash table with optional auto-resizing, allowing reads that are concurrent with reads, and reads/writes that are concurrent with writes to separate buckets. 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 --- include/qemu/qht.h | 183 ++++++++++++ util/Makefile.objs | 1 + util/qht.c | 837 +++++++++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 1021 insertions(+) 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..aec60aa --- /dev/null +++ b/include/qemu/qht.h @@ -0,0 +1,183 @@ +/* + * Copyright (C) 2016, Emilio G. Cota + * + * License: GNU GPL, version 2 or later. + * See the COPYING file in the top-level directory. + */ +#ifndef QEMU_QHT_H +#define QEMU_QHT_H + +#include "qemu/osdep.h" +#include "qemu/seqlock.h" +#include "qemu/thread.h" +#include "qemu/qdist.h" + +struct qht { + struct qht_map *map; + QemuMutex lock; /* serializes setters of ht->map */ + unsigned int mode; +}; + +/** + * struct qht_stats - Statistics of a QHT + * @head_buckets: number of head buckets + * @used_head_buckets: number of non-empty head buckets + * @entries: total number of entries + * @chain: frequency distribution representing the number of buckets in each + * chain, excluding empty chains. + * @occupancy: frequency distribution representing chain occupancy rate. + * Valid range: from 0.0 (empty) to 1.0 (full occupancy). + * + * An entry is a pointer-hash pair. + * Each bucket can host several entries. + * Chains are chains of buckets, whose first link is always a head bucket. + */ +struct qht_stats { + size_t head_buckets; + size_t used_head_buckets; + size_t entries; + struct qdist chain; + struct qdist occupancy; +}; + +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_AUTO_RESIZE 0x1 /* auto-resize when heavily loaded */ + +/** + * qht_init - Initialize a QHT + * @ht: QHT to be initialized + * @n_elems: number of entries the hash table should be optimized for. + * @mode: bitmask with OR'ed QHT_MODE_* + */ +void qht_init(struct qht *ht, size_t n_elems, unsigned int mode); + +/** + * qht_destroy - destroy a previously initialized QHT + * @ht: QHT to be destroyed + * + * Call only when there are no readers/writers left. + */ +void qht_destroy(struct qht *ht); + +/** + * qht_insert - Insert a pointer into the hash table + * @ht: QHT to insert to + * @p: pointer to be inserted + * @hash: hash corresponding to @p + * + * Attempting to insert a NULL @p is a bug. + * Inserting the same pointer @p with different @hash values is a bug. + * + * Returns true on sucess. + * Returns false if the @p-@hash pair already exists in the hash table. + */ +bool qht_insert(struct qht *ht, void *p, uint32_t hash); + +/** + * qht_lookup - Look up a pointer in a QHT + * @ht: QHT to be looked up + * @func: function to compare existing pointers against @userp + * @userp: pointer to pass to @func + * @hash: hash of the pointer to be looked up + * + * Needs to be called under an RCU read-critical section. + * + * The user-provided @func compares pointers in QHT against @userp. + * If the function returns true, a match has been found. + * + * Returns the corresponding pointer when a match is found. + * Returns NULL otherwise. + */ +void *qht_lookup(struct qht *ht, qht_lookup_func_t func, const void *userp, + uint32_t hash); + +/** + * qht_remove - remove a pointer from the hash table + * @ht: QHT to remove from + * @p: pointer to be removed + * @hash: hash corresponding to @p + * + * Attempting to remove a NULL @p is a bug. + * + * Just-removed @p pointers cannot be immediately freed; they need to remain + * valid until the end of the RCU grace period in which qht_remove() is called. + * This guarantees that concurrent lookups will always compare against valid + * data. + * + * Returns true on success. + * Returns false if the @p-@hash pair was not found. + */ +bool qht_remove(struct qht *ht, const void *p, uint32_t hash); + +/** + * qht_reset - reset a QHT + * @ht: QHT to be reset + * + * All entries in the hash table are reset. No resizing is performed. + * + * If concurrent readers may exist, the objects pointed to by the hash table + * must remain valid for the existing RCU grace period -- see qht_remove(). + * See also: qht_reset_size() + */ +void qht_reset(struct qht *ht); + +/** + * qht_reset_size - reset and resize a QHT + * @ht: QHT to be reset and resized + * @n_elems: number of entries the resized hash table should be optimized for. + * + * Returns true if the resize was necessary and therefore performed. + * Returns false otherwise. + * + * If concurrent readers may exist, the objects pointed to by the hash table + * must remain valid for the existing RCU grace period -- see qht_remove(). + * See also: qht_reset(), qht_resize(). + */ +bool qht_reset_size(struct qht *ht, size_t n_elems); + +/** + * qht_resize - resize a QHT + * @ht: QHT to be resized + * @n_elems: number of entries the resized hash table should be optimized for + * + * Returns true on success. + * Returns false if the resize was not necessary and therefore not performed. + * See also: qht_reset_size(). + */ +bool qht_resize(struct qht *ht, size_t n_elems); + +/** + * qht_iter - Iterate over a QHT + * @ht: QHT to be iterated over + * @func: function to be called for each entry in QHT + * @userp: additional pointer to be passed to @func + * + * Each time it is called, user-provided @func is passed a pointer-hash pair, + * plus @userp. + */ +void qht_iter(struct qht *ht, qht_iter_func_t func, void *userp); + +/** + * qht_statistics_init - Gather statistics from a QHT + * @ht: QHT to gather statistics from + * @stats: pointer to a struct qht_stats to be filled in + * + * Does NOT need to be called under an RCU read-critical section, + * since it does not dereference any pointers stored in the hash table. + * + * When done with @stats, pass the struct to qht_statistics_destroy(). + * Failing to do this will leak memory. + */ +void qht_statistics_init(struct qht *ht, struct qht_stats *stats); + +/** + * qht_statistics_destroy - Destroy a struct qht_stats + * @stats: stuct qht_stats to be destroyed + * + * See also: qht_statistics_init(). + */ +void qht_statistics_destroy(struct qht_stats *stats); + +#endif /* QEMU_QHT_H */ diff --git a/util/Makefile.objs b/util/Makefile.objs index 702435e..45f8794 100644 --- a/util/Makefile.objs +++ b/util/Makefile.objs @@ -33,3 +33,4 @@ util-obj-y += timed-average.o util-obj-y += base64.o util-obj-y += log.o util-obj-y += qdist.o +util-obj-y += qht.o diff --git a/util/qht.c b/util/qht.c new file mode 100644 index 0000000..ca5a620 --- /dev/null +++ b/util/qht.c @@ -0,0 +1,837 @@ +/* + * qht.c - QEMU Hash Table, designed to scale for read-mostly workloads. + * + * Copyright (C) 2016, Emilio G. Cota + * + * License: GNU GPL, version 2 or later. + * See the COPYING file in the top-level directory. + * + * Assumptions: + * - NULL cannot be inserted/removed as a pointer value. + * - Trying to insert an already-existing hash-pointer pair is OK. However, + * it is not OK to insert into the same hash table different hash-pointer + * pairs that have the same pointer value, but not the hashes. + * - Lookups are performed under an RCU read-critical section; removals + * must wait for a grace period to elapse before freeing removed objects. + * + * Features: + * - Reads (i.e. lookups and iterators) can be concurrent with other reads. + * Lookups that are concurrent with writes to the same bucket will retry + * via a seqlock; iterators acquire all bucket locks and therefore can be + * concurrent with lookups and are serialized wrt writers. + * - Writes (i.e. insertions/removals) can be concurrent with writes to + * different buckets; writes to the same bucket are serialized through a lock. + * - Optional auto-resizing: the hash table resizes up if the load surpasses + * a certain threshold. Resizing is done concurrently with readers; writes + * are serialized with the resize operation. + * + * 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 two advantages: + * - Failed lookups fail fast, and touch a minimum number of cache lines. + * - Resizing the hash table with concurrent lookups is easy. + * + * 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. + * + * On removals, we move the last valid item in the chain to the position of the + * just-removed entry. This makes lookups slightly faster, since the moment an + * invalid entry is found, the (failed) lookup is over. + * + * Resizing is done by taking all bucket spinlocks (so that no other writers can + * race with us) and then copying all entries into a new hash map. Then, the + * ht->map pointer is set, and the old map is freed once no RCU readers can see + * it anymore. + * + * Writers check for concurrent resizes by comparing ht->map before and after + * acquiring their bucket lock. If they don't match, a resize has occured + * while the bucket spinlock was being acquired. + * + * 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/ + */ +#include "qemu/qht.h" +#include "qemu/atomic.h" +#include "qemu/rcu.h" + +//#define QHT_DEBUG + +/* + * We want to avoid false sharing of cache lines. Most systems have 64-byte + * cache lines so we go with it for simplicity. + * + * Note that systems with smaller cache lines will be fine (the struct is + * almost 64-bytes); systems with larger cache lines might suffer from + * some false sharing. + */ +#define QHT_BUCKET_ALIGN 64 + +/* define these to keep sizeof(qht_bucket) within QHT_BUCKET_ALIGN */ +#if HOST_LONG_BITS == 32 +#define QHT_BUCKET_ENTRIES 6 +#else /* 64-bit */ +#define QHT_BUCKET_ENTRIES 4 +#endif + +/* + * Note: reading partially-updated pointers in @pointers could lead to + * segfaults. We thus access them with atomic_read/set; this guarantees + * that the compiler makes all those accesses atomic. We also need the + * volatile-like behavior in atomic_read, since otherwise the compiler + * might refetch the pointer. + * atomic_read's are of course not necessary when the bucket lock is held. + * + * If both ht->lock and b->lock are grabbed, ht->lock should always + * be grabbed first. + */ +struct qht_bucket { + QemuSpin lock; + QemuSeqLock sequence; + uint32_t hashes[QHT_BUCKET_ENTRIES]; + void *pointers[QHT_BUCKET_ENTRIES]; + struct qht_bucket *next; +} QEMU_ALIGNED(QHT_BUCKET_ALIGN); + +QEMU_BUILD_BUG_ON(sizeof(struct qht_bucket) > QHT_BUCKET_ALIGN); + +/** + * struct qht_map - structure to track an array of buckets + * @rcu: used by RCU. Keep it as the top field in the struct to help valgrind + * find the whole struct. + * @buckets: array of head buckets. It is constant once the map is created. + * @n_buckets: number of head buckets. It is constant once the map is created. + * @n_added_buckets: number of added (i.e. "non-head") buckets + * @n_added_buckets_threshold: threshold to trigger an upward resize once the + * number of added buckets surpasses it. + * + * Buckets are tracked in what we call a "map", i.e. this structure. + */ +struct qht_map { + struct rcu_head rcu; + struct qht_bucket *buckets; + size_t n_buckets; + size_t n_added_buckets; + size_t n_added_buckets_threshold; +}; + +/* trigger a resize when n_added_buckets > n_buckets / div */ +#define QHT_NR_ADDED_BUCKETS_THRESHOLD_DIV 8 + +static void qht_do_resize(struct qht *ht, struct qht_map *new); +static void qht_grow_maybe(struct qht *ht); + +#ifdef QHT_DEBUG + +#define qht_debug_assert(X) do { assert(X); } while (0) + +static void qht_bucket_debug__locked(struct qht_bucket *b) +{ + bool seen_empty = false; + bool corrupt = false; + int i; + + do { + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { + if (b->pointers[i] == NULL) { + seen_empty = true; + continue; + } + if (seen_empty) { + fprintf(stderr, "%s: b: %p, pos: %i, hash: 0x%x, p: %p\n", + __func__, b, i, b->hashes[i], b->pointers[i]); + corrupt = true; + } + } + b = b->next; + } while (b); + qht_debug_assert(!corrupt); +} + +static void qht_map_debug__all_locked(struct qht_map *map) +{ + int i; + + for (i = 0; i < map->n_buckets; i++) { + qht_bucket_debug__locked(&map->buckets[i]); + } +} +#else + +#define qht_debug_assert(X) do { (void)(X); } while (0) + +static inline void qht_bucket_debug__locked(struct qht_bucket *b) +{ } + +static inline void qht_map_debug__all_locked(struct qht_map *map) +{ } +#endif /* QHT_DEBUG */ + +static inline size_t qht_elems_to_buckets(size_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_spin_init(&b->lock); + seqlock_init(&b->sequence); +} + +static inline +struct qht_bucket *qht_map_to_bucket(struct qht_map *map, uint32_t hash) +{ + return &map->buckets[hash & (map->n_buckets - 1)]; +} + +/* acquire all bucket locks from a map */ +static void qht_map_lock_buckets(struct qht_map *map) +{ + size_t i; + + for (i = 0; i < map->n_buckets; i++) { + struct qht_bucket *b = &map->buckets[i]; + + qemu_spin_lock(&b->lock); + } +} + +static void qht_map_unlock_buckets(struct qht_map *map) +{ + size_t i; + + for (i = 0; i < map->n_buckets; i++) { + struct qht_bucket *b = &map->buckets[i]; + + qemu_spin_unlock(&b->lock); + } +} + +/* + * Call with at least a bucket lock held. + * @map should be the value read before acquiring the lock (or locks). + */ +static inline bool qht_map_is_stale__locked(struct qht *ht, struct qht_map *map) +{ + return map != ht->map; +} + +/* + * Grab all bucket locks, and set @pmap after making sure the map isn't stale. + * + * Pairs with qht_map_unlock_buckets(), hence the pass-by-reference. + * + * Note: callers cannot have ht->lock held. + */ +static inline +void qht_map_lock_buckets__no_stale(struct qht *ht, struct qht_map **pmap) +{ + struct qht_map *map; + + map = atomic_rcu_read(&ht->map); + qht_map_lock_buckets(map); + if (likely(!qht_map_is_stale__locked(ht, map))) { + *pmap = map; + return; + } + qht_map_unlock_buckets(map); + + /* we raced with a resize; acquire ht->lock to see the updated ht->map */ + qemu_mutex_lock(&ht->lock); + map = ht->map; + qht_map_lock_buckets(map); + qemu_mutex_unlock(&ht->lock); + *pmap = map; + return; +} + +/* + * Get a head bucket and lock it, making sure its parent map is not stale. + * @pmap is filled with a pointer to the bucket's parent map. + * + * Unlock with qemu_spin_unlock(&b->lock). + * + * Note: callers cannot have ht->lock held. + */ +static inline +struct qht_bucket *qht_bucket_lock__no_stale(struct qht *ht, uint32_t hash, + struct qht_map **pmap) +{ + struct qht_bucket *b; + struct qht_map *map; + + map = atomic_rcu_read(&ht->map); + b = qht_map_to_bucket(map, hash); + + qemu_spin_lock(&b->lock); + if (likely(!qht_map_is_stale__locked(ht, map))) { + *pmap = map; + return b; + } + qemu_spin_unlock(&b->lock); + + /* we raced with a resize; acquire ht->lock to see the updated ht->map */ + qemu_mutex_lock(&ht->lock); + map = ht->map; + b = qht_map_to_bucket(map, hash); + qemu_spin_lock(&b->lock); + qemu_mutex_unlock(&ht->lock); + *pmap = map; + return b; +} + +static inline bool qht_map_needs_resize(struct qht_map *map) +{ + return atomic_read(&map->n_added_buckets) > map->n_added_buckets_threshold; +} + +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) +{ + size_t i; + + for (i = 0; i < map->n_buckets; 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(size_t n_buckets) +{ + struct qht_map *map; + size_t i; + + map = g_malloc(sizeof(*map)); + map->n_buckets = n_buckets; + + map->n_added_buckets = 0; + map->n_added_buckets_threshold = n_buckets / + QHT_NR_ADDED_BUCKETS_THRESHOLD_DIV; + + /* let tiny hash tables to at least add one non-head bucket */ + if (unlikely(map->n_added_buckets_threshold == 0)) { + map->n_added_buckets_threshold = 1; + } + + map->buckets = qemu_memalign(QHT_BUCKET_ALIGN, + sizeof(*map->buckets) * n_buckets); + for (i = 0; i < n_buckets; i++) { + qht_head_init(&map->buckets[i]); + } + return map; +} + +void qht_init(struct qht *ht, size_t n_elems, unsigned int mode) +{ + struct qht_map *map; + size_t n_buckets = qht_elems_to_buckets(n_elems); + + ht->mode = mode; + qemu_mutex_init(&ht->lock); + map = qht_map_create(n_buckets); + atomic_rcu_set(&ht->map, map); +} + +/* call only when there are no readers/writers left */ +void qht_destroy(struct qht *ht) +{ + qht_map_destroy(ht->map); + memset(ht, 0, sizeof(*ht)); +} + +static void qht_bucket_reset__locked(struct qht_bucket *head) +{ + struct qht_bucket *b = head; + int i; + + seqlock_write_begin(&head->sequence); + do { + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { + if (b->pointers[i] == NULL) { + goto done; + } + b->hashes[i] = 0; + atomic_set(&b->pointers[i], NULL); + } + b = b->next; + } while (b); + done: + seqlock_write_end(&head->sequence); +} + +/* call with all bucket locks held */ +static void qht_map_reset__all_locked(struct qht_map *map) +{ + size_t i; + + for (i = 0; i < map->n_buckets; i++) { + qht_bucket_reset__locked(&map->buckets[i]); + } + qht_map_debug__all_locked(map); +} + +void qht_reset(struct qht *ht) +{ + struct qht_map *map; + + qht_map_lock_buckets__no_stale(ht, &map); + qht_map_reset__all_locked(map); + qht_map_unlock_buckets(map); +} + +bool qht_reset_size(struct qht *ht, size_t n_elems) +{ + struct qht_map *new; + struct qht_map *map; + size_t n_buckets; + bool resize = false; + + n_buckets = qht_elems_to_buckets(n_elems); + + qemu_mutex_lock(&ht->lock); + map = ht->map; + if (n_buckets != map->n_buckets) { + new = qht_map_create(n_buckets); + resize = true; + } + + qht_map_lock_buckets(map); + qht_map_reset__all_locked(map); + if (resize) { + qht_do_resize(ht, new); + } + qht_map_unlock_buckets(map); + qemu_mutex_unlock(&ht->lock); + + return resize; +} + +static inline +void *qht_do_lookup(struct qht_bucket *head, qht_lookup_func_t func, + const void *userp, uint32_t hash) +{ + struct qht_bucket *b = head; + int i; + + do { + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { + if (b->hashes[i] == hash) { + void *p = atomic_read(&b->pointers[i]); + + if (likely(p) && likely(func(p, userp))) { + return p; + } + } + } + b = atomic_rcu_read(&b->next); + } while (b); + + return NULL; +} + +static __attribute__((noinline)) +void *qht_lookup__slowpath(struct qht_bucket *b, qht_lookup_func_t func, + const void *userp, uint32_t hash) +{ + unsigned int version; + void *ret; + + do { + version = seqlock_read_begin(&b->sequence); + ret = qht_do_lookup(b, func, userp, hash); + } while (seqlock_read_retry(&b->sequence, version)); + return ret; +} + +void *qht_lookup(struct qht *ht, qht_lookup_func_t func, const void *userp, + uint32_t hash) +{ + struct qht_bucket *b; + struct qht_map *map; + unsigned int version; + void *ret; + + map = atomic_rcu_read(&ht->map); + b = qht_map_to_bucket(map, hash); + + version = seqlock_read_begin(&b->sequence); + ret = qht_do_lookup(b, func, userp, hash); + if (likely(!seqlock_read_retry(&b->sequence, version))) { + return ret; + } + /* + * Removing the do/while from the fastpath gives a 4% perf. increase when + * running a 100%-lookup microbenchmark. + */ + return qht_lookup__slowpath(b, func, userp, hash); +} + +/* call with head->lock held */ +static bool qht_insert__locked(struct qht *ht, struct qht_map *map, + struct qht_bucket *head, void *p, uint32_t hash, + bool *needs_resize) +{ + struct qht_bucket *b = head; + struct qht_bucket *prev = NULL; + struct qht_bucket *new = NULL; + int i; + + for (;;) { + if (b == NULL) { + b = qemu_memalign(QHT_BUCKET_ALIGN, sizeof(*b)); + memset(b, 0, sizeof(*b)); + new = b; + atomic_inc(&map->n_added_buckets); + if (unlikely(qht_map_needs_resize(map)) && needs_resize) { + *needs_resize = true; + } + } + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { + if (b->pointers[i]) { + if (unlikely(b->pointers[i] == p)) { + return false; + } + continue; + } + /* found an empty key: acquire the seqlock and write */ + seqlock_write_begin(&head->sequence); + if (new) { + atomic_rcu_set(&prev->next, b); + } + b->hashes[i] = hash; + atomic_set(&b->pointers[i], p); + seqlock_write_end(&head->sequence); + return true; + } + prev = b; + b = b->next; + } +} + +bool qht_insert(struct qht *ht, void *p, uint32_t hash) +{ + struct qht_bucket *b; + struct qht_map *map; + bool needs_resize = false; + bool ret; + + /* NULL pointers are not supported */ + qht_debug_assert(p); + + b = qht_bucket_lock__no_stale(ht, hash, &map); + ret = qht_insert__locked(ht, map, b, p, hash, &needs_resize); + qht_bucket_debug__locked(b); + qemu_spin_unlock(&b->lock); + + if (unlikely(needs_resize) && ht->mode & QHT_MODE_AUTO_RESIZE) { + qht_grow_maybe(ht); + } + return ret; +} + +static inline bool qht_entry_is_last(struct qht_bucket *b, int pos) +{ + if (pos == QHT_BUCKET_ENTRIES - 1) { + if (b->next == NULL) { + return true; + } + return b->next->pointers[0] == NULL; + } + return b->pointers[pos + 1] == NULL; +} + +static void +qht_entry_move(struct qht_bucket *to, int i, struct qht_bucket *from, int j) +{ + qht_debug_assert(!(to == from && i == j)); + qht_debug_assert(to->pointers[i]); + qht_debug_assert(from->pointers[j]); + + to->hashes[i] = from->hashes[j]; + atomic_set(&to->pointers[i], from->pointers[j]); + + from->hashes[j] = 0; + atomic_set(&from->pointers[j], NULL); +} + +/* + * Find the last valid entry in @head, and swap it with @orig[pos], which has + * just been invalidated. + */ +static inline void qht_bucket_fill_hole(struct qht_bucket *orig, int pos) +{ + struct qht_bucket *b = orig; + struct qht_bucket *prev = NULL; + int i; + + if (qht_entry_is_last(orig, pos)) { + orig->hashes[pos] = 0; + atomic_set(&orig->pointers[pos], NULL); + return; + } + do { + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { + if (b->pointers[i]) { + continue; + } + if (i > 0) { + return qht_entry_move(orig, pos, b, i - 1); + } + qht_debug_assert(prev); + return qht_entry_move(orig, pos, prev, QHT_BUCKET_ENTRIES - 1); + } + prev = b; + b = b->next; + } while (b); + /* no free entries other than orig[pos], so swap it with the last one */ + qht_entry_move(orig, pos, prev, QHT_BUCKET_ENTRIES - 1); +} + +/* call with b->lock held */ +static inline +bool qht_remove__locked(struct qht_map *map, struct qht_bucket *head, + const void *p, uint32_t hash) +{ + struct qht_bucket *b = head; + int i; + + do { + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { + void *q = b->pointers[i]; + + if (unlikely(q == NULL)) { + return false; + } + if (q == p) { + qht_debug_assert(b->hashes[i] == hash); + seqlock_write_begin(&head->sequence); + qht_bucket_fill_hole(b, i); + seqlock_write_end(&head->sequence); + return true; + } + } + b = b->next; + } while (b); + return false; +} + +bool qht_remove(struct qht *ht, const void *p, uint32_t hash) +{ + struct qht_bucket *b; + struct qht_map *map; + bool ret; + + /* NULL pointers are not supported */ + qht_debug_assert(p); + + b = qht_bucket_lock__no_stale(ht, hash, &map); + ret = qht_remove__locked(map, b, p, hash); + qht_bucket_debug__locked(b); + qemu_spin_unlock(&b->lock); + return ret; +} + +static inline void qht_bucket_iter(struct qht *ht, struct qht_bucket *b, + qht_iter_func_t func, void *userp) +{ + int i; + + do { + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { + if (b->pointers[i] == NULL) { + return; + } + func(ht, b->pointers[i], b->hashes[i], userp); + } + b = b->next; + } while (b); +} + +/* call with all of the map's locks held */ +static inline void qht_map_iter__all_locked(struct qht *ht, struct qht_map *map, + qht_iter_func_t func, void *userp) +{ + size_t i; + + for (i = 0; i < map->n_buckets; i++) { + qht_bucket_iter(ht, &map->buckets[i], func, userp); + } +} + +void qht_iter(struct qht *ht, qht_iter_func_t func, void *userp) +{ + struct qht_map *map; + + map = atomic_rcu_read(&ht->map); + qht_map_lock_buckets(map); + /* Note: ht here is merely for carrying ht->mode; ht->map won't be read */ + qht_map_iter__all_locked(ht, map, func, userp); + qht_map_unlock_buckets(map); +} + +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__locked(ht, new, b, p, hash, NULL); +} + +/* + * Call with ht->lock and all bucket locks held. + * + * Creating the @new map here would add unnecessary delay while all the locks + * are held--holding up the bucket locks is particularly bad, since no writes + * can occur while these are held. Thus, we let callers create the new map, + * hopefully without the bucket locks held. + */ +static void qht_do_resize(struct qht *ht, struct qht_map *new) +{ + struct qht_map *old; + + old = ht->map; + g_assert_cmpuint(new->n_buckets, !=, old->n_buckets); + + qht_map_iter__all_locked(ht, old, qht_map_copy, new); + qht_map_debug__all_locked(new); + + atomic_rcu_set(&ht->map, new); + call_rcu1(&old->rcu, qht_map_reclaim); +} + +bool qht_resize(struct qht *ht, size_t n_elems) +{ + size_t n_buckets = qht_elems_to_buckets(n_elems); + size_t ret = false; + + qemu_mutex_lock(&ht->lock); + if (n_buckets != ht->map->n_buckets) { + struct qht_map *new; + struct qht_map *old = ht->map; + + new = qht_map_create(n_buckets); + qht_map_lock_buckets(old); + qht_do_resize(ht, new); + qht_map_unlock_buckets(old); + ret = true; + } + qemu_mutex_unlock(&ht->lock); + + return ret; +} + +static __attribute__((noinline)) void qht_grow_maybe(struct qht *ht) +{ + struct qht_map *map; + + /* + * If the lock is taken it probably means there's an ongoing resize, + * so bail out. + */ + if (qemu_mutex_trylock(&ht->lock)) { + return; + } + map = ht->map; + /* another thread might have just performed the resize we were after */ + if (qht_map_needs_resize(map)) { + struct qht_map *new = qht_map_create(map->n_buckets * 2); + + qht_map_lock_buckets(map); + qht_do_resize(ht, new); + qht_map_unlock_buckets(map); + } + qemu_mutex_unlock(&ht->lock); +} + +/* pass @stats to qht_statistics_destroy() when done */ +void qht_statistics_init(struct qht *ht, struct qht_stats *stats) +{ + struct qht_map *map; + int i; + + map = atomic_rcu_read(&ht->map); + + stats->head_buckets = map->n_buckets; + stats->used_head_buckets = 0; + stats->entries = 0; + qdist_init(&stats->chain); + qdist_init(&stats->occupancy); + + for (i = 0; i < map->n_buckets; i++) { + struct qht_bucket *head = &map->buckets[i]; + struct qht_bucket *b; + unsigned int version; + size_t buckets; + size_t entries; + int j; + + do { + version = seqlock_read_begin(&head->sequence); + buckets = 0; + entries = 0; + b = head; + do { + for (j = 0; j < QHT_BUCKET_ENTRIES; j++) { + if (atomic_read(&b->pointers[j]) == NULL) { + break; + } + entries++; + } + buckets++; + b = atomic_rcu_read(&b->next); + } while (b); + } while (seqlock_read_retry(&head->sequence, version)); + + if (entries) { + qdist_inc(&stats->chain, buckets); + qdist_inc(&stats->occupancy, + (double)entries / QHT_BUCKET_ENTRIES / buckets); + stats->used_head_buckets++; + stats->entries += entries; + } else { + qdist_inc(&stats->occupancy, 0); + } + } +} + +void qht_statistics_destroy(struct qht_stats *stats) +{ + qdist_destroy(&stats->occupancy); + qdist_destroy(&stats->chain); +}