Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2006 Oracle. All rights reserved. |
| 3 | * |
| 4 | * This software is available to you under a choice of one of two |
| 5 | * licenses. You may choose to be licensed under the terms of the GNU |
| 6 | * General Public License (GPL) Version 2, available from the file |
| 7 | * COPYING in the main directory of this source tree, or the |
| 8 | * OpenIB.org BSD license below: |
| 9 | * |
| 10 | * Redistribution and use in source and binary forms, with or |
| 11 | * without modification, are permitted provided that the following |
| 12 | * conditions are met: |
| 13 | * |
| 14 | * - Redistributions of source code must retain the above |
| 15 | * copyright notice, this list of conditions and the following |
| 16 | * disclaimer. |
| 17 | * |
| 18 | * - Redistributions in binary form must reproduce the above |
| 19 | * copyright notice, this list of conditions and the following |
| 20 | * disclaimer in the documentation and/or other materials |
| 21 | * provided with the distribution. |
| 22 | * |
| 23 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, |
| 24 | * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF |
| 25 | * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND |
| 26 | * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS |
| 27 | * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN |
| 28 | * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN |
| 29 | * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 30 | * SOFTWARE. |
| 31 | * |
| 32 | */ |
| 33 | #include <linux/kernel.h> |
| 34 | #include <linux/in.h> |
| 35 | #include <linux/vmalloc.h> |
| 36 | |
| 37 | #include "rds.h" |
| 38 | #include "ib.h" |
| 39 | |
| 40 | /* |
| 41 | * Set the selected protocol version |
| 42 | */ |
| 43 | static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version) |
| 44 | { |
| 45 | conn->c_version = version; |
| 46 | } |
| 47 | |
| 48 | /* |
| 49 | * Set up flow control |
| 50 | */ |
| 51 | static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits) |
| 52 | { |
| 53 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 54 | |
| 55 | if (rds_ib_sysctl_flow_control && credits != 0) { |
| 56 | /* We're doing flow control */ |
| 57 | ic->i_flowctl = 1; |
| 58 | rds_ib_send_add_credits(conn, credits); |
| 59 | } else { |
| 60 | ic->i_flowctl = 0; |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | /* |
| 65 | * Tune RNR behavior. Without flow control, we use a rather |
| 66 | * low timeout, but not the absolute minimum - this should |
| 67 | * be tunable. |
| 68 | * |
| 69 | * We already set the RNR retry count to 7 (which is the |
| 70 | * smallest infinite number :-) above. |
| 71 | * If flow control is off, we want to change this back to 0 |
| 72 | * so that we learn quickly when our credit accounting is |
| 73 | * buggy. |
| 74 | * |
| 75 | * Caller passes in a qp_attr pointer - don't waste stack spacv |
| 76 | * by allocation this twice. |
| 77 | */ |
| 78 | static void |
| 79 | rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr) |
| 80 | { |
| 81 | int ret; |
| 82 | |
| 83 | attr->min_rnr_timer = IB_RNR_TIMER_000_32; |
| 84 | ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER); |
| 85 | if (ret) |
| 86 | printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret); |
| 87 | } |
| 88 | |
| 89 | /* |
| 90 | * Connection established. |
| 91 | * We get here for both outgoing and incoming connection. |
| 92 | */ |
| 93 | void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event) |
| 94 | { |
| 95 | const struct rds_ib_connect_private *dp = NULL; |
| 96 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 97 | struct rds_ib_device *rds_ibdev; |
| 98 | struct ib_qp_attr qp_attr; |
| 99 | int err; |
| 100 | |
Andy Grover | 9ddbcfa | 2009-07-17 13:13:23 +0000 | [diff] [blame] | 101 | if (event->param.conn.private_data_len >= sizeof(*dp)) { |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 102 | dp = event->param.conn.private_data; |
| 103 | |
Andy Grover | 02a6a25 | 2009-07-17 13:13:24 +0000 | [diff] [blame] | 104 | /* make sure it isn't empty data */ |
| 105 | if (dp->dp_protocol_major) { |
| 106 | rds_ib_set_protocol(conn, |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 107 | RDS_PROTOCOL(dp->dp_protocol_major, |
Andy Grover | 02a6a25 | 2009-07-17 13:13:24 +0000 | [diff] [blame] | 108 | dp->dp_protocol_minor)); |
| 109 | rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit)); |
| 110 | } |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 111 | } |
| 112 | |
| 113 | printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n", |
Andy Grover | 8dacd57 | 2009-07-17 13:13:25 +0000 | [diff] [blame] | 114 | &conn->c_faddr, |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 115 | RDS_PROTOCOL_MAJOR(conn->c_version), |
| 116 | RDS_PROTOCOL_MINOR(conn->c_version), |
| 117 | ic->i_flowctl ? ", flow control" : ""); |
| 118 | |
Andy Grover | e11d912 | 2009-07-17 13:13:29 +0000 | [diff] [blame] | 119 | /* |
| 120 | * Init rings and fill recv. this needs to wait until protocol negotiation |
| 121 | * is complete, since ring layout is different from 3.0 to 3.1. |
| 122 | */ |
| 123 | rds_ib_send_init_ring(ic); |
| 124 | rds_ib_recv_init_ring(ic); |
| 125 | /* Post receive buffers - as a side effect, this will update |
| 126 | * the posted credit count. */ |
| 127 | rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1); |
| 128 | |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 129 | /* Tune RNR behavior */ |
| 130 | rds_ib_tune_rnr(ic, &qp_attr); |
| 131 | |
| 132 | qp_attr.qp_state = IB_QPS_RTS; |
| 133 | err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE); |
| 134 | if (err) |
| 135 | printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err); |
| 136 | |
| 137 | /* update ib_device with this local ipaddr & conn */ |
| 138 | rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client); |
| 139 | err = rds_ib_update_ipaddr(rds_ibdev, conn->c_laddr); |
| 140 | if (err) |
| 141 | printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", err); |
Andy Grover | 745cbcc | 2009-04-01 08:20:19 +0000 | [diff] [blame] | 142 | rds_ib_add_conn(rds_ibdev, conn); |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 143 | |
| 144 | /* If the peer gave us the last packet it saw, process this as if |
| 145 | * we had received a regular ACK. */ |
| 146 | if (dp && dp->dp_ack_seq) |
| 147 | rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL); |
| 148 | |
| 149 | rds_connect_complete(conn); |
| 150 | } |
| 151 | |
| 152 | static void rds_ib_cm_fill_conn_param(struct rds_connection *conn, |
| 153 | struct rdma_conn_param *conn_param, |
| 154 | struct rds_ib_connect_private *dp, |
| 155 | u32 protocol_version) |
| 156 | { |
| 157 | memset(conn_param, 0, sizeof(struct rdma_conn_param)); |
| 158 | /* XXX tune these? */ |
| 159 | conn_param->responder_resources = 1; |
| 160 | conn_param->initiator_depth = 1; |
Andy Grover | 3ba23ad | 2009-07-17 13:13:22 +0000 | [diff] [blame] | 161 | conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7); |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 162 | conn_param->rnr_retry_count = 7; |
| 163 | |
| 164 | if (dp) { |
| 165 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 166 | |
| 167 | memset(dp, 0, sizeof(*dp)); |
| 168 | dp->dp_saddr = conn->c_laddr; |
| 169 | dp->dp_daddr = conn->c_faddr; |
| 170 | dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version); |
| 171 | dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version); |
| 172 | dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS); |
| 173 | dp->dp_ack_seq = rds_ib_piggyb_ack(ic); |
| 174 | |
| 175 | /* Advertise flow control */ |
| 176 | if (ic->i_flowctl) { |
| 177 | unsigned int credits; |
| 178 | |
| 179 | credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)); |
| 180 | dp->dp_credit = cpu_to_be32(credits); |
| 181 | atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits); |
| 182 | } |
| 183 | |
| 184 | conn_param->private_data = dp; |
| 185 | conn_param->private_data_len = sizeof(*dp); |
| 186 | } |
| 187 | } |
| 188 | |
| 189 | static void rds_ib_cq_event_handler(struct ib_event *event, void *data) |
| 190 | { |
| 191 | rdsdebug("event %u data %p\n", event->event, data); |
| 192 | } |
| 193 | |
| 194 | static void rds_ib_qp_event_handler(struct ib_event *event, void *data) |
| 195 | { |
| 196 | struct rds_connection *conn = data; |
| 197 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 198 | |
| 199 | rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event); |
| 200 | |
| 201 | switch (event->event) { |
| 202 | case IB_EVENT_COMM_EST: |
| 203 | rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST); |
| 204 | break; |
| 205 | default: |
Andy Grover | fdf6e6b | 2009-07-17 13:13:31 +0000 | [diff] [blame^] | 206 | rds_ib_conn_error(conn, "RDS/IB: Fatal QP Event %u " |
| 207 | "- connection %pI4->%pI4, reconnecting\n", |
| 208 | event->event, &conn->c_laddr, &conn->c_faddr); |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 209 | break; |
| 210 | } |
| 211 | } |
| 212 | |
| 213 | /* |
| 214 | * This needs to be very careful to not leave IS_ERR pointers around for |
| 215 | * cleanup to trip over. |
| 216 | */ |
| 217 | static int rds_ib_setup_qp(struct rds_connection *conn) |
| 218 | { |
| 219 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 220 | struct ib_device *dev = ic->i_cm_id->device; |
| 221 | struct ib_qp_init_attr attr; |
| 222 | struct rds_ib_device *rds_ibdev; |
| 223 | int ret; |
| 224 | |
| 225 | /* rds_ib_add_one creates a rds_ib_device object per IB device, |
| 226 | * and allocates a protection domain, memory range and FMR pool |
| 227 | * for each. If that fails for any reason, it will not register |
| 228 | * the rds_ibdev at all. |
| 229 | */ |
| 230 | rds_ibdev = ib_get_client_data(dev, &rds_ib_client); |
| 231 | if (rds_ibdev == NULL) { |
| 232 | if (printk_ratelimit()) |
| 233 | printk(KERN_NOTICE "RDS/IB: No client_data for device %s\n", |
| 234 | dev->name); |
| 235 | return -EOPNOTSUPP; |
| 236 | } |
| 237 | |
| 238 | if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1) |
| 239 | rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1); |
| 240 | if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1) |
| 241 | rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1); |
| 242 | |
| 243 | /* Protection domain and memory range */ |
| 244 | ic->i_pd = rds_ibdev->pd; |
| 245 | ic->i_mr = rds_ibdev->mr; |
| 246 | |
| 247 | ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler, |
| 248 | rds_ib_cq_event_handler, conn, |
| 249 | ic->i_send_ring.w_nr + 1, 0); |
| 250 | if (IS_ERR(ic->i_send_cq)) { |
| 251 | ret = PTR_ERR(ic->i_send_cq); |
| 252 | ic->i_send_cq = NULL; |
| 253 | rdsdebug("ib_create_cq send failed: %d\n", ret); |
| 254 | goto out; |
| 255 | } |
| 256 | |
| 257 | ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler, |
| 258 | rds_ib_cq_event_handler, conn, |
| 259 | ic->i_recv_ring.w_nr, 0); |
| 260 | if (IS_ERR(ic->i_recv_cq)) { |
| 261 | ret = PTR_ERR(ic->i_recv_cq); |
| 262 | ic->i_recv_cq = NULL; |
| 263 | rdsdebug("ib_create_cq recv failed: %d\n", ret); |
| 264 | goto out; |
| 265 | } |
| 266 | |
| 267 | ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); |
| 268 | if (ret) { |
| 269 | rdsdebug("ib_req_notify_cq send failed: %d\n", ret); |
| 270 | goto out; |
| 271 | } |
| 272 | |
| 273 | ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); |
| 274 | if (ret) { |
| 275 | rdsdebug("ib_req_notify_cq recv failed: %d\n", ret); |
| 276 | goto out; |
| 277 | } |
| 278 | |
| 279 | /* XXX negotiate max send/recv with remote? */ |
| 280 | memset(&attr, 0, sizeof(attr)); |
| 281 | attr.event_handler = rds_ib_qp_event_handler; |
| 282 | attr.qp_context = conn; |
| 283 | /* + 1 to allow for the single ack message */ |
| 284 | attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1; |
| 285 | attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1; |
| 286 | attr.cap.max_send_sge = rds_ibdev->max_sge; |
| 287 | attr.cap.max_recv_sge = RDS_IB_RECV_SGE; |
| 288 | attr.sq_sig_type = IB_SIGNAL_REQ_WR; |
| 289 | attr.qp_type = IB_QPT_RC; |
| 290 | attr.send_cq = ic->i_send_cq; |
| 291 | attr.recv_cq = ic->i_recv_cq; |
| 292 | |
| 293 | /* |
| 294 | * XXX this can fail if max_*_wr is too large? Are we supposed |
| 295 | * to back off until we get a value that the hardware can support? |
| 296 | */ |
| 297 | ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr); |
| 298 | if (ret) { |
| 299 | rdsdebug("rdma_create_qp failed: %d\n", ret); |
| 300 | goto out; |
| 301 | } |
| 302 | |
| 303 | ic->i_send_hdrs = ib_dma_alloc_coherent(dev, |
| 304 | ic->i_send_ring.w_nr * |
| 305 | sizeof(struct rds_header), |
| 306 | &ic->i_send_hdrs_dma, GFP_KERNEL); |
| 307 | if (ic->i_send_hdrs == NULL) { |
| 308 | ret = -ENOMEM; |
| 309 | rdsdebug("ib_dma_alloc_coherent send failed\n"); |
| 310 | goto out; |
| 311 | } |
| 312 | |
| 313 | ic->i_recv_hdrs = ib_dma_alloc_coherent(dev, |
| 314 | ic->i_recv_ring.w_nr * |
| 315 | sizeof(struct rds_header), |
| 316 | &ic->i_recv_hdrs_dma, GFP_KERNEL); |
| 317 | if (ic->i_recv_hdrs == NULL) { |
| 318 | ret = -ENOMEM; |
| 319 | rdsdebug("ib_dma_alloc_coherent recv failed\n"); |
| 320 | goto out; |
| 321 | } |
| 322 | |
| 323 | ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header), |
| 324 | &ic->i_ack_dma, GFP_KERNEL); |
| 325 | if (ic->i_ack == NULL) { |
| 326 | ret = -ENOMEM; |
| 327 | rdsdebug("ib_dma_alloc_coherent ack failed\n"); |
| 328 | goto out; |
| 329 | } |
| 330 | |
| 331 | ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work)); |
| 332 | if (ic->i_sends == NULL) { |
| 333 | ret = -ENOMEM; |
| 334 | rdsdebug("send allocation failed\n"); |
| 335 | goto out; |
| 336 | } |
Andy Grover | e11d912 | 2009-07-17 13:13:29 +0000 | [diff] [blame] | 337 | memset(ic->i_sends, 0, ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work)); |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 338 | |
| 339 | ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work)); |
| 340 | if (ic->i_recvs == NULL) { |
| 341 | ret = -ENOMEM; |
| 342 | rdsdebug("recv allocation failed\n"); |
| 343 | goto out; |
| 344 | } |
Andy Grover | e11d912 | 2009-07-17 13:13:29 +0000 | [diff] [blame] | 345 | memset(ic->i_recvs, 0, ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work)); |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 346 | |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 347 | rds_ib_recv_init_ack(ic); |
| 348 | |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 349 | rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr, |
| 350 | ic->i_send_cq, ic->i_recv_cq); |
| 351 | |
| 352 | out: |
| 353 | return ret; |
| 354 | } |
| 355 | |
Andy Grover | 9ddbcfa | 2009-07-17 13:13:23 +0000 | [diff] [blame] | 356 | static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event) |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 357 | { |
Andy Grover | 9ddbcfa | 2009-07-17 13:13:23 +0000 | [diff] [blame] | 358 | const struct rds_ib_connect_private *dp = event->param.conn.private_data; |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 359 | u16 common; |
| 360 | u32 version = 0; |
| 361 | |
Andy Grover | 9ddbcfa | 2009-07-17 13:13:23 +0000 | [diff] [blame] | 362 | /* |
| 363 | * rdma_cm private data is odd - when there is any private data in the |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 364 | * request, we will be given a pretty large buffer without telling us the |
| 365 | * original size. The only way to tell the difference is by looking at |
| 366 | * the contents, which are initialized to zero. |
| 367 | * If the protocol version fields aren't set, this is a connection attempt |
| 368 | * from an older version. This could could be 3.0 or 2.0 - we can't tell. |
Andy Grover | 9ddbcfa | 2009-07-17 13:13:23 +0000 | [diff] [blame] | 369 | * We really should have changed this for OFED 1.3 :-( |
| 370 | */ |
| 371 | |
| 372 | /* Be paranoid. RDS always has privdata */ |
| 373 | if (!event->param.conn.private_data_len) { |
| 374 | printk(KERN_NOTICE "RDS incoming connection has no private data, " |
| 375 | "rejecting\n"); |
| 376 | return 0; |
| 377 | } |
| 378 | |
| 379 | /* Even if len is crap *now* I still want to check it. -ASG */ |
| 380 | if (event->param.conn.private_data_len < sizeof (*dp) |
| 381 | || dp->dp_protocol_major == 0) |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 382 | return RDS_PROTOCOL_3_0; |
| 383 | |
| 384 | common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS; |
| 385 | if (dp->dp_protocol_major == 3 && common) { |
| 386 | version = RDS_PROTOCOL_3_0; |
| 387 | while ((common >>= 1) != 0) |
| 388 | version++; |
| 389 | } else if (printk_ratelimit()) { |
| 390 | printk(KERN_NOTICE "RDS: Connection from %pI4 using " |
| 391 | "incompatible protocol version %u.%u\n", |
| 392 | &dp->dp_saddr, |
| 393 | dp->dp_protocol_major, |
| 394 | dp->dp_protocol_minor); |
| 395 | } |
| 396 | return version; |
| 397 | } |
| 398 | |
| 399 | int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, |
| 400 | struct rdma_cm_event *event) |
| 401 | { |
| 402 | __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id; |
| 403 | __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id; |
| 404 | const struct rds_ib_connect_private *dp = event->param.conn.private_data; |
| 405 | struct rds_ib_connect_private dp_rep; |
| 406 | struct rds_connection *conn = NULL; |
| 407 | struct rds_ib_connection *ic = NULL; |
| 408 | struct rdma_conn_param conn_param; |
| 409 | u32 version; |
| 410 | int err, destroy = 1; |
| 411 | |
| 412 | /* Check whether the remote protocol version matches ours. */ |
Andy Grover | 9ddbcfa | 2009-07-17 13:13:23 +0000 | [diff] [blame] | 413 | version = rds_ib_protocol_compatible(event); |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 414 | if (!version) |
| 415 | goto out; |
| 416 | |
| 417 | rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid " |
| 418 | "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr, |
| 419 | RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version), |
| 420 | (unsigned long long)be64_to_cpu(lguid), |
| 421 | (unsigned long long)be64_to_cpu(fguid)); |
| 422 | |
| 423 | conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport, |
| 424 | GFP_KERNEL); |
| 425 | if (IS_ERR(conn)) { |
| 426 | rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn)); |
| 427 | conn = NULL; |
| 428 | goto out; |
| 429 | } |
| 430 | |
| 431 | /* |
| 432 | * The connection request may occur while the |
| 433 | * previous connection exist, e.g. in case of failover. |
| 434 | * But as connections may be initiated simultaneously |
| 435 | * by both hosts, we have a random backoff mechanism - |
| 436 | * see the comment above rds_queue_reconnect() |
| 437 | */ |
| 438 | mutex_lock(&conn->c_cm_lock); |
| 439 | if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) { |
| 440 | if (rds_conn_state(conn) == RDS_CONN_UP) { |
| 441 | rdsdebug("incoming connect while connecting\n"); |
| 442 | rds_conn_drop(conn); |
| 443 | rds_ib_stats_inc(s_ib_listen_closed_stale); |
| 444 | } else |
| 445 | if (rds_conn_state(conn) == RDS_CONN_CONNECTING) { |
| 446 | /* Wait and see - our connect may still be succeeding */ |
| 447 | rds_ib_stats_inc(s_ib_connect_raced); |
| 448 | } |
| 449 | mutex_unlock(&conn->c_cm_lock); |
| 450 | goto out; |
| 451 | } |
| 452 | |
| 453 | ic = conn->c_transport_data; |
| 454 | |
| 455 | rds_ib_set_protocol(conn, version); |
| 456 | rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit)); |
| 457 | |
| 458 | /* If the peer gave us the last packet it saw, process this as if |
| 459 | * we had received a regular ACK. */ |
| 460 | if (dp->dp_ack_seq) |
| 461 | rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL); |
| 462 | |
| 463 | BUG_ON(cm_id->context); |
| 464 | BUG_ON(ic->i_cm_id); |
| 465 | |
| 466 | ic->i_cm_id = cm_id; |
| 467 | cm_id->context = conn; |
| 468 | |
| 469 | /* We got halfway through setting up the ib_connection, if we |
| 470 | * fail now, we have to take the long route out of this mess. */ |
| 471 | destroy = 0; |
| 472 | |
| 473 | err = rds_ib_setup_qp(conn); |
| 474 | if (err) { |
| 475 | rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err); |
| 476 | goto out; |
| 477 | } |
| 478 | |
| 479 | rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version); |
| 480 | |
| 481 | /* rdma_accept() calls rdma_reject() internally if it fails */ |
| 482 | err = rdma_accept(cm_id, &conn_param); |
| 483 | mutex_unlock(&conn->c_cm_lock); |
| 484 | if (err) { |
| 485 | rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err); |
| 486 | goto out; |
| 487 | } |
| 488 | |
| 489 | return 0; |
| 490 | |
| 491 | out: |
| 492 | rdma_reject(cm_id, NULL, 0); |
| 493 | return destroy; |
| 494 | } |
| 495 | |
| 496 | |
| 497 | int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id) |
| 498 | { |
| 499 | struct rds_connection *conn = cm_id->context; |
| 500 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 501 | struct rdma_conn_param conn_param; |
| 502 | struct rds_ib_connect_private dp; |
| 503 | int ret; |
| 504 | |
| 505 | /* If the peer doesn't do protocol negotiation, we must |
| 506 | * default to RDSv3.0 */ |
| 507 | rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0); |
| 508 | ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */ |
| 509 | |
| 510 | ret = rds_ib_setup_qp(conn); |
| 511 | if (ret) { |
| 512 | rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret); |
| 513 | goto out; |
| 514 | } |
| 515 | |
| 516 | rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION); |
| 517 | |
| 518 | ret = rdma_connect(cm_id, &conn_param); |
| 519 | if (ret) |
| 520 | rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret); |
| 521 | |
| 522 | out: |
| 523 | /* Beware - returning non-zero tells the rdma_cm to destroy |
| 524 | * the cm_id. We should certainly not do it as long as we still |
| 525 | * "own" the cm_id. */ |
| 526 | if (ret) { |
| 527 | if (ic->i_cm_id == cm_id) |
| 528 | ret = 0; |
| 529 | } |
| 530 | return ret; |
| 531 | } |
| 532 | |
| 533 | int rds_ib_conn_connect(struct rds_connection *conn) |
| 534 | { |
| 535 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 536 | struct sockaddr_in src, dest; |
| 537 | int ret; |
| 538 | |
| 539 | /* XXX I wonder what affect the port space has */ |
| 540 | /* delegate cm event handler to rdma_transport */ |
| 541 | ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn, |
| 542 | RDMA_PS_TCP); |
| 543 | if (IS_ERR(ic->i_cm_id)) { |
| 544 | ret = PTR_ERR(ic->i_cm_id); |
| 545 | ic->i_cm_id = NULL; |
| 546 | rdsdebug("rdma_create_id() failed: %d\n", ret); |
| 547 | goto out; |
| 548 | } |
| 549 | |
| 550 | rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn); |
| 551 | |
| 552 | src.sin_family = AF_INET; |
| 553 | src.sin_addr.s_addr = (__force u32)conn->c_laddr; |
| 554 | src.sin_port = (__force u16)htons(0); |
| 555 | |
| 556 | dest.sin_family = AF_INET; |
| 557 | dest.sin_addr.s_addr = (__force u32)conn->c_faddr; |
| 558 | dest.sin_port = (__force u16)htons(RDS_PORT); |
| 559 | |
| 560 | ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src, |
| 561 | (struct sockaddr *)&dest, |
| 562 | RDS_RDMA_RESOLVE_TIMEOUT_MS); |
| 563 | if (ret) { |
| 564 | rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id, |
| 565 | ret); |
| 566 | rdma_destroy_id(ic->i_cm_id); |
| 567 | ic->i_cm_id = NULL; |
| 568 | } |
| 569 | |
| 570 | out: |
| 571 | return ret; |
| 572 | } |
| 573 | |
| 574 | /* |
| 575 | * This is so careful about only cleaning up resources that were built up |
| 576 | * so that it can be called at any point during startup. In fact it |
| 577 | * can be called multiple times for a given connection. |
| 578 | */ |
| 579 | void rds_ib_conn_shutdown(struct rds_connection *conn) |
| 580 | { |
| 581 | struct rds_ib_connection *ic = conn->c_transport_data; |
| 582 | int err = 0; |
| 583 | |
| 584 | rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id, |
| 585 | ic->i_pd, ic->i_send_cq, ic->i_recv_cq, |
| 586 | ic->i_cm_id ? ic->i_cm_id->qp : NULL); |
| 587 | |
| 588 | if (ic->i_cm_id) { |
| 589 | struct ib_device *dev = ic->i_cm_id->device; |
| 590 | |
| 591 | rdsdebug("disconnecting cm %p\n", ic->i_cm_id); |
| 592 | err = rdma_disconnect(ic->i_cm_id); |
| 593 | if (err) { |
| 594 | /* Actually this may happen quite frequently, when |
| 595 | * an outgoing connect raced with an incoming connect. |
| 596 | */ |
| 597 | rdsdebug("failed to disconnect, cm: %p err %d\n", |
| 598 | ic->i_cm_id, err); |
| 599 | } |
| 600 | |
| 601 | wait_event(rds_ib_ring_empty_wait, |
| 602 | rds_ib_ring_empty(&ic->i_send_ring) && |
| 603 | rds_ib_ring_empty(&ic->i_recv_ring)); |
| 604 | |
| 605 | if (ic->i_send_hdrs) |
| 606 | ib_dma_free_coherent(dev, |
| 607 | ic->i_send_ring.w_nr * |
| 608 | sizeof(struct rds_header), |
| 609 | ic->i_send_hdrs, |
| 610 | ic->i_send_hdrs_dma); |
| 611 | |
| 612 | if (ic->i_recv_hdrs) |
| 613 | ib_dma_free_coherent(dev, |
| 614 | ic->i_recv_ring.w_nr * |
| 615 | sizeof(struct rds_header), |
| 616 | ic->i_recv_hdrs, |
| 617 | ic->i_recv_hdrs_dma); |
| 618 | |
| 619 | if (ic->i_ack) |
| 620 | ib_dma_free_coherent(dev, sizeof(struct rds_header), |
| 621 | ic->i_ack, ic->i_ack_dma); |
| 622 | |
| 623 | if (ic->i_sends) |
| 624 | rds_ib_send_clear_ring(ic); |
| 625 | if (ic->i_recvs) |
| 626 | rds_ib_recv_clear_ring(ic); |
| 627 | |
| 628 | if (ic->i_cm_id->qp) |
| 629 | rdma_destroy_qp(ic->i_cm_id); |
| 630 | if (ic->i_send_cq) |
| 631 | ib_destroy_cq(ic->i_send_cq); |
| 632 | if (ic->i_recv_cq) |
| 633 | ib_destroy_cq(ic->i_recv_cq); |
| 634 | rdma_destroy_id(ic->i_cm_id); |
| 635 | |
| 636 | /* |
| 637 | * Move connection back to the nodev list. |
| 638 | */ |
Andy Grover | 745cbcc | 2009-04-01 08:20:19 +0000 | [diff] [blame] | 639 | if (ic->rds_ibdev) |
| 640 | rds_ib_remove_conn(ic->rds_ibdev, conn); |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 641 | |
| 642 | ic->i_cm_id = NULL; |
| 643 | ic->i_pd = NULL; |
| 644 | ic->i_mr = NULL; |
| 645 | ic->i_send_cq = NULL; |
| 646 | ic->i_recv_cq = NULL; |
| 647 | ic->i_send_hdrs = NULL; |
| 648 | ic->i_recv_hdrs = NULL; |
| 649 | ic->i_ack = NULL; |
| 650 | } |
| 651 | BUG_ON(ic->rds_ibdev); |
| 652 | |
| 653 | /* Clear pending transmit */ |
| 654 | if (ic->i_rm) { |
| 655 | rds_message_put(ic->i_rm); |
| 656 | ic->i_rm = NULL; |
| 657 | } |
| 658 | |
| 659 | /* Clear the ACK state */ |
| 660 | clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); |
Andy Grover | 8cbd960 | 2009-04-01 08:20:20 +0000 | [diff] [blame] | 661 | #ifdef KERNEL_HAS_ATOMIC64 |
| 662 | atomic64_set(&ic->i_ack_next, 0); |
| 663 | #else |
| 664 | ic->i_ack_next = 0; |
| 665 | #endif |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 666 | ic->i_ack_recv = 0; |
| 667 | |
| 668 | /* Clear flow control state */ |
| 669 | ic->i_flowctl = 0; |
| 670 | atomic_set(&ic->i_credits, 0); |
| 671 | |
| 672 | rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); |
| 673 | rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); |
| 674 | |
| 675 | if (ic->i_ibinc) { |
| 676 | rds_inc_put(&ic->i_ibinc->ii_inc); |
| 677 | ic->i_ibinc = NULL; |
| 678 | } |
| 679 | |
| 680 | vfree(ic->i_sends); |
| 681 | ic->i_sends = NULL; |
| 682 | vfree(ic->i_recvs); |
| 683 | ic->i_recvs = NULL; |
| 684 | } |
| 685 | |
| 686 | int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp) |
| 687 | { |
| 688 | struct rds_ib_connection *ic; |
| 689 | unsigned long flags; |
| 690 | |
| 691 | /* XXX too lazy? */ |
| 692 | ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL); |
| 693 | if (ic == NULL) |
| 694 | return -ENOMEM; |
| 695 | |
| 696 | INIT_LIST_HEAD(&ic->ib_node); |
| 697 | mutex_init(&ic->i_recv_mutex); |
Andy Grover | 8cbd960 | 2009-04-01 08:20:20 +0000 | [diff] [blame] | 698 | #ifndef KERNEL_HAS_ATOMIC64 |
| 699 | spin_lock_init(&ic->i_ack_lock); |
| 700 | #endif |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 701 | |
| 702 | /* |
| 703 | * rds_ib_conn_shutdown() waits for these to be emptied so they |
| 704 | * must be initialized before it can be called. |
| 705 | */ |
| 706 | rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); |
| 707 | rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); |
| 708 | |
| 709 | ic->conn = conn; |
| 710 | conn->c_transport_data = ic; |
| 711 | |
| 712 | spin_lock_irqsave(&ib_nodev_conns_lock, flags); |
| 713 | list_add_tail(&ic->ib_node, &ib_nodev_conns); |
| 714 | spin_unlock_irqrestore(&ib_nodev_conns_lock, flags); |
| 715 | |
| 716 | |
| 717 | rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data); |
| 718 | return 0; |
| 719 | } |
| 720 | |
Andy Grover | 745cbcc | 2009-04-01 08:20:19 +0000 | [diff] [blame] | 721 | /* |
| 722 | * Free a connection. Connection must be shut down and not set for reconnect. |
| 723 | */ |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 724 | void rds_ib_conn_free(void *arg) |
| 725 | { |
| 726 | struct rds_ib_connection *ic = arg; |
Andy Grover | 745cbcc | 2009-04-01 08:20:19 +0000 | [diff] [blame] | 727 | spinlock_t *lock_ptr; |
| 728 | |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 729 | rdsdebug("ic %p\n", ic); |
Andy Grover | 745cbcc | 2009-04-01 08:20:19 +0000 | [diff] [blame] | 730 | |
| 731 | /* |
| 732 | * Conn is either on a dev's list or on the nodev list. |
| 733 | * A race with shutdown() or connect() would cause problems |
| 734 | * (since rds_ibdev would change) but that should never happen. |
| 735 | */ |
| 736 | lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock; |
| 737 | |
| 738 | spin_lock_irq(lock_ptr); |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 739 | list_del(&ic->ib_node); |
Andy Grover | 745cbcc | 2009-04-01 08:20:19 +0000 | [diff] [blame] | 740 | spin_unlock_irq(lock_ptr); |
| 741 | |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 742 | kfree(ic); |
| 743 | } |
| 744 | |
| 745 | |
| 746 | /* |
| 747 | * An error occurred on the connection |
| 748 | */ |
| 749 | void |
| 750 | __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...) |
| 751 | { |
| 752 | va_list ap; |
| 753 | |
| 754 | rds_conn_drop(conn); |
| 755 | |
| 756 | va_start(ap, fmt); |
| 757 | vprintk(fmt, ap); |
| 758 | va_end(ap); |
| 759 | } |