From 7b442f1a43714455fad06c527b6fbc10f82af857 Mon Sep 17 00:00:00 2001 From: Isaac Huang Date: Wed, 17 Nov 2010 07:14:46 -0700 Subject: [PATCH] b=20153 add IB bonding failover support to o2iblnd O2iblnd changes to support failover events from an IB bonding IPoIB interface. Mostly to recreate device specific resources, e.g. listener CMID. i=isaac i=liang --- lnet/autoconf/lustre-lnet.m4 | 18 + lnet/klnds/o2iblnd/o2iblnd.c | 736 ++++++++++++++++++++++++--------- lnet/klnds/o2iblnd/o2iblnd.h | 90 +++- lnet/klnds/o2iblnd/o2iblnd_cb.c | 147 ++++++- lnet/klnds/o2iblnd/o2iblnd_modparams.c | 22 +- 5 files changed, 775 insertions(+), 238 deletions(-) diff --git a/lnet/autoconf/lustre-lnet.m4 b/lnet/autoconf/lustre-lnet.m4 index d941cd6..ca7f785 100644 --- a/lnet/autoconf/lustre-lnet.m4 +++ b/lnet/autoconf/lustre-lnet.m4 @@ -42,6 +42,23 @@ fi ]) # +# LN_FUNC_DEV_GET_BY_NAME_2ARG +# +AC_DEFUN([LN_FUNC_DEV_GET_BY_NAME_2ARG], +[AC_MSG_CHECKING([if dev_get_by_name has two args]) +LB_LINUX_TRY_COMPILE([ + #include +],[ + dev_get_by_name(NULL, NULL); +],[ + AC_MSG_RESULT([yes]) + AC_DEFINE(HAVE_DEV_GET_BY_NAME_2ARG, 1, [dev_get_by_name has 2 args]) +],[ + AC_MSG_RESULT([no]) +]) +]) + +# # LN_CONFIG_AFFINITY # # check if cpu affinity is available/wanted @@ -539,6 +556,7 @@ LB_LINUX_TRY_COMPILE([ # AC_DEFUN([LN_PROG_LINUX], [ +LN_FUNC_DEV_GET_BY_NAME_2ARG LN_CONFIG_AFFINITY LN_CONFIG_BACKOFF LN_CONFIG_QUADRICS diff --git a/lnet/klnds/o2iblnd/o2iblnd.c b/lnet/klnds/o2iblnd/o2iblnd.c index 7ebee52..216dbff 100644 --- a/lnet/klnds/o2iblnd/o2iblnd.c +++ b/lnet/klnds/o2iblnd/o2iblnd.c @@ -668,23 +668,18 @@ kiblnd_translate_mtu(int value) } static void -kiblnd_setup_mtu(struct rdma_cm_id *cmid) +kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid) { - unsigned long flags; int mtu; /* XXX There is no path record for iWARP, set by netdev->change_mtu? */ if (cmid->route.path_rec == NULL) return; - cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); - mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu); LASSERT (mtu >= 0); if (mtu != 0) cmid->route.path_rec->mtu = mtu; - - cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); } kib_conn_t * @@ -698,7 +693,9 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, * she must dispose of 'cmid'. (Actually I'd block forever if I tried * to destroy 'cmid' here since I'm called from the CM which still has * its ref on 'cmid'). */ + cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock; kib_net_t *net = peer->ibp_ni->ni_data; + kib_dev_t *dev = net->ibn_dev; struct ib_qp_init_attr *init_qp_attr; kib_conn_t *conn; struct ib_cq *cq; @@ -745,6 +742,34 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, } memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars)); + cfs_write_lock_irqsave(glock, flags); + if (dev->ibd_failover) { + cfs_write_unlock_irqrestore(glock, flags); + CERROR("%s: failover in progress\n", dev->ibd_ifname); + goto failed_2; + } + + if (dev->ibd_hdev->ibh_ibdev != cmid->device) { + /* wakeup failover thread and teardown connection */ + if (kiblnd_dev_can_failover(dev)) { + cfs_list_add_tail(&dev->ibd_fail_list, + &kiblnd_data.kib_failed_devs); + cfs_waitq_signal(&kiblnd_data.kib_failover_waitq); + } + + cfs_write_unlock_irqrestore(glock, flags); + CERROR("cmid HCA(%s), kib_dev(%s) need failover\n", + cmid->device->name, dev->ibd_ifname); + goto failed_2; + } + + kiblnd_hdev_addref_locked(dev->ibd_hdev); + conn->ibc_hdev = dev->ibd_hdev; + + kiblnd_setup_mtu_locked(cmid); + + cfs_write_unlock_irqrestore(glock, flags); + LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t)); if (conn->ibc_rxs == NULL) { CERROR("Cannot allocate RX buffers\n"); @@ -782,8 +807,6 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, goto failed_2; } - kiblnd_setup_mtu(cmid); - memset(init_qp_attr, 0, sizeof(*init_qp_attr)); init_qp_attr->event_handler = kiblnd_qp_event; init_qp_attr->qp_context = conn; @@ -796,7 +819,7 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid, init_qp_attr->send_cq = cq; init_qp_attr->recv_cq = cq; - rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr); + rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr); if (rc != 0) { CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n", rc, init_qp_attr->cap.max_send_wr, @@ -911,6 +934,9 @@ kiblnd_destroy_conn (kib_conn_t *conn) if (conn->ibc_connvars != NULL) LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars)); + if (conn->ibc_hdev != NULL) + kiblnd_hdev_decref(conn->ibc_hdev); + /* See CAVEAT EMPTOR above in kiblnd_create_conn */ if (conn->ibc_state != IBLND_CONN_INIT) { kib_net_t *net = peer->ibp_ni->ni_data; @@ -1118,8 +1144,6 @@ kiblnd_free_pages (kib_pages_t *p) int npages = p->ibp_npages; int i; - LASSERT (p->ibp_device == NULL); - for (i = 0; i < npages; i++) if (p->ibp_pages[i] != NULL) __free_page(p->ibp_pages[i]); @@ -1162,21 +1186,19 @@ kiblnd_unmap_rx_descs(kib_conn_t *conn) int i; LASSERT (conn->ibc_rxs != NULL); - LASSERT (conn->ibc_rx_pages->ibp_device != NULL); + LASSERT (conn->ibc_hdev != NULL); for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) { rx = &conn->ibc_rxs[i]; LASSERT (rx->rx_nob >= 0); /* not posted */ - kiblnd_dma_unmap_single(conn->ibc_rx_pages->ibp_device, + kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev, KIBLND_UNMAP_ADDR(rx, rx_msgunmap, rx->rx_msgaddr), IBLND_MSG_SIZE, DMA_FROM_DEVICE); } - conn->ibc_rx_pages->ibp_device = NULL; - kiblnd_free_pages(conn->ibc_rx_pages); conn->ibc_rx_pages = NULL; @@ -1199,10 +1221,10 @@ kiblnd_map_rx_descs(kib_conn_t *conn) rx->rx_conn = conn; rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off); - rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_cmid->device, + rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev, rx->rx_msg, IBLND_MSG_SIZE, DMA_FROM_DEVICE); - LASSERT (!kiblnd_dma_mapping_error(conn->ibc_cmid->device, + LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev, rx->rx_msgaddr)); KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr); @@ -1219,27 +1241,55 @@ kiblnd_map_rx_descs(kib_conn_t *conn) LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version)); } } - - conn->ibc_rx_pages->ibp_device = conn->ibc_cmid->device; } static void kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo) { + kib_hca_dev_t *hdev = tpo->tpo_hdev; kib_tx_t *tx; int i; LASSERT (tpo->tpo_pool.po_allocated == 0); - LASSERT (tpo->tpo_tx_pages->ibp_device != NULL); + + if (hdev == NULL) + return; for (i = 0; i < tpo->tpo_pool.po_size; i++) { tx = &tpo->tpo_tx_descs[i]; - kiblnd_dma_unmap_single(tpo->tpo_tx_pages->ibp_device, + kiblnd_dma_unmap_single(hdev->ibh_ibdev, KIBLND_UNMAP_ADDR(tx, tx_msgunmap, tx->tx_msgaddr), IBLND_MSG_SIZE, DMA_TO_DEVICE); } - tpo->tpo_tx_pages->ibp_device = NULL; + + kiblnd_hdev_decref(hdev); + tpo->tpo_hdev = NULL; +} + +static kib_hca_dev_t * +kiblnd_current_hdev(kib_dev_t *dev) +{ + kib_hca_dev_t *hdev; + unsigned long flags; + int i = 0; + + cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + while (dev->ibd_failover) { + cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + if (i++ % 50 == 0) + CDEBUG(D_NET, "Wait for dev(%s) failover\n", dev->ibd_ifname); + cfs_schedule_timeout(cfs_time_seconds(1) / 100); + + cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + } + + kiblnd_hdev_addref_locked(dev->ibd_hdev); + hdev = dev->ibd_hdev; + + cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + + return hdev; } static void @@ -1248,6 +1298,7 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo) kib_pages_t *txpgs = tpo->tpo_tx_pages; kib_pool_t *pool = &tpo->tpo_pool; kib_net_t *net = pool->po_owner->ps_net; + kib_dev_t *dev = net->ibn_dev; struct page *page; kib_tx_t *tx; int page_offset; @@ -1255,8 +1306,6 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo) int i; LASSERT (net != NULL); - LASSERT (net->ibn_dev->ibd_cmid != NULL && - net->ibn_dev->ibd_cmid->device != NULL); /* pre-mapped messages are not bigger than 1 page */ CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE); @@ -1264,7 +1313,7 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo) /* No fancy arithmetic when we do the buffer calculations */ CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0); - txpgs->ibp_device = net->ibn_dev->ibd_cmid->device; + tpo->tpo_hdev = kiblnd_current_hdev(dev); for (ipage = page_offset = i = 0; i < pool->po_size; i++) { page = txpgs->ibp_pages[ipage]; @@ -1274,9 +1323,9 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo) page_offset); tx->tx_msgaddr = kiblnd_dma_map_single( - txpgs->ibp_device, tx->tx_msg, + tpo->tpo_hdev->ibh_ibdev, tx->tx_msg, IBLND_MSG_SIZE, DMA_TO_DEVICE); - LASSERT (!kiblnd_dma_mapping_error(txpgs->ibp_device, + LASSERT (!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev, tx->tx_msgaddr)); KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr); @@ -1294,43 +1343,43 @@ kiblnd_map_tx_pool(kib_tx_pool_t *tpo) } struct ib_mr * -kiblnd_find_dma_mr(kib_net_t *net, __u64 addr, __u64 size) +kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size) { __u64 index; - LASSERT (net->ibn_dev->ibd_mrs[0] != NULL); + LASSERT (hdev->ibh_mrs[0] != NULL); - if (net->ibn_dev->ibd_nmrs == 1) - return net->ibn_dev->ibd_mrs[0]; + if (hdev->ibh_nmrs == 1) + return hdev->ibh_mrs[0]; - index = addr >> net->ibn_dev->ibd_mr_shift; + index = addr >> hdev->ibh_mr_shift; - if (index < net->ibn_dev->ibd_nmrs && - index == ((addr + size - 1) >> net->ibn_dev->ibd_mr_shift)) - return net->ibn_dev->ibd_mrs[index]; + if (index < hdev->ibh_nmrs && + index == ((addr + size - 1) >> hdev->ibh_mr_shift)) + return hdev->ibh_mrs[index]; return NULL; } struct ib_mr * -kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd) +kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd) { struct ib_mr *prev_mr; struct ib_mr *mr; int i; - LASSERT (net->ibn_dev->ibd_mrs[0] != NULL); + LASSERT (hdev->ibh_mrs[0] != NULL); if (*kiblnd_tunables.kib_map_on_demand > 0 && *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags) return NULL; - if (net->ibn_dev->ibd_nmrs == 1) - return net->ibn_dev->ibd_mrs[0]; + if (hdev->ibh_nmrs == 1) + return hdev->ibh_mrs[0]; for (i = 0, mr = prev_mr = NULL; i < rd->rd_nfrags; i++) { - mr = kiblnd_find_dma_mr(net, + mr = kiblnd_find_dma_mr(hdev, rd->rd_frags[i].rf_addr, rd->rd_frags[i].rf_nob); if (prev_mr == NULL) @@ -1354,6 +1403,9 @@ kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool) if (pool->fpo_fmr_pool != NULL) ib_destroy_fmr_pool(pool->fpo_fmr_pool); + if (pool->fpo_hdev != NULL) + kiblnd_hdev_decref(pool->fpo_hdev); + LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t)); } @@ -1373,6 +1425,7 @@ int kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo) { /* FMR pool for RDMA */ + kib_dev_t *dev = fps->fps_net->ibn_dev; kib_fmr_pool_t *fpo; struct ib_fmr_pool_param param = { .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE, @@ -1386,19 +1439,18 @@ kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo) .cache = !!*kiblnd_tunables.kib_fmr_cache}; int rc; - LASSERT (fps->fps_net->ibn_dev != NULL && - fps->fps_net->ibn_dev->ibd_pd != NULL); - LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t)); if (fpo == NULL) return -ENOMEM; memset(fpo, 0, sizeof(kib_fmr_pool_t)); - fpo->fpo_fmr_pool = ib_create_fmr_pool(fps->fps_net->ibn_dev->ibd_pd, ¶m); + fpo->fpo_hdev = kiblnd_current_hdev(dev); + fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, ¶m); if (IS_ERR(fpo->fpo_fmr_pool)) { CERROR("Failed to create FMR pool: %ld\n", PTR_ERR(fpo->fpo_fmr_pool)); rc = PTR_ERR(fpo->fpo_fmr_pool); + kiblnd_hdev_decref(fpo->fpo_hdev); LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t)); return rc; } @@ -1411,8 +1463,28 @@ kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo) } static void +kiblnd_fail_fmr_pool_set(kib_fmr_poolset_t *fps, cfs_list_t *zombies) +{ + cfs_spin_lock(&fps->fps_lock); + + while (!cfs_list_empty(&fps->fps_pool_list)) { + kib_fmr_pool_t *fpo = cfs_list_entry(fps->fps_pool_list.next, + kib_fmr_pool_t, fpo_list); + fpo->fpo_failed = 1; + cfs_list_del(&fpo->fpo_list); + if (fpo->fpo_map_count == 0) + cfs_list_add(&fpo->fpo_list, zombies); + else + cfs_list_add(&fpo->fpo_list, &fps->fps_failed_pool_list); + } + + cfs_spin_unlock(&fps->fps_lock); +} + +static void kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps) { + kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list); kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list); } @@ -1427,6 +1499,7 @@ kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net) fps->fps_net = net; cfs_spin_lock_init(&fps->fps_lock); CFS_INIT_LIST_HEAD(&fps->fps_pool_list); + CFS_INIT_LIST_HEAD(&fps->fps_failed_pool_list); rc = kiblnd_create_fmr_pool(fps, &fpo); if (rc == 0) cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list); @@ -1434,12 +1507,23 @@ kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net) return rc; } +static int +kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now) +{ + if (fpo->fpo_map_count != 0) /* still in use */ + return 0; + if (fpo->fpo_failed) + return 1; + return cfs_time_aftereq(now, fpo->fpo_deadline); +} + void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status) { CFS_LIST_HEAD (zombies); kib_fmr_pool_t *fpo = fmr->fmr_pool; kib_fmr_poolset_t *fps = fpo->fpo_owner; + cfs_time_t now = cfs_time_current(); kib_fmr_pool_t *tmp; int rc; @@ -1462,8 +1546,7 @@ kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status) if (fps->fps_pool_list.next == &fpo->fpo_list) continue; - if (fpo->fpo_map_count == 0 && /* no more reference */ - cfs_time_aftereq(cfs_time_current(), fpo->fpo_deadline)) { + if (kiblnd_fmr_pool_is_idle(fpo, now)) { cfs_list_move(&fpo->fpo_list, &zombies); fps->fps_version ++; } @@ -1569,21 +1652,41 @@ kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size) } void -kiblnd_destroy_pool_list(kib_poolset_t *ps, cfs_list_t *head) +kiblnd_destroy_pool_list(cfs_list_t *head) { kib_pool_t *pool; while (!cfs_list_empty(head)) { pool = cfs_list_entry(head->next, kib_pool_t, po_list); cfs_list_del(&pool->po_list); - ps->ps_pool_destroy(pool); + + LASSERT (pool->po_owner != NULL); + pool->po_owner->ps_pool_destroy(pool); + } +} + +static void +kiblnd_fail_pool_set(kib_poolset_t *ps, cfs_list_t *zombies) +{ + cfs_spin_lock(&ps->ps_lock); + while (!cfs_list_empty(&ps->ps_pool_list)) { + kib_pool_t *po = cfs_list_entry(ps->ps_pool_list.next, + kib_pool_t, po_list); + po->po_failed = 1; + cfs_list_del(&po->po_list); + if (po->po_allocated == 0) + cfs_list_add(&po->po_list, zombies); + else + cfs_list_add(&po->po_list, &ps->ps_failed_pool_list); } + cfs_spin_unlock(&ps->ps_lock); } static void kiblnd_fini_pool_set(kib_poolset_t *ps) { - kiblnd_destroy_pool_list(ps, &ps->ps_pool_list); + kiblnd_destroy_pool_list(&ps->ps_failed_pool_list); + kiblnd_destroy_pool_list(&ps->ps_pool_list); } static int @@ -1608,6 +1711,7 @@ kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net, strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN); cfs_spin_lock_init(&ps->ps_lock); CFS_INIT_LIST_HEAD(&ps->ps_pool_list); + CFS_INIT_LIST_HEAD(&ps->ps_failed_pool_list); rc = ps->ps_pool_create(ps, size, &pool); if (rc == 0) @@ -1618,6 +1722,16 @@ kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net, return rc; } +static int +kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now) +{ + if (pool->po_allocated != 0) /* still in use */ + return 0; + if (pool->po_failed) + return 1; + return cfs_time_aftereq(now, pool->po_deadline); +} + void kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node) { @@ -1640,14 +1754,13 @@ kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node) if (ps->ps_pool_list.next == &pool->po_list) continue; - if (pool->po_allocated == 0 && - cfs_time_aftereq(now, pool->po_deadline)) + if (kiblnd_pool_is_idle(pool, now)) cfs_list_move(&pool->po_list, &zombies); } cfs_spin_unlock(&ps->ps_lock); if (!cfs_list_empty(&zombies)) - kiblnd_destroy_pool_list(ps, &zombies); + kiblnd_destroy_pool_list(&zombies); } cfs_list_t * @@ -1728,8 +1841,8 @@ kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr) } int -kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd, - __u64 *iova, kib_phys_mr_t **pp_pmr) +kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev, + kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr) { kib_phys_mr_t *pmr; cfs_list_t *node; @@ -1743,12 +1856,17 @@ kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd, } pmr = container_of(node, kib_phys_mr_t, pmr_list); + if (pmr->pmr_pool->ppo_hdev != hdev) { + kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node); + return -EAGAIN; + } + for (i = 0; i < rd->rd_nfrags; i ++) { pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr; pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob; } - pmr->pmr_mr = ib_reg_phys_mr(pps->pps_poolset.ps_net->ibn_dev->ibd_pd, + pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd, pmr->pmr_ipb, rd->rd_nfrags, IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE, @@ -1793,6 +1911,9 @@ kiblnd_destroy_pmr_pool(kib_pool_t *pool) } kiblnd_fini_pool(pool); + if (ppo->ppo_hdev != NULL) + kiblnd_hdev_decref(ppo->ppo_hdev); + LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t)); } @@ -1834,6 +1955,7 @@ kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po) return -ENOMEM; } + ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev); *pp_po = pool; return 0; } @@ -1847,8 +1969,7 @@ kiblnd_destroy_tx_pool(kib_pool_t *pool) LASSERT (pool->po_allocated == 0); if (tpo->tpo_tx_pages != NULL) { - if (tpo->tpo_tx_pages->ibp_device != NULL) - kiblnd_unmap_tx_pool(tpo); + kiblnd_unmap_tx_pool(tpo); kiblnd_free_pages(tpo->tpo_tx_pages); } @@ -1995,6 +2116,7 @@ kiblnd_net_init_pools(kib_net_t *net) kib_fmr_poolset_t *fps = &net->ibn_fmr_ps; kib_pmr_poolset_t *pps = &net->ibn_pmr_ps; kib_tx_poolset_t *tps = &net->ibn_tx_ps; + unsigned long flags; int rc; if (*kiblnd_tunables.kib_fmr_pool_size < @@ -2013,10 +2135,12 @@ kiblnd_net_init_pools(kib_net_t *net) return -EINVAL; } + cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags); if (*kiblnd_tunables.kib_map_on_demand > 0 || - net->ibn_dev->ibd_nmrs > 1) { /* premapping can fail if ibd_nmr > 1, - * so we always create FMR/PMR pool and - * map-on-demand if premapping failed */ + net->ibn_dev->ibd_hdev->ibh_nmrs > 1) { + /* premapping can fail if ibd_nmr > 1, so we always create + * FMR/PMR pool and map-on-demand if premapping failed */ + cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); rc = kiblnd_init_fmr_pool_set(fps, net); if (rc == 0) { net->ibn_with_fmr = 1; @@ -2031,6 +2155,8 @@ kiblnd_net_init_pools(kib_net_t *net) } if (rc != 0) return rc; + } else { + cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); } rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(), @@ -2047,36 +2173,17 @@ kiblnd_net_init_pools(kib_net_t *net) return rc; } -void -kiblnd_dev_cleanup(kib_dev_t *ibdev) -{ - int i; - - if (ibdev->ibd_mrs == NULL) - return; - - for (i = 0; i < ibdev->ibd_nmrs; i++) { - if (ibdev->ibd_mrs[i] == NULL) - break; - - ib_dereg_mr(ibdev->ibd_mrs[i]); - } - - LIBCFS_FREE(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs); - ibdev->ibd_mrs = NULL; -} - static int -kiblnd_dev_get_attr(kib_dev_t *ibdev) +kiblnd_hdev_get_attr(kib_hca_dev_t *hdev) { struct ib_device_attr *attr; int rc; /* It's safe to assume a HCA can handle a page size * matching that of the native system */ - ibdev->ibd_page_shift = PAGE_SHIFT; - ibdev->ibd_page_size = 1 << PAGE_SHIFT; - ibdev->ibd_page_mask = ~((__u64)ibdev->ibd_page_size - 1); + hdev->ibh_page_shift = PAGE_SHIFT; + hdev->ibh_page_size = 1 << PAGE_SHIFT; + hdev->ibh_page_mask = ~((__u64)hdev->ibh_page_size - 1); LIBCFS_ALLOC(attr, sizeof(*attr)); if (attr == NULL) { @@ -2084,9 +2191,9 @@ kiblnd_dev_get_attr(kib_dev_t *ibdev) return -ENOMEM; } - rc = ib_query_device(ibdev->ibd_cmid->device, attr); + rc = ib_query_device(hdev->ibh_ibdev, attr); if (rc == 0) - ibdev->ibd_mr_size = attr->max_mr_size; + hdev->ibh_mr_size = attr->max_mr_size; LIBCFS_FREE(attr, sizeof(*attr)); @@ -2099,29 +2206,63 @@ kiblnd_dev_get_attr(kib_dev_t *ibdev) /* XXX We can't trust this value returned by Chelsio driver, it's wrong * and we have reported the bug, remove these in the future when Chelsio * bug got fixed. */ - if (rdma_node_get_transport(ibdev->ibd_cmid->device->node_type) == + if (rdma_node_get_transport(hdev->ibh_ibdev->node_type) == RDMA_TRANSPORT_IWARP) - ibdev->ibd_mr_size = (1ULL << 32) - 1; + hdev->ibh_mr_size = (1ULL << 32) - 1; #endif - if (ibdev->ibd_mr_size == ~0ULL) { - ibdev->ibd_mr_shift = 64; + if (hdev->ibh_mr_size == ~0ULL) { + hdev->ibh_mr_shift = 64; return 0; } - for (ibdev->ibd_mr_shift = 0; - ibdev->ibd_mr_shift < 64; ibdev->ibd_mr_shift ++) { - if (ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) || - ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) - 1) + for (hdev->ibh_mr_shift = 0; + hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) { + if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) || + hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1) return 0; } - CERROR("Invalid mr size: "LPX64"\n", ibdev->ibd_mr_size); + CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size); return -EINVAL; } +void +kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev) +{ + int i; + + if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL) + return; + + for (i = 0; i < hdev->ibh_nmrs; i++) { + if (hdev->ibh_mrs[i] == NULL) + break; + + ib_dereg_mr(hdev->ibh_mrs[i]); + } + + LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs); + hdev->ibh_mrs = NULL; + hdev->ibh_nmrs = 0; +} + +void +kiblnd_hdev_destroy(kib_hca_dev_t *hdev) +{ + kiblnd_hdev_cleanup_mrs(hdev); + + if (hdev->ibh_pd != NULL) + ib_dealloc_pd(hdev->ibh_pd); + + if (hdev->ibh_cmid != NULL) + rdma_destroy_id(hdev->ibh_cmid); + + LIBCFS_FREE(hdev, sizeof(*hdev)); +} + int -kiblnd_dev_setup(kib_dev_t *ibdev) +kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev) { struct ib_mr *mr; int i; @@ -2131,38 +2272,38 @@ kiblnd_dev_setup(kib_dev_t *ibdev) int acflags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE; - rc = kiblnd_dev_get_attr(ibdev); + rc = kiblnd_hdev_get_attr(hdev); if (rc != 0) return rc; - if (ibdev->ibd_mr_shift == 64) { - LIBCFS_ALLOC(ibdev->ibd_mrs, 1 * sizeof(*ibdev->ibd_mrs)); - if (ibdev->ibd_mrs == NULL) { + if (hdev->ibh_mr_shift == 64) { + LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs)); + if (hdev->ibh_mrs == NULL) { CERROR("Failed to allocate MRs table\n"); return -ENOMEM; } - ibdev->ibd_mrs[0] = NULL; - ibdev->ibd_nmrs = 1; + hdev->ibh_mrs[0] = NULL; + hdev->ibh_nmrs = 1; - mr = ib_get_dma_mr(ibdev->ibd_pd, acflags); + mr = ib_get_dma_mr(hdev->ibh_pd, acflags); if (IS_ERR(mr)) { CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr)); - kiblnd_dev_cleanup(ibdev); + kiblnd_hdev_cleanup_mrs(hdev); return PTR_ERR(mr); } - ibdev->ibd_mrs[0] = mr; + hdev->ibh_mrs[0] = mr; goto out; } - mr_size = (1ULL << ibdev->ibd_mr_shift); + mr_size = (1ULL << hdev->ibh_mr_shift); mm_size = (unsigned long)high_memory - PAGE_OFFSET; - ibdev->ibd_nmrs = (int)((mm_size + mr_size - 1) >> ibdev->ibd_mr_shift); + hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift); - if (ibdev->ibd_mr_shift < 32 || ibdev->ibd_nmrs > 1024) { + if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) { /* it's 4T..., assume we will re-code at that time */ CERROR("Can't support memory size: x"LPX64 " with MR size: x"LPX64"\n", mm_size, mr_size); @@ -2170,63 +2311,305 @@ kiblnd_dev_setup(kib_dev_t *ibdev) } /* create an array of MRs to cover all memory */ - LIBCFS_ALLOC(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs); - if (ibdev->ibd_mrs == NULL) { + LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs); + if (hdev->ibh_mrs == NULL) { CERROR("Failed to allocate MRs' table\n"); return -ENOMEM; } - memset(ibdev->ibd_mrs, 0, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs); + memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs); - for (i = 0; i < ibdev->ibd_nmrs; i++) { + for (i = 0; i < hdev->ibh_nmrs; i++) { struct ib_phys_buf ipb; __u64 iova; - ipb.size = ibdev->ibd_mr_size; + ipb.size = hdev->ibh_mr_size; ipb.addr = i * mr_size; iova = ipb.addr; - mr = ib_reg_phys_mr(ibdev->ibd_pd, &ipb, 1, acflags, &iova); + mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova); if (IS_ERR(mr)) { CERROR("Failed ib_reg_phys_mr addr "LPX64 " size "LPX64" : %ld\n", ipb.addr, ipb.size, PTR_ERR(mr)); - kiblnd_dev_cleanup(ibdev); + kiblnd_hdev_cleanup_mrs(hdev); return PTR_ERR(mr); } LASSERT (iova == ipb.addr); - ibdev->ibd_mrs[i] = mr; + hdev->ibh_mrs[i] = mr; } out: CDEBUG(D_CONSOLE, "Register global MR array, MR size: " LPX64", array size: %d\n", - ibdev->ibd_mr_size, ibdev->ibd_nmrs); + hdev->ibh_mr_size, hdev->ibh_nmrs); + return 0; +} - cfs_list_add_tail(&ibdev->ibd_list, - &kiblnd_data.kib_devs); +static int +kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event) +{ /* DUMMY */ return 0; } +static int +kiblnd_dev_need_failover(kib_dev_t *dev) +{ + struct rdma_cm_id *cmid; + struct sockaddr_in srcaddr; + struct sockaddr_in dstaddr; + int rc; + + if (dev->ibd_hdev == NULL || /* initializing */ + dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */ + *kiblnd_tunables.kib_dev_failover > 1) /* debugging */ + return 1; + + /* XXX: it's UGLY, but I don't have better way to find + * ib-bonding HCA failover because: + * + * a. no reliable CM event for HCA failover... + * b. no OFED API to get ib_device for current net_device... + * + * We have only two choices at this point: + * + * a. rdma_bind_addr(), it will conflict with listener cmid + * b. rdma_resolve_addr() to zero addr */ + cmid = rdma_create_id(kiblnd_dummy_callback, + dev, RDMA_PS_TCP); + if (IS_ERR(cmid)) { + rc = PTR_ERR(cmid); + CERROR("Failed to create cmid for failover: %d\n", rc); + return rc; + } + + memset(&srcaddr, 0, sizeof(srcaddr)); + srcaddr.sin_family = AF_INET; + srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip); + + memset(&dstaddr, 0, sizeof(dstaddr)); + dstaddr.sin_family = AF_INET; + rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr, + (struct sockaddr *)&dstaddr, 1); + if (rc != 0) { + CERROR("Failed to bind %s to device: %d\n", + dev->ibd_ifname, rc); + rdma_destroy_id(cmid); + return rc; + } + + LASSERT (cmid->device != NULL); + if (dev->ibd_hdev->ibh_ibdev == cmid->device) { + /* don't need device failover */ + rdma_destroy_id(cmid); + return 0; + } + + return 1; +} + +int +kiblnd_dev_failover(kib_dev_t *dev) +{ + CFS_LIST_HEAD (zombie_tpo); + CFS_LIST_HEAD (zombie_ppo); + CFS_LIST_HEAD (zombie_fpo); + struct rdma_cm_id *cmid = NULL; + kib_hca_dev_t *hdev = NULL; + kib_hca_dev_t *old; + struct ib_pd *pd; + kib_net_t *net; + struct sockaddr_in addr; + unsigned long flags; + int rc = 0; + + LASSERT (*kiblnd_tunables.kib_dev_failover > 1 || + dev->ibd_can_failover || + dev->ibd_hdev == NULL); + + rc = kiblnd_dev_need_failover(dev); + if (rc <= 0) + goto out; + + if (dev->ibd_hdev != NULL && + dev->ibd_hdev->ibh_cmid != NULL) { + /* XXX it's not good to close old listener at here, + * because we can fail to create new listener. + * But we have to close it now, otherwise rdma_bind_addr + * will return EADDRINUSE... How crap! */ + cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + + cmid = dev->ibd_hdev->ibh_cmid; + /* make next schedule of kiblnd_dev_need_failover + * will return 1 for me */ + dev->ibd_hdev->ibh_cmid = NULL; + cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + + rdma_destroy_id(cmid); + } + + cmid = rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP); + if (IS_ERR(cmid)) { + rc = PTR_ERR(cmid); + CERROR("Failed to create cmid for failover: %d\n", rc); + goto out; + } + + memset(&addr, 0, sizeof(addr)); + addr.sin_family = AF_INET; + addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip); + addr.sin_port = htons(*kiblnd_tunables.kib_service); + + /* Bind to failover device or port */ + rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr); + if (rc != 0) { + CERROR("Failed to bind %s to device: %d\n", + dev->ibd_ifname, rc); + rdma_destroy_id(cmid); + goto out; + } + + LIBCFS_ALLOC(hdev, sizeof(*hdev)); + if (hdev == NULL) { + CERROR("Failed to allocate kib_hca_dev\n"); + rdma_destroy_id(cmid); + rc = -ENOMEM; + goto out; + } + memset(hdev, 0, sizeof(*hdev)); + atomic_set(&hdev->ibh_ref, 1); + hdev->ibh_dev = dev; + hdev->ibh_cmid = cmid; + hdev->ibh_ibdev = cmid->device; + + pd = ib_alloc_pd(cmid->device); + if (IS_ERR(pd)) { + rc = PTR_ERR(pd); + CERROR("Can't allocate PD: %d\n", rc); + goto out; + } + + hdev->ibh_pd = pd; + + rc = rdma_listen(cmid, 0); + if (rc != 0) { + CERROR("Can't start new listener: %d\n", rc); + goto out; + } + + rc = kiblnd_hdev_setup_mrs(hdev); + if (rc != 0) { + CERROR("Can't setup device: %d\n", rc); + goto out; + } + + cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); + + old = dev->ibd_hdev; + dev->ibd_hdev = hdev; /* take over the refcount */ + hdev = old; + + cfs_list_for_each_entry(net, &dev->ibd_nets, ibn_list) { + kiblnd_fail_pool_set(&net->ibn_tx_ps.tps_poolset, &zombie_tpo); + if (net->ibn_with_pmr) + kiblnd_fail_pool_set(&net->ibn_pmr_ps.pps_poolset, &zombie_ppo); + if (net->ibn_with_fmr) + kiblnd_fail_fmr_pool_set(&net->ibn_fmr_ps, &zombie_fpo); + } + + cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + out: + if (!cfs_list_empty(&zombie_tpo)) + kiblnd_destroy_pool_list(&zombie_tpo); + if (!cfs_list_empty(&zombie_ppo)) + kiblnd_destroy_pool_list(&zombie_ppo); + if (!cfs_list_empty(&zombie_fpo)) + kiblnd_destroy_fmr_pool_list(&zombie_fpo); + if (hdev != NULL) + kiblnd_hdev_decref(hdev); + + if (rc != 0) + dev->ibd_failed_failover++; + else + dev->ibd_failed_failover = 0; + + return rc; +} + void kiblnd_destroy_dev (kib_dev_t *dev) { LASSERT (dev->ibd_nnets == 0); + LASSERT (cfs_list_empty(&dev->ibd_nets)); - if (!cfs_list_empty(&dev->ibd_list)) /* on kib_devs? */ - cfs_list_del_init(&dev->ibd_list); + cfs_list_del(&dev->ibd_fail_list); + cfs_list_del(&dev->ibd_list); - kiblnd_dev_cleanup(dev); + if (dev->ibd_hdev != NULL) + kiblnd_hdev_decref(dev->ibd_hdev); - if (dev->ibd_pd != NULL) - ib_dealloc_pd(dev->ibd_pd); + LIBCFS_FREE(dev, sizeof(*dev)); +} - if (dev->ibd_cmid != NULL) - rdma_destroy_id(dev->ibd_cmid); +kib_dev_t * +kiblnd_create_dev(char *ifname) +{ + struct net_device *netdev; + kib_dev_t *dev; + __u32 netmask; + __u32 ip; + int up; + int rc; - LIBCFS_FREE(dev, sizeof(*dev)); + rc = libcfs_ipif_query(ifname, &up, &ip, &netmask); + if (rc != 0) { + CERROR("Can't query IPoIB interface %s: %d\n", + ifname, rc); + return NULL; + } + + if (!up) { + CERROR("Can't query IPoIB interface %s: it's down\n", ifname); + return NULL; + } + + LIBCFS_ALLOC(dev, sizeof(*dev)); + if (dev == NULL) + return NULL; + + memset(dev, 0, sizeof(*dev)); +#ifdef HAVE_DEV_GET_BY_NAME_2ARG + if ((netdev = dev_get_by_name(&init_net, ifname)) == NULL) { +#else + if ((netdev = dev_get_by_name(ifname)) == NULL) { +#endif + CERROR("Can't find netdev of IF: %s\n", ifname); + LIBCFS_FREE(dev, sizeof(*dev)); + return NULL; + } + + dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER); + dev_put(netdev); + + CFS_INIT_LIST_HEAD(&dev->ibd_nets); + CFS_INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */ + CFS_INIT_LIST_HEAD(&dev->ibd_fail_list); + dev->ibd_ifip = ip; + strcpy(&dev->ibd_ifname[0], ifname); + + /* initialize the device */ + rc = kiblnd_dev_failover(dev); + if (rc != 0) { + CERROR("Can't initialize device: %d\n", rc); + LIBCFS_FREE(dev, sizeof(*dev)); + return NULL; + } + + cfs_list_add_tail(&dev->ibd_list, + &kiblnd_data.kib_devs); + return dev; } void @@ -2256,6 +2639,7 @@ kiblnd_base_shutdown (void) kiblnd_data.kib_shutdown = 1; cfs_waitq_broadcast(&kiblnd_data.kib_sched_waitq); cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq); + cfs_waitq_broadcast(&kiblnd_data.kib_failover_waitq); i = 2; while (cfs_atomic_read(&kiblnd_data.kib_nthreads) != 0) { @@ -2325,8 +2709,11 @@ kiblnd_shutdown (lnet_ni_t *ni) kiblnd_ni_fini_pools(net); + cfs_write_lock_irqsave(g_lock, flags); LASSERT (net->ibn_dev->ibd_nnets > 0); net->ibn_dev->ibd_nnets--; + cfs_list_del(&net->ibn_list); + cfs_write_unlock_irqrestore(g_lock, flags); /* fall through */ @@ -2368,6 +2755,7 @@ kiblnd_base_startup (void) cfs_rwlock_init(&kiblnd_data.kib_global_lock); CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs); + CFS_INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs); kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE; LIBCFS_ALLOC(kiblnd_data.kib_peers, @@ -2387,6 +2775,7 @@ kiblnd_base_startup (void) cfs_spin_lock_init(&kiblnd_data.kib_sched_lock); CFS_INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns); cfs_waitq_init(&kiblnd_data.kib_sched_waitq); + cfs_waitq_init(&kiblnd_data.kib_failover_waitq); kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR; @@ -2409,6 +2798,14 @@ kiblnd_base_startup (void) goto failed; } + if (*kiblnd_tunables.kib_dev_failover != 0) + rc = kiblnd_thread_start(kiblnd_failover_thread, NULL); + + if (rc != 0) { + CERROR("Can't spawn o2iblnd failover thread: %d\n", rc); + goto failed; + } + /* flag everything initialised */ kiblnd_data.kib_init = IBLND_INIT_ALL; /*****************************************************/ @@ -2428,6 +2825,7 @@ kiblnd_startup (lnet_ni_t *ni) kib_net_t *net; cfs_list_t *tmp; struct timeval tv; + unsigned long flags; int rc; LASSERT (ni->ni_lnd == &the_o2iblnd); @@ -2481,92 +2879,26 @@ kiblnd_startup (lnet_ni_t *ni) ibdev = NULL; } - if (ibdev == NULL) { - __u32 ip; - __u32 netmask; - int up; - struct rdma_cm_id *id; - struct ib_pd *pd; - struct sockaddr_in addr; - - rc = libcfs_ipif_query(ifname, &up, &ip, &netmask); - if (rc != 0) { - CERROR("Can't query IPoIB interface %s: %d\n", - ifname, rc); - goto failed; - } - - if (!up) { - CERROR("Can't query IPoIB interface %s: it's down\n", - ifname); - goto failed; - } - - LIBCFS_ALLOC(ibdev, sizeof(*ibdev)); - if (ibdev == NULL) - goto failed; - - memset(ibdev, 0, sizeof(*ibdev)); + if (ibdev == NULL) + ibdev = kiblnd_create_dev(ifname); - CFS_INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */ - ibdev->ibd_ifip = ip; - strcpy(&ibdev->ibd_ifname[0], ifname); - - id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP); - if (IS_ERR(id)) { - CERROR("Can't create listen ID: %ld\n", PTR_ERR(id)); - goto failed; - } - - ibdev->ibd_cmid = id; - - memset(&addr, 0, sizeof(addr)); - addr.sin_family = AF_INET; - addr.sin_port = htons(*kiblnd_tunables.kib_service); - addr.sin_addr.s_addr = htonl(ip); - - rc = rdma_bind_addr(id, (struct sockaddr *)&addr); - if (rc != 0) { - CERROR("Can't bind to %s: %d\n", ifname, rc); - goto failed; - } - - /* Binding should have assigned me an IB device */ - LASSERT (id->device != NULL); - CDEBUG(D_CONSOLE, "Listener bound to %s:%u.%u.%u.%u:%d:%s\n", - ifname, HIPQUAD(ip), *kiblnd_tunables.kib_service, - id->device->name); - - pd = ib_alloc_pd(id->device); - if (IS_ERR(pd)) { - CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd)); - goto failed; - } - - ibdev->ibd_pd = pd; - - rc = rdma_listen(id, 256); - if (rc != 0) { - CERROR("Can't start listener: %d\n", rc); - goto failed; - } - - rc = kiblnd_dev_setup(ibdev); - if (rc != 0) { - CERROR("Can't setup device: %d\n", rc); - goto failed; - } - } + if (ibdev == NULL) + goto failed; - ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip); net->ibn_dev = ibdev; + ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip); rc = kiblnd_net_init_pools(net); if (rc != 0) { CERROR("Failed to initialize NI pools: %d\n", rc); goto failed; } + + cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags); ibdev->ibd_nnets++; + cfs_list_add_tail(&net->ibn_list, &ibdev->ibd_nets); + cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags); + net->ibn_init = IBLND_INIT_ALL; return 0; diff --git a/lnet/klnds/o2iblnd/o2iblnd.h b/lnet/klnds/o2iblnd/o2iblnd.h index f946762..8f7db5a 100644 --- a/lnet/klnds/o2iblnd/o2iblnd.h +++ b/lnet/klnds/o2iblnd/o2iblnd.h @@ -98,6 +98,7 @@ typedef int gfp_t; typedef struct { + int *kib_dev_failover; /* HCA failover */ unsigned int *kib_service; /* IB service number */ int *kib_min_reconnect_interval; /* first failed connection retry... */ int *kib_max_reconnect_interval; /* ...exponentially increasing to this */ @@ -184,30 +185,44 @@ kiblnd_concurrent_sends_v1(void) #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)) +struct kib_hca_dev; + typedef struct { cfs_list_t ibd_list; /* chain on kib_devs */ + cfs_list_t ibd_fail_list; /* chain on kib_failed_devs */ __u32 ibd_ifip; /* IPoIB interface IP */ char ibd_ifname[32]; /* IPoIB interface name */ int ibd_nnets; /* # nets extant */ - struct rdma_cm_id *ibd_cmid; /* IB listener (bound to 1 device) */ - struct ib_pd *ibd_pd; /* PD for the device */ - int ibd_page_shift; /* page shift of current HCA */ - int ibd_page_size; /* page size of current HCA */ - __u64 ibd_page_mask; /* page mask of current HCA */ - int ibd_mr_shift; /* bits shift of max MR size */ - __u64 ibd_mr_size; /* size of MR */ - - int ibd_nmrs; /* # of global MRs */ - struct ib_mr **ibd_mrs; /* MR for non RDMA I/O */ + cfs_time_t ibd_next_failover; + int ibd_failed_failover; /* # failover failures */ + unsigned int ibd_failover; /* failover in progress */ + unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */ + cfs_list_t ibd_nets; + struct kib_hca_dev *ibd_hdev; } kib_dev_t; +typedef struct kib_hca_dev +{ + struct rdma_cm_id *ibh_cmid; /* listener cmid */ + struct ib_device *ibh_ibdev; /* IB device */ + int ibh_page_shift; /* page shift of current HCA */ + int ibh_page_size; /* page size of current HCA */ + __u64 ibh_page_mask; /* page mask of current HCA */ + int ibh_mr_shift; /* bits shift of max MR size */ + __u64 ibh_mr_size; /* size of MR */ + int ibh_nmrs; /* # of global MRs */ + struct ib_mr **ibh_mrs; /* global MR */ + struct ib_pd *ibh_pd; /* PD */ + kib_dev_t *ibh_dev; /* owner */ + cfs_atomic_t ibh_ref; /* refcount */ +} kib_hca_dev_t; + #define IBLND_POOL_DEADLINE 300 /* # of seconds to keep pool alive */ typedef struct { - struct ib_device *ibp_device; /* device for mapping */ int ibp_npages; /* # pages */ struct page *ibp_pages[0]; /* page array */ } kib_pages_t; @@ -243,6 +258,7 @@ typedef struct kib_poolset struct kib_net *ps_net; /* network it belongs to */ char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */ cfs_list_t ps_pool_list; /* list of pools */ + cfs_list_t ps_failed_pool_list; /* failed pool list */ cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */ int ps_increasing; /* is allocating new pool */ int ps_pool_size; /* new pool size */ @@ -260,6 +276,7 @@ typedef struct kib_pool kib_poolset_t *po_owner; /* pool_set of this pool */ cfs_time_t po_deadline; /* deadline of this pool */ int po_allocated; /* # of elements in use */ + int po_failed; /* pool is created on failed HCA */ int po_size; /* # of pre-allocated elements */ } kib_pool_t; @@ -270,6 +287,7 @@ typedef struct { typedef struct { kib_pool_t tpo_pool; /* pool */ + struct kib_hca_dev *tpo_hdev; /* device for this pool */ struct kib_tx *tpo_tx_descs; /* all the tx descriptors */ kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */ } kib_tx_pool_t; @@ -279,6 +297,7 @@ typedef struct { } kib_pmr_poolset_t; typedef struct kib_pmr_pool { + struct kib_hca_dev *ppo_hdev; /* device for this pool */ kib_pool_t ppo_pool; /* pool */ } kib_pmr_pool_t; @@ -287,6 +306,7 @@ typedef struct cfs_spinlock_t fps_lock; /* serialize */ struct kib_net *fps_net; /* IB network */ cfs_list_t fps_pool_list; /* FMR pool list */ + cfs_list_t fps_failed_pool_list; /* FMR pool list */ __u64 fps_version; /* validity stamp */ int fps_increasing; /* is allocating new pool */ cfs_time_t fps_next_retry; /* time stamp for retry if failed to allocate */ @@ -295,9 +315,11 @@ typedef struct typedef struct { cfs_list_t 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 */ cfs_time_t fpo_deadline; /* deadline of this pool */ + int fpo_failed; /* fmr pool is failed */ int fpo_map_count; /* # of mapped FMR */ } kib_fmr_pool_t; @@ -308,6 +330,7 @@ typedef struct { typedef struct kib_net { + cfs_list_t ibn_list; /* chain on kib_dev_t::ibd_nets */ __u64 ibn_incarnation; /* my epoch */ int ibn_init; /* initialisation state */ int ibn_shutdown; /* shutting down? */ @@ -329,6 +352,7 @@ typedef struct int kib_init; /* initialisation state */ int kib_shutdown; /* shut down? */ cfs_list_t kib_devs; /* IB devices extant */ + cfs_list_t kib_failed_devs; /* list head of failed devices */ cfs_atomic_t kib_nthreads; /* # live threads */ cfs_rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */ @@ -344,6 +368,7 @@ typedef struct cfs_waitq_t kib_sched_waitq; /* schedulers sleep here */ cfs_list_t kib_sched_conns; /* conns to check for rx completions */ cfs_spinlock_t kib_sched_lock; /* serialise */ + cfs_waitq_t kib_failover_waitq; /* schedulers sleep here */ struct ib_qp_attr kib_error_qpa; /* QP->ERROR */ } kib_data_t; @@ -529,6 +554,7 @@ typedef struct kib_connvars typedef struct kib_conn { struct kib_peer *ibc_peer; /* owning peer */ + kib_hca_dev_t *ibc_hdev; /* HCA bound on */ cfs_list_t ibc_list; /* stash on peer's conn list */ cfs_list_t ibc_sched_list; /* schedule for attention */ __u16 ibc_version; /* version of connection */ @@ -585,6 +611,38 @@ typedef struct kib_peer extern kib_data_t kiblnd_data; +extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev); + +static inline void +kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev) +{ + LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0); + cfs_atomic_inc(&hdev->ibh_ref); +} + +static inline void +kiblnd_hdev_decref(kib_hca_dev_t *hdev) +{ + LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0); + if (cfs_atomic_dec_and_test(&hdev->ibh_ref)) + kiblnd_hdev_destroy(hdev); +} + +static inline int +kiblnd_dev_can_failover(kib_dev_t *dev) +{ + if (!cfs_list_empty(&dev->ibd_fail_list)) /* already scheduled */ + return 0; + + if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */ + return 0; + + if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */ + return 1; + + return dev->ibd_can_failover; +} + #define kiblnd_conn_addref(conn) \ do { \ CDEBUG(D_NET, "conn[%p] (%d)++\n", \ @@ -922,9 +980,9 @@ static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev, #endif -struct ib_mr *kiblnd_find_rd_dma_mr(kib_net_t *net, +struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd); -struct ib_mr *kiblnd_find_dma_mr(kib_net_t *net, +struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size); void kiblnd_map_rx_descs(kib_conn_t *conn); void kiblnd_unmap_rx_descs(kib_conn_t *conn); @@ -938,8 +996,8 @@ int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages, __u64 iov, kib_fmr_t *fmr); void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status); -int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd, - __u64 *iova, kib_phys_mr_t **pp_pmr); +int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev, + kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr); void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr); int kiblnd_startup (lnet_ni_t *ni); @@ -953,6 +1011,7 @@ void kiblnd_tunables_fini(void); int kiblnd_connd (void *arg); int kiblnd_scheduler(void *arg); int kiblnd_thread_start (int (*fn)(void *arg), void *arg); +int kiblnd_failover_thread (void *arg); int kiblnd_alloc_pages (kib_pages_t **pp, int npages); void kiblnd_free_pages (kib_pages_t *p); @@ -961,6 +1020,7 @@ int kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event); 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); void kiblnd_destroy_peer (kib_peer_t *peer); void kiblnd_destroy_dev (kib_dev_t *dev); diff --git a/lnet/klnds/o2iblnd/o2iblnd_cb.c b/lnet/klnds/o2iblnd/o2iblnd_cb.c index 167005f..577db3e 100644 --- a/lnet/klnds/o2iblnd/o2iblnd_cb.c +++ b/lnet/klnds/o2iblnd/o2iblnd_cb.c @@ -154,7 +154,7 @@ kiblnd_post_rx (kib_rx_t *rx, int credit) credit == IBLND_POSTRX_PEER_CREDIT || credit == IBLND_POSTRX_RSRVD_CREDIT); - mr = kiblnd_find_dma_mr(net, rx->rx_msgaddr, IBLND_MSG_SIZE); + mr = kiblnd_find_dma_mr(conn->ibc_hdev, rx->rx_msgaddr, IBLND_MSG_SIZE); LASSERT (mr != NULL); rx->rx_sge.lkey = mr->lkey; @@ -538,8 +538,8 @@ kiblnd_kvaddr_to_page (unsigned long vaddr) static int kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) { - kib_dev_t *ibdev = net->ibn_dev; - __u64 *pages = tx->tx_pages; + kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev; + __u64 *pages = tx->tx_pages; int npages; int size; int rc; @@ -547,9 +547,9 @@ kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) for (i = 0, npages = 0; i < rd->rd_nfrags; i++) { for (size = 0; size < rd->rd_frags[i].rf_nob; - size += ibdev->ibd_page_size) { + size += hdev->ibh_page_size) { pages[npages ++] = (rd->rd_frags[i].rf_addr & - ibdev->ibd_page_mask) + size; + hdev->ibh_page_mask) + size; } } @@ -563,7 +563,7 @@ kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) * the rkey */ rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.fmr.fmr_pfmr->fmr->rkey : tx->tx_u.fmr.fmr_pfmr->fmr->lkey; - rd->rd_frags[0].rf_addr &= ~ibdev->ibd_page_mask; + rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask; rd->rd_frags[0].rf_nob = nob; rd->rd_nfrags = 1; @@ -573,12 +573,13 @@ kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) static int kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob) { + kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev; __u64 iova; int rc; - iova = rd->rd_frags[0].rf_addr & ~net->ibn_dev->ibd_page_mask; + iova = rd->rd_frags[0].rf_addr & ~hdev->ibh_page_mask; - rc = kiblnd_pmr_pool_map(&net->ibn_pmr_ps, rd, &iova, &tx->tx_u.pmr); + rc = kiblnd_pmr_pool_map(&net->ibn_pmr_ps, hdev, rd, &iova, &tx->tx_u.pmr); if (rc != 0) { CERROR("Failed to create MR by phybuf: %d\n", rc); return rc; @@ -611,7 +612,7 @@ kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx) } if (tx->tx_nfrags != 0) { - kiblnd_dma_unmap_sg(net->ibn_dev->ibd_cmid->device, + kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev, tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir); tx->tx_nfrags = 0; } @@ -621,6 +622,7 @@ int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, int nfrags) { + kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev; kib_net_t *net = ni->ni_data; struct ib_mr *mr = NULL; __u32 nob; @@ -632,19 +634,19 @@ kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, tx->tx_nfrags = nfrags; rd->rd_nfrags = - kiblnd_dma_map_sg(net->ibn_dev->ibd_cmid->device, + kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir); for (i = 0, nob = 0; i < rd->rd_nfrags; i++) { rd->rd_frags[i].rf_nob = kiblnd_sg_dma_len( - net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]); + hdev->ibh_ibdev, &tx->tx_frags[i]); rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address( - net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]); + hdev->ibh_ibdev, &tx->tx_frags[i]); nob += rd->rd_frags[i].rf_nob; } /* looking for pre-mapping MR */ - mr = kiblnd_find_rd_dma_mr(net, rd); + mr = kiblnd_find_rd_dma_mr(hdev, rd); if (mr != NULL) { /* found pre-mapping MR */ rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey; @@ -835,11 +837,17 @@ kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit) cfs_list_add(&tx->tx_list, &conn->ibc_active_txs); /* I'm still holding ibc_lock! */ - if (conn->ibc_state != IBLND_CONN_ESTABLISHED) + if (conn->ibc_state != IBLND_CONN_ESTABLISHED) { rc = -ECONNABORTED; - else + } else if (tx->tx_pool->tpo_pool.po_failed || + conn->ibc_hdev != tx->tx_pool->tpo_hdev) { + /* close_conn will launch failover */ + rc = -ENETDOWN; + } else { rc = ib_post_send(conn->ibc_cmid->qp, tx->tx_wrq, &bad_wrq); + } + conn->ibc_last_send = jiffies; if (rc == 0) @@ -1005,20 +1013,19 @@ kiblnd_tx_complete (kib_tx_t *tx, int status) void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob) { - kib_net_t *net = ni->ni_data; + kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev; struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq]; struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq]; int nob = offsetof (kib_msg_t, ibm_u) + body_nob; struct ib_mr *mr; - LASSERT (net != NULL); LASSERT (tx->tx_nwrq >= 0); LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1); LASSERT (nob <= IBLND_MSG_SIZE); kiblnd_init_msg(tx->tx_msg, type, body_nob); - mr = kiblnd_find_dma_mr(net, tx->tx_msgaddr, nob); + mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob); LASSERT (mr != NULL); sge->lkey = mr->lkey; @@ -1768,8 +1775,9 @@ kiblnd_close_conn_locked (kib_conn_t *conn, int error) * connection to be finished off by the connd. Otherwise the connd is * already dealing with it (either to set it up or tear it down). * Caller holds kib_global_lock exclusively in irq context */ - unsigned long flags; kib_peer_t *peer = conn->ibc_peer; + kib_dev_t *dev; + unsigned long flags; LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED); @@ -1795,6 +1803,7 @@ kiblnd_close_conn_locked (kib_conn_t *conn, int error) cfs_list_empty(&conn->ibc_active_txs) ? "" : "(waiting)"); } + dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev; cfs_list_del(&conn->ibc_list); /* connd (see below) takes over ibc_list's ref */ @@ -1808,6 +1817,13 @@ kiblnd_close_conn_locked (kib_conn_t *conn, int error) kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING); + if (error != 0 && + kiblnd_dev_can_failover(dev)) { + cfs_list_add_tail(&dev->ibd_fail_list, + &kiblnd_data.kib_failed_devs); + cfs_waitq_signal(&kiblnd_data.kib_failover_waitq); + } + cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); cfs_list_add_tail (&conn->ibc_list, &kiblnd_data.kib_connd_conns); @@ -3107,7 +3123,7 @@ kiblnd_qp_event(struct ib_event *event, void *arg) CDEBUG(D_NET, "%s established\n", libcfs_nid2str(conn->ibc_peer->ibp_nid)); return; - + default: CERROR("%s: Async QP event type %d\n", libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event); @@ -3311,3 +3327,94 @@ kiblnd_scheduler(void *arg) kiblnd_thread_fini(); return (0); } + +int +kiblnd_failover_thread(void *arg) +{ + cfs_rwlock_t *glock = &kiblnd_data.kib_global_lock; + kib_dev_t *dev; + cfs_waitlink_t wait; + unsigned long flags; + int rc; + + LASSERT (*kiblnd_tunables.kib_dev_failover != 0); + + cfs_daemonize ("kiblnd_failover"); + cfs_block_allsigs (); + + cfs_waitlink_init(&wait); + cfs_write_lock_irqsave(glock, flags); + + while (!kiblnd_data.kib_shutdown) { + int do_failover = 0; + int long_sleep; + + cfs_list_for_each_entry(dev, &kiblnd_data.kib_failed_devs, + ibd_fail_list) { + if (cfs_time_before(cfs_time_current(), + dev->ibd_next_failover)) + continue; + do_failover = 1; + break; + } + + if (do_failover) { + cfs_list_del_init(&dev->ibd_fail_list); + dev->ibd_failover = 1; + cfs_write_unlock_irqrestore(glock, flags); + + rc = kiblnd_dev_failover(dev); + + cfs_write_lock_irqsave(glock, flags); + + LASSERT (dev->ibd_failover); + dev->ibd_failover = 0; + if (rc >= 0) { /* Device is OK or failover succeed */ + dev->ibd_next_failover = cfs_time_shift(3); + continue; + } + + /* failed to failover, retry later */ + dev->ibd_next_failover = + cfs_time_shift(min(dev->ibd_failed_failover, 10)); + if (kiblnd_dev_can_failover(dev)) { + cfs_list_add_tail(&dev->ibd_fail_list, + &kiblnd_data.kib_failed_devs); + } + + continue; + } + + /* long sleep if no more pending failover */ + long_sleep = cfs_list_empty(&kiblnd_data.kib_failed_devs); + + cfs_set_current_state(CFS_TASK_INTERRUPTIBLE); + cfs_waitq_add(&kiblnd_data.kib_failover_waitq, &wait); + cfs_write_unlock_irqrestore(glock, flags); + + rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) : + cfs_time_seconds(1)); + cfs_set_current_state(CFS_TASK_RUNNING); + cfs_waitq_del(&kiblnd_data.kib_failover_waitq, &wait); + cfs_write_lock_irqsave(glock, flags); + + if (!long_sleep || rc != 0) + continue; + + /* have a long sleep, routine check all active devices, + * we need checking like this because if there is not active + * connection on the dev and no SEND from local, we may listen + * on wrong HCA for ever while there is a bonding failover */ + cfs_list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) { + if (kiblnd_dev_can_failover(dev)) { + cfs_list_add_tail(&dev->ibd_fail_list, + &kiblnd_data.kib_failed_devs); + } + } + } + + cfs_write_unlock_irqrestore(glock, flags); + + kiblnd_thread_fini(); + return 0; +} diff --git a/lnet/klnds/o2iblnd/o2iblnd_modparams.c b/lnet/klnds/o2iblnd/o2iblnd_modparams.c index 4aec961..3b95cd3 100644 --- a/lnet/klnds/o2iblnd/o2iblnd_modparams.c +++ b/lnet/klnds/o2iblnd/o2iblnd_modparams.c @@ -120,7 +120,17 @@ static int pmr_pool_size = 512; CFS_MODULE_PARM(pmr_pool_size, "i", int, 0444, "size of the MR cache pmr pool"); +/* + * 0: disable failover + * 1: enable failover if necessary + * 2: force to failover (for debug) + */ +static int dev_failover = 0; +CFS_MODULE_PARM(dev_failover, "i", int, 0444, + "HCA failover for bonding (0 off, 1 on, other values reserved)"); + kib_tunables_t kiblnd_tunables = { + .kib_dev_failover = &dev_failover, .kib_service = &service, .kib_cksum = &cksum, .kib_timeout = &timeout, @@ -169,7 +179,8 @@ enum { O2IBLND_FMR_POOL_SIZE, O2IBLND_FMR_FLUSH_TRIGGER, O2IBLND_FMR_CACHE, - O2IBLND_PMR_POOL_SIZE + O2IBLND_PMR_POOL_SIZE, + O2IBLND_DEV_FAILOVER }; #else @@ -193,6 +204,7 @@ enum { #define O2IBLND_FMR_FLUSH_TRIGGER CTL_UNNUMBERED #define O2IBLND_FMR_CACHE CTL_UNNUMBERED #define O2IBLND_PMR_POOL_SIZE CTL_UNNUMBERED +#define O2IBLND_DEV_FAILOVER CTL_UNNUMBERED #endif @@ -358,6 +370,14 @@ static cfs_sysctl_table_t kiblnd_ctl_table[] = { .mode = 0444, .proc_handler = &proc_dointvec }, + { + .ctl_name = O2IBLND_DEV_FAILOVER, + .procname = "dev_failover", + .data = &dev_failover, + .maxlen = sizeof(int), + .mode = 0444, + .proc_handler = &proc_dointvec + }, {0} }; -- 1.8.3.1