X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fo2iblnd%2Fo2iblnd.h;h=3a8d5f6ed948ed05911a4862334bcf65567822f6;hp=9183c536f3336e528b9c2586bb048964ee0ab4f5;hb=3e95945ff970426a5904c859feed46176e678305;hpb=0f34ed0c1caf5cbe8965af730466b11760de8782 diff --git a/lnet/klnds/o2iblnd/o2iblnd.h b/lnet/klnds/o2iblnd/o2iblnd.h index 9183c53..3a8d5f6 100644 --- a/lnet/klnds/o2iblnd/o2iblnd.h +++ b/lnet/klnds/o2iblnd/o2iblnd.h @@ -27,7 +27,7 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2014, Intel Corporation. + * Copyright (c) 2011, 2015, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -99,21 +99,10 @@ typedef struct int *kib_timeout; /* comms timeout (seconds) */ int *kib_keepalive; /* keepalive timeout (seconds) */ int *kib_ntx; /* # tx descs */ - int *kib_credits; /* # concurrent sends */ - int *kib_peertxcredits; /* # concurrent sends to 1 peer */ - int *kib_peerrtrcredits; /* # per-peer router buffer credits */ - int *kib_peercredits_hiw; /* # when eagerly to return credits */ - int *kib_peertimeout; /* seconds to consider peer dead */ char **kib_default_ipif; /* default IPoIB interface */ int *kib_retry_count; int *kib_rnr_retry_count; - int *kib_concurrent_sends; /* send work queue sizing */ int *kib_ib_mtu; /* IB MTU */ - int *kib_map_on_demand; /* map-on-demand if RD has more fragments - * than this value, 0 disable map-on-demand */ - int *kib_fmr_pool_size; /* # FMRs in pool */ - int *kib_fmr_flush_trigger; /* When to trigger FMR flush */ - int *kib_fmr_cache; /* enable FMR pool cache? */ #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM struct ctl_table_header *kib_sysctl; /* sysctl interface */ #endif @@ -131,45 +120,30 @@ extern kib_tunables_t kiblnd_tunables; #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */ #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */ -#define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \ - IBLND_MSG_QUEUE_SIZE_V1 : \ - *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */ -#define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \ - IBLND_CREDIT_HIGHWATER_V1 : \ - *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */ +/* when eagerly to return credits */ +#define IBLND_CREDITS_HIGHWATER(t, v) ((v) == IBLND_MSG_VERSION_1 ? \ + IBLND_CREDIT_HIGHWATER_V1 : \ + t->lnd_peercredits_hiw) -#ifdef HAVE_RDMA_CREATE_ID_4ARG -#define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt) +#ifdef HAVE_RDMA_CREATE_ID_5ARG +# define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(current->nsproxy->net_ns, \ + cb, dev, \ + ps, qpt) #else -#define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps) +# ifdef HAVE_RDMA_CREATE_ID_4ARG +# define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, \ + ps, qpt) +# else +# define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps) +# endif #endif -static inline int -kiblnd_concurrent_sends_v1(void) -{ - if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2) - return IBLND_MSG_QUEUE_SIZE_V1 * 2; - - if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2) - return IBLND_MSG_QUEUE_SIZE_V1 / 2; - - return *kiblnd_tunables.kib_concurrent_sends; -} - -#define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \ - kiblnd_concurrent_sends_v1() : \ - *kiblnd_tunables.kib_concurrent_sends) /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */ #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1) #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0) #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */ #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */ -#define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \ - *kiblnd_tunables.kib_map_on_demand : \ - IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */ -#define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \ - IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS) /************************/ /* derived constants... */ @@ -179,18 +153,19 @@ kiblnd_concurrent_sends_v1(void) #define IBLND_FMR_POOL 256 #define IBLND_FMR_POOL_FLUSH 192 -/* TX messages (shared by all connections) */ -#define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx) - /* RX messages (per connection) */ -#define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v)) -#define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE) -#define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE) +#define IBLND_RX_MSGS(c) \ + ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version)) +#define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE) +#define IBLND_RX_MSG_PAGES(c) \ + ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE) /* WRs and CQEs (per connection) */ -#define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v) -#define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v)) -#define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v)) +#define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c) +#define IBLND_SEND_WRS(c) \ + ((c->ibc_max_frags + 1) * kiblnd_concurrent_sends(c->ibc_version, \ + c->ibc_peer->ibp_ni)) +#define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + IBLND_SEND_WRS(c)) struct kib_hca_dev; @@ -331,26 +306,47 @@ typedef struct int fps_cpt; /* CPT id */ int fps_pool_size; int fps_flush_trigger; + int fps_cache; /* is allocating new pool */ int fps_increasing; /* time stamp for retry if failed to allocate */ cfs_time_t fps_next_retry; } kib_fmr_poolset_t; +struct kib_fast_reg_descriptor { /* For fast registration */ + struct list_head frd_list; + struct ib_send_wr frd_inv_wr; + struct ib_send_wr frd_fastreg_wr; + struct ib_mr *frd_mr; + struct ib_fast_reg_page_list *frd_frpl; + bool frd_valid; +}; + typedef struct { struct list_head fpo_list; /* chain on pool list */ struct kib_hca_dev *fpo_hdev; /* device for this pool */ kib_fmr_poolset_t *fpo_owner; /* owner of this pool */ - struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */ + union { + struct { + struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */ + } fmr; + struct { /* For fast registration */ + struct list_head fpo_pool_list; + int fpo_pool_size; + } fast_reg; + }; cfs_time_t fpo_deadline; /* deadline of this pool */ int fpo_failed; /* fmr pool is failed */ int fpo_map_count; /* # of mapped FMR */ + int fpo_is_fmr; } kib_fmr_pool_t; typedef struct { - struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */ - kib_fmr_pool_t *fmr_pool; /* pool of FMR */ + kib_fmr_pool_t *fmr_pool; /* pool of FMR */ + struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */ + struct kib_fast_reg_descriptor *fmr_frd; + u32 fmr_key; } kib_fmr_t; typedef struct kib_net @@ -411,6 +407,15 @@ typedef struct struct list_head kib_connd_conns; /* connections with zero refcount */ struct list_head kib_connd_zombies; + /* connections to reconnect */ + struct list_head kib_reconn_list; + /* peers wait for reconnection */ + struct list_head kib_reconn_wait; + /* + * The second that peers are pulled out from \a kib_reconn_wait + * for reconnection. + */ + unsigned int kib_reconn_sec; /* connection daemon sleeps here */ wait_queue_head_t kib_connd_waitq; spinlock_t kib_connd_lock; /* serialise */ @@ -539,8 +544,10 @@ typedef struct { #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */ #define IBLND_REJECT_CONN_STALE 5 /* stale peer */ -#define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */ -#define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */ +/* peer's rdma frags doesn't match mine */ +#define IBLND_REJECT_RDMA_FRAGS 6 +/* peer's msg queue size doesn't match mine */ +#define IBLND_REJECT_MSG_QUEUE_SIZE 7 /***********************************************************************/ @@ -639,6 +646,8 @@ typedef struct kib_conn struct list_head ibc_sched_list; /* version of connection */ __u16 ibc_version; + /* reconnect later */ + __u16 ibc_reconnect:1; /* which instance of the peer */ __u64 ibc_incarnation; /* # users */ @@ -657,14 +666,16 @@ typedef struct kib_conn int ibc_reserved_credits; /* set on comms error */ int ibc_comms_error; + /* connections queue depth */ + __u16 ibc_queue_depth; + /* connections max frags */ + __u16 ibc_max_frags; /* receive buffers owned */ - unsigned short ibc_nrx; - /** rejected by connection race */ - unsigned short ibc_conn_race:1; + unsigned int ibc_nrx:16; /* scheduled for attention */ - unsigned short ibc_scheduled:1; + unsigned int ibc_scheduled:1; /* CQ callback fired */ - unsigned short ibc_ready:1; + unsigned int ibc_ready:1; /* time of last send */ unsigned long ibc_last_send; /** link chain for kiblnd_check_conns only */ @@ -712,30 +723,93 @@ typedef struct kib_peer lnet_nid_t ibp_nid; /* LNet interface */ lnet_ni_t *ibp_ni; - /* # users */ - atomic_t ibp_refcount; /* all active connections */ struct list_head ibp_conns; /* msgs waiting for a conn */ struct list_head ibp_tx_queue; - /* version of peer */ - __u16 ibp_version; /* incarnation of peer */ __u64 ibp_incarnation; - /* current active connection attempts */ - int ibp_connecting; + /* when (in jiffies) I was last alive */ + cfs_time_t ibp_last_alive; + /* # users */ + atomic_t ibp_refcount; + /* version of peer */ + __u16 ibp_version; /* current passive connection attempts */ - int ibp_accepting; + unsigned short ibp_accepting; + /* current active connection attempts */ + unsigned short ibp_connecting; + /* reconnect this peer later */ + unsigned short ibp_reconnecting:1; + /* # consecutive reconnection attempts to this peer */ + unsigned int ibp_reconnected; /* errno on closing this peer */ int ibp_error; - /* when (in jiffies) I was last alive */ - cfs_time_t ibp_last_alive; + /* max map_on_demand */ + __u16 ibp_max_frags; + /* max_peer_credits */ + __u16 ibp_queue_depth; } kib_peer_t; +#ifndef HAVE_IB_INC_RKEY +/** + * ib_inc_rkey - increments the key portion of the given rkey. Can be used + * for calculating a new rkey for type 2 memory windows. + * @rkey - the rkey to increment. + */ +static inline u32 ib_inc_rkey(u32 rkey) +{ + const u32 mask = 0x000000ff; + return ((rkey + 1) & mask) | (rkey & ~mask); +} +#endif + extern kib_data_t kiblnd_data; extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev); +int kiblnd_msg_queue_size(int version, struct lnet_ni *ni); + +/* max # of fragments configured by user */ +static inline int +kiblnd_cfg_rdma_frags(struct lnet_ni *ni) +{ + struct lnet_ioctl_config_o2iblnd_tunables *tunables; + int mod; + + tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib; + mod = tunables->lnd_map_on_demand; + return mod != 0 ? mod : IBLND_MAX_RDMA_FRAGS; +} + +static inline int +kiblnd_rdma_frags(int version, struct lnet_ni *ni) +{ + return version == IBLND_MSG_VERSION_1 ? + IBLND_MAX_RDMA_FRAGS : + kiblnd_cfg_rdma_frags(ni); +} + +static inline int +kiblnd_concurrent_sends(int version, struct lnet_ni *ni) +{ + struct lnet_ioctl_config_o2iblnd_tunables *tunables; + int concurrent_sends; + + tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib; + concurrent_sends = tunables->lnd_concurrent_sends; + + if (version == IBLND_MSG_VERSION_1) { + if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2) + return IBLND_MSG_QUEUE_SIZE_V1 * 2; + + if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2) + return IBLND_MSG_QUEUE_SIZE_V1 / 2; + } + + return concurrent_sends; +} + static inline void kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev) { @@ -807,6 +881,20 @@ do { \ kiblnd_destroy_peer(peer); \ } while (0) +static inline bool +kiblnd_peer_connecting(kib_peer_t *peer) +{ + return peer->ibp_connecting != 0 || + peer->ibp_reconnecting != 0 || + peer->ibp_accepting != 0; +} + +static inline bool +kiblnd_peer_idle(kib_peer_t *peer) +{ + return !kiblnd_peer_connecting(peer) && list_empty(&peer->ibp_conns); +} + static inline struct list_head * kiblnd_nid2peerlist (lnet_nid_t nid) { @@ -844,10 +932,14 @@ kiblnd_send_keepalive(kib_conn_t *conn) static inline int kiblnd_need_noop(kib_conn_t *conn) { - LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED); + lnet_ni_t *ni = conn->ibc_peer->ibp_ni; + struct lnet_ioctl_config_o2iblnd_tunables *tunables; + + LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED); + tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib; if (conn->ibc_outstanding_credits < - IBLND_CREDITS_HIGHWATER(conn->ibc_version) && + IBLND_CREDITS_HIGHWATER(tunables, conn->ibc_version) && !kiblnd_send_keepalive(conn)) return 0; /* No need to send NOOP */ @@ -902,10 +994,12 @@ kiblnd_queue2str(kib_conn_t *conn, struct list_head *q) /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the * lowest bits of the work request id to stash the work item type. */ -#define IBLND_WID_TX 0 -#define IBLND_WID_RDMA 1 -#define IBLND_WID_RX 2 -#define IBLND_WID_MASK 3UL +#define IBLND_WID_INVAL 0 +#define IBLND_WID_TX 1 +#define IBLND_WID_RX 2 +#define IBLND_WID_RDMA 3 +#define IBLND_WID_MR 4 +#define IBLND_WID_MASK 7UL static inline __u64 kiblnd_ptr2wreqid (void *ptr, int type) @@ -1053,17 +1147,18 @@ static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev, #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data) #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len) -struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, - kib_rdma_desc_t *rd); +struct ib_mr *kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd, + int negotiated_nfrags); void kiblnd_map_rx_descs(kib_conn_t *conn); void kiblnd_unmap_rx_descs(kib_conn_t *conn); void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node); struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps); -int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, - int npages, __u64 iov, kib_fmr_t *fmr); +int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages, + __u32 nob, __u64 iov, bool is_rx, kib_fmr_t *fmr); void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status); +int kiblnd_tunables_setup(struct lnet_ni *ni); int kiblnd_tunables_init(void); void kiblnd_tunables_fini(void); @@ -1079,9 +1174,9 @@ int kiblnd_cm_callback(struct rdma_cm_id *cmid, int kiblnd_translate_mtu(int value); int kiblnd_dev_failover(kib_dev_t *dev); -int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid); +int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid); void kiblnd_destroy_peer (kib_peer_t *peer); -void kiblnd_connect_peer(kib_peer_t *peer); +bool kiblnd_reconnect_peer(kib_peer_t *peer); void kiblnd_destroy_dev (kib_dev_t *dev); void kiblnd_unlink_peer_locked (kib_peer_t *peer); kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid); @@ -1089,9 +1184,9 @@ int kiblnd_close_stale_conns_locked (kib_peer_t *peer, int version, __u64 incarnation); int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why); -kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid, - int state, int version); -void kiblnd_destroy_conn (kib_conn_t *conn); +kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, + int state, int version); +void kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn); void kiblnd_close_conn (kib_conn_t *conn, int error); void kiblnd_close_conn_locked (kib_conn_t *conn, int error); @@ -1109,6 +1204,6 @@ int kiblnd_post_rx (kib_rx_t *rx, int credit); int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg); int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed, - unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov, + unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov, unsigned int offset, unsigned int mlen, unsigned int rlen);