From 9797bfcce1c9fcbf75e78dc17ecdc9b0da3ae945 Mon Sep 17 00:00:00 2001 From: Yongseok Koh Date: Wed, 9 May 2018 04:09:06 -0700 Subject: [PATCH] net/mlx4: add new memory region support This is the new design of Memory Region (MR) for mlx PMD, in order to: - Accommodate the new memory hotplug model. - Support non-contiguous Mempool. There are multiple layers for MR search. L0 is to look up the last-hit entry which is pointed by mr_ctrl->mru (Most Recently Used). If L0 misses, L1 is to look up the address in a fixed-sized array by linear search. L0/L1 is in an inline function - mlx4_mr_lookup_cache(). If L1 misses, the bottom-half function is called to look up the address from the bigger local cache of the queue. This is L2 - mlx4_mr_addr2mr_bh() and it is not an inline function. Data structure for L2 is the Binary Tree. If L2 misses, the search falls into the slowest path which takes locks in order to access global device cache (priv->mr.cache) which is also a B-tree and caches the original MR list (priv->mr.mr_list) of the device. Unless the global cache is overflowed, it is all-inclusive of the MR list. This is L3 - mlx4_mr_lookup_dev(). The size of the L3 cache table is limited and can't be expanded on the fly due to deadlock. Refer to the comments in the code for the details - mr_lookup_dev(). If L3 is overflowed, the list will have to be searched directly bypassing the cache although it is slower. If L3 misses, a new MR for the address should be created - mlx4_mr_create(). When it creates a new MR, it tries to register adjacent memsegs as much as possible which are virtually contiguous around the address. This must take two locks - memory_hotplug_lock and priv->mr.rwlock. Due to memory_hotplug_lock, there can't be any allocation/free of memory inside. In the free callback of the memory hotplug event, freed space is searched from the MR list and corresponding bits are cleared from the bitmap of MRs. This can fragment a MR and the MR will have multiple search entries in the caches. Once there's a change by the event, the global cache must be rebuilt and all the per-queue caches will be flushed as well. If memory is frequently freed in run-time, that may cause jitter on dataplane processing in the worst case by incurring MR cache flush and rebuild. But, it would be the least probable scenario. To guarantee the most optimal performance, it is highly recommended to use an EAL option - '--socket-mem'. Then, the reserved memory will be pinned and won't be freed dynamically. And it is also recommended to configure per-lcore cache of Mempool. Even though there're many MRs for a device or MRs are highly fragmented, the cache of Mempool will be much helpful to reduce misses on per-queue caches anyway. '--legacy-mem' is also supported. Signed-off-by: Yongseok Koh --- doc/guides/nics/mlx4.rst | 6 + drivers/net/mlx4/mlx4.c | 26 + drivers/net/mlx4/mlx4.h | 14 + drivers/net/mlx4/mlx4_mr.c | 1144 ++++++++++++++++++++++++++++++++++ drivers/net/mlx4/mlx4_mr.h | 122 ++++ drivers/net/mlx4/mlx4_rxq.c | 14 +- drivers/net/mlx4/mlx4_rxtx.c | 35 +- drivers/net/mlx4/mlx4_rxtx.h | 71 ++- drivers/net/mlx4/mlx4_txq.c | 8 + 9 files changed, 1409 insertions(+), 31 deletions(-) create mode 100644 drivers/net/mlx4/mlx4_mr.h diff --git a/doc/guides/nics/mlx4.rst b/doc/guides/nics/mlx4.rst index 0bac6d374b..a6317547a0 100644 --- a/doc/guides/nics/mlx4.rst +++ b/doc/guides/nics/mlx4.rst @@ -370,6 +370,12 @@ Performance tuning The XXX can be different on different systems. Make sure to configure according to the setpci output. +6. To minimize overhead of searching Memory Regions: + + - '--socket-mem' is recommended to pin memory by predictable amount. + - Configure per-lcore cache when creating Mempools for packet buffer. + - Refrain from dynamically allocating/freeing memory in run-time. + Usage example ------------- diff --git a/drivers/net/mlx4/mlx4.c b/drivers/net/mlx4/mlx4.c index 47451b6515..abed2f5dcf 100644 --- a/drivers/net/mlx4/mlx4.c +++ b/drivers/net/mlx4/mlx4.c @@ -44,9 +44,15 @@ #include "mlx4.h" #include "mlx4_glue.h" #include "mlx4_flow.h" +#include "mlx4_mr.h" #include "mlx4_rxtx.h" #include "mlx4_utils.h" +struct mlx4_dev_list mlx4_mem_event_cb_list = + LIST_HEAD_INITIALIZER(mlx4_mem_event_cb_list); + +rte_rwlock_t mlx4_mem_event_rwlock = RTE_RWLOCK_INITIALIZER; + /** Configuration structure for device arguments. */ struct mlx4_conf { struct { @@ -92,6 +98,20 @@ mlx4_dev_configure(struct rte_eth_dev *dev) if (ret) ERROR("%p: interrupt handler installation failed", (void *)dev); + /* + * Once the device is added to the list of memory event callback, its + * global MR cache table cannot be expanded on the fly because of + * deadlock. If it overflows, lookup should be done by searching MR list + * linearly, which is slow. + */ + if (mlx4_mr_btree_init(&priv->mr.cache, MLX4_MR_BTREE_CACHE_N * 2, + dev->device->numa_node)) { + /* rte_errno is already set. */ + return -rte_errno; + } + rte_rwlock_write_lock(&mlx4_mem_event_rwlock); + LIST_INSERT_HEAD(&mlx4_mem_event_cb_list, priv, mem_event_cb); + rte_rwlock_write_unlock(&mlx4_mem_event_rwlock); exit: return ret; } @@ -125,6 +145,9 @@ mlx4_dev_start(struct rte_eth_dev *dev) (void *)dev, strerror(-ret)); goto err; } +#ifndef NDEBUG + mlx4_mr_dump_dev(dev); +#endif ret = mlx4_rxq_intr_enable(priv); if (ret) { ERROR("%p: interrupt handler installation failed", @@ -200,6 +223,7 @@ mlx4_dev_close(struct rte_eth_dev *dev) mlx4_rx_queue_release(dev->data->rx_queues[i]); for (i = 0; i != dev->data->nb_tx_queues; ++i) mlx4_tx_queue_release(dev->data->tx_queues[i]); + mlx4_mr_release(dev); if (priv->pd != NULL) { assert(priv->ctx != NULL); claim_zero(mlx4_glue->dealloc_pd(priv->pd)); @@ -964,6 +988,8 @@ rte_mlx4_pmd_init(void) } mlx4_glue->fork_init(); rte_pci_register(&mlx4_driver); + rte_mem_event_callback_register("MLX4_MEM_EVENT_CB", + mlx4_mr_mem_event_cb, NULL); } RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__); diff --git a/drivers/net/mlx4/mlx4.h b/drivers/net/mlx4/mlx4.h index e0e1b5d4cf..300cb4d7ab 100644 --- a/drivers/net/mlx4/mlx4.h +++ b/drivers/net/mlx4/mlx4.h @@ -23,6 +23,9 @@ #include #include #include +#include + +#include "mlx4_mr.h" #ifndef IBV_RX_HASH_INNER /** This is not necessarily defined by supported RDMA core versions. */ @@ -66,8 +69,12 @@ struct rxq; struct txq; struct rte_flow; +LIST_HEAD(mlx4_dev_list, priv); +LIST_HEAD(mlx4_mr_list, mlx4_mr); + /** Private data structure. */ struct priv { + LIST_ENTRY(priv) mem_event_cb; /* Called by memory event callback. */ struct rte_eth_dev *dev; /**< Ethernet device. */ struct ibv_context *ctx; /**< Verbs context. */ struct ibv_device_attr device_attr; /**< Device properties. */ @@ -86,6 +93,13 @@ struct priv { uint64_t hw_rss_sup; /**< Supported RSS hash fields (Verbs format). */ struct rte_intr_handle intr_handle; /**< Port interrupt handle. */ struct mlx4_drop *drop; /**< Shared resources for drop flow rules. */ + struct { + uint32_t dev_gen; /* Generation number to flush local caches. */ + rte_rwlock_t rwlock; /* MR Lock. */ + struct mlx4_mr_btree cache; /* Global MR cache table. */ + struct mlx4_mr_list mr_list; /* Registered MR list. */ + struct mlx4_mr_list mr_free_list; /* Freed MR list. */ + } mr; LIST_HEAD(, mlx4_rss) rss; /**< Shared targets for Rx flow rules. */ LIST_HEAD(, rte_flow) flows; /**< Configured flow rule handles. */ struct ether_addr mac[MLX4_MAX_MAC_ADDRESSES]; diff --git a/drivers/net/mlx4/mlx4_mr.c b/drivers/net/mlx4/mlx4_mr.c index 3c87f6849c..d23d3c613e 100644 --- a/drivers/net/mlx4/mlx4_mr.c +++ b/drivers/net/mlx4/mlx4_mr.c @@ -30,8 +30,1152 @@ #include #include #include +#include #include "mlx4_glue.h" +#include "mlx4_mr.h" #include "mlx4_rxtx.h" #include "mlx4_utils.h" +struct mr_find_contig_memsegs_data { + uintptr_t addr; + uintptr_t start; + uintptr_t end; + const struct rte_memseg_list *msl; +}; + +struct mr_update_mp_data { + struct rte_eth_dev *dev; + struct mlx4_mr_ctrl *mr_ctrl; + int ret; +}; + +/** + * Expand B-tree table to a given size. Can't be called with holding + * memory_hotplug_lock or priv->mr.rwlock due to rte_realloc(). + * + * @param bt + * Pointer to B-tree structure. + * @param n + * Number of entries for expansion. + * + * @return + * 0 on success, -1 on failure. + */ +static int +mr_btree_expand(struct mlx4_mr_btree *bt, int n) +{ + void *mem; + int ret = 0; + + if (n <= bt->size) + return ret; + /* + * Downside of directly using rte_realloc() is that SOCKET_ID_ANY is + * used inside if there's no room to expand. Because this is a quite + * rare case and a part of very slow path, it is very acceptable. + * Initially cache_bh[] will be given practically enough space and once + * it is expanded, expansion wouldn't be needed again ever. + */ + mem = rte_realloc(bt->table, n * sizeof(struct mlx4_mr_cache), 0); + if (mem == NULL) { + /* Not an error, B-tree search will be skipped. */ + WARN("failed to expand MR B-tree (%p) table", (void *)bt); + ret = -1; + } else { + DEBUG("expanded MR B-tree table (size=%u)", n); + bt->table = mem; + bt->size = n; + } + return ret; +} + +/** + * Look up LKey from given B-tree lookup table, store the last index and return + * searched LKey. + * + * @param bt + * Pointer to B-tree structure. + * @param[out] idx + * Pointer to index. Even on search failure, returns index where it stops + * searching so that index can be used when inserting a new entry. + * @param addr + * Search key. + * + * @return + * Searched LKey on success, UINT32_MAX on no match. + */ +static uint32_t +mr_btree_lookup(struct mlx4_mr_btree *bt, uint16_t *idx, uintptr_t addr) +{ + struct mlx4_mr_cache *lkp_tbl; + uint16_t n; + uint16_t base = 0; + + assert(bt != NULL); + lkp_tbl = *bt->table; + n = bt->len; + /* First entry must be NULL for comparison. */ + assert(bt->len > 0 || (lkp_tbl[0].start == 0 && + lkp_tbl[0].lkey == UINT32_MAX)); + /* Binary search. */ + do { + register uint16_t delta = n >> 1; + + if (addr < lkp_tbl[base + delta].start) { + n = delta; + } else { + base += delta; + n -= delta; + } + } while (n > 1); + assert(addr >= lkp_tbl[base].start); + *idx = base; + if (addr < lkp_tbl[base].end) + return lkp_tbl[base].lkey; + /* Not found. */ + return UINT32_MAX; +} + +/** + * Insert an entry to B-tree lookup table. + * + * @param bt + * Pointer to B-tree structure. + * @param entry + * Pointer to new entry to insert. + * + * @return + * 0 on success, -1 on failure. + */ +static int +mr_btree_insert(struct mlx4_mr_btree *bt, struct mlx4_mr_cache *entry) +{ + struct mlx4_mr_cache *lkp_tbl; + uint16_t idx = 0; + size_t shift; + + assert(bt != NULL); + assert(bt->len <= bt->size); + assert(bt->len > 0); + lkp_tbl = *bt->table; + /* Find out the slot for insertion. */ + if (mr_btree_lookup(bt, &idx, entry->start) != UINT32_MAX) { + DEBUG("abort insertion to B-tree(%p): already exist at" + " idx=%u [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x", + (void *)bt, idx, entry->start, entry->end, entry->lkey); + /* Already exist, return. */ + return 0; + } + /* If table is full, return error. */ + if (unlikely(bt->len == bt->size)) { + bt->overflow = 1; + return -1; + } + /* Insert entry. */ + ++idx; + shift = (bt->len - idx) * sizeof(struct mlx4_mr_cache); + if (shift) + memmove(&lkp_tbl[idx + 1], &lkp_tbl[idx], shift); + lkp_tbl[idx] = *entry; + bt->len++; + DEBUG("inserted B-tree(%p)[%u]," + " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x", + (void *)bt, idx, entry->start, entry->end, entry->lkey); + return 0; +} + +/** + * Initialize B-tree and allocate memory for lookup table. + * + * @param bt + * Pointer to B-tree structure. + * @param n + * Number of entries to allocate. + * @param socket + * NUMA socket on which memory must be allocated. + * + * @return + * 0 on success, a negative errno value otherwise and rte_errno is set. + */ +int +mlx4_mr_btree_init(struct mlx4_mr_btree *bt, int n, int socket) +{ + if (bt == NULL) { + rte_errno = EINVAL; + return -rte_errno; + } + memset(bt, 0, sizeof(*bt)); + bt->table = rte_calloc_socket("B-tree table", + n, sizeof(struct mlx4_mr_cache), + 0, socket); + if (bt->table == NULL) { + rte_errno = ENOMEM; + ERROR("failed to allocate memory for btree cache on socket %d", + socket); + return -rte_errno; + } + bt->size = n; + /* First entry must be NULL for binary search. */ + (*bt->table)[bt->len++] = (struct mlx4_mr_cache) { + .lkey = UINT32_MAX, + }; + DEBUG("initialized B-tree %p with table %p", + (void *)bt, (void *)bt->table); + return 0; +} + +/** + * Free B-tree resources. + * + * @param bt + * Pointer to B-tree structure. + */ +void +mlx4_mr_btree_free(struct mlx4_mr_btree *bt) +{ + if (bt == NULL) + return; + DEBUG("freeing B-tree %p with table %p", (void *)bt, (void *)bt->table); + rte_free(bt->table); + memset(bt, 0, sizeof(*bt)); +} + +#ifndef NDEBUG +/** + * Dump all the entries in a B-tree + * + * @param bt + * Pointer to B-tree structure. + */ +void +mlx4_mr_btree_dump(struct mlx4_mr_btree *bt) +{ + int idx; + struct mlx4_mr_cache *lkp_tbl; + + if (bt == NULL) + return; + lkp_tbl = *bt->table; + for (idx = 0; idx < bt->len; ++idx) { + struct mlx4_mr_cache *entry = &lkp_tbl[idx]; + + DEBUG("B-tree(%p)[%u]," + " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x", + (void *)bt, idx, entry->start, entry->end, entry->lkey); + } +} +#endif + +/** + * Find virtually contiguous memory chunk in a given MR. + * + * @param dev + * Pointer to MR structure. + * @param[out] entry + * Pointer to returning MR cache entry. If not found, this will not be + * updated. + * @param start_idx + * Start index of the memseg bitmap. + * + * @return + * Next index to go on lookup. + */ +static int +mr_find_next_chunk(struct mlx4_mr *mr, struct mlx4_mr_cache *entry, + int base_idx) +{ + uintptr_t start = 0; + uintptr_t end = 0; + uint32_t idx = 0; + + for (idx = base_idx; idx < mr->ms_bmp_n; ++idx) { + if (rte_bitmap_get(mr->ms_bmp, idx)) { + const struct rte_memseg_list *msl; + const struct rte_memseg *ms; + + msl = mr->msl; + ms = rte_fbarray_get(&msl->memseg_arr, + mr->ms_base_idx + idx); + assert(msl->page_sz == ms->hugepage_sz); + if (!start) + start = ms->addr_64; + end = ms->addr_64 + ms->hugepage_sz; + } else if (start) { + /* Passed the end of a fragment. */ + break; + } + } + if (start) { + /* Found one chunk. */ + entry->start = start; + entry->end = end; + entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey); + } + return idx; +} + +/** + * Insert a MR to the global B-tree cache. It may fail due to low-on-memory. + * Then, this entry will have to be searched by mr_lookup_dev_list() in + * mlx4_mr_create() on miss. + * + * @param dev + * Pointer to Ethernet device. + * @param mr + * Pointer to MR to insert. + * + * @return + * 0 on success, -1 on failure. + */ +static int +mr_insert_dev_cache(struct rte_eth_dev *dev, struct mlx4_mr *mr) +{ + struct priv *priv = dev->data->dev_private; + unsigned int n; + + DEBUG("port %u inserting MR(%p) to global cache", + dev->data->port_id, (void *)mr); + for (n = 0; n < mr->ms_bmp_n; ) { + struct mlx4_mr_cache entry = { 0, }; + + /* Find a contiguous chunk and advance the index. */ + n = mr_find_next_chunk(mr, &entry, n); + if (!entry.end) + break; + if (mr_btree_insert(&priv->mr.cache, &entry) < 0) { + /* + * Overflowed, but the global table cannot be expanded + * because of deadlock. + */ + return -1; + } + } + return 0; +} + +/** + * Look up address in the original global MR list. + * + * @param dev + * Pointer to Ethernet device. + * @param[out] entry + * Pointer to returning MR cache entry. If no match, this will not be updated. + * @param addr + * Search key. + * + * @return + * Found MR on match, NULL otherwise. + */ +static struct mlx4_mr * +mr_lookup_dev_list(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry, + uintptr_t addr) +{ + struct priv *priv = dev->data->dev_private; + struct mlx4_mr *mr; + + /* Iterate all the existing MRs. */ + LIST_FOREACH(mr, &priv->mr.mr_list, mr) { + unsigned int n; + + if (mr->ms_n == 0) + continue; + for (n = 0; n < mr->ms_bmp_n; ) { + struct mlx4_mr_cache ret = { 0, }; + + n = mr_find_next_chunk(mr, &ret, n); + if (addr >= ret.start && addr < ret.end) { + /* Found. */ + *entry = ret; + return mr; + } + } + } + return NULL; +} + +/** + * Look up address on device. + * + * @param dev + * Pointer to Ethernet device. + * @param[out] entry + * Pointer to returning MR cache entry. If no match, this will not be updated. + * @param addr + * Search key. + * + * @return + * Searched LKey on success, UINT32_MAX on failure and rte_errno is set. + */ +static uint32_t +mr_lookup_dev(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry, + uintptr_t addr) +{ + struct priv *priv = dev->data->dev_private; + uint16_t idx; + uint32_t lkey = UINT32_MAX; + struct mlx4_mr *mr; + + /* + * If the global cache has overflowed since it failed to expand the + * B-tree table, it can't have all the existing MRs. Then, the address + * has to be searched by traversing the original MR list instead, which + * is very slow path. Otherwise, the global cache is all inclusive. + */ + if (!unlikely(priv->mr.cache.overflow)) { + lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr); + if (lkey != UINT32_MAX) + *entry = (*priv->mr.cache.table)[idx]; + } else { + /* Falling back to the slowest path. */ + mr = mr_lookup_dev_list(dev, entry, addr); + if (mr != NULL) + lkey = entry->lkey; + } + assert(lkey == UINT32_MAX || (addr >= entry->start && + addr < entry->end)); + return lkey; +} + +/** + * Free MR resources. MR lock must not be held to avoid a deadlock. rte_free() + * can raise memory free event and the callback function will spin on the lock. + * + * @param mr + * Pointer to MR to free. + */ +static void +mr_free(struct mlx4_mr *mr) +{ + if (mr == NULL) + return; + DEBUG("freeing MR(%p):", (void *)mr); + if (mr->ibv_mr != NULL) + claim_zero(mlx4_glue->dereg_mr(mr->ibv_mr)); + if (mr->ms_bmp != NULL) + rte_bitmap_free(mr->ms_bmp); + rte_free(mr); +} + +/** + * Releass resources of detached MR having no online entry. + * + * @param dev + * Pointer to Ethernet device. + */ +static void +mlx4_mr_garbage_collect(struct rte_eth_dev *dev) +{ + struct priv *priv = dev->data->dev_private; + struct mlx4_mr *mr_next; + struct mlx4_mr_list free_list = LIST_HEAD_INITIALIZER(free_list); + + /* + * MR can't be freed with holding the lock because rte_free() could call + * memory free callback function. This will be a deadlock situation. + */ + rte_rwlock_write_lock(&priv->mr.rwlock); + /* Detach the whole free list and release it after unlocking. */ + free_list = priv->mr.mr_free_list; + LIST_INIT(&priv->mr.mr_free_list); + rte_rwlock_write_unlock(&priv->mr.rwlock); + /* Release resources. */ + mr_next = LIST_FIRST(&free_list); + while (mr_next != NULL) { + struct mlx4_mr *mr = mr_next; + + mr_next = LIST_NEXT(mr, mr); + mr_free(mr); + } +} + +/* Called during rte_memseg_contig_walk() by mlx4_mr_create(). */ +static int +mr_find_contig_memsegs_cb(const struct rte_memseg_list *msl, + const struct rte_memseg *ms, size_t len, void *arg) +{ + struct mr_find_contig_memsegs_data *data = arg; + + if (data->addr < ms->addr_64 || data->addr >= ms->addr_64 + len) + return 0; + /* Found, save it and stop walking. */ + data->start = ms->addr_64; + data->end = ms->addr_64 + len; + data->msl = msl; + return 1; +} + +/** + * Create a new global Memroy Region (MR) for a missing virtual address. + * Register entire virtually contiguous memory chunk around the address. + * + * @param dev + * Pointer to Ethernet device. + * @param[out] entry + * Pointer to returning MR cache entry, found in the global cache or newly + * created. If failed to create one, this will not be updated. + * @param addr + * Target virtual address to register. + * + * @return + * Searched LKey on success, UINT32_MAX on failure and rte_errno is set. + */ +static uint32_t +mlx4_mr_create(struct rte_eth_dev *dev, struct mlx4_mr_cache *entry, + uintptr_t addr) +{ + struct priv *priv = dev->data->dev_private; + struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config; + const struct rte_memseg_list *msl; + const struct rte_memseg *ms; + struct mlx4_mr *mr = NULL; + size_t len; + uint32_t ms_n; + uint32_t bmp_size; + void *bmp_mem; + int ms_idx_shift = -1; + unsigned int n; + struct mr_find_contig_memsegs_data data = { + .addr = addr, + }; + struct mr_find_contig_memsegs_data data_re; + + DEBUG("port %u creating a MR using address (%p)", + dev->data->port_id, (void *)addr); + /* + * Release detached MRs if any. This can't be called with holding either + * memory_hotplug_lock or priv->mr.rwlock. MRs on the free list have + * been detached by the memory free event but it couldn't be released + * inside the callback due to deadlock. As a result, releasing resources + * is quite opportunistic. + */ + mlx4_mr_garbage_collect(dev); + /* + * Find out a contiguous virtual address chunk in use, to which the + * given address belongs, in order to register maximum range. In the + * best case where mempools are not dynamically recreated and + * '--socket-mem' is speicified as an EAL option, it is very likely to + * have only one MR(LKey) per a socket and per a hugepage-size even + * though the system memory is highly fragmented. + */ + if (!rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data)) { + WARN("port %u unable to find virtually contiguous" + " chunk for address (%p)." + " rte_memseg_contig_walk() failed.", + dev->data->port_id, (void *)addr); + rte_errno = ENXIO; + goto err_nolock; + } +alloc_resources: + /* Addresses must be page-aligned. */ + assert(rte_is_aligned((void *)data.start, data.msl->page_sz)); + assert(rte_is_aligned((void *)data.end, data.msl->page_sz)); + msl = data.msl; + ms = rte_mem_virt2memseg((void *)data.start, msl); + len = data.end - data.start; + assert(msl->page_sz == ms->hugepage_sz); + /* Number of memsegs in the range. */ + ms_n = len / msl->page_sz; + DEBUG("port %u extending %p to [0x%" PRIxPTR ", 0x%" PRIxPTR ")," + " page_sz=0x%" PRIx64 ", ms_n=%u", + dev->data->port_id, (void *)addr, + data.start, data.end, msl->page_sz, ms_n); + /* Size of memory for bitmap. */ + bmp_size = rte_bitmap_get_memory_footprint(ms_n); + mr = rte_zmalloc_socket(NULL, + RTE_ALIGN_CEIL(sizeof(*mr), + RTE_CACHE_LINE_SIZE) + + bmp_size, + RTE_CACHE_LINE_SIZE, msl->socket_id); + if (mr == NULL) { + WARN("port %u unable to allocate memory for a new MR of" + " address (%p).", + dev->data->port_id, (void *)addr); + rte_errno = ENOMEM; + goto err_nolock; + } + mr->msl = msl; + /* + * Save the index of the first memseg and initialize memseg bitmap. To + * see if a memseg of ms_idx in the memseg-list is still valid, check: + * rte_bitmap_get(mr->bmp, ms_idx - mr->ms_base_idx) + */ + mr->ms_base_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms); + bmp_mem = RTE_PTR_ALIGN_CEIL(mr + 1, RTE_CACHE_LINE_SIZE); + mr->ms_bmp = rte_bitmap_init(ms_n, bmp_mem, bmp_size); + if (mr->ms_bmp == NULL) { + WARN("port %u unable to initialize bitamp for a new MR of" + " address (%p).", + dev->data->port_id, (void *)addr); + rte_errno = EINVAL; + goto err_nolock; + } + /* + * Should recheck whether the extended contiguous chunk is still valid. + * Because memory_hotplug_lock can't be held if there's any memory + * related calls in a critical path, resource allocation above can't be + * locked. If the memory has been changed at this point, try again with + * just single page. If not, go on with the big chunk atomically from + * here. + */ + rte_rwlock_read_lock(&mcfg->memory_hotplug_lock); + data_re = data; + if (len > msl->page_sz && + !rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data_re)) { + WARN("port %u unable to find virtually contiguous" + " chunk for address (%p)." + " rte_memseg_contig_walk() failed.", + dev->data->port_id, (void *)addr); + rte_errno = ENXIO; + goto err_memlock; + } + if (data.start != data_re.start || data.end != data_re.end) { + /* + * The extended contiguous chunk has been changed. Try again + * with single memseg instead. + */ + data.start = RTE_ALIGN_FLOOR(addr, msl->page_sz); + data.end = data.start + msl->page_sz; + rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock); + mr_free(mr); + goto alloc_resources; + } + assert(data.msl == data_re.msl); + rte_rwlock_write_lock(&priv->mr.rwlock); + /* + * Check the address is really missing. If other thread already created + * one or it is not found due to overflow, abort and return. + */ + if (mr_lookup_dev(dev, entry, addr) != UINT32_MAX) { + /* + * Insert to the global cache table. It may fail due to + * low-on-memory. Then, this entry will have to be searched + * here again. + */ + mr_btree_insert(&priv->mr.cache, entry); + DEBUG("port %u found MR for %p on final lookup, abort", + dev->data->port_id, (void *)addr); + rte_rwlock_write_unlock(&priv->mr.rwlock); + rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock); + /* + * Must be unlocked before calling rte_free() because + * mlx4_mr_mem_event_free_cb() can be called inside. + */ + mr_free(mr); + return entry->lkey; + } + /* + * Trim start and end addresses for verbs MR. Set bits for registering + * memsegs but exclude already registered ones. Bitmap can be + * fragmented. + */ + for (n = 0; n < ms_n; ++n) { + uintptr_t start; + struct mlx4_mr_cache ret = { 0, }; + + start = data_re.start + n * msl->page_sz; + /* Exclude memsegs already registered by other MRs. */ + if (mr_lookup_dev(dev, &ret, start) == UINT32_MAX) { + /* + * Start from the first unregistered memseg in the + * extended range. + */ + if (ms_idx_shift == -1) { + mr->ms_base_idx += n; + data.start = start; + ms_idx_shift = n; + } + data.end = start + msl->page_sz; + rte_bitmap_set(mr->ms_bmp, n - ms_idx_shift); + ++mr->ms_n; + } + } + len = data.end - data.start; + mr->ms_bmp_n = len / msl->page_sz; + assert(ms_idx_shift + mr->ms_bmp_n <= ms_n); + /* + * Finally create a verbs MR for the memory chunk. ibv_reg_mr() can be + * called with holding the memory lock because it doesn't use + * mlx4_alloc_buf_extern() which eventually calls rte_malloc_socket() + * through mlx4_alloc_verbs_buf(). + */ + mr->ibv_mr = mlx4_glue->reg_mr(priv->pd, (void *)data.start, len, + IBV_ACCESS_LOCAL_WRITE); + if (mr->ibv_mr == NULL) { + WARN("port %u fail to create a verbs MR for address (%p)", + dev->data->port_id, (void *)addr); + rte_errno = EINVAL; + goto err_mrlock; + } + assert((uintptr_t)mr->ibv_mr->addr == data.start); + assert(mr->ibv_mr->length == len); + LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr); + DEBUG("port %u MR CREATED (%p) for %p:\n" + " [0x%" PRIxPTR ", 0x%" PRIxPTR ")," + " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u", + dev->data->port_id, (void *)mr, (void *)addr, + data.start, data.end, rte_cpu_to_be_32(mr->ibv_mr->lkey), + mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n); + /* Insert to the global cache table. */ + mr_insert_dev_cache(dev, mr); + /* Fill in output data. */ + mr_lookup_dev(dev, entry, addr); + /* Lookup can't fail. */ + assert(entry->lkey != UINT32_MAX); + rte_rwlock_write_unlock(&priv->mr.rwlock); + rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock); + return entry->lkey; +err_mrlock: + rte_rwlock_write_unlock(&priv->mr.rwlock); +err_memlock: + rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock); +err_nolock: + /* + * In case of error, as this can be called in a datapath, a warning + * message per an error is preferable instead. Must be unlocked before + * calling rte_free() because mlx4_mr_mem_event_free_cb() can be called + * inside. + */ + mr_free(mr); + return UINT32_MAX; +} + +/** + * Rebuild the global B-tree cache of device from the original MR list. + * + * @param dev + * Pointer to Ethernet device. + */ +static void +mr_rebuild_dev_cache(struct rte_eth_dev *dev) +{ + struct priv *priv = dev->data->dev_private; + struct mlx4_mr *mr; + + DEBUG("port %u rebuild dev cache[]", dev->data->port_id); + /* Flush cache to rebuild. */ + priv->mr.cache.len = 1; + priv->mr.cache.overflow = 0; + /* Iterate all the existing MRs. */ + LIST_FOREACH(mr, &priv->mr.mr_list, mr) + if (mr_insert_dev_cache(dev, mr) < 0) + return; +} + +/** + * Callback for memory free event. Iterate freed memsegs and check whether it + * belongs to an existing MR. If found, clear the bit from bitmap of MR. As a + * result, the MR would be fragmented. If it becomes empty, the MR will be freed + * later by mlx4_mr_garbage_collect(). + * + * The global cache must be rebuilt if there's any change and this event has to + * be propagated to dataplane threads to flush the local caches. + * + * @param dev + * Pointer to Ethernet device. + * @param addr + * Address of freed memory. + * @param len + * Size of freed memory. + */ +static void +mlx4_mr_mem_event_free_cb(struct rte_eth_dev *dev, const void *addr, size_t len) +{ + struct priv *priv = dev->data->dev_private; + const struct rte_memseg_list *msl; + struct mlx4_mr *mr; + int ms_n; + int i; + int rebuild = 0; + + DEBUG("port %u free callback: addr=%p, len=%zu", + dev->data->port_id, addr, len); + msl = rte_mem_virt2memseg_list(addr); + /* addr and len must be page-aligned. */ + assert((uintptr_t)addr == RTE_ALIGN((uintptr_t)addr, msl->page_sz)); + assert(len == RTE_ALIGN(len, msl->page_sz)); + ms_n = len / msl->page_sz; + rte_rwlock_write_lock(&priv->mr.rwlock); + /* Clear bits of freed memsegs from MR. */ + for (i = 0; i < ms_n; ++i) { + const struct rte_memseg *ms; + struct mlx4_mr_cache entry; + uintptr_t start; + int ms_idx; + uint32_t pos; + + /* Find MR having this memseg. */ + start = (uintptr_t)addr + i * msl->page_sz; + mr = mr_lookup_dev_list(dev, &entry, start); + if (mr == NULL) + continue; + ms = rte_mem_virt2memseg((void *)start, msl); + assert(ms != NULL); + assert(msl->page_sz == ms->hugepage_sz); + ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms); + pos = ms_idx - mr->ms_base_idx; + assert(rte_bitmap_get(mr->ms_bmp, pos)); + assert(pos < mr->ms_bmp_n); + DEBUG("port %u MR(%p): clear bitmap[%u] for addr %p", + dev->data->port_id, (void *)mr, pos, (void *)start); + rte_bitmap_clear(mr->ms_bmp, pos); + if (--mr->ms_n == 0) { + LIST_REMOVE(mr, mr); + LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr); + DEBUG("port %u remove MR(%p) from list", + dev->data->port_id, (void *)mr); + } + /* + * MR is fragmented or will be freed. the global cache must be + * rebuilt. + */ + rebuild = 1; + } + if (rebuild) { + mr_rebuild_dev_cache(dev); + /* + * Flush local caches by propagating invalidation across cores. + * rte_smp_wmb() is enough to synchronize this event. If one of + * freed memsegs is seen by other core, that means the memseg + * has been allocated by allocator, which will come after this + * free call. Therefore, this store instruction (incrementing + * generation below) will be guaranteed to be seen by other core + * before the core sees the newly allocated memory. + */ + ++priv->mr.dev_gen; + DEBUG("broadcasting local cache flush, gen=%d", + priv->mr.dev_gen); + rte_smp_wmb(); + } + rte_rwlock_write_unlock(&priv->mr.rwlock); +#ifndef NDEBUG + if (rebuild) + mlx4_mr_dump_dev(dev); +#endif +} + +/** + * Callback for memory event. + * + * @param event_type + * Memory event type. + * @param addr + * Address of memory. + * @param len + * Size of memory. + */ +void +mlx4_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr, + size_t len, void *arg __rte_unused) +{ + struct priv *priv; + + switch (event_type) { + case RTE_MEM_EVENT_FREE: + rte_rwlock_read_lock(&mlx4_mem_event_rwlock); + /* Iterate all the existing mlx4 devices. */ + LIST_FOREACH(priv, &mlx4_mem_event_cb_list, mem_event_cb) + mlx4_mr_mem_event_free_cb(priv->dev, addr, len); + rte_rwlock_read_unlock(&mlx4_mem_event_rwlock); + break; + case RTE_MEM_EVENT_ALLOC: + default: + break; + } +} + +/** + * Look up address in the global MR cache table. If not found, create a new MR. + * Insert the found/created entry to local bottom-half cache table. + * + * @param dev + * Pointer to Ethernet device. + * @param mr_ctrl + * Pointer to per-queue MR control structure. + * @param[out] entry + * Pointer to returning MR cache entry, found in the global cache or newly + * created. If failed to create one, this is not written. + * @param addr + * Search key. + * + * @return + * Searched LKey on success, UINT32_MAX on no match. + */ +static uint32_t +mlx4_mr_lookup_dev(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl, + struct mlx4_mr_cache *entry, uintptr_t addr) +{ + struct priv *priv = dev->data->dev_private; + struct mlx4_mr_btree *bt = &mr_ctrl->cache_bh; + uint16_t idx; + uint32_t lkey; + + /* If local cache table is full, try to double it. */ + if (unlikely(bt->len == bt->size)) + mr_btree_expand(bt, bt->size << 1); + /* Look up in the global cache. */ + rte_rwlock_read_lock(&priv->mr.rwlock); + lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr); + if (lkey != UINT32_MAX) { + /* Found. */ + *entry = (*priv->mr.cache.table)[idx]; + rte_rwlock_read_unlock(&priv->mr.rwlock); + /* + * Update local cache. Even if it fails, return the found entry + * to update top-half cache. Next time, this entry will be found + * in the global cache. + */ + mr_btree_insert(bt, entry); + return lkey; + } + rte_rwlock_read_unlock(&priv->mr.rwlock); + /* First time to see the address? Create a new MR. */ + lkey = mlx4_mr_create(dev, entry, addr); + /* + * Update the local cache if successfully created a new global MR. Even + * if failed to create one, there's no action to take in this datapath + * code. As returning LKey is invalid, this will eventually make HW + * fail. + */ + if (lkey != UINT32_MAX) + mr_btree_insert(bt, entry); + return lkey; +} + +/** + * Bottom-half of LKey search on datapath. Firstly search in cache_bh[] and if + * misses, search in the global MR cache table and update the new entry to + * per-queue local caches. + * + * @param dev + * Pointer to Ethernet device. + * @param mr_ctrl + * Pointer to per-queue MR control structure. + * @param addr + * Search key. + * + * @return + * Searched LKey on success, UINT32_MAX on no match. + */ +static uint32_t +mlx4_mr_addr2mr_bh(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl, + uintptr_t addr) +{ + uint32_t lkey; + uint16_t bh_idx = 0; + /* Victim in top-half cache to replace with new entry. */ + struct mlx4_mr_cache *repl = &mr_ctrl->cache[mr_ctrl->head]; + + /* Binary-search MR translation table. */ + lkey = mr_btree_lookup(&mr_ctrl->cache_bh, &bh_idx, addr); + /* Update top-half cache. */ + if (likely(lkey != UINT32_MAX)) { + *repl = (*mr_ctrl->cache_bh.table)[bh_idx]; + } else { + /* + * If missed in local lookup table, search in the global cache + * and local cache_bh[] will be updated inside if possible. + * Top-half cache entry will also be updated. + */ + lkey = mlx4_mr_lookup_dev(dev, mr_ctrl, repl, addr); + if (unlikely(lkey == UINT32_MAX)) + return UINT32_MAX; + } + /* Update the most recently used entry. */ + mr_ctrl->mru = mr_ctrl->head; + /* Point to the next victim, the oldest. */ + mr_ctrl->head = (mr_ctrl->head + 1) % MLX4_MR_CACHE_N; + return lkey; +} + +/** + * Bottom-half of LKey search on Rx. + * + * @param rxq + * Pointer to Rx queue structure. + * @param addr + * Search key. + * + * @return + * Searched LKey on success, UINT32_MAX on no match. + */ +uint32_t +mlx4_rx_addr2mr_bh(struct rxq *rxq, uintptr_t addr) +{ + struct mlx4_mr_ctrl *mr_ctrl = &rxq->mr_ctrl; + struct priv *priv = rxq->priv; + + DEBUG("Rx queue %u: miss on top-half, mru=%u, head=%u, addr=%p", + rxq->stats.idx, mr_ctrl->mru, mr_ctrl->head, (void *)addr); + return mlx4_mr_addr2mr_bh(priv->dev, mr_ctrl, addr); +} + +/** + * Bottom-half of LKey search on Tx. + * + * @param txq + * Pointer to Tx queue structure. + * @param addr + * Search key. + * + * @return + * Searched LKey on success, UINT32_MAX on no match. + */ +uint32_t +mlx4_tx_addr2mr_bh(struct txq *txq, uintptr_t addr) +{ + struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl; + struct priv *priv = txq->priv; + + DEBUG("Tx queue %u: miss on top-half, mru=%u, head=%u, addr=%p", + txq->stats.idx, mr_ctrl->mru, mr_ctrl->head, (void *)addr); + return mlx4_mr_addr2mr_bh(priv->dev, mr_ctrl, addr); +} + +/** + * Flush all of the local cache entries. + * + * @param mr_ctrl + * Pointer to per-queue MR control structure. + */ +void +mlx4_mr_flush_local_cache(struct mlx4_mr_ctrl *mr_ctrl) +{ + /* Reset the most-recently-used index. */ + mr_ctrl->mru = 0; + /* Reset the linear search array. */ + mr_ctrl->head = 0; + memset(mr_ctrl->cache, 0, sizeof(mr_ctrl->cache)); + /* Reset the B-tree table. */ + mr_ctrl->cache_bh.len = 1; + mr_ctrl->cache_bh.overflow = 0; + /* Update the generation number. */ + mr_ctrl->cur_gen = *mr_ctrl->dev_gen_ptr; + DEBUG("mr_ctrl(%p): flushed, cur_gen=%d", + (void *)mr_ctrl, mr_ctrl->cur_gen); +} + +/* Called during rte_mempool_mem_iter() by mlx4_mr_update_mp(). */ +static void +mlx4_mr_update_mp_cb(struct rte_mempool *mp __rte_unused, void *opaque, + struct rte_mempool_memhdr *memhdr, + unsigned mem_idx __rte_unused) +{ + struct mr_update_mp_data *data = opaque; + uint32_t lkey; + + /* Stop iteration if failed in the previous walk. */ + if (data->ret < 0) + return; + /* Register address of the chunk and update local caches. */ + lkey = mlx4_mr_addr2mr_bh(data->dev, data->mr_ctrl, + (uintptr_t)memhdr->addr); + if (lkey == UINT32_MAX) + data->ret = -1; +} + +/** + * Register entire memory chunks in a Mempool. + * + * @param dev + * Pointer to Ethernet device. + * @param mr_ctrl + * Pointer to per-queue MR control structure. + * @param mp + * Pointer to registering Mempool. + * + * @return + * 0 on success, -1 on failure. + */ +int +mlx4_mr_update_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl, + struct rte_mempool *mp) +{ + struct mr_update_mp_data data = { + .dev = dev, + .mr_ctrl = mr_ctrl, + .ret = 0, + }; + + rte_mempool_mem_iter(mp, mlx4_mr_update_mp_cb, &data); + return data.ret; +} + +#ifndef NDEBUG +/** + * Dump all the created MRs and the global cache entries. + * + * @param dev + * Pointer to Ethernet device. + */ +void +mlx4_mr_dump_dev(struct rte_eth_dev *dev) +{ + struct priv *priv = dev->data->dev_private; + struct mlx4_mr *mr; + int mr_n = 0; + int chunk_n = 0; + + rte_rwlock_read_lock(&priv->mr.rwlock); + /* Iterate all the existing MRs. */ + LIST_FOREACH(mr, &priv->mr.mr_list, mr) { + unsigned int n; + + DEBUG("port %u MR[%u], LKey = 0x%x, ms_n = %u, ms_bmp_n = %u", + dev->data->port_id, mr_n++, + rte_cpu_to_be_32(mr->ibv_mr->lkey), + mr->ms_n, mr->ms_bmp_n); + if (mr->ms_n == 0) + continue; + for (n = 0; n < mr->ms_bmp_n; ) { + struct mlx4_mr_cache ret = { 0, }; + + n = mr_find_next_chunk(mr, &ret, n); + if (!ret.end) + break; + DEBUG(" chunk[%u], [0x%" PRIxPTR ", 0x%" PRIxPTR ")", + chunk_n++, ret.start, ret.end); + } + } + DEBUG("port %u dumping global cache", dev->data->port_id); + mlx4_mr_btree_dump(&priv->mr.cache); + rte_rwlock_read_unlock(&priv->mr.rwlock); +} +#endif + +/** + * Release all the created MRs and resources. Remove device from memory callback + * list. + * + * @param dev + * Pointer to Ethernet device. + */ +void +mlx4_mr_release(struct rte_eth_dev *dev) +{ + struct priv *priv = dev->data->dev_private; + struct mlx4_mr *mr_next = LIST_FIRST(&priv->mr.mr_list); + + /* Remove from memory callback device list. */ + rte_rwlock_write_lock(&mlx4_mem_event_rwlock); + LIST_REMOVE(priv, mem_event_cb); + rte_rwlock_write_unlock(&mlx4_mem_event_rwlock); +#ifndef NDEBUG + mlx4_mr_dump_dev(dev); +#endif + rte_rwlock_write_lock(&priv->mr.rwlock); + /* Detach from MR list and move to free list. */ + while (mr_next != NULL) { + struct mlx4_mr *mr = mr_next; + + mr_next = LIST_NEXT(mr, mr); + LIST_REMOVE(mr, mr); + LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr); + } + LIST_INIT(&priv->mr.mr_list); + /* Free global cache. */ + mlx4_mr_btree_free(&priv->mr.cache); + rte_rwlock_write_unlock(&priv->mr.rwlock); + /* Free all remaining MRs. */ + mlx4_mr_garbage_collect(dev); +} diff --git a/drivers/net/mlx4/mlx4_mr.h b/drivers/net/mlx4/mlx4_mr.h new file mode 100644 index 0000000000..37a365a8b5 --- /dev/null +++ b/drivers/net/mlx4/mlx4_mr.h @@ -0,0 +1,122 @@ +/* SPDX-License-Identifier: BSD-3-Clause + * Copyright 2018 6WIND S.A. + * Copyright 2018 Mellanox Technologies, Ltd + */ + +#ifndef RTE_PMD_MLX4_MR_H_ +#define RTE_PMD_MLX4_MR_H_ + +#include +#include +#include + +/* Verbs headers do not support -pedantic. */ +#ifdef PEDANTIC +#pragma GCC diagnostic ignored "-Wpedantic" +#endif +#include +#ifdef PEDANTIC +#pragma GCC diagnostic error "-Wpedantic" +#endif + +#include +#include +#include +#include + +/* Size of per-queue MR cache array for linear search. */ +#define MLX4_MR_CACHE_N 8 + +/* Size of MR cache table for binary search. */ +#define MLX4_MR_BTREE_CACHE_N 256 + +/* Memory Region object. */ +struct mlx4_mr { + LIST_ENTRY(mlx4_mr) mr; /**< Pointer to the prev/next entry. */ + struct ibv_mr *ibv_mr; /* Verbs Memory Region. */ + const struct rte_memseg_list *msl; + int ms_base_idx; /* Start index of msl->memseg_arr[]. */ + int ms_n; /* Number of memsegs in use. */ + uint32_t ms_bmp_n; /* Number of bits in memsegs bit-mask. */ + struct rte_bitmap *ms_bmp; /* Bit-mask of memsegs belonged to MR. */ +}; + +/* Cache entry for Memory Region. */ +struct mlx4_mr_cache { + uintptr_t start; /* Start address of MR. */ + uintptr_t end; /* End address of MR. */ + uint32_t lkey; /* rte_cpu_to_be_32(ibv_mr->lkey). */ +} __rte_packed; + +/* MR Cache table for Binary search. */ +struct mlx4_mr_btree { + uint16_t len; /* Number of entries. */ + uint16_t size; /* Total number of entries. */ + int overflow; /* Mark failure of table expansion. */ + struct mlx4_mr_cache (*table)[]; +} __rte_packed; + +/* Per-queue MR control descriptor. */ +struct mlx4_mr_ctrl { + uint32_t *dev_gen_ptr; /* Generation number of device to poll. */ + uint32_t cur_gen; /* Generation number saved to flush caches. */ + uint16_t mru; /* Index of last hit entry in top-half cache. */ + uint16_t head; /* Index of the oldest entry in top-half cache. */ + struct mlx4_mr_cache cache[MLX4_MR_CACHE_N]; /* Cache for top-half. */ + struct mlx4_mr_btree cache_bh; /* Cache for bottom-half. */ +} __rte_packed; + +extern struct mlx4_dev_list mlx4_mem_event_cb_list; +extern rte_rwlock_t mlx4_mem_event_rwlock; + +/* First entry must be NULL for comparison. */ +#define mlx4_mr_btree_len(bt) ((bt)->len - 1) + +int mlx4_mr_btree_init(struct mlx4_mr_btree *bt, int n, int socket); +void mlx4_mr_btree_free(struct mlx4_mr_btree *bt); +void mlx4_mr_btree_dump(struct mlx4_mr_btree *bt); +void mlx4_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr, + size_t len, void *arg); +int mlx4_mr_update_mp(struct rte_eth_dev *dev, struct mlx4_mr_ctrl *mr_ctrl, + struct rte_mempool *mp); +void mlx4_mr_dump_dev(struct rte_eth_dev *dev); +void mlx4_mr_release(struct rte_eth_dev *dev); + +/** + * Look up LKey from given lookup table by linear search. Firstly look up the + * last-hit entry. If miss, the entire array is searched. If found, update the + * last-hit index and return LKey. + * + * @param lkp_tbl + * Pointer to lookup table. + * @param[in,out] cached_idx + * Pointer to last-hit index. + * @param n + * Size of lookup table. + * @param addr + * Search key. + * + * @return + * Searched LKey on success, UINT32_MAX on no match. + */ +static __rte_always_inline uint32_t +mlx4_mr_lookup_cache(struct mlx4_mr_cache *lkp_tbl, uint16_t *cached_idx, + uint16_t n, uintptr_t addr) +{ + uint16_t idx; + + if (likely(addr >= lkp_tbl[*cached_idx].start && + addr < lkp_tbl[*cached_idx].end)) + return lkp_tbl[*cached_idx].lkey; + for (idx = 0; idx < n && lkp_tbl[idx].start != 0; ++idx) { + if (addr >= lkp_tbl[idx].start && + addr < lkp_tbl[idx].end) { + /* Found. */ + *cached_idx = idx; + return lkp_tbl[idx].lkey; + } + } + return UINT32_MAX; +} + +#endif /* RTE_PMD_MLX4_MR_H_ */ diff --git a/drivers/net/mlx4/mlx4_rxq.c b/drivers/net/mlx4/mlx4_rxq.c index 18b9d52e3b..87688c1c7e 100644 --- a/drivers/net/mlx4/mlx4_rxq.c +++ b/drivers/net/mlx4/mlx4_rxq.c @@ -488,6 +488,7 @@ mlx4_rxq_attach(struct rxq *rxq) } struct priv *priv = rxq->priv; + struct rte_eth_dev *dev = priv->dev; const uint32_t elts_n = 1 << rxq->elts_n; const uint32_t sges_n = 1 << rxq->sges_n; struct rte_mbuf *(*elts)[elts_n] = rxq->elts; @@ -552,6 +553,11 @@ mlx4_rxq_attach(struct rxq *rxq) msg = "failed to obtain device information from WQ/CQ objects"; goto error; } + /* Pre-register Rx mempool. */ + DEBUG("port %u Rx queue %u registering mp %s having %u chunks", + priv->dev->data->port_id, rxq->stats.idx, + rxq->mp->name, rxq->mp->nb_mem_chunks); + mlx4_mr_update_mp(dev, &rxq->mr_ctrl, rxq->mp); wqes = (volatile struct mlx4_wqe_data_seg (*)[]) ((uintptr_t)dv_rwq.buf.buf + dv_rwq.rq.offset); for (i = 0; i != RTE_DIM(*elts); ++i) { @@ -583,7 +589,7 @@ mlx4_rxq_attach(struct rxq *rxq) .addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(buf, uintptr_t)), .byte_count = rte_cpu_to_be_32(buf->data_len), - .lkey = UINT32_MAX, + .lkey = mlx4_rx_mb2mr(rxq, buf), }; (*elts)[i] = buf; } @@ -855,6 +861,11 @@ mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc, 1 << rxq->sges_n); goto error; } + if (mlx4_mr_btree_init(&rxq->mr_ctrl.cache_bh, + MLX4_MR_BTREE_CACHE_N, socket)) { + /* rte_errno is already set. */ + goto error; + } if (dev->data->dev_conf.intr_conf.rxq) { rxq->channel = mlx4_glue->create_comp_channel(priv->ctx); if (rxq->channel == NULL) { @@ -912,5 +923,6 @@ mlx4_rx_queue_release(void *dpdk_rxq) assert(!rxq->rq_db); if (rxq->channel) claim_zero(mlx4_glue->destroy_comp_channel(rxq->channel)); + mlx4_mr_btree_free(&rxq->mr_ctrl.cache_bh); rte_free(rxq); } diff --git a/drivers/net/mlx4/mlx4_rxtx.c b/drivers/net/mlx4/mlx4_rxtx.c index 21ffd435aa..44efeb4470 100644 --- a/drivers/net/mlx4/mlx4_rxtx.c +++ b/drivers/net/mlx4/mlx4_rxtx.c @@ -343,24 +343,6 @@ mlx4_txq_complete(struct txq *txq, const unsigned int elts_m, txq->elts_tail = elts_tail; } -/** - * Get memory pool (MP) from mbuf. If mbuf is indirect, the pool from which - * the cloned mbuf is allocated is returned instead. - * - * @param buf - * Pointer to mbuf. - * - * @return - * Memory pool where data is located for given mbuf. - */ -static struct rte_mempool * -mlx4_txq_mb2mp(struct rte_mbuf *buf) -{ - if (unlikely(RTE_MBUF_INDIRECT(buf))) - return rte_mbuf_from_indirect(buf)->pool; - return buf->pool; -} - /** * Write Tx data segment to the SQ. * @@ -378,7 +360,7 @@ mlx4_fill_tx_data_seg(volatile struct mlx4_wqe_data_seg *dseg, uint32_t lkey, uintptr_t addr, rte_be32_t byte_count) { dseg->addr = rte_cpu_to_be_64(addr); - dseg->lkey = rte_cpu_to_be_32(lkey); + dseg->lkey = lkey; #if RTE_CACHE_LINE_SIZE < 64 /* * Need a barrier here before writing the byte_count @@ -437,7 +419,7 @@ mlx4_tx_burst_segs(struct rte_mbuf *buf, struct txq *txq, goto txbb_tail_segs; txbb_head_seg: /* Memory region key (big endian) for this memory pool. */ - lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(sbuf)); + lkey = mlx4_tx_mb2mr(txq, sbuf); if (unlikely(lkey == (uint32_t)-1)) { DEBUG("%p: unable to get MP <-> MR association", (void *)txq); @@ -449,7 +431,7 @@ txbb_head_seg: dseg = (volatile struct mlx4_wqe_data_seg *) sq->buf; dseg->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(sbuf, uintptr_t)); - dseg->lkey = rte_cpu_to_be_32(lkey); + dseg->lkey = lkey; /* * This data segment starts at the beginning of a new * TXBB, so we need to postpone its byte_count writing @@ -469,7 +451,7 @@ txbb_tail_segs: /* Jump to default if there are more than two segments remaining. */ switch (nb_segs) { default: - lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(sbuf)); + lkey = mlx4_tx_mb2mr(txq, sbuf); if (unlikely(lkey == (uint32_t)-1)) { DEBUG("%p: unable to get MP <-> MR association", (void *)txq); @@ -485,7 +467,7 @@ txbb_tail_segs: nb_segs--; /* fallthrough */ case 2: - lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(sbuf)); + lkey = mlx4_tx_mb2mr(txq, sbuf); if (unlikely(lkey == (uint32_t)-1)) { DEBUG("%p: unable to get MP <-> MR association", (void *)txq); @@ -501,7 +483,7 @@ txbb_tail_segs: nb_segs--; /* fallthrough */ case 1: - lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(sbuf)); + lkey = mlx4_tx_mb2mr(txq, sbuf); if (unlikely(lkey == (uint32_t)-1)) { DEBUG("%p: unable to get MP <-> MR association", (void *)txq); @@ -611,7 +593,7 @@ mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n) elt->buf = NULL; break; } - lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(buf)); + lkey = mlx4_tx_mb2mr(txq, buf); if (unlikely(lkey == (uint32_t)-1)) { /* MR does not exist. */ DEBUG("%p: unable to get MP <-> MR association", @@ -966,6 +948,9 @@ mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) * changes. */ scat->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t)); + /* If there's only one MR, no need to replace LKey in WQE. */ + if (unlikely(mlx4_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1)) + scat->lkey = mlx4_rx_mb2mr(rxq, rep); if (len > seg->data_len) { len -= seg->data_len; ++pkt->nb_segs; diff --git a/drivers/net/mlx4/mlx4_rxtx.h b/drivers/net/mlx4/mlx4_rxtx.h index 2f9d3798b0..4c025e3a01 100644 --- a/drivers/net/mlx4/mlx4_rxtx.h +++ b/drivers/net/mlx4/mlx4_rxtx.h @@ -25,6 +25,7 @@ #include "mlx4.h" #include "mlx4_prm.h" +#include "mlx4_mr.h" /** Rx queue counters. */ struct mlx4_rxq_stats { @@ -46,6 +47,7 @@ struct rxq { uint16_t port_id; /**< Port ID for incoming packets. */ uint16_t sges_n; /**< Number of segments per packet (log2 value). */ uint16_t elts_n; /**< Mbuf queue size (log2 value). */ + struct mlx4_mr_ctrl mr_ctrl; /* MR control descriptor. */ struct rte_mbuf *(*elts)[]; /**< Rx elements. */ volatile struct mlx4_wqe_data_seg (*wqes)[]; /**< HW queue entries. */ volatile uint32_t *rq_db; /**< RQ doorbell record. */ @@ -100,6 +102,7 @@ struct txq { int elts_comp_cd; /**< Countdown for next completion. */ unsigned int elts_comp_cd_init; /**< Initial value for countdown. */ unsigned int elts_n; /**< (*elts)[] length. */ + struct mlx4_mr_ctrl mr_ctrl; /* MR control descriptor. */ struct txq_elt (*elts)[]; /**< Tx elements. */ struct mlx4_txq_stats stats; /**< Tx queue counters. */ uint32_t max_inline; /**< Max inline send size. */ @@ -155,12 +158,70 @@ int mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, const struct rte_eth_txconf *conf); void mlx4_tx_queue_release(void *dpdk_txq); -static inline uint32_t -mlx4_txq_mp2mr(struct txq *txq, struct rte_mempool *mp) +/* mlx4_mr.c */ + +void mlx4_mr_flush_local_cache(struct mlx4_mr_ctrl *mr_ctrl); +uint32_t mlx4_rx_addr2mr_bh(struct rxq *rxq, uintptr_t addr); +uint32_t mlx4_tx_addr2mr_bh(struct txq *txq, uintptr_t addr); + +/** + * Query LKey from a packet buffer for Rx. No need to flush local caches for Rx + * as mempool is pre-configured and static. + * + * @param rxq + * Pointer to Rx queue structure. + * @param addr + * Address to search. + * + * @return + * Searched LKey on success, UINT32_MAX on no match. + */ +static __rte_always_inline uint32_t +mlx4_rx_addr2mr(struct rxq *rxq, uintptr_t addr) +{ + struct mlx4_mr_ctrl *mr_ctrl = &rxq->mr_ctrl; + uint32_t lkey; + + /* Linear search on MR cache array. */ + lkey = mlx4_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru, + MLX4_MR_CACHE_N, addr); + if (likely(lkey != UINT32_MAX)) + return lkey; + /* Take slower bottom-half (Binary Search) on miss. */ + return mlx4_rx_addr2mr_bh(rxq, addr); +} + +#define mlx4_rx_mb2mr(rxq, mb) mlx4_rx_addr2mr(rxq, (uintptr_t)((mb)->buf_addr)) + +/** + * Query LKey from a packet buffer for Tx. If not found, add the mempool. + * + * @param txq + * Pointer to Tx queue structure. + * @param addr + * Address to search. + * + * @return + * Searched LKey on success, UINT32_MAX on no match. + */ +static __rte_always_inline uint32_t +mlx4_tx_addr2mr(struct txq *txq, uintptr_t addr) { - (void)txq; - (void)mp; - return UINT32_MAX; + struct mlx4_mr_ctrl *mr_ctrl = &txq->mr_ctrl; + uint32_t lkey; + + /* Check generation bit to see if there's any change on existing MRs. */ + if (unlikely(*mr_ctrl->dev_gen_ptr != mr_ctrl->cur_gen)) + mlx4_mr_flush_local_cache(mr_ctrl); + /* Linear search on MR cache array. */ + lkey = mlx4_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru, + MLX4_MR_CACHE_N, addr); + if (likely(lkey != UINT32_MAX)) + return lkey; + /* Take slower bottom-half (binary search) on miss. */ + return mlx4_tx_addr2mr_bh(txq, addr); } +#define mlx4_tx_mb2mr(rxq, mb) mlx4_tx_addr2mr(rxq, (uintptr_t)((mb)->buf_addr)) + #endif /* MLX4_RXTX_H_ */ diff --git a/drivers/net/mlx4/mlx4_txq.c b/drivers/net/mlx4/mlx4_txq.c index bd65dcc515..a358732ccd 100644 --- a/drivers/net/mlx4/mlx4_txq.c +++ b/drivers/net/mlx4/mlx4_txq.c @@ -316,6 +316,13 @@ mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc, /* Save first wqe pointer in the first element. */ (&(*txq->elts)[0])->wqe = (volatile struct mlx4_wqe_ctrl_seg *)txq->msq.buf; + if (mlx4_mr_btree_init(&txq->mr_ctrl.cache_bh, + MLX4_MR_BTREE_CACHE_N, socket)) { + /* rte_errno is already set. */ + goto error; + } + /* Save pointer of global generation number to check memory event. */ + txq->mr_ctrl.dev_gen_ptr = &priv->mr.dev_gen; DEBUG("%p: adding Tx queue %p to list", (void *)dev, (void *)txq); dev->data->tx_queues[idx] = txq; return 0; @@ -356,5 +363,6 @@ mlx4_tx_queue_release(void *dpdk_txq) claim_zero(mlx4_glue->destroy_qp(txq->qp)); if (txq->cq) claim_zero(mlx4_glue->destroy_cq(txq->cq)); + mlx4_mr_btree_free(&txq->mr_ctrl.cache_bh); rte_free(txq); } -- 2.20.1