Back to home page

OSCL-LXR
 
 

     

0001 /*
0002  * Copyright (c) 2006, 2019 Oracle and/or its affiliates. All rights reserved.
0003  *
0004  * This software is available to you under a choice of one of two
0005  * licenses.  You may choose to be licensed under the terms of the GNU
0006  * General Public License (GPL) Version 2, available from the file
0007  * COPYING in the main directory of this source tree, or the
0008  * OpenIB.org BSD license below:
0009  *
0010  *     Redistribution and use in source and binary forms, with or
0011  *     without modification, are permitted provided that the following
0012  *     conditions are met:
0013  *
0014  *      - Redistributions of source code must retain the above
0015  *        copyright notice, this list of conditions and the following
0016  *        disclaimer.
0017  *
0018  *      - Redistributions in binary form must reproduce the above
0019  *        copyright notice, this list of conditions and the following
0020  *        disclaimer in the documentation and/or other materials
0021  *        provided with the distribution.
0022  *
0023  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
0024  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
0025  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
0026  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
0027  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
0028  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
0029  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
0030  * SOFTWARE.
0031  *
0032  */
0033 #include <linux/kernel.h>
0034 #include <linux/in.h>
0035 #include <linux/slab.h>
0036 #include <linux/vmalloc.h>
0037 #include <linux/ratelimit.h>
0038 #include <net/addrconf.h>
0039 #include <rdma/ib_cm.h>
0040 
0041 #include "rds_single_path.h"
0042 #include "rds.h"
0043 #include "ib.h"
0044 #include "ib_mr.h"
0045 
0046 /*
0047  * Set the selected protocol version
0048  */
0049 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
0050 {
0051     conn->c_version = version;
0052 }
0053 
0054 /*
0055  * Set up flow control
0056  */
0057 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
0058 {
0059     struct rds_ib_connection *ic = conn->c_transport_data;
0060 
0061     if (rds_ib_sysctl_flow_control && credits != 0) {
0062         /* We're doing flow control */
0063         ic->i_flowctl = 1;
0064         rds_ib_send_add_credits(conn, credits);
0065     } else {
0066         ic->i_flowctl = 0;
0067     }
0068 }
0069 
0070 /*
0071  * Connection established.
0072  * We get here for both outgoing and incoming connection.
0073  */
0074 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
0075 {
0076     struct rds_ib_connection *ic = conn->c_transport_data;
0077     const union rds_ib_conn_priv *dp = NULL;
0078     __be64 ack_seq = 0;
0079     __be32 credit = 0;
0080     u8 major = 0;
0081     u8 minor = 0;
0082     int err;
0083 
0084     dp = event->param.conn.private_data;
0085     if (conn->c_isv6) {
0086         if (event->param.conn.private_data_len >=
0087             sizeof(struct rds6_ib_connect_private)) {
0088             major = dp->ricp_v6.dp_protocol_major;
0089             minor = dp->ricp_v6.dp_protocol_minor;
0090             credit = dp->ricp_v6.dp_credit;
0091             /* dp structure start is not guaranteed to be 8 bytes
0092              * aligned.  Since dp_ack_seq is 64-bit extended load
0093              * operations can be used so go through get_unaligned
0094              * to avoid unaligned errors.
0095              */
0096             ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq);
0097         }
0098     } else if (event->param.conn.private_data_len >=
0099            sizeof(struct rds_ib_connect_private)) {
0100         major = dp->ricp_v4.dp_protocol_major;
0101         minor = dp->ricp_v4.dp_protocol_minor;
0102         credit = dp->ricp_v4.dp_credit;
0103         ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq);
0104     }
0105 
0106     /* make sure it isn't empty data */
0107     if (major) {
0108         rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor));
0109         rds_ib_set_flow_control(conn, be32_to_cpu(credit));
0110     }
0111 
0112     if (conn->c_version < RDS_PROTOCOL_VERSION) {
0113         if (conn->c_version != RDS_PROTOCOL_COMPAT_VERSION) {
0114             pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n",
0115                   &conn->c_laddr, &conn->c_faddr,
0116                   RDS_PROTOCOL_MAJOR(conn->c_version),
0117                   RDS_PROTOCOL_MINOR(conn->c_version));
0118             rds_conn_destroy(conn);
0119             return;
0120         }
0121     }
0122 
0123     pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c,%d> version %u.%u%s\n",
0124           ic->i_active_side ? "Active" : "Passive",
0125           &conn->c_laddr, &conn->c_faddr, conn->c_tos,
0126           RDS_PROTOCOL_MAJOR(conn->c_version),
0127           RDS_PROTOCOL_MINOR(conn->c_version),
0128           ic->i_flowctl ? ", flow control" : "");
0129 
0130     /* receive sl from the peer */
0131     ic->i_sl = ic->i_cm_id->route.path_rec->sl;
0132 
0133     atomic_set(&ic->i_cq_quiesce, 0);
0134 
0135     /* Init rings and fill recv. this needs to wait until protocol
0136      * negotiation is complete, since ring layout is different
0137      * from 3.1 to 4.1.
0138      */
0139     rds_ib_send_init_ring(ic);
0140     rds_ib_recv_init_ring(ic);
0141     /* Post receive buffers - as a side effect, this will update
0142      * the posted credit count. */
0143     rds_ib_recv_refill(conn, 1, GFP_KERNEL);
0144 
0145     /* update ib_device with this local ipaddr */
0146     err = rds_ib_update_ipaddr(ic->rds_ibdev, &conn->c_laddr);
0147     if (err)
0148         printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
0149             err);
0150 
0151     /* If the peer gave us the last packet it saw, process this as if
0152      * we had received a regular ACK. */
0153     if (dp) {
0154         if (ack_seq)
0155             rds_send_drop_acked(conn, be64_to_cpu(ack_seq),
0156                         NULL);
0157     }
0158 
0159     conn->c_proposed_version = conn->c_version;
0160     rds_connect_complete(conn);
0161 }
0162 
0163 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
0164                       struct rdma_conn_param *conn_param,
0165                       union rds_ib_conn_priv *dp,
0166                       u32 protocol_version,
0167                       u32 max_responder_resources,
0168                       u32 max_initiator_depth,
0169                       bool isv6)
0170 {
0171     struct rds_ib_connection *ic = conn->c_transport_data;
0172     struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
0173 
0174     memset(conn_param, 0, sizeof(struct rdma_conn_param));
0175 
0176     conn_param->responder_resources =
0177         min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
0178     conn_param->initiator_depth =
0179         min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
0180     conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
0181     conn_param->rnr_retry_count = 7;
0182 
0183     if (dp) {
0184         memset(dp, 0, sizeof(*dp));
0185         if (isv6) {
0186             dp->ricp_v6.dp_saddr = conn->c_laddr;
0187             dp->ricp_v6.dp_daddr = conn->c_faddr;
0188             dp->ricp_v6.dp_protocol_major =
0189                 RDS_PROTOCOL_MAJOR(protocol_version);
0190             dp->ricp_v6.dp_protocol_minor =
0191                 RDS_PROTOCOL_MINOR(protocol_version);
0192             dp->ricp_v6.dp_protocol_minor_mask =
0193                 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
0194             dp->ricp_v6.dp_ack_seq =
0195                 cpu_to_be64(rds_ib_piggyb_ack(ic));
0196             dp->ricp_v6.dp_cmn.ricpc_dp_toss = conn->c_tos;
0197 
0198             conn_param->private_data = &dp->ricp_v6;
0199             conn_param->private_data_len = sizeof(dp->ricp_v6);
0200         } else {
0201             dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3];
0202             dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3];
0203             dp->ricp_v4.dp_protocol_major =
0204                 RDS_PROTOCOL_MAJOR(protocol_version);
0205             dp->ricp_v4.dp_protocol_minor =
0206                 RDS_PROTOCOL_MINOR(protocol_version);
0207             dp->ricp_v4.dp_protocol_minor_mask =
0208                 cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
0209             dp->ricp_v4.dp_ack_seq =
0210                 cpu_to_be64(rds_ib_piggyb_ack(ic));
0211             dp->ricp_v4.dp_cmn.ricpc_dp_toss = conn->c_tos;
0212 
0213             conn_param->private_data = &dp->ricp_v4;
0214             conn_param->private_data_len = sizeof(dp->ricp_v4);
0215         }
0216 
0217         /* Advertise flow control */
0218         if (ic->i_flowctl) {
0219             unsigned int credits;
0220 
0221             credits = IB_GET_POST_CREDITS
0222                 (atomic_read(&ic->i_credits));
0223             if (isv6)
0224                 dp->ricp_v6.dp_credit = cpu_to_be32(credits);
0225             else
0226                 dp->ricp_v4.dp_credit = cpu_to_be32(credits);
0227             atomic_sub(IB_SET_POST_CREDITS(credits),
0228                    &ic->i_credits);
0229         }
0230     }
0231 }
0232 
0233 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
0234 {
0235     rdsdebug("event %u (%s) data %p\n",
0236          event->event, ib_event_msg(event->event), data);
0237 }
0238 
0239 /* Plucking the oldest entry from the ring can be done concurrently with
0240  * the thread refilling the ring.  Each ring operation is protected by
0241  * spinlocks and the transient state of refilling doesn't change the
0242  * recording of which entry is oldest.
0243  *
0244  * This relies on IB only calling one cq comp_handler for each cq so that
0245  * there will only be one caller of rds_recv_incoming() per RDS connection.
0246  */
0247 static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
0248 {
0249     struct rds_connection *conn = context;
0250     struct rds_ib_connection *ic = conn->c_transport_data;
0251 
0252     rdsdebug("conn %p cq %p\n", conn, cq);
0253 
0254     rds_ib_stats_inc(s_ib_evt_handler_call);
0255 
0256     tasklet_schedule(&ic->i_recv_tasklet);
0257 }
0258 
0259 static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
0260              struct ib_wc *wcs)
0261 {
0262     int nr, i;
0263     struct ib_wc *wc;
0264 
0265     while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
0266         for (i = 0; i < nr; i++) {
0267             wc = wcs + i;
0268             rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
0269                  (unsigned long long)wc->wr_id, wc->status,
0270                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
0271 
0272             if (wc->wr_id <= ic->i_send_ring.w_nr ||
0273                 wc->wr_id == RDS_IB_ACK_WR_ID)
0274                 rds_ib_send_cqe_handler(ic, wc);
0275             else
0276                 rds_ib_mr_cqe_handler(ic, wc);
0277 
0278         }
0279     }
0280 }
0281 
0282 static void rds_ib_tasklet_fn_send(unsigned long data)
0283 {
0284     struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
0285     struct rds_connection *conn = ic->conn;
0286 
0287     rds_ib_stats_inc(s_ib_tasklet_call);
0288 
0289     /* if cq has been already reaped, ignore incoming cq event */
0290     if (atomic_read(&ic->i_cq_quiesce))
0291         return;
0292 
0293     poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
0294     ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
0295     poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
0296 
0297     if (rds_conn_up(conn) &&
0298         (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
0299         test_bit(0, &conn->c_map_queued)))
0300         rds_send_xmit(&ic->conn->c_path[0]);
0301 }
0302 
0303 static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
0304              struct ib_wc *wcs,
0305              struct rds_ib_ack_state *ack_state)
0306 {
0307     int nr, i;
0308     struct ib_wc *wc;
0309 
0310     while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
0311         for (i = 0; i < nr; i++) {
0312             wc = wcs + i;
0313             rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
0314                  (unsigned long long)wc->wr_id, wc->status,
0315                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
0316 
0317             rds_ib_recv_cqe_handler(ic, wc, ack_state);
0318         }
0319     }
0320 }
0321 
0322 static void rds_ib_tasklet_fn_recv(unsigned long data)
0323 {
0324     struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
0325     struct rds_connection *conn = ic->conn;
0326     struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
0327     struct rds_ib_ack_state state;
0328 
0329     if (!rds_ibdev)
0330         rds_conn_drop(conn);
0331 
0332     rds_ib_stats_inc(s_ib_tasklet_call);
0333 
0334     /* if cq has been already reaped, ignore incoming cq event */
0335     if (atomic_read(&ic->i_cq_quiesce))
0336         return;
0337 
0338     memset(&state, 0, sizeof(state));
0339     poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
0340     ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
0341     poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
0342 
0343     if (state.ack_next_valid)
0344         rds_ib_set_ack(ic, state.ack_next, state.ack_required);
0345     if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
0346         rds_send_drop_acked(conn, state.ack_recv, NULL);
0347         ic->i_ack_recv = state.ack_recv;
0348     }
0349 
0350     if (rds_conn_up(conn))
0351         rds_ib_attempt_ack(ic);
0352 }
0353 
0354 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
0355 {
0356     struct rds_connection *conn = data;
0357     struct rds_ib_connection *ic = conn->c_transport_data;
0358 
0359     rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
0360          ib_event_msg(event->event));
0361 
0362     switch (event->event) {
0363     case IB_EVENT_COMM_EST:
0364         rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
0365         break;
0366     default:
0367         rdsdebug("Fatal QP Event %u (%s) - connection %pI6c->%pI6c, reconnecting\n",
0368              event->event, ib_event_msg(event->event),
0369              &conn->c_laddr, &conn->c_faddr);
0370         rds_conn_drop(conn);
0371         break;
0372     }
0373 }
0374 
0375 static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
0376 {
0377     struct rds_connection *conn = context;
0378     struct rds_ib_connection *ic = conn->c_transport_data;
0379 
0380     rdsdebug("conn %p cq %p\n", conn, cq);
0381 
0382     rds_ib_stats_inc(s_ib_evt_handler_call);
0383 
0384     tasklet_schedule(&ic->i_send_tasklet);
0385 }
0386 
0387 static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
0388 {
0389     int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
0390     int index = rds_ibdev->dev->num_comp_vectors - 1;
0391     int i;
0392 
0393     for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
0394         if (rds_ibdev->vector_load[i] < min) {
0395             index = i;
0396             min = rds_ibdev->vector_load[i];
0397         }
0398     }
0399 
0400     rds_ibdev->vector_load[index]++;
0401     return index;
0402 }
0403 
0404 static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
0405 {
0406     rds_ibdev->vector_load[index]--;
0407 }
0408 
0409 static void rds_dma_hdr_free(struct ib_device *dev, struct rds_header *hdr,
0410         dma_addr_t dma_addr, enum dma_data_direction dir)
0411 {
0412     ib_dma_unmap_single(dev, dma_addr, sizeof(*hdr), dir);
0413     kfree(hdr);
0414 }
0415 
0416 static struct rds_header *rds_dma_hdr_alloc(struct ib_device *dev,
0417         dma_addr_t *dma_addr, enum dma_data_direction dir)
0418 {
0419     struct rds_header *hdr;
0420 
0421     hdr = kzalloc_node(sizeof(*hdr), GFP_KERNEL, ibdev_to_node(dev));
0422     if (!hdr)
0423         return NULL;
0424 
0425     *dma_addr = ib_dma_map_single(dev, hdr, sizeof(*hdr),
0426                       DMA_BIDIRECTIONAL);
0427     if (ib_dma_mapping_error(dev, *dma_addr)) {
0428         kfree(hdr);
0429         return NULL;
0430     }
0431 
0432     return hdr;
0433 }
0434 
0435 /* Free the DMA memory used to store struct rds_header.
0436  *
0437  * @dev: the RDS IB device
0438  * @hdrs: pointer to the array storing DMA memory pointers
0439  * @dma_addrs: pointer to the array storing DMA addresses
0440  * @num_hdars: number of headers to free.
0441  */
0442 static void rds_dma_hdrs_free(struct rds_ib_device *dev,
0443         struct rds_header **hdrs, dma_addr_t *dma_addrs, u32 num_hdrs,
0444         enum dma_data_direction dir)
0445 {
0446     u32 i;
0447 
0448     for (i = 0; i < num_hdrs; i++)
0449         rds_dma_hdr_free(dev->dev, hdrs[i], dma_addrs[i], dir);
0450     kvfree(hdrs);
0451     kvfree(dma_addrs);
0452 }
0453 
0454 
0455 /* Allocate DMA coherent memory to be used to store struct rds_header for
0456  * sending/receiving packets.  The pointers to the DMA memory and the
0457  * associated DMA addresses are stored in two arrays.
0458  *
0459  * @dev: the RDS IB device
0460  * @dma_addrs: pointer to the array for storing DMA addresses
0461  * @num_hdrs: number of headers to allocate
0462  *
0463  * It returns the pointer to the array storing the DMA memory pointers.  On
0464  * error, NULL pointer is returned.
0465  */
0466 static struct rds_header **rds_dma_hdrs_alloc(struct rds_ib_device *dev,
0467         dma_addr_t **dma_addrs, u32 num_hdrs,
0468         enum dma_data_direction dir)
0469 {
0470     struct rds_header **hdrs;
0471     dma_addr_t *hdr_daddrs;
0472     u32 i;
0473 
0474     hdrs = kvmalloc_node(sizeof(*hdrs) * num_hdrs, GFP_KERNEL,
0475                  ibdev_to_node(dev->dev));
0476     if (!hdrs)
0477         return NULL;
0478 
0479     hdr_daddrs = kvmalloc_node(sizeof(*hdr_daddrs) * num_hdrs, GFP_KERNEL,
0480                    ibdev_to_node(dev->dev));
0481     if (!hdr_daddrs) {
0482         kvfree(hdrs);
0483         return NULL;
0484     }
0485 
0486     for (i = 0; i < num_hdrs; i++) {
0487         hdrs[i] = rds_dma_hdr_alloc(dev->dev, &hdr_daddrs[i], dir);
0488         if (!hdrs[i]) {
0489             rds_dma_hdrs_free(dev, hdrs, hdr_daddrs, i, dir);
0490             return NULL;
0491         }
0492     }
0493 
0494     *dma_addrs = hdr_daddrs;
0495     return hdrs;
0496 }
0497 
0498 /*
0499  * This needs to be very careful to not leave IS_ERR pointers around for
0500  * cleanup to trip over.
0501  */
0502 static int rds_ib_setup_qp(struct rds_connection *conn)
0503 {
0504     struct rds_ib_connection *ic = conn->c_transport_data;
0505     struct ib_device *dev = ic->i_cm_id->device;
0506     struct ib_qp_init_attr attr;
0507     struct ib_cq_init_attr cq_attr = {};
0508     struct rds_ib_device *rds_ibdev;
0509     unsigned long max_wrs;
0510     int ret, fr_queue_space;
0511 
0512     /*
0513      * It's normal to see a null device if an incoming connection races
0514      * with device removal, so we don't print a warning.
0515      */
0516     rds_ibdev = rds_ib_get_client_data(dev);
0517     if (!rds_ibdev)
0518         return -EOPNOTSUPP;
0519 
0520     /* The fr_queue_space is currently set to 512, to add extra space on
0521      * completion queue and send queue. This extra space is used for FRWR
0522      * registration and invalidation work requests
0523      */
0524     fr_queue_space = RDS_IB_DEFAULT_FR_WR;
0525 
0526     /* add the conn now so that connection establishment has the dev */
0527     rds_ib_add_conn(rds_ibdev, conn);
0528 
0529     max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_send_wr + 1 ?
0530         rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_send_wr;
0531     if (ic->i_send_ring.w_nr != max_wrs)
0532         rds_ib_ring_resize(&ic->i_send_ring, max_wrs);
0533 
0534     max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_recv_wr + 1 ?
0535         rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_recv_wr;
0536     if (ic->i_recv_ring.w_nr != max_wrs)
0537         rds_ib_ring_resize(&ic->i_recv_ring, max_wrs);
0538 
0539     /* Protection domain and memory range */
0540     ic->i_pd = rds_ibdev->pd;
0541 
0542     ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
0543     cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
0544     cq_attr.comp_vector = ic->i_scq_vector;
0545     ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
0546                      rds_ib_cq_event_handler, conn,
0547                      &cq_attr);
0548     if (IS_ERR(ic->i_send_cq)) {
0549         ret = PTR_ERR(ic->i_send_cq);
0550         ic->i_send_cq = NULL;
0551         ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
0552         rdsdebug("ib_create_cq send failed: %d\n", ret);
0553         goto rds_ibdev_out;
0554     }
0555 
0556     ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
0557     cq_attr.cqe = ic->i_recv_ring.w_nr;
0558     cq_attr.comp_vector = ic->i_rcq_vector;
0559     ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
0560                      rds_ib_cq_event_handler, conn,
0561                      &cq_attr);
0562     if (IS_ERR(ic->i_recv_cq)) {
0563         ret = PTR_ERR(ic->i_recv_cq);
0564         ic->i_recv_cq = NULL;
0565         ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
0566         rdsdebug("ib_create_cq recv failed: %d\n", ret);
0567         goto send_cq_out;
0568     }
0569 
0570     ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
0571     if (ret) {
0572         rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
0573         goto recv_cq_out;
0574     }
0575 
0576     ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
0577     if (ret) {
0578         rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
0579         goto recv_cq_out;
0580     }
0581 
0582     /* XXX negotiate max send/recv with remote? */
0583     memset(&attr, 0, sizeof(attr));
0584     attr.event_handler = rds_ib_qp_event_handler;
0585     attr.qp_context = conn;
0586     /* + 1 to allow for the single ack message */
0587     attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
0588     attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
0589     attr.cap.max_send_sge = rds_ibdev->max_sge;
0590     attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
0591     attr.sq_sig_type = IB_SIGNAL_REQ_WR;
0592     attr.qp_type = IB_QPT_RC;
0593     attr.send_cq = ic->i_send_cq;
0594     attr.recv_cq = ic->i_recv_cq;
0595 
0596     /*
0597      * XXX this can fail if max_*_wr is too large?  Are we supposed
0598      * to back off until we get a value that the hardware can support?
0599      */
0600     ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
0601     if (ret) {
0602         rdsdebug("rdma_create_qp failed: %d\n", ret);
0603         goto recv_cq_out;
0604     }
0605 
0606     ic->i_send_hdrs = rds_dma_hdrs_alloc(rds_ibdev, &ic->i_send_hdrs_dma,
0607                          ic->i_send_ring.w_nr,
0608                          DMA_TO_DEVICE);
0609     if (!ic->i_send_hdrs) {
0610         ret = -ENOMEM;
0611         rdsdebug("DMA send hdrs alloc failed\n");
0612         goto qp_out;
0613     }
0614 
0615     ic->i_recv_hdrs = rds_dma_hdrs_alloc(rds_ibdev, &ic->i_recv_hdrs_dma,
0616                          ic->i_recv_ring.w_nr,
0617                          DMA_FROM_DEVICE);
0618     if (!ic->i_recv_hdrs) {
0619         ret = -ENOMEM;
0620         rdsdebug("DMA recv hdrs alloc failed\n");
0621         goto send_hdrs_dma_out;
0622     }
0623 
0624     ic->i_ack = rds_dma_hdr_alloc(rds_ibdev->dev, &ic->i_ack_dma,
0625                       DMA_TO_DEVICE);
0626     if (!ic->i_ack) {
0627         ret = -ENOMEM;
0628         rdsdebug("DMA ack header alloc failed\n");
0629         goto recv_hdrs_dma_out;
0630     }
0631 
0632     ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work),
0633                           ic->i_send_ring.w_nr),
0634                    ibdev_to_node(dev));
0635     if (!ic->i_sends) {
0636         ret = -ENOMEM;
0637         rdsdebug("send allocation failed\n");
0638         goto ack_dma_out;
0639     }
0640 
0641     ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work),
0642                           ic->i_recv_ring.w_nr),
0643                    ibdev_to_node(dev));
0644     if (!ic->i_recvs) {
0645         ret = -ENOMEM;
0646         rdsdebug("recv allocation failed\n");
0647         goto sends_out;
0648     }
0649 
0650     rds_ib_recv_init_ack(ic);
0651 
0652     rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
0653          ic->i_send_cq, ic->i_recv_cq);
0654 
0655     goto out;
0656 
0657 sends_out:
0658     vfree(ic->i_sends);
0659 
0660 ack_dma_out:
0661     rds_dma_hdr_free(rds_ibdev->dev, ic->i_ack, ic->i_ack_dma,
0662              DMA_TO_DEVICE);
0663     ic->i_ack = NULL;
0664 
0665 recv_hdrs_dma_out:
0666     rds_dma_hdrs_free(rds_ibdev, ic->i_recv_hdrs, ic->i_recv_hdrs_dma,
0667               ic->i_recv_ring.w_nr, DMA_FROM_DEVICE);
0668     ic->i_recv_hdrs = NULL;
0669     ic->i_recv_hdrs_dma = NULL;
0670 
0671 send_hdrs_dma_out:
0672     rds_dma_hdrs_free(rds_ibdev, ic->i_send_hdrs, ic->i_send_hdrs_dma,
0673               ic->i_send_ring.w_nr, DMA_TO_DEVICE);
0674     ic->i_send_hdrs = NULL;
0675     ic->i_send_hdrs_dma = NULL;
0676 
0677 qp_out:
0678     rdma_destroy_qp(ic->i_cm_id);
0679 recv_cq_out:
0680     ib_destroy_cq(ic->i_recv_cq);
0681     ic->i_recv_cq = NULL;
0682 send_cq_out:
0683     ib_destroy_cq(ic->i_send_cq);
0684     ic->i_send_cq = NULL;
0685 rds_ibdev_out:
0686     rds_ib_remove_conn(rds_ibdev, conn);
0687 out:
0688     rds_ib_dev_put(rds_ibdev);
0689 
0690     return ret;
0691 }
0692 
0693 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6)
0694 {
0695     const union rds_ib_conn_priv *dp = event->param.conn.private_data;
0696     u8 data_len, major, minor;
0697     u32 version = 0;
0698     __be16 mask;
0699     u16 common;
0700 
0701     /*
0702      * rdma_cm private data is odd - when there is any private data in the
0703      * request, we will be given a pretty large buffer without telling us the
0704      * original size. The only way to tell the difference is by looking at
0705      * the contents, which are initialized to zero.
0706      * If the protocol version fields aren't set, this is a connection attempt
0707      * from an older version. This could be 3.0 or 2.0 - we can't tell.
0708      * We really should have changed this for OFED 1.3 :-(
0709      */
0710 
0711     /* Be paranoid. RDS always has privdata */
0712     if (!event->param.conn.private_data_len) {
0713         printk(KERN_NOTICE "RDS incoming connection has no private data, "
0714             "rejecting\n");
0715         return 0;
0716     }
0717 
0718     if (isv6) {
0719         data_len = sizeof(struct rds6_ib_connect_private);
0720         major = dp->ricp_v6.dp_protocol_major;
0721         minor = dp->ricp_v6.dp_protocol_minor;
0722         mask = dp->ricp_v6.dp_protocol_minor_mask;
0723     } else {
0724         data_len = sizeof(struct rds_ib_connect_private);
0725         major = dp->ricp_v4.dp_protocol_major;
0726         minor = dp->ricp_v4.dp_protocol_minor;
0727         mask = dp->ricp_v4.dp_protocol_minor_mask;
0728     }
0729 
0730     /* Even if len is crap *now* I still want to check it. -ASG */
0731     if (event->param.conn.private_data_len < data_len || major == 0)
0732         return RDS_PROTOCOL_4_0;
0733 
0734     common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS;
0735     if (major == 4 && common) {
0736         version = RDS_PROTOCOL_4_0;
0737         while ((common >>= 1) != 0)
0738             version++;
0739     } else if (RDS_PROTOCOL_COMPAT_VERSION ==
0740            RDS_PROTOCOL(major, minor)) {
0741         version = RDS_PROTOCOL_COMPAT_VERSION;
0742     } else {
0743         if (isv6)
0744             printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n",
0745                        &dp->ricp_v6.dp_saddr, major, minor);
0746         else
0747             printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
0748                        &dp->ricp_v4.dp_saddr, major, minor);
0749     }
0750     return version;
0751 }
0752 
0753 #if IS_ENABLED(CONFIG_IPV6)
0754 /* Given an IPv6 address, find the net_device which hosts that address and
0755  * return its index.  This is used by the rds_ib_cm_handle_connect() code to
0756  * find the interface index of where an incoming request comes from when
0757  * the request is using a link local address.
0758  *
0759  * Note one problem in this search.  It is possible that two interfaces have
0760  * the same link local address.  Unfortunately, this cannot be solved unless
0761  * the underlying layer gives us the interface which an incoming RDMA connect
0762  * request comes from.
0763  */
0764 static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr)
0765 {
0766     struct net_device *dev;
0767     int idx = 0;
0768 
0769     rcu_read_lock();
0770     for_each_netdev_rcu(net, dev) {
0771         if (ipv6_chk_addr(net, addr, dev, 1)) {
0772             idx = dev->ifindex;
0773             break;
0774         }
0775     }
0776     rcu_read_unlock();
0777 
0778     return idx;
0779 }
0780 #endif
0781 
0782 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
0783                  struct rdma_cm_event *event, bool isv6)
0784 {
0785     __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
0786     __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
0787     const struct rds_ib_conn_priv_cmn *dp_cmn;
0788     struct rds_connection *conn = NULL;
0789     struct rds_ib_connection *ic = NULL;
0790     struct rdma_conn_param conn_param;
0791     const union rds_ib_conn_priv *dp;
0792     union rds_ib_conn_priv dp_rep;
0793     struct in6_addr s_mapped_addr;
0794     struct in6_addr d_mapped_addr;
0795     const struct in6_addr *saddr6;
0796     const struct in6_addr *daddr6;
0797     int destroy = 1;
0798     u32 ifindex = 0;
0799     u32 version;
0800     int err = 1;
0801 
0802     /* Check whether the remote protocol version matches ours. */
0803     version = rds_ib_protocol_compatible(event, isv6);
0804     if (!version) {
0805         err = RDS_RDMA_REJ_INCOMPAT;
0806         goto out;
0807     }
0808 
0809     dp = event->param.conn.private_data;
0810     if (isv6) {
0811 #if IS_ENABLED(CONFIG_IPV6)
0812         dp_cmn = &dp->ricp_v6.dp_cmn;
0813         saddr6 = &dp->ricp_v6.dp_saddr;
0814         daddr6 = &dp->ricp_v6.dp_daddr;
0815         /* If either address is link local, need to find the
0816          * interface index in order to create a proper RDS
0817          * connection.
0818          */
0819         if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) {
0820             /* Using init_net for now ..  */
0821             ifindex = __rds_find_ifindex(&init_net, daddr6);
0822             /* No index found...  Need to bail out. */
0823             if (ifindex == 0) {
0824                 err = -EOPNOTSUPP;
0825                 goto out;
0826             }
0827         } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) {
0828             /* Use our address to find the correct index. */
0829             ifindex = __rds_find_ifindex(&init_net, daddr6);
0830             /* No index found...  Need to bail out. */
0831             if (ifindex == 0) {
0832                 err = -EOPNOTSUPP;
0833                 goto out;
0834             }
0835         }
0836 #else
0837         err = -EOPNOTSUPP;
0838         goto out;
0839 #endif
0840     } else {
0841         dp_cmn = &dp->ricp_v4.dp_cmn;
0842         ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr);
0843         ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr);
0844         saddr6 = &s_mapped_addr;
0845         daddr6 = &d_mapped_addr;
0846     }
0847 
0848     rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid 0x%llx, tos:%d\n",
0849          saddr6, daddr6, RDS_PROTOCOL_MAJOR(version),
0850          RDS_PROTOCOL_MINOR(version),
0851          (unsigned long long)be64_to_cpu(lguid),
0852          (unsigned long long)be64_to_cpu(fguid), dp_cmn->ricpc_dp_toss);
0853 
0854     /* RDS/IB is not currently netns aware, thus init_net */
0855     conn = rds_conn_create(&init_net, daddr6, saddr6,
0856                    &rds_ib_transport, dp_cmn->ricpc_dp_toss,
0857                    GFP_KERNEL, ifindex);
0858     if (IS_ERR(conn)) {
0859         rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
0860         conn = NULL;
0861         goto out;
0862     }
0863 
0864     /*
0865      * The connection request may occur while the
0866      * previous connection exist, e.g. in case of failover.
0867      * But as connections may be initiated simultaneously
0868      * by both hosts, we have a random backoff mechanism -
0869      * see the comment above rds_queue_reconnect()
0870      */
0871     mutex_lock(&conn->c_cm_lock);
0872     if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
0873         if (rds_conn_state(conn) == RDS_CONN_UP) {
0874             rdsdebug("incoming connect while connecting\n");
0875             rds_conn_drop(conn);
0876             rds_ib_stats_inc(s_ib_listen_closed_stale);
0877         } else
0878         if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
0879             /* Wait and see - our connect may still be succeeding */
0880             rds_ib_stats_inc(s_ib_connect_raced);
0881         }
0882         goto out;
0883     }
0884 
0885     ic = conn->c_transport_data;
0886 
0887     rds_ib_set_protocol(conn, version);
0888     rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit));
0889 
0890     /* If the peer gave us the last packet it saw, process this as if
0891      * we had received a regular ACK. */
0892     if (dp_cmn->ricpc_ack_seq)
0893         rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq),
0894                     NULL);
0895 
0896     BUG_ON(cm_id->context);
0897     BUG_ON(ic->i_cm_id);
0898 
0899     ic->i_cm_id = cm_id;
0900     cm_id->context = conn;
0901 
0902     /* We got halfway through setting up the ib_connection, if we
0903      * fail now, we have to take the long route out of this mess. */
0904     destroy = 0;
0905 
0906     err = rds_ib_setup_qp(conn);
0907     if (err) {
0908         rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
0909         goto out;
0910     }
0911 
0912     rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
0913                   event->param.conn.responder_resources,
0914                   event->param.conn.initiator_depth, isv6);
0915 
0916     rdma_set_min_rnr_timer(cm_id, IB_RNR_TIMER_000_32);
0917     /* rdma_accept() calls rdma_reject() internally if it fails */
0918     if (rdma_accept(cm_id, &conn_param))
0919         rds_ib_conn_error(conn, "rdma_accept failed\n");
0920 
0921 out:
0922     if (conn)
0923         mutex_unlock(&conn->c_cm_lock);
0924     if (err)
0925         rdma_reject(cm_id, &err, sizeof(int),
0926                 IB_CM_REJ_CONSUMER_DEFINED);
0927     return destroy;
0928 }
0929 
0930 
0931 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
0932 {
0933     struct rds_connection *conn = cm_id->context;
0934     struct rds_ib_connection *ic = conn->c_transport_data;
0935     struct rdma_conn_param conn_param;
0936     union rds_ib_conn_priv dp;
0937     int ret;
0938 
0939     /* If the peer doesn't do protocol negotiation, we must
0940      * default to RDSv3.0 */
0941     rds_ib_set_protocol(conn, RDS_PROTOCOL_4_1);
0942     ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
0943 
0944     ret = rds_ib_setup_qp(conn);
0945     if (ret) {
0946         rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
0947         goto out;
0948     }
0949 
0950     rds_ib_cm_fill_conn_param(conn, &conn_param, &dp,
0951                   conn->c_proposed_version,
0952                   UINT_MAX, UINT_MAX, isv6);
0953     ret = rdma_connect_locked(cm_id, &conn_param);
0954     if (ret)
0955         rds_ib_conn_error(conn, "rdma_connect_locked failed (%d)\n",
0956                   ret);
0957 
0958 out:
0959     /* Beware - returning non-zero tells the rdma_cm to destroy
0960      * the cm_id. We should certainly not do it as long as we still
0961      * "own" the cm_id. */
0962     if (ret) {
0963         if (ic->i_cm_id == cm_id)
0964             ret = 0;
0965     }
0966     ic->i_active_side = true;
0967     return ret;
0968 }
0969 
0970 int rds_ib_conn_path_connect(struct rds_conn_path *cp)
0971 {
0972     struct rds_connection *conn = cp->cp_conn;
0973     struct sockaddr_storage src, dest;
0974     rdma_cm_event_handler handler;
0975     struct rds_ib_connection *ic;
0976     int ret;
0977 
0978     ic = conn->c_transport_data;
0979 
0980     /* XXX I wonder what affect the port space has */
0981     /* delegate cm event handler to rdma_transport */
0982 #if IS_ENABLED(CONFIG_IPV6)
0983     if (conn->c_isv6)
0984         handler = rds6_rdma_cm_event_handler;
0985     else
0986 #endif
0987         handler = rds_rdma_cm_event_handler;
0988     ic->i_cm_id = rdma_create_id(&init_net, handler, conn,
0989                      RDMA_PS_TCP, IB_QPT_RC);
0990     if (IS_ERR(ic->i_cm_id)) {
0991         ret = PTR_ERR(ic->i_cm_id);
0992         ic->i_cm_id = NULL;
0993         rdsdebug("rdma_create_id() failed: %d\n", ret);
0994         goto out;
0995     }
0996 
0997     rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
0998 
0999     if (ipv6_addr_v4mapped(&conn->c_faddr)) {
1000         struct sockaddr_in *sin;
1001 
1002         sin = (struct sockaddr_in *)&src;
1003         sin->sin_family = AF_INET;
1004         sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
1005         sin->sin_port = 0;
1006 
1007         sin = (struct sockaddr_in *)&dest;
1008         sin->sin_family = AF_INET;
1009         sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
1010         sin->sin_port = htons(RDS_PORT);
1011     } else {
1012         struct sockaddr_in6 *sin6;
1013 
1014         sin6 = (struct sockaddr_in6 *)&src;
1015         sin6->sin6_family = AF_INET6;
1016         sin6->sin6_addr = conn->c_laddr;
1017         sin6->sin6_port = 0;
1018         sin6->sin6_scope_id = conn->c_dev_if;
1019 
1020         sin6 = (struct sockaddr_in6 *)&dest;
1021         sin6->sin6_family = AF_INET6;
1022         sin6->sin6_addr = conn->c_faddr;
1023         sin6->sin6_port = htons(RDS_CM_PORT);
1024         sin6->sin6_scope_id = conn->c_dev_if;
1025     }
1026 
1027     ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
1028                 (struct sockaddr *)&dest,
1029                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
1030     if (ret) {
1031         rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
1032              ret);
1033         rdma_destroy_id(ic->i_cm_id);
1034         ic->i_cm_id = NULL;
1035     }
1036 
1037 out:
1038     return ret;
1039 }
1040 
1041 /*
1042  * This is so careful about only cleaning up resources that were built up
1043  * so that it can be called at any point during startup.  In fact it
1044  * can be called multiple times for a given connection.
1045  */
1046 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
1047 {
1048     struct rds_connection *conn = cp->cp_conn;
1049     struct rds_ib_connection *ic = conn->c_transport_data;
1050     int err = 0;
1051 
1052     rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
1053          ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
1054          ic->i_cm_id ? ic->i_cm_id->qp : NULL);
1055 
1056     if (ic->i_cm_id) {
1057         rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
1058         err = rdma_disconnect(ic->i_cm_id);
1059         if (err) {
1060             /* Actually this may happen quite frequently, when
1061              * an outgoing connect raced with an incoming connect.
1062              */
1063             rdsdebug("failed to disconnect, cm: %p err %d\n",
1064                 ic->i_cm_id, err);
1065         }
1066 
1067         /* kick off "flush_worker" for all pools in order to reap
1068          * all FRMR registrations that are still marked "FRMR_IS_INUSE"
1069          */
1070         rds_ib_flush_mrs();
1071 
1072         /*
1073          * We want to wait for tx and rx completion to finish
1074          * before we tear down the connection, but we have to be
1075          * careful not to get stuck waiting on a send ring that
1076          * only has unsignaled sends in it.  We've shutdown new
1077          * sends before getting here so by waiting for signaled
1078          * sends to complete we're ensured that there will be no
1079          * more tx processing.
1080          */
1081         wait_event(rds_ib_ring_empty_wait,
1082                rds_ib_ring_empty(&ic->i_recv_ring) &&
1083                (atomic_read(&ic->i_signaled_sends) == 0) &&
1084                (atomic_read(&ic->i_fastreg_inuse_count) == 0) &&
1085                (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
1086         tasklet_kill(&ic->i_send_tasklet);
1087         tasklet_kill(&ic->i_recv_tasklet);
1088 
1089         atomic_set(&ic->i_cq_quiesce, 1);
1090 
1091         /* first destroy the ib state that generates callbacks */
1092         if (ic->i_cm_id->qp)
1093             rdma_destroy_qp(ic->i_cm_id);
1094         if (ic->i_send_cq) {
1095             if (ic->rds_ibdev)
1096                 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
1097             ib_destroy_cq(ic->i_send_cq);
1098         }
1099 
1100         if (ic->i_recv_cq) {
1101             if (ic->rds_ibdev)
1102                 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
1103             ib_destroy_cq(ic->i_recv_cq);
1104         }
1105 
1106         if (ic->rds_ibdev) {
1107             /* then free the resources that ib callbacks use */
1108             if (ic->i_send_hdrs) {
1109                 rds_dma_hdrs_free(ic->rds_ibdev,
1110                           ic->i_send_hdrs,
1111                           ic->i_send_hdrs_dma,
1112                           ic->i_send_ring.w_nr,
1113                           DMA_TO_DEVICE);
1114                 ic->i_send_hdrs = NULL;
1115                 ic->i_send_hdrs_dma = NULL;
1116             }
1117 
1118             if (ic->i_recv_hdrs) {
1119                 rds_dma_hdrs_free(ic->rds_ibdev,
1120                           ic->i_recv_hdrs,
1121                           ic->i_recv_hdrs_dma,
1122                           ic->i_recv_ring.w_nr,
1123                           DMA_FROM_DEVICE);
1124                 ic->i_recv_hdrs = NULL;
1125                 ic->i_recv_hdrs_dma = NULL;
1126             }
1127 
1128             if (ic->i_ack) {
1129                 rds_dma_hdr_free(ic->rds_ibdev->dev, ic->i_ack,
1130                          ic->i_ack_dma, DMA_TO_DEVICE);
1131                 ic->i_ack = NULL;
1132             }
1133         } else {
1134             WARN_ON(ic->i_send_hdrs);
1135             WARN_ON(ic->i_send_hdrs_dma);
1136             WARN_ON(ic->i_recv_hdrs);
1137             WARN_ON(ic->i_recv_hdrs_dma);
1138             WARN_ON(ic->i_ack);
1139         }
1140 
1141         if (ic->i_sends)
1142             rds_ib_send_clear_ring(ic);
1143         if (ic->i_recvs)
1144             rds_ib_recv_clear_ring(ic);
1145 
1146         rdma_destroy_id(ic->i_cm_id);
1147 
1148         /*
1149          * Move connection back to the nodev list.
1150          */
1151         if (ic->rds_ibdev)
1152             rds_ib_remove_conn(ic->rds_ibdev, conn);
1153 
1154         ic->i_cm_id = NULL;
1155         ic->i_pd = NULL;
1156         ic->i_send_cq = NULL;
1157         ic->i_recv_cq = NULL;
1158     }
1159     BUG_ON(ic->rds_ibdev);
1160 
1161     /* Clear pending transmit */
1162     if (ic->i_data_op) {
1163         struct rds_message *rm;
1164 
1165         rm = container_of(ic->i_data_op, struct rds_message, data);
1166         rds_message_put(rm);
1167         ic->i_data_op = NULL;
1168     }
1169 
1170     /* Clear the ACK state */
1171     clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1172 #ifdef KERNEL_HAS_ATOMIC64
1173     atomic64_set(&ic->i_ack_next, 0);
1174 #else
1175     ic->i_ack_next = 0;
1176 #endif
1177     ic->i_ack_recv = 0;
1178 
1179     /* Clear flow control state */
1180     ic->i_flowctl = 0;
1181     atomic_set(&ic->i_credits, 0);
1182 
1183     /* Re-init rings, but retain sizes. */
1184     rds_ib_ring_init(&ic->i_send_ring, ic->i_send_ring.w_nr);
1185     rds_ib_ring_init(&ic->i_recv_ring, ic->i_recv_ring.w_nr);
1186 
1187     if (ic->i_ibinc) {
1188         rds_inc_put(&ic->i_ibinc->ii_inc);
1189         ic->i_ibinc = NULL;
1190     }
1191 
1192     vfree(ic->i_sends);
1193     ic->i_sends = NULL;
1194     vfree(ic->i_recvs);
1195     ic->i_recvs = NULL;
1196     ic->i_active_side = false;
1197 }
1198 
1199 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
1200 {
1201     struct rds_ib_connection *ic;
1202     unsigned long flags;
1203     int ret;
1204 
1205     /* XXX too lazy? */
1206     ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
1207     if (!ic)
1208         return -ENOMEM;
1209 
1210     ret = rds_ib_recv_alloc_caches(ic, gfp);
1211     if (ret) {
1212         kfree(ic);
1213         return ret;
1214     }
1215 
1216     INIT_LIST_HEAD(&ic->ib_node);
1217     tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
1218              (unsigned long)ic);
1219     tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
1220              (unsigned long)ic);
1221     mutex_init(&ic->i_recv_mutex);
1222 #ifndef KERNEL_HAS_ATOMIC64
1223     spin_lock_init(&ic->i_ack_lock);
1224 #endif
1225     atomic_set(&ic->i_signaled_sends, 0);
1226     atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
1227 
1228     /*
1229      * rds_ib_conn_shutdown() waits for these to be emptied so they
1230      * must be initialized before it can be called.
1231      */
1232     rds_ib_ring_init(&ic->i_send_ring, 0);
1233     rds_ib_ring_init(&ic->i_recv_ring, 0);
1234 
1235     ic->conn = conn;
1236     conn->c_transport_data = ic;
1237 
1238     spin_lock_irqsave(&ib_nodev_conns_lock, flags);
1239     list_add_tail(&ic->ib_node, &ib_nodev_conns);
1240     spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
1241 
1242 
1243     rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
1244     return 0;
1245 }
1246 
1247 /*
1248  * Free a connection. Connection must be shut down and not set for reconnect.
1249  */
1250 void rds_ib_conn_free(void *arg)
1251 {
1252     struct rds_ib_connection *ic = arg;
1253     spinlock_t  *lock_ptr;
1254 
1255     rdsdebug("ic %p\n", ic);
1256 
1257     /*
1258      * Conn is either on a dev's list or on the nodev list.
1259      * A race with shutdown() or connect() would cause problems
1260      * (since rds_ibdev would change) but that should never happen.
1261      */
1262     lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1263 
1264     spin_lock_irq(lock_ptr);
1265     list_del(&ic->ib_node);
1266     spin_unlock_irq(lock_ptr);
1267 
1268     rds_ib_recv_free_caches(ic);
1269 
1270     kfree(ic);
1271 }
1272 
1273 
1274 /*
1275  * An error occurred on the connection
1276  */
1277 void
1278 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1279 {
1280     va_list ap;
1281 
1282     rds_conn_drop(conn);
1283 
1284     va_start(ap, fmt);
1285     vprintk(fmt, ap);
1286     va_end(ap);
1287 }
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!