From 95b60f078899a9680688bab909d02169e1fd632e Mon Sep 17 00:00:00 2001 From: Nicholas Bellinger Date: Tue, 5 Nov 2013 13:16:12 -0800 Subject: ib_isert: Add support for completion interrupt coalescing This patch adds support for completion interrupt coalescing that allows only every ISERT_COMP_BATCH_COUNT (8) to set IB_SEND_SIGNALED, thus avoiding completion interrupts for every posted iser_tx_desc. The batch processing is done using a per isert_conn llist that once IB_SEND_SIGNALED has been set is saved to tx_desc->comp_llnode_batch, and completion processing of previously posted iser_tx_descs is done in a single shot from within isert_send_completion() code. Note this is only done for response PDUs from ISCSI_OP_SCSI_CMD, and all other control type of PDU responses will force an implicit batch drain to occur. Cc: Or Gerlitz Cc: Sagi Grimberg Cc: Kent Overstreet Cc: Roland Dreier Signed-off-by: Nicholas Bellinger --- drivers/infiniband/ulp/isert/ib_isert.c | 63 ++++++++++++++++++++++++++------- drivers/infiniband/ulp/isert/ib_isert.h | 6 ++++ 2 files changed, 57 insertions(+), 12 deletions(-) (limited to 'drivers/infiniband/ulp') diff --git a/drivers/infiniband/ulp/isert/ib_isert.c b/drivers/infiniband/ulp/isert/ib_isert.c index 3591855cc5b5..27708c3430d5 100644 --- a/drivers/infiniband/ulp/isert/ib_isert.c +++ b/drivers/infiniband/ulp/isert/ib_isert.c @@ -22,6 +22,7 @@ #include #include #include +#include #include #include #include @@ -489,6 +490,7 @@ isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event) kref_init(&isert_conn->conn_kref); kref_get(&isert_conn->conn_kref); mutex_init(&isert_conn->conn_mutex); + mutex_init(&isert_conn->conn_comp_mutex); spin_lock_init(&isert_conn->conn_lock); cma_id->context = isert_conn; @@ -843,14 +845,32 @@ isert_init_tx_hdrs(struct isert_conn *isert_conn, } static void -isert_init_send_wr(struct isert_cmd *isert_cmd, struct ib_send_wr *send_wr) +isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd, + struct ib_send_wr *send_wr, bool coalesce) { + struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc; + isert_cmd->rdma_wr.iser_ib_op = ISER_IB_SEND; send_wr->wr_id = (unsigned long)&isert_cmd->tx_desc; send_wr->opcode = IB_WR_SEND; - send_wr->send_flags = IB_SEND_SIGNALED; - send_wr->sg_list = &isert_cmd->tx_desc.tx_sg[0]; + send_wr->sg_list = &tx_desc->tx_sg[0]; send_wr->num_sge = isert_cmd->tx_desc.num_sge; + /* + * Coalesce send completion interrupts by only setting IB_SEND_SIGNALED + * bit for every ISERT_COMP_BATCH_COUNT number of ib_post_send() calls. + */ + mutex_lock(&isert_conn->conn_comp_mutex); + if (coalesce && + ++isert_conn->conn_comp_batch < ISERT_COMP_BATCH_COUNT) { + llist_add(&tx_desc->comp_llnode, &isert_conn->conn_comp_llist); + mutex_unlock(&isert_conn->conn_comp_mutex); + return; + } + isert_conn->conn_comp_batch = 0; + tx_desc->comp_llnode_batch = llist_del_all(&isert_conn->conn_comp_llist); + mutex_unlock(&isert_conn->conn_comp_mutex); + + send_wr->send_flags = IB_SEND_SIGNALED; } static int @@ -1582,8 +1602,8 @@ isert_response_completion(struct iser_tx_desc *tx_desc, } static void -isert_send_completion(struct iser_tx_desc *tx_desc, - struct isert_conn *isert_conn) +__isert_send_completion(struct iser_tx_desc *tx_desc, + struct isert_conn *isert_conn) { struct ib_device *ib_dev = isert_conn->conn_cm_id->device; struct isert_cmd *isert_cmd = tx_desc->isert_cmd; @@ -1623,6 +1643,24 @@ isert_send_completion(struct iser_tx_desc *tx_desc, } } +static void +isert_send_completion(struct iser_tx_desc *tx_desc, + struct isert_conn *isert_conn) +{ + struct llist_node *llnode = tx_desc->comp_llnode_batch; + struct iser_tx_desc *t; + /* + * Drain coalesced completion llist starting from comp_llnode_batch + * setup in isert_init_send_wr(), and then complete trailing tx_desc. + */ + while (llnode) { + t = llist_entry(llnode, struct iser_tx_desc, comp_llnode); + llnode = llist_next(llnode); + __isert_send_completion(t, isert_conn); + } + __isert_send_completion(tx_desc, isert_conn); +} + static void isert_cq_comp_err(struct iser_tx_desc *tx_desc, struct isert_conn *isert_conn) { @@ -1793,7 +1831,7 @@ isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd) isert_cmd->tx_desc.num_sge = 2; } - isert_init_send_wr(isert_cmd, send_wr); + isert_init_send_wr(isert_conn, isert_cmd, send_wr, true); pr_debug("Posting SCSI Response IB_WR_SEND >>>>>>>>>>>>>>>>>>>>>>\n"); @@ -1813,7 +1851,7 @@ isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn, &isert_cmd->tx_desc.iscsi_header, nopout_response); isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc); - isert_init_send_wr(isert_cmd, send_wr); + isert_init_send_wr(isert_conn, isert_cmd, send_wr, false); pr_debug("Posting NOPIN Response IB_WR_SEND >>>>>>>>>>>>>>>>>>>>>>\n"); @@ -1831,7 +1869,7 @@ isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn) iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *) &isert_cmd->tx_desc.iscsi_header); isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc); - isert_init_send_wr(isert_cmd, send_wr); + isert_init_send_wr(isert_conn, isert_cmd, send_wr, false); pr_debug("Posting Logout Response IB_WR_SEND >>>>>>>>>>>>>>>>>>>>>>\n"); @@ -1849,7 +1887,7 @@ isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn) iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *) &isert_cmd->tx_desc.iscsi_header); isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc); - isert_init_send_wr(isert_cmd, send_wr); + isert_init_send_wr(isert_conn, isert_cmd, send_wr, false); pr_debug("Posting Task Management Response IB_WR_SEND >>>>>>>>>>>>>>>>>>>>>>\n"); @@ -1881,7 +1919,7 @@ isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn) tx_dsg->lkey = isert_conn->conn_mr->lkey; isert_cmd->tx_desc.num_sge = 2; - isert_init_send_wr(isert_cmd, send_wr); + isert_init_send_wr(isert_conn, isert_cmd, send_wr, false); pr_debug("Posting Reject IB_WR_SEND >>>>>>>>>>>>>>>>>>>>>>\n"); @@ -1921,7 +1959,7 @@ isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn) tx_dsg->lkey = isert_conn->conn_mr->lkey; isert_cmd->tx_desc.num_sge = 2; } - isert_init_send_wr(isert_cmd, send_wr); + isert_init_send_wr(isert_conn, isert_cmd, send_wr, false); pr_debug("Posting Text Response IB_WR_SEND >>>>>>>>>>>>>>>>>>>>>>\n"); @@ -2309,7 +2347,8 @@ isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd) iscsit_build_rsp_pdu(cmd, conn, false, (struct iscsi_scsi_rsp *) &isert_cmd->tx_desc.iscsi_header); isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc); - isert_init_send_wr(isert_cmd, &isert_cmd->tx_desc.send_wr); + isert_init_send_wr(isert_conn, isert_cmd, + &isert_cmd->tx_desc.send_wr, true); atomic_inc(&isert_conn->post_send_buf_count); diff --git a/drivers/infiniband/ulp/isert/ib_isert.h b/drivers/infiniband/ulp/isert/ib_isert.h index 631f2090f0b8..691f90ff2d83 100644 --- a/drivers/infiniband/ulp/isert/ib_isert.h +++ b/drivers/infiniband/ulp/isert/ib_isert.h @@ -43,6 +43,8 @@ struct iser_tx_desc { struct ib_sge tx_sg[2]; int num_sge; struct isert_cmd *isert_cmd; + struct llist_node *comp_llnode_batch; + struct llist_node comp_llnode; struct ib_send_wr send_wr; } __packed; @@ -121,6 +123,10 @@ struct isert_conn { int conn_frwr_pool_size; /* lock to protect frwr_pool */ spinlock_t conn_lock; +#define ISERT_COMP_BATCH_COUNT 8 + int conn_comp_batch; + struct llist_head conn_comp_llist; + struct mutex conn_comp_mutex; }; #define ISERT_MAX_CQ 64 -- cgit v1.2.3 From 4a9a6c8d538eff6d627fa14409c7c25aff652541 Mon Sep 17 00:00:00 2001 From: Nicholas Bellinger Date: Wed, 6 Nov 2013 21:05:19 -0800 Subject: target: Drop left-over se_lun->lun_cmd_list shutdown code Now with percpu refcounting for se_lun in place, go ahead and drop the legacy per se_cmd accounting for se_lun shutdown. This includes __transport_clear_lun_from_sessions(), the associated transport_lun_wait_for_tasks() logic, along with a handful of now unused se_cmd structure members and ->transport_state bits. Cc: Kent Overstreet Signed-off-by: Nicholas Bellinger --- drivers/infiniband/ulp/srpt/ib_srpt.c | 10 -- drivers/target/target_core_tpg.c | 15 +-- drivers/target/target_core_transport.c | 219 +-------------------------------- drivers/target/target_core_xcopy.c | 2 - include/target/target_core_base.h | 13 +- 5 files changed, 7 insertions(+), 252 deletions(-) (limited to 'drivers/infiniband/ulp') diff --git a/drivers/infiniband/ulp/srpt/ib_srpt.c b/drivers/infiniband/ulp/srpt/ib_srpt.c index 653ac6bfc57a..16087966cb7d 100644 --- a/drivers/infiniband/ulp/srpt/ib_srpt.c +++ b/drivers/infiniband/ulp/srpt/ib_srpt.c @@ -1352,11 +1352,8 @@ static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx) /* XXX(hch): this is a horrible layering violation.. */ spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags); - ioctx->cmd.transport_state |= CMD_T_LUN_STOP; ioctx->cmd.transport_state &= ~CMD_T_ACTIVE; spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags); - - complete(&ioctx->cmd.transport_lun_stop_comp); break; case SRPT_STATE_CMD_RSP_SENT: /* @@ -1364,9 +1361,6 @@ static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx) * not been received in time. */ srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx); - spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags); - ioctx->cmd.transport_state |= CMD_T_LUN_STOP; - spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags); target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd); break; case SRPT_STATE_MGMT_RSP_SENT: @@ -1476,7 +1470,6 @@ static void srpt_handle_rdma_err_comp(struct srpt_rdma_ch *ch, { struct se_cmd *cmd; enum srpt_command_state state; - unsigned long flags; cmd = &ioctx->cmd; state = srpt_get_cmd_state(ioctx); @@ -1496,9 +1489,6 @@ static void srpt_handle_rdma_err_comp(struct srpt_rdma_ch *ch, __func__, __LINE__, state); break; case SRPT_RDMA_WRITE_LAST: - spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags); - ioctx->cmd.transport_state |= CMD_T_LUN_STOP; - spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags); break; default: printk(KERN_ERR "%s[%d]: opcode = %u\n", __func__, diff --git a/drivers/target/target_core_tpg.c b/drivers/target/target_core_tpg.c index f321af04ef03..f697f8baec54 100644 --- a/drivers/target/target_core_tpg.c +++ b/drivers/target/target_core_tpg.c @@ -654,9 +654,7 @@ static int core_tpg_setup_virtual_lun0(struct se_portal_group *se_tpg) atomic_set(&lun->lun_acl_count, 0); init_completion(&lun->lun_shutdown_comp); INIT_LIST_HEAD(&lun->lun_acl_list); - INIT_LIST_HEAD(&lun->lun_cmd_list); spin_lock_init(&lun->lun_acl_lock); - spin_lock_init(&lun->lun_cmd_lock); spin_lock_init(&lun->lun_sep_lock); init_completion(&lun->lun_ref_comp); @@ -706,9 +704,7 @@ int core_tpg_register( atomic_set(&lun->lun_acl_count, 0); init_completion(&lun->lun_shutdown_comp); INIT_LIST_HEAD(&lun->lun_acl_list); - INIT_LIST_HEAD(&lun->lun_cmd_list); spin_lock_init(&lun->lun_acl_lock); - spin_lock_init(&lun->lun_cmd_lock); spin_lock_init(&lun->lun_sep_lock); init_completion(&lun->lun_ref_comp); } @@ -849,14 +845,6 @@ int core_tpg_post_addlun( return 0; } -static void core_tpg_shutdown_lun( - struct se_portal_group *tpg, - struct se_lun *lun) -{ - core_clear_lun_from_tpg(lun, tpg); - transport_clear_lun_ref(lun); -} - struct se_lun *core_tpg_pre_dellun( struct se_portal_group *tpg, u32 unpacked_lun) @@ -891,7 +879,8 @@ int core_tpg_post_dellun( struct se_portal_group *tpg, struct se_lun *lun) { - core_tpg_shutdown_lun(tpg, lun); + core_clear_lun_from_tpg(lun, tpg); + transport_clear_lun_ref(lun); core_dev_unexport(lun->lun_se_dev, tpg, lun); diff --git a/drivers/target/target_core_transport.c b/drivers/target/target_core_transport.c index ad143d7293ec..e0669ac9278b 100644 --- a/drivers/target/target_core_transport.c +++ b/drivers/target/target_core_transport.c @@ -505,23 +505,6 @@ static int transport_cmd_check_stop(struct se_cmd *cmd, bool remove_from_lists, if (write_pending) cmd->t_state = TRANSPORT_WRITE_PENDING; - /* - * Determine if IOCTL context caller in requesting the stopping of this - * command for LUN shutdown purposes. - */ - if (cmd->transport_state & CMD_T_LUN_STOP) { - pr_debug("%s:%d CMD_T_LUN_STOP for ITT: 0x%08x\n", - __func__, __LINE__, cmd->se_tfo->get_task_tag(cmd)); - - cmd->transport_state &= ~CMD_T_ACTIVE; - if (remove_from_lists) - target_remove_from_state_list(cmd); - spin_unlock_irqrestore(&cmd->t_state_lock, flags); - - complete(&cmd->transport_lun_stop_comp); - return 1; - } - if (remove_from_lists) { target_remove_from_state_list(cmd); @@ -1078,13 +1061,10 @@ void transport_init_se_cmd( int task_attr, unsigned char *sense_buffer) { - INIT_LIST_HEAD(&cmd->se_lun_node); INIT_LIST_HEAD(&cmd->se_delayed_node); INIT_LIST_HEAD(&cmd->se_qf_node); INIT_LIST_HEAD(&cmd->se_cmd_list); INIT_LIST_HEAD(&cmd->state_list); - init_completion(&cmd->transport_lun_fe_stop_comp); - init_completion(&cmd->transport_lun_stop_comp); init_completion(&cmd->t_transport_stop_comp); init_completion(&cmd->cmd_wait_comp); init_completion(&cmd->task_stop_comp); @@ -1705,29 +1685,14 @@ void target_execute_cmd(struct se_cmd *cmd) /* * If the received CDB has aleady been aborted stop processing it here. */ - if (transport_check_aborted_status(cmd, 1)) { - complete(&cmd->transport_lun_stop_comp); + if (transport_check_aborted_status(cmd, 1)) return; - } - /* - * Determine if IOCTL context caller in requesting the stopping of this - * command for LUN shutdown purposes. - */ - spin_lock_irq(&cmd->t_state_lock); - if (cmd->transport_state & CMD_T_LUN_STOP) { - pr_debug("%s:%d CMD_T_LUN_STOP for ITT: 0x%08x\n", - __func__, __LINE__, cmd->se_tfo->get_task_tag(cmd)); - - cmd->transport_state &= ~CMD_T_ACTIVE; - spin_unlock_irq(&cmd->t_state_lock); - complete(&cmd->transport_lun_stop_comp); - return; - } /* * Determine if frontend context caller is requesting the stopping of * this command for frontend exceptions. */ + spin_lock_irq(&cmd->t_state_lock); if (cmd->transport_state & CMD_T_STOP) { pr_debug("%s:%d CMD_T_STOP for ITT: 0x%08x\n", __func__, __LINE__, @@ -2390,149 +2355,6 @@ void target_wait_for_sess_cmds(struct se_session *se_sess) } EXPORT_SYMBOL(target_wait_for_sess_cmds); -/* transport_lun_wait_for_tasks(): - * - * Called from ConfigFS context to stop the passed struct se_cmd to allow - * an struct se_lun to be successfully shutdown. - */ -static int transport_lun_wait_for_tasks(struct se_cmd *cmd, struct se_lun *lun) -{ - unsigned long flags; - int ret = 0; - - /* - * If the frontend has already requested this struct se_cmd to - * be stopped, we can safely ignore this struct se_cmd. - */ - spin_lock_irqsave(&cmd->t_state_lock, flags); - if (cmd->transport_state & CMD_T_STOP) { - cmd->transport_state &= ~CMD_T_LUN_STOP; - - pr_debug("ConfigFS ITT[0x%08x] - CMD_T_STOP, skipping\n", - cmd->se_tfo->get_task_tag(cmd)); - spin_unlock_irqrestore(&cmd->t_state_lock, flags); - transport_cmd_check_stop(cmd, false, false); - return -EPERM; - } - cmd->transport_state |= CMD_T_LUN_FE_STOP; - spin_unlock_irqrestore(&cmd->t_state_lock, flags); - - // XXX: audit task_flags checks. - spin_lock_irqsave(&cmd->t_state_lock, flags); - if ((cmd->transport_state & CMD_T_BUSY) && - (cmd->transport_state & CMD_T_SENT)) { - if (!target_stop_cmd(cmd, &flags)) - ret++; - } - spin_unlock_irqrestore(&cmd->t_state_lock, flags); - - pr_debug("ConfigFS: cmd: %p stop tasks ret:" - " %d\n", cmd, ret); - if (!ret) { - pr_debug("ConfigFS: ITT[0x%08x] - stopping cmd....\n", - cmd->se_tfo->get_task_tag(cmd)); - wait_for_completion(&cmd->transport_lun_stop_comp); - pr_debug("ConfigFS: ITT[0x%08x] - stopped cmd....\n", - cmd->se_tfo->get_task_tag(cmd)); - } - - return 0; -} - -static void __transport_clear_lun_from_sessions(struct se_lun *lun) -{ - struct se_cmd *cmd = NULL; - unsigned long lun_flags, cmd_flags; - /* - * Do exception processing and return CHECK_CONDITION status to the - * Initiator Port. - */ - spin_lock_irqsave(&lun->lun_cmd_lock, lun_flags); - while (!list_empty(&lun->lun_cmd_list)) { - cmd = list_first_entry(&lun->lun_cmd_list, - struct se_cmd, se_lun_node); - list_del_init(&cmd->se_lun_node); - - spin_lock(&cmd->t_state_lock); - pr_debug("SE_LUN[%d] - Setting cmd->transport" - "_lun_stop for ITT: 0x%08x\n", - cmd->se_lun->unpacked_lun, - cmd->se_tfo->get_task_tag(cmd)); - cmd->transport_state |= CMD_T_LUN_STOP; - spin_unlock(&cmd->t_state_lock); - - spin_unlock_irqrestore(&lun->lun_cmd_lock, lun_flags); - - if (!cmd->se_lun) { - pr_err("ITT: 0x%08x, [i,t]_state: %u/%u\n", - cmd->se_tfo->get_task_tag(cmd), - cmd->se_tfo->get_cmd_state(cmd), cmd->t_state); - BUG(); - } - /* - * If the Storage engine still owns the iscsi_cmd_t, determine - * and/or stop its context. - */ - pr_debug("SE_LUN[%d] - ITT: 0x%08x before transport" - "_lun_wait_for_tasks()\n", cmd->se_lun->unpacked_lun, - cmd->se_tfo->get_task_tag(cmd)); - - if (transport_lun_wait_for_tasks(cmd, cmd->se_lun) < 0) { - spin_lock_irqsave(&lun->lun_cmd_lock, lun_flags); - continue; - } - - pr_debug("SE_LUN[%d] - ITT: 0x%08x after transport_lun" - "_wait_for_tasks(): SUCCESS\n", - cmd->se_lun->unpacked_lun, - cmd->se_tfo->get_task_tag(cmd)); - - spin_lock_irqsave(&cmd->t_state_lock, cmd_flags); - if (!(cmd->transport_state & CMD_T_DEV_ACTIVE)) { - spin_unlock_irqrestore(&cmd->t_state_lock, cmd_flags); - goto check_cond; - } - cmd->transport_state &= ~CMD_T_DEV_ACTIVE; - target_remove_from_state_list(cmd); - spin_unlock_irqrestore(&cmd->t_state_lock, cmd_flags); - - /* - * The Storage engine stopped this struct se_cmd before it was - * send to the fabric frontend for delivery back to the - * Initiator Node. Return this SCSI CDB back with an - * CHECK_CONDITION status. - */ -check_cond: - transport_send_check_condition_and_sense(cmd, - TCM_NON_EXISTENT_LUN, 0); - /* - * If the fabric frontend is waiting for this iscsi_cmd_t to - * be released, notify the waiting thread now that LU has - * finished accessing it. - */ - spin_lock_irqsave(&cmd->t_state_lock, cmd_flags); - if (cmd->transport_state & CMD_T_LUN_FE_STOP) { - pr_debug("SE_LUN[%d] - Detected FE stop for" - " struct se_cmd: %p ITT: 0x%08x\n", - lun->unpacked_lun, - cmd, cmd->se_tfo->get_task_tag(cmd)); - - spin_unlock_irqrestore(&cmd->t_state_lock, - cmd_flags); - transport_cmd_check_stop(cmd, false, false); - complete(&cmd->transport_lun_fe_stop_comp); - spin_lock_irqsave(&lun->lun_cmd_lock, lun_flags); - continue; - } - pr_debug("SE_LUN[%d] - ITT: 0x%08x finished processing\n", - lun->unpacked_lun, cmd->se_tfo->get_task_tag(cmd)); - - spin_unlock_irqrestore(&cmd->t_state_lock, cmd_flags); - spin_lock_irqsave(&lun->lun_cmd_lock, lun_flags); - } - spin_unlock_irqrestore(&lun->lun_cmd_lock, lun_flags); -} - static int transport_clear_lun_ref_thread(void *p) { struct se_lun *lun = p; @@ -2583,43 +2405,6 @@ bool transport_wait_for_tasks(struct se_cmd *cmd) spin_unlock_irqrestore(&cmd->t_state_lock, flags); return false; } - /* - * If we are already stopped due to an external event (ie: LUN shutdown) - * sleep until the connection can have the passed struct se_cmd back. - * The cmd->transport_lun_stopped_sem will be upped by - * transport_clear_lun_from_sessions() once the ConfigFS context caller - * has completed its operation on the struct se_cmd. - */ - if (cmd->transport_state & CMD_T_LUN_STOP) { - pr_debug("wait_for_tasks: Stopping" - " wait_for_completion(&cmd->t_tasktransport_lun_fe" - "_stop_comp); for ITT: 0x%08x\n", - cmd->se_tfo->get_task_tag(cmd)); - /* - * There is a special case for WRITES where a FE exception + - * LUN shutdown means ConfigFS context is still sleeping on - * transport_lun_stop_comp in transport_lun_wait_for_tasks(). - * We go ahead and up transport_lun_stop_comp just to be sure - * here. - */ - spin_unlock_irqrestore(&cmd->t_state_lock, flags); - complete(&cmd->transport_lun_stop_comp); - wait_for_completion(&cmd->transport_lun_fe_stop_comp); - spin_lock_irqsave(&cmd->t_state_lock, flags); - - target_remove_from_state_list(cmd); - /* - * At this point, the frontend who was the originator of this - * struct se_cmd, now owns the structure and can be released through - * normal means below. - */ - pr_debug("wait_for_tasks: Stopped" - " wait_for_completion(&cmd->t_tasktransport_lun_fe_" - "stop_comp); for ITT: 0x%08x\n", - cmd->se_tfo->get_task_tag(cmd)); - - cmd->transport_state &= ~CMD_T_LUN_STOP; - } if (!(cmd->transport_state & CMD_T_ACTIVE)) { spin_unlock_irqrestore(&cmd->t_state_lock, flags); diff --git a/drivers/target/target_core_xcopy.c b/drivers/target/target_core_xcopy.c index 7a870f6add7b..b12fc43b5463 100644 --- a/drivers/target/target_core_xcopy.c +++ b/drivers/target/target_core_xcopy.c @@ -574,10 +574,8 @@ static int target_xcopy_init_pt_lun( return -ENOMEM; } init_completion(&pt_cmd->se_lun->lun_shutdown_comp); - INIT_LIST_HEAD(&pt_cmd->se_lun->lun_cmd_list); INIT_LIST_HEAD(&pt_cmd->se_lun->lun_acl_list); spin_lock_init(&pt_cmd->se_lun->lun_acl_lock); - spin_lock_init(&pt_cmd->se_lun->lun_cmd_lock); spin_lock_init(&pt_cmd->se_lun->lun_sep_lock); init_completion(&pt_cmd->se_lun->lun_ref_comp); diff --git a/include/target/target_core_base.h b/include/target/target_core_base.h index 4d16fd9330a6..1ad0a4240b00 100644 --- a/include/target/target_core_base.h +++ b/include/target/target_core_base.h @@ -443,7 +443,6 @@ struct se_cmd { /* Used for sense data */ void *sense_buffer; struct list_head se_delayed_node; - struct list_head se_lun_node; struct list_head se_qf_node; struct se_device *se_dev; struct se_dev_entry *se_deve; @@ -471,15 +470,11 @@ struct se_cmd { #define CMD_T_SENT (1 << 4) #define CMD_T_STOP (1 << 5) #define CMD_T_FAILED (1 << 6) -#define CMD_T_LUN_STOP (1 << 7) -#define CMD_T_LUN_FE_STOP (1 << 8) -#define CMD_T_DEV_ACTIVE (1 << 9) -#define CMD_T_REQUEST_STOP (1 << 10) -#define CMD_T_BUSY (1 << 11) +#define CMD_T_DEV_ACTIVE (1 << 7) +#define CMD_T_REQUEST_STOP (1 << 8) +#define CMD_T_BUSY (1 << 9) spinlock_t t_state_lock; struct completion t_transport_stop_comp; - struct completion transport_lun_fe_stop_comp; - struct completion transport_lun_stop_comp; struct work_struct work; @@ -751,10 +746,8 @@ struct se_lun { u32 unpacked_lun; atomic_t lun_acl_count; spinlock_t lun_acl_lock; - spinlock_t lun_cmd_lock; spinlock_t lun_sep_lock; struct completion lun_shutdown_comp; - struct list_head lun_cmd_list; struct list_head lun_acl_list; struct se_device *lun_se_dev; struct se_port *lun_sep; -- cgit v1.2.3 From f01b9f73392b48c6cda7c2c66594c73137c776da Mon Sep 17 00:00:00 2001 From: Vu Pham Date: Mon, 11 Nov 2013 19:04:29 +0200 Subject: iser-target: Avoid using FRMR for single dma entry requests This patch changes isert_reg_rdma_frwr() to not use FRMR for single dma entry requests from small I/Os, in order to avoid the associated memory registration overhead. Using DMA MR is sufficient here for the single dma entry requests, and addresses a >= v3.12 performance regression. Signed-off-by: Vu Pham Cc: # v3.12+ Signed-off-by: Nicholas Bellinger --- drivers/infiniband/ulp/isert/ib_isert.c | 32 ++++++++++++++++++++------------ 1 file changed, 20 insertions(+), 12 deletions(-) (limited to 'drivers/infiniband/ulp') diff --git a/drivers/infiniband/ulp/isert/ib_isert.c b/drivers/infiniband/ulp/isert/ib_isert.c index 27708c3430d5..bbd86e82e042 100644 --- a/drivers/infiniband/ulp/isert/ib_isert.c +++ b/drivers/infiniband/ulp/isert/ib_isert.c @@ -2297,18 +2297,26 @@ isert_reg_rdma_frwr(struct iscsi_conn *conn, struct iscsi_cmd *cmd, data_len = min(data_left, rdma_write_max); wr->cur_rdma_length = data_len; - spin_lock_irqsave(&isert_conn->conn_lock, flags); - fr_desc = list_first_entry(&isert_conn->conn_frwr_pool, - struct fast_reg_descriptor, list); - list_del(&fr_desc->list); - spin_unlock_irqrestore(&isert_conn->conn_lock, flags); - wr->fr_desc = fr_desc; - - ret = isert_fast_reg_mr(fr_desc, isert_cmd, isert_conn, - ib_sge, offset, data_len); - if (ret) { - list_add_tail(&fr_desc->list, &isert_conn->conn_frwr_pool); - goto unmap_sg; + /* if there is a single dma entry, dma mr is sufficient */ + if (count == 1) { + ib_sge->addr = ib_sg_dma_address(ib_dev, &sg_start[0]); + ib_sge->length = ib_sg_dma_len(ib_dev, &sg_start[0]); + ib_sge->lkey = isert_conn->conn_mr->lkey; + wr->fr_desc = NULL; + } else { + spin_lock_irqsave(&isert_conn->conn_lock, flags); + fr_desc = list_first_entry(&isert_conn->conn_frwr_pool, + struct fast_reg_descriptor, list); + list_del(&fr_desc->list); + spin_unlock_irqrestore(&isert_conn->conn_lock, flags); + wr->fr_desc = fr_desc; + + ret = isert_fast_reg_mr(fr_desc, isert_cmd, isert_conn, + ib_sge, offset, data_len); + if (ret) { + list_add_tail(&fr_desc->list, &isert_conn->conn_frwr_pool); + goto unmap_sg; + } } return 0; -- cgit v1.2.3 From 04d9cd1224e5bc9d6146bab2866cdc81deb9b509 Mon Sep 17 00:00:00 2001 From: Nicholas Bellinger Date: Tue, 12 Nov 2013 18:05:07 -0800 Subject: ib_isert: Avoid duplicate iscsit_increment_maxcmdsn call This patch avoids a duplicate iscsit_increment_maxcmdsn() call for ISER_IB_RDMA_WRITE within isert_map_rdma() + isert_reg_rdma_frwr(), which will already be occuring once during isert_put_datain() -> iscsit_build_rsp_pdu() operation. It also removes the local conn->stat_sn assignment + increment, and changes the third parameter to iscsit_build_rsp_pdu() to signal this should be done by iscsi_target_mode code. Tested-by: Moussa Ba Cc: # v3.10+ Signed-off-by: Nicholas Bellinger --- drivers/infiniband/ulp/isert/ib_isert.c | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-) (limited to 'drivers/infiniband/ulp') diff --git a/drivers/infiniband/ulp/isert/ib_isert.c b/drivers/infiniband/ulp/isert/ib_isert.c index bbd86e82e042..56610751eb28 100644 --- a/drivers/infiniband/ulp/isert/ib_isert.c +++ b/drivers/infiniband/ulp/isert/ib_isert.c @@ -2029,8 +2029,6 @@ isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd, if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) { data_left = se_cmd->data_length; - iscsit_increment_maxcmdsn(cmd, conn->sess); - cmd->stat_sn = conn->stat_sn++; } else { sg_off = cmd->write_data_done / PAGE_SIZE; data_left = se_cmd->data_length - cmd->write_data_done; @@ -2242,8 +2240,6 @@ isert_reg_rdma_frwr(struct iscsi_conn *conn, struct iscsi_cmd *cmd, if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) { data_left = se_cmd->data_length; - iscsit_increment_maxcmdsn(cmd, conn->sess); - cmd->stat_sn = conn->stat_sn++; } else { sg_off = cmd->write_data_done / PAGE_SIZE; data_left = se_cmd->data_length - cmd->write_data_done; @@ -2352,7 +2348,7 @@ isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd) * Build isert_conn->tx_desc for iSCSI response PDU and attach */ isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc); - iscsit_build_rsp_pdu(cmd, conn, false, (struct iscsi_scsi_rsp *) + iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *) &isert_cmd->tx_desc.iscsi_header); isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc); isert_init_send_wr(isert_conn, isert_cmd, -- cgit v1.2.3