hash: support read/write concurrency
authorYipeng Wang <yipeng1.wang@intel.com>
Tue, 10 Jul 2018 16:59:58 +0000 (09:59 -0700)
committerThomas Monjalon <thomas@monjalon.net>
Thu, 12 Jul 2018 21:03:50 +0000 (23:03 +0200)
The existing implementation of librte_hash does not support read-write
concurrency. This commit implements read-write safety using rte_rwlock
and rte_rwlock TM version if hardware transactional memory is available.

Both multi-writer and read-write concurrency is protected by rte_rwlock
now. The x86 specific header file is removed since the x86 specific RTM
function is not called directly by rte hash now.

Signed-off-by: Yipeng Wang <yipeng1.wang@intel.com>
Acked-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
lib/librte_hash/meson.build
lib/librte_hash/rte_cuckoo_hash.c
lib/librte_hash/rte_cuckoo_hash.h
lib/librte_hash/rte_cuckoo_hash_x86.h [deleted file]
lib/librte_hash/rte_hash.h

index e139e1d..efc06ed 100644 (file)
@@ -6,7 +6,6 @@ headers = files('rte_cmp_arm64.h',
        'rte_cmp_x86.h',
        'rte_crc_arm64.h',
        'rte_cuckoo_hash.h',
-       'rte_cuckoo_hash_x86.h',
        'rte_fbk_hash.h',
        'rte_hash_crc.h',
        'rte_hash.h',
index 1d12fdb..ee8dc03 100644 (file)
@@ -31,9 +31,6 @@
 #include "rte_hash.h"
 #include "rte_cuckoo_hash.h"
 
-#if defined(RTE_ARCH_X86)
-#include "rte_cuckoo_hash_x86.h"
-#endif
 
 TAILQ_HEAD(rte_hash_list, rte_tailq_entry);
 
@@ -93,8 +90,10 @@ rte_hash_create(const struct rte_hash_parameters *params)
        void *buckets = NULL;
        char ring_name[RTE_RING_NAMESIZE];
        unsigned num_key_slots;
-       unsigned hw_trans_mem_support = 0;
        unsigned i;
+       unsigned int hw_trans_mem_support = 0, multi_writer_support = 0;
+       unsigned int readwrite_concur_support = 0;
+
        rte_hash_function default_hash_func = (rte_hash_function)rte_jhash;
 
        hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
@@ -117,8 +116,16 @@ rte_hash_create(const struct rte_hash_parameters *params)
        if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_TRANS_MEM_SUPPORT)
                hw_trans_mem_support = 1;
 
+       if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD)
+               multi_writer_support = 1;
+
+       if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY) {
+               readwrite_concur_support = 1;
+               multi_writer_support = 1;
+       }
+
        /* Store all keys and leave the first entry as a dummy entry for lookup_bulk */
-       if (hw_trans_mem_support)
+       if (multi_writer_support)
                /*
                 * Increase number of slots by total number of indices
                 * that can be stored in the lcore caches
@@ -232,7 +239,7 @@ rte_hash_create(const struct rte_hash_parameters *params)
        h->cmp_jump_table_idx = KEY_OTHER_BYTES;
 #endif
 
-       if (hw_trans_mem_support) {
+       if (multi_writer_support) {
                h->local_free_slots = rte_zmalloc_socket(NULL,
                                sizeof(struct lcore_cache) * RTE_MAX_LCORE,
                                RTE_CACHE_LINE_SIZE, params->socket_id);
@@ -260,6 +267,8 @@ rte_hash_create(const struct rte_hash_parameters *params)
        h->key_store = k;
        h->free_slots = r;
        h->hw_trans_mem_support = hw_trans_mem_support;
+       h->multi_writer_support = multi_writer_support;
+       h->readwrite_concur_support = readwrite_concur_support;
 
 #if defined(RTE_ARCH_X86)
        if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
@@ -270,24 +279,17 @@ rte_hash_create(const struct rte_hash_parameters *params)
 #endif
                h->sig_cmp_fn = RTE_HASH_COMPARE_SCALAR;
 
-       /* Turn on multi-writer only with explicit flat from user and TM
+       /* Turn on multi-writer only with explicit flag from user and TM
         * support.
         */
-       if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) {
-               if (h->hw_trans_mem_support) {
-                       h->add_key = ADD_KEY_MULTIWRITER_TM;
-               } else {
-                       h->add_key = ADD_KEY_MULTIWRITER;
-                       h->multiwriter_lock = rte_malloc(NULL,
-                                                       sizeof(rte_spinlock_t),
-                                                       RTE_CACHE_LINE_SIZE);
-                       if (h->multiwriter_lock == NULL)
-                               goto err_unlock;
-
-                       rte_spinlock_init(h->multiwriter_lock);
-               }
-       } else
-               h->add_key = ADD_KEY_SINGLEWRITER;
+       if (h->multi_writer_support) {
+               h->readwrite_lock = rte_malloc(NULL, sizeof(rte_rwlock_t),
+                                               RTE_CACHE_LINE_SIZE);
+               if (h->readwrite_lock == NULL)
+                       goto err_unlock;
+
+               rte_rwlock_init(h->readwrite_lock);
+       }
 
        /* Populate free slots ring. Entry zero is reserved for key misses. */
        for (i = 1; i < num_key_slots; i++)
@@ -337,11 +339,10 @@ rte_hash_free(struct rte_hash *h)
 
        rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 
-       if (h->hw_trans_mem_support)
+       if (h->multi_writer_support) {
                rte_free(h->local_free_slots);
-
-       if (h->add_key == ADD_KEY_MULTIWRITER)
-               rte_free(h->multiwriter_lock);
+               rte_free(h->readwrite_lock);
+       }
        rte_ring_free(h->free_slots);
        rte_free(h->key_store);
        rte_free(h->buckets);
@@ -368,6 +369,44 @@ rte_hash_secondary_hash(const hash_sig_t primary_hash)
        return primary_hash ^ ((tag + 1) * alt_bits_xor);
 }
 
+/* Read write locks implemented using rte_rwlock */
+static inline void
+__hash_rw_writer_lock(const struct rte_hash *h)
+{
+       if (h->multi_writer_support && h->hw_trans_mem_support)
+               rte_rwlock_write_lock_tm(h->readwrite_lock);
+       else if (h->multi_writer_support)
+               rte_rwlock_write_lock(h->readwrite_lock);
+}
+
+
+static inline void
+__hash_rw_reader_lock(const struct rte_hash *h)
+{
+       if (h->readwrite_concur_support && h->hw_trans_mem_support)
+               rte_rwlock_read_lock_tm(h->readwrite_lock);
+       else if (h->readwrite_concur_support)
+               rte_rwlock_read_lock(h->readwrite_lock);
+}
+
+static inline void
+__hash_rw_writer_unlock(const struct rte_hash *h)
+{
+       if (h->multi_writer_support && h->hw_trans_mem_support)
+               rte_rwlock_write_unlock_tm(h->readwrite_lock);
+       else if (h->multi_writer_support)
+               rte_rwlock_write_unlock(h->readwrite_lock);
+}
+
+static inline void
+__hash_rw_reader_unlock(const struct rte_hash *h)
+{
+       if (h->readwrite_concur_support && h->hw_trans_mem_support)
+               rte_rwlock_read_unlock_tm(h->readwrite_lock);
+       else if (h->readwrite_concur_support)
+               rte_rwlock_read_unlock(h->readwrite_lock);
+}
+
 void
 rte_hash_reset(struct rte_hash *h)
 {
@@ -377,6 +416,7 @@ rte_hash_reset(struct rte_hash *h)
        if (h == NULL)
                return;
 
+       __hash_rw_writer_lock(h);
        memset(h->buckets, 0, h->num_buckets * sizeof(struct rte_hash_bucket));
        memset(h->key_store, 0, h->key_entry_size * (h->entries + 1));
 
@@ -385,7 +425,7 @@ rte_hash_reset(struct rte_hash *h)
                rte_pause();
 
        /* Repopulate the free slots ring. Entry zero is reserved for key misses */
-       if (h->hw_trans_mem_support)
+       if (h->multi_writer_support)
                tot_ring_cnt = h->entries + (RTE_MAX_LCORE - 1) *
                                        (LCORE_CACHE_SIZE - 1);
        else
@@ -394,77 +434,12 @@ rte_hash_reset(struct rte_hash *h)
        for (i = 1; i < tot_ring_cnt + 1; i++)
                rte_ring_sp_enqueue(h->free_slots, (void *)((uintptr_t) i));
 
-       if (h->hw_trans_mem_support) {
+       if (h->multi_writer_support) {
                /* Reset local caches per lcore */
                for (i = 0; i < RTE_MAX_LCORE; i++)
                        h->local_free_slots[i].len = 0;
        }
-}
-
-/* Search for an entry that can be pushed to its alternative location */
-static inline int
-make_space_bucket(const struct rte_hash *h, struct rte_hash_bucket *bkt,
-               unsigned int *nr_pushes)
-{
-       unsigned i, j;
-       int ret;
-       uint32_t next_bucket_idx;
-       struct rte_hash_bucket *next_bkt[RTE_HASH_BUCKET_ENTRIES];
-
-       /*
-        * Push existing item (search for bucket with space in
-        * alternative locations) to its alternative location
-        */
-       for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
-               /* Search for space in alternative locations */
-               next_bucket_idx = bkt->sig_alt[i] & h->bucket_bitmask;
-               next_bkt[i] = &h->buckets[next_bucket_idx];
-               for (j = 0; j < RTE_HASH_BUCKET_ENTRIES; j++) {
-                       if (next_bkt[i]->key_idx[j] == EMPTY_SLOT)
-                               break;
-               }
-
-               if (j != RTE_HASH_BUCKET_ENTRIES)
-                       break;
-       }
-
-       /* Alternative location has spare room (end of recursive function) */
-       if (i != RTE_HASH_BUCKET_ENTRIES) {
-               next_bkt[i]->sig_alt[j] = bkt->sig_current[i];
-               next_bkt[i]->sig_current[j] = bkt->sig_alt[i];
-               next_bkt[i]->key_idx[j] = bkt->key_idx[i];
-               return i;
-       }
-
-       /* Pick entry that has not been pushed yet */
-       for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++)
-               if (bkt->flag[i] == 0)
-                       break;
-
-       /* All entries have been pushed, so entry cannot be added */
-       if (i == RTE_HASH_BUCKET_ENTRIES || ++(*nr_pushes) > RTE_HASH_MAX_PUSHES)
-               return -ENOSPC;
-
-       /* Set flag to indicate that this entry is going to be pushed */
-       bkt->flag[i] = 1;
-
-       /* Need room in alternative bucket to insert the pushed entry */
-       ret = make_space_bucket(h, next_bkt[i], nr_pushes);
-       /*
-        * After recursive function.
-        * Clear flags and insert the pushed entry
-        * in its alternative location if successful,
-        * or return error
-        */
-       bkt->flag[i] = 0;
-       if (ret >= 0) {
-               next_bkt[i]->sig_alt[ret] = bkt->sig_current[i];
-               next_bkt[i]->sig_current[ret] = bkt->sig_alt[i];
-               next_bkt[i]->key_idx[ret] = bkt->key_idx[i];
-               return i;
-       } else
-               return ret;
-
+       __hash_rw_writer_unlock(h);
 }
 
 /*
@@ -477,7 +452,7 @@ enqueue_slot_back(const struct rte_hash *h,
                struct lcore_cache *cached_free_slots,
                void *slot_id)
 {
-       if (h->hw_trans_mem_support) {
+       if (h->multi_writer_support) {
                cached_free_slots->objs[cached_free_slots->len] = slot_id;
                cached_free_slots->len++;
        } else
@@ -511,13 +486,207 @@ search_and_update(const struct rte_hash *h, void *data, const void *key,
        return -1;
 }
 
+/* Only tries to insert at one bucket (@prim_bkt) without trying to push
+ * buckets around.
+ * return 1 if matching existing key, return 0 if succeeds, return -1 for no
+ * empty entry.
+ */
+static inline int32_t
+rte_hash_cuckoo_insert_mw(const struct rte_hash *h,
+               struct rte_hash_bucket *prim_bkt,
+               struct rte_hash_bucket *sec_bkt,
+               const struct rte_hash_key *key, void *data,
+               hash_sig_t sig, hash_sig_t alt_hash, uint32_t new_idx,
+               int32_t *ret_val)
+{
+       unsigned int i;
+       struct rte_hash_bucket *cur_bkt = prim_bkt;
+       int32_t ret;
+
+       __hash_rw_writer_lock(h);
+       /* Check if key was inserted after last check but before this
+        * protected region in case of inserting duplicated keys.
+        */
+       ret = search_and_update(h, data, key, cur_bkt, sig, alt_hash);
+       if (ret != -1) {
+               __hash_rw_writer_unlock(h);
+               *ret_val = ret;
+               return 1;
+       }
+       ret = search_and_update(h, data, key, sec_bkt, alt_hash, sig);
+       if (ret != -1) {
+               __hash_rw_writer_unlock(h);
+               *ret_val = ret;
+               return 1;
+       }
+
+       /* Insert new entry if there is room in the primary
+        * bucket.
+        */
+       for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
+               /* Check if slot is available */
+               if (likely(prim_bkt->key_idx[i] == EMPTY_SLOT)) {
+                       prim_bkt->sig_current[i] = sig;
+                       prim_bkt->sig_alt[i] = alt_hash;
+                       prim_bkt->key_idx[i] = new_idx;
+                       break;
+               }
+       }
+       __hash_rw_writer_unlock(h);
+
+       if (i != RTE_HASH_BUCKET_ENTRIES)
+               return 0;
+
+       /* no empty entry */
+       return -1;
+}
+
+/* Shift buckets along provided cuckoo_path (@leaf and @leaf_slot) and fill
+ * the path head with new entry (sig, alt_hash, new_idx)
+ * return 1 if matched key found, return -1 if cuckoo path invalided and fail,
+ * return 0 if succeeds.
+ */
+static inline int
+rte_hash_cuckoo_move_insert_mw(const struct rte_hash *h,
+                       struct rte_hash_bucket *bkt,
+                       struct rte_hash_bucket *alt_bkt,
+                       const struct rte_hash_key *key, void *data,
+                       struct queue_node *leaf, uint32_t leaf_slot,
+                       hash_sig_t sig, hash_sig_t alt_hash, uint32_t new_idx,
+                       int32_t *ret_val)
+{
+       uint32_t prev_alt_bkt_idx;
+       struct rte_hash_bucket *cur_bkt = bkt;
+       struct queue_node *prev_node, *curr_node = leaf;
+       struct rte_hash_bucket *prev_bkt, *curr_bkt = leaf->bkt;
+       uint32_t prev_slot, curr_slot = leaf_slot;
+       int32_t ret;
+
+       __hash_rw_writer_lock(h);
+
+       /* In case empty slot was gone before entering protected region */
+       if (curr_bkt->key_idx[curr_slot] != EMPTY_SLOT) {
+               __hash_rw_writer_unlock(h);
+               return -1;
+       }
+
+       /* Check if key was inserted after last check but before this
+        * protected region.
+        */
+       ret = search_and_update(h, data, key, cur_bkt, sig, alt_hash);
+       if (ret != -1) {
+               __hash_rw_writer_unlock(h);
+               *ret_val = ret;
+               return 1;
+       }
+
+       ret = search_and_update(h, data, key, alt_bkt, alt_hash, sig);
+       if (ret != -1) {
+               __hash_rw_writer_unlock(h);
+               *ret_val = ret;
+               return 1;
+       }
+
+       while (likely(curr_node->prev != NULL)) {
+               prev_node = curr_node->prev;
+               prev_bkt = prev_node->bkt;
+               prev_slot = curr_node->prev_slot;
+
+               prev_alt_bkt_idx =
+                       prev_bkt->sig_alt[prev_slot] & h->bucket_bitmask;
+
+               if (unlikely(&h->buckets[prev_alt_bkt_idx]
+                               != curr_bkt)) {
+                       /* revert it to empty, otherwise duplicated keys */
+                       curr_bkt->key_idx[curr_slot] = EMPTY_SLOT;
+                       __hash_rw_writer_unlock(h);
+                       return -1;
+               }
+
+               /* Need to swap current/alt sig to allow later
+                * Cuckoo insert to move elements back to its
+                * primary bucket if available
+                */
+               curr_bkt->sig_alt[curr_slot] =
+                        prev_bkt->sig_current[prev_slot];
+               curr_bkt->sig_current[curr_slot] =
+                       prev_bkt->sig_alt[prev_slot];
+               curr_bkt->key_idx[curr_slot] =
+                       prev_bkt->key_idx[prev_slot];
+
+               curr_slot = prev_slot;
+               curr_node = prev_node;
+               curr_bkt = curr_node->bkt;
+       }
+
+       curr_bkt->sig_current[curr_slot] = sig;
+       curr_bkt->sig_alt[curr_slot] = alt_hash;
+       curr_bkt->key_idx[curr_slot] = new_idx;
+
+       __hash_rw_writer_unlock(h);
+
+       return 0;
+
+}
+
+/*
+ * Make space for new key, using bfs Cuckoo Search and Multi-Writer safe
+ * Cuckoo
+ */
+static inline int
+rte_hash_cuckoo_make_space_mw(const struct rte_hash *h,
+                       struct rte_hash_bucket *bkt,
+                       struct rte_hash_bucket *sec_bkt,
+                       const struct rte_hash_key *key, void *data,
+                       hash_sig_t sig, hash_sig_t alt_hash,
+                       uint32_t new_idx, int32_t *ret_val)
+{
+       unsigned int i;
+       struct queue_node queue[RTE_HASH_BFS_QUEUE_MAX_LEN];
+       struct queue_node *tail, *head;
+       struct rte_hash_bucket *curr_bkt, *alt_bkt;
+
+       tail = queue;
+       head = queue + 1;
+       tail->bkt = bkt;
+       tail->prev = NULL;
+       tail->prev_slot = -1;
+
+       /* Cuckoo bfs Search */
+       while (likely(tail != head && head <
+                                       queue + RTE_HASH_BFS_QUEUE_MAX_LEN -
+                                       RTE_HASH_BUCKET_ENTRIES)) {
+               curr_bkt = tail->bkt;
+               for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
+                       if (curr_bkt->key_idx[i] == EMPTY_SLOT) {
+                               int32_t ret = rte_hash_cuckoo_move_insert_mw(h,
+                                               bkt, sec_bkt, key, data,
+                                               tail, i, sig, alt_hash,
+                                               new_idx, ret_val);
+                               if (likely(ret != -1))
+                                       return ret;
+                       }
+
+                       /* Enqueue new node and keep prev node info */
+                       alt_bkt = &(h->buckets[curr_bkt->sig_alt[i]
+                                                   & h->bucket_bitmask]);
+                       head->bkt = alt_bkt;
+                       head->prev = tail;
+                       head->prev_slot = i;
+                       head++;
+               }
+               tail++;
+       }
+
+       return -ENOSPC;
+}
+
 static inline int32_t
 __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
                                                hash_sig_t sig, void *data)
 {
        hash_sig_t alt_hash;
        uint32_t prim_bucket_idx, sec_bucket_idx;
-       unsigned i;
        struct rte_hash_bucket *prim_bkt, *sec_bkt;
        struct rte_hash_key *new_k, *keys = h->key_store;
        void *slot_id = NULL;
@@ -526,10 +695,7 @@ __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
        unsigned n_slots;
        unsigned lcore_id;
        struct lcore_cache *cached_free_slots = NULL;
-       unsigned int nr_pushes = 0;
-
-       if (h->add_key == ADD_KEY_MULTIWRITER)
-               rte_spinlock_lock(h->multiwriter_lock);
+       int32_t ret_val;
 
        prim_bucket_idx = sig & h->bucket_bitmask;
        prim_bkt = &h->buckets[prim_bucket_idx];
@@ -540,8 +706,24 @@ __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
        sec_bkt = &h->buckets[sec_bucket_idx];
        rte_prefetch0(sec_bkt);
 
-       /* Get a new slot for storing the new key */
-       if (h->hw_trans_mem_support) {
+       /* Check if key is already inserted in primary location */
+       __hash_rw_writer_lock(h);
+       ret = search_and_update(h, data, key, prim_bkt, sig, alt_hash);
+       if (ret != -1) {
+               __hash_rw_writer_unlock(h);
+               return ret;
+       }
+
+       /* Check if key is already inserted in secondary location */
+       ret = search_and_update(h, data, key, sec_bkt, alt_hash, sig);
+       if (ret != -1) {
+               __hash_rw_writer_unlock(h);
+               return ret;
+       }
+       __hash_rw_writer_unlock(h);
+
+       /* Did not find a match, so get a new slot for storing the new key */
+       if (h->multi_writer_support) {
                lcore_id = rte_lcore_id();
                cached_free_slots = &h->local_free_slots[lcore_id];
                /* Try to get a free slot from the local cache */
@@ -551,8 +733,7 @@ __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
                                        cached_free_slots->objs,
                                        LCORE_CACHE_SIZE, NULL);
                        if (n_slots == 0) {
-                               ret = -ENOSPC;
-                               goto failure;
+                               return -ENOSPC;
                        }
 
                        cached_free_slots->len += n_slots;
@@ -563,92 +744,50 @@ __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
                slot_id = cached_free_slots->objs[cached_free_slots->len];
        } else {
                if (rte_ring_sc_dequeue(h->free_slots, &slot_id) != 0) {
-                       ret = -ENOSPC;
-                       goto failure;
+                       return -ENOSPC;
                }
        }
 
        new_k = RTE_PTR_ADD(keys, (uintptr_t)slot_id * h->key_entry_size);
-       rte_prefetch0(new_k);
        new_idx = (uint32_t)((uintptr_t) slot_id);
-
-       /* Check if key is already inserted in primary location */
-       ret = search_and_update(h, data, key, prim_bkt, sig, alt_hash);
-       if (ret != -1)
-               goto failure;
-
-       /* Check if key is already inserted in secondary location */
-       ret = search_and_update(h, data, key, sec_bkt, alt_hash, sig);
-       if (ret != -1)
-               goto failure;
-
        /* Copy key */
        rte_memcpy(new_k->key, key, h->key_len);
        new_k->pdata = data;
 
-#if defined(RTE_ARCH_X86) /* currently only x86 support HTM */
-       if (h->add_key == ADD_KEY_MULTIWRITER_TM) {
-               ret = rte_hash_cuckoo_insert_mw_tm(prim_bkt,
-                               sig, alt_hash, new_idx);
-               if (ret >= 0)
-                       return new_idx - 1;
 
-               /* Primary bucket full, need to make space for new entry */
-               ret = rte_hash_cuckoo_make_space_mw_tm(h, prim_bkt, sig,
-                                                       alt_hash, new_idx);
+       /* Find an empty slot and insert */
+       ret = rte_hash_cuckoo_insert_mw(h, prim_bkt, sec_bkt, key, data,
+                                       sig, alt_hash, new_idx, &ret_val);
+       if (ret == 0)
+               return new_idx - 1;
+       else if (ret == 1) {
+               enqueue_slot_back(h, cached_free_slots, slot_id);
+               return ret_val;
+       }
 
-               if (ret >= 0)
-                       return new_idx - 1;
+       /* Primary bucket full, need to make space for new entry */
+       ret = rte_hash_cuckoo_make_space_mw(h, prim_bkt, sec_bkt, key, data,
+                                       sig, alt_hash, new_idx, &ret_val);
+       if (ret == 0)
+               return new_idx - 1;
+       else if (ret == 1) {
+               enqueue_slot_back(h, cached_free_slots, slot_id);
+               return ret_val;
+       }
 
-               /* Also search secondary bucket to get better occupancy */
-               ret = rte_hash_cuckoo_make_space_mw_tm(h, sec_bkt, sig,
-                                                       alt_hash, new_idx);
+       /* Also search secondary bucket to get better occupancy */
+       ret = rte_hash_cuckoo_make_space_mw(h, sec_bkt, prim_bkt, key, data,
+                                       alt_hash, sig, new_idx, &ret_val);
 
-               if (ret >= 0)
-                       return new_idx - 1;
+       if (ret == 0)
+               return new_idx - 1;
+       else if (ret == 1) {
+               enqueue_slot_back(h, cached_free_slots, slot_id);
+               return ret_val;
        } else {
-#endif
-               for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
-                       /* Check if slot is available */
-                       if (likely(prim_bkt->key_idx[i] == EMPTY_SLOT)) {
-                               prim_bkt->sig_current[i] = sig;
-                               prim_bkt->sig_alt[i] = alt_hash;
-                               prim_bkt->key_idx[i] = new_idx;
-                               break;
-                       }
-               }
-
-               if (i != RTE_HASH_BUCKET_ENTRIES) {
-                       if (h->add_key == ADD_KEY_MULTIWRITER)
-                               rte_spinlock_unlock(h->multiwriter_lock);
-                       return new_idx - 1;
-               }
-
-               /* Primary bucket full, need to make space for new entry
-                * After recursive function.
-                * Insert the new entry in the position of the pushed entry
-                * if successful or return error and
-                * store the new slot back in the ring
-                */
-               ret = make_space_bucket(h, prim_bkt, &nr_pushes);
-               if (ret >= 0) {
-                       prim_bkt->sig_current[ret] = sig;
-                       prim_bkt->sig_alt[ret] = alt_hash;
-                       prim_bkt->key_idx[ret] = new_idx;
-                       if (h->add_key == ADD_KEY_MULTIWRITER)
-                               rte_spinlock_unlock(h->multiwriter_lock);
-                       return new_idx - 1;
-               }
-#if defined(RTE_ARCH_X86)
+               enqueue_slot_back(h, cached_free_slots, slot_id);
+               return ret;
        }
-#endif
-       /* Error in addition, store new slot back in the ring and return error */
-       enqueue_slot_back(h, cached_free_slots, (void *)((uintptr_t) new_idx));
-
-failure:
-       if (h->add_key == ADD_KEY_MULTIWRITER)
-               rte_spinlock_unlock(h->multiwriter_lock);
-       return ret;
 }
 
 int32_t
@@ -733,12 +872,14 @@ __rte_hash_lookup_with_hash(const struct rte_hash *h, const void *key,
        bucket_idx = sig & h->bucket_bitmask;
        bkt = &h->buckets[bucket_idx];
 
+       __hash_rw_reader_lock(h);
 
        /* Check if key is in primary location */
        ret = search_one_bucket(h, key, sig, data, bkt);
-       if (ret != -1)
+       if (ret != -1) {
+               __hash_rw_reader_unlock(h);
                return ret;
-
+       }
        /* Calculate secondary hash */
        alt_hash = rte_hash_secondary_hash(sig);
        bucket_idx = alt_hash & h->bucket_bitmask;
@@ -746,9 +887,11 @@ __rte_hash_lookup_with_hash(const struct rte_hash *h, const void *key,
 
        /* Check if key is in secondary location */
        ret = search_one_bucket(h, key, alt_hash, data, bkt);
-       if (ret != -1)
+       if (ret != -1) {
+               __hash_rw_reader_unlock(h);
                return ret;
-
+       }
+       __hash_rw_reader_unlock(h);
        return -ENOENT;
 }
 
@@ -790,7 +933,7 @@ remove_entry(const struct rte_hash *h, struct rte_hash_bucket *bkt, unsigned i)
 
        bkt->sig_current[i] = NULL_SIGNATURE;
        bkt->sig_alt[i] = NULL_SIGNATURE;
-       if (h->hw_trans_mem_support) {
+       if (h->multi_writer_support) {
                lcore_id = rte_lcore_id();
                cached_free_slots = &h->local_free_slots[lcore_id];
                /* Cache full, need to free it. */
@@ -854,10 +997,13 @@ __rte_hash_del_key_with_hash(const struct rte_hash *h, const void *key,
        bucket_idx = sig & h->bucket_bitmask;
        bkt = &h->buckets[bucket_idx];
 
+       __hash_rw_writer_lock(h);
        /* look for key in primary bucket */
        ret = search_and_remove(h, key, bkt, sig);
-       if (ret != -1)
+       if (ret != -1) {
+               __hash_rw_writer_unlock(h);
                return ret;
+       }
 
        /* Calculate secondary hash */
        alt_hash = rte_hash_secondary_hash(sig);
@@ -866,9 +1012,12 @@ __rte_hash_del_key_with_hash(const struct rte_hash *h, const void *key,
 
        /* look for key in secondary bucket */
        ret = search_and_remove(h, key, bkt, alt_hash);
-       if (ret != -1)
+       if (ret != -1) {
+               __hash_rw_writer_unlock(h);
                return ret;
+       }
 
+       __hash_rw_writer_unlock(h);
        return -ENOENT;
 }
 
@@ -1010,6 +1159,7 @@ __rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
                rte_prefetch0(secondary_bkt[i]);
        }
 
+       __hash_rw_reader_lock(h);
        /* Compare signatures and prefetch key slot of first hit */
        for (i = 0; i < num_keys; i++) {
                compare_signatures(&prim_hitmask[i], &sec_hitmask[i],
@@ -1092,6 +1242,8 @@ next_key:
                continue;
        }
 
+       __hash_rw_reader_unlock(h);
+
        if (hit_mask != NULL)
                *hit_mask = hits;
 }
@@ -1150,7 +1302,7 @@ rte_hash_iterate(const struct rte_hash *h, const void **key, void **data, uint32
                bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
                idx = *next % RTE_HASH_BUCKET_ENTRIES;
        }
-
+       __hash_rw_reader_lock(h);
        /* Get position of entry in key table */
        position = h->buckets[bucket_idx].key_idx[idx];
        next_key = (struct rte_hash_key *) ((char *)h->key_store +
@@ -1159,6 +1311,8 @@ rte_hash_iterate(const struct rte_hash *h, const void **key, void **data, uint32
        *key = next_key->key;
        *data = next_key->pdata;
 
+       __hash_rw_reader_unlock(h);
+
        /* Increment iterator */
        (*next)++;
 
index bd6ad1b..b43f467 100644 (file)
@@ -88,11 +88,6 @@ const rte_hash_cmp_eq_t cmp_jump_table[NUM_KEY_CMP_CASES] = {
 
 #endif
 
-enum add_key_case {
-       ADD_KEY_SINGLEWRITER = 0,
-       ADD_KEY_MULTIWRITER,
-       ADD_KEY_MULTIWRITER_TM,
-};
 
 /** Number of items per bucket. */
 #define RTE_HASH_BUCKET_ENTRIES                8
@@ -159,18 +154,20 @@ struct rte_hash {
 
        struct rte_ring *free_slots;
        /**< Ring that stores all indexes of the free slots in the key table */
-       uint8_t hw_trans_mem_support;
-       /**< Hardware transactional memory support */
+
        struct lcore_cache *local_free_slots;
        /**< Local cache per lcore, storing some indexes of the free slots */
-       enum add_key_case add_key; /**< Multi-writer hash add behavior */
-
-       rte_spinlock_t *multiwriter_lock; /**< Multi-writer spinlock for w/o TM */
 
        /* Fields used in lookup */
 
        uint32_t key_len __rte_cache_aligned;
        /**< Length of hash key. */
+       uint8_t hw_trans_mem_support;
+       /**< If hardware transactional memory is used. */
+       uint8_t multi_writer_support;
+       /**< If multi-writer support is enabled. */
+       uint8_t readwrite_concur_support;
+       /**< If read-write concurrency support is enabled */
        rte_hash_function hash_func;    /**< Function used to calculate hash. */
        uint32_t hash_func_init_val;    /**< Init value used by hash_func. */
        rte_hash_cmp_eq_t rte_hash_custom_cmp_eq;
@@ -188,6 +185,7 @@ struct rte_hash {
        /**< Table with buckets storing all the hash values and key indexes
         * to the key table.
         */
+       rte_rwlock_t *readwrite_lock; /**< Read-write lock thread-safety. */
 } __rte_cache_aligned;
 
 struct queue_node {
diff --git a/lib/librte_hash/rte_cuckoo_hash_x86.h b/lib/librte_hash/rte_cuckoo_hash_x86.h
deleted file mode 100644 (file)
index 981d7bd..0000000
+++ /dev/null
@@ -1,167 +0,0 @@
-/* SPDX-License-Identifier: BSD-3-Clause
- * Copyright(c) 2016 Intel Corporation
- */
-
-/* rte_cuckoo_hash_x86.h
- * This file holds all x86 specific Cuckoo Hash functions
- */
-
-/* Only tries to insert at one bucket (@prim_bkt) without trying to push
- * buckets around
- */
-static inline unsigned
-rte_hash_cuckoo_insert_mw_tm(struct rte_hash_bucket *prim_bkt,
-               hash_sig_t sig, hash_sig_t alt_hash, uint32_t new_idx)
-{
-       unsigned i, status;
-       unsigned try = 0;
-
-       while (try < RTE_HASH_TSX_MAX_RETRY) {
-               status = rte_xbegin();
-               if (likely(status == RTE_XBEGIN_STARTED)) {
-                       /* Insert new entry if there is room in the primary
-                       * bucket.
-                       */
-                       for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
-                               /* Check if slot is available */
-                               if (likely(prim_bkt->key_idx[i] == EMPTY_SLOT)) {
-                                       prim_bkt->sig_current[i] = sig;
-                                       prim_bkt->sig_alt[i] = alt_hash;
-                                       prim_bkt->key_idx[i] = new_idx;
-                                       break;
-                               }
-                       }
-                       rte_xend();
-
-                       if (i != RTE_HASH_BUCKET_ENTRIES)
-                               return 0;
-
-                       break; /* break off try loop if transaction commits */
-               } else {
-                       /* If we abort we give up this cuckoo path. */
-                       try++;
-                       rte_pause();
-               }
-       }
-
-       return -1;
-}
-
-/* Shift buckets along provided cuckoo_path (@leaf and @leaf_slot) and fill
- * the path head with new entry (sig, alt_hash, new_idx)
- */
-static inline int
-rte_hash_cuckoo_move_insert_mw_tm(const struct rte_hash *h,
-                       struct queue_node *leaf, uint32_t leaf_slot,
-                       hash_sig_t sig, hash_sig_t alt_hash, uint32_t new_idx)
-{
-       unsigned try = 0;
-       unsigned status;
-       uint32_t prev_alt_bkt_idx;
-
-       struct queue_node *prev_node, *curr_node = leaf;
-       struct rte_hash_bucket *prev_bkt, *curr_bkt = leaf->bkt;
-       uint32_t prev_slot, curr_slot = leaf_slot;
-
-       while (try < RTE_HASH_TSX_MAX_RETRY) {
-               status = rte_xbegin();
-               if (likely(status == RTE_XBEGIN_STARTED)) {
-                       /* In case empty slot was gone before entering TSX */
-                       if (curr_bkt->key_idx[curr_slot] != EMPTY_SLOT)
-                               rte_xabort(RTE_XABORT_CUCKOO_PATH_INVALIDED);
-                       while (likely(curr_node->prev != NULL)) {
-                               prev_node = curr_node->prev;
-                               prev_bkt = prev_node->bkt;
-                               prev_slot = curr_node->prev_slot;
-
-                               prev_alt_bkt_idx
-                                       = prev_bkt->sig_alt[prev_slot]
-                                           & h->bucket_bitmask;
-
-                               if (unlikely(&h->buckets[prev_alt_bkt_idx]
-                                            != curr_bkt)) {
-                                       rte_xabort(RTE_XABORT_CUCKOO_PATH_INVALIDED);
-                               }
-
-                               /* Need to swap current/alt sig to allow later
-                                * Cuckoo insert to move elements back to its
-                                * primary bucket if available
-                                */
-                               curr_bkt->sig_alt[curr_slot] =
-                                   prev_bkt->sig_current[prev_slot];
-                               curr_bkt->sig_current[curr_slot] =
-                                   prev_bkt->sig_alt[prev_slot];
-                               curr_bkt->key_idx[curr_slot]
-                                   = prev_bkt->key_idx[prev_slot];
-
-                               curr_slot = prev_slot;
-                               curr_node = prev_node;
-                               curr_bkt = curr_node->bkt;
-                       }
-
-                       curr_bkt->sig_current[curr_slot] = sig;
-                       curr_bkt->sig_alt[curr_slot] = alt_hash;
-                       curr_bkt->key_idx[curr_slot] = new_idx;
-
-                       rte_xend();
-
-                       return 0;
-               }
-
-               /* If we abort we give up this cuckoo path, since most likely it's
-                * no longer valid as TSX detected data conflict
-                */
-               try++;
-               rte_pause();
-       }
-
-       return -1;
-}
-
-/*
- * Make space for new key, using bfs Cuckoo Search and Multi-Writer safe
- * Cuckoo
- */
-static inline int
-rte_hash_cuckoo_make_space_mw_tm(const struct rte_hash *h,
-                       struct rte_hash_bucket *bkt,
-                       hash_sig_t sig, hash_sig_t alt_hash,
-                       uint32_t new_idx)
-{
-       unsigned i;
-       struct queue_node queue[RTE_HASH_BFS_QUEUE_MAX_LEN];
-       struct queue_node *tail, *head;
-       struct rte_hash_bucket *curr_bkt, *alt_bkt;
-
-       tail = queue;
-       head = queue + 1;
-       tail->bkt = bkt;
-       tail->prev = NULL;
-       tail->prev_slot = -1;
-
-       /* Cuckoo bfs Search */
-       while (likely(tail != head && head <
-                                       queue + RTE_HASH_BFS_QUEUE_MAX_LEN -
-                                       RTE_HASH_BUCKET_ENTRIES)) {
-               curr_bkt = tail->bkt;
-               for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
-                       if (curr_bkt->key_idx[i] == EMPTY_SLOT) {
-                               if (likely(rte_hash_cuckoo_move_insert_mw_tm(h,
-                                               tail, i, sig,
-                                               alt_hash, new_idx) == 0))
-                                       return 0;
-                       }
-
-                       /* Enqueue new node and keep prev node info */
-                       alt_bkt = &(h->buckets[curr_bkt->sig_alt[i]
-                                                   & h->bucket_bitmask]);
-                       head->bkt = alt_bkt;
-                       head->prev = tail;
-                       head->prev_slot = i;
-                       head++;
-               }
-               tail++;
-       }
-
-       return -ENOSPC;
-}
index f71ca9f..ecb49e4 100644 (file)
@@ -34,6 +34,9 @@ extern "C" {
 /** Default behavior of insertion, single writer/multi writer */
 #define RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD 0x02
 
+/** Flag to support reader writer concurrency */
+#define RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY 0x04
+
 /** Signature of key that is stored internally. */
 typedef uint32_t hash_sig_t;