Commit | Line | Data |
---|---|---|
ec16227e | 1 | /* |
eee2fa6a | 2 | * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved. |
ec16227e AG |
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> | |
5a0e3ad6 | 35 | #include <linux/slab.h> |
ec16227e | 36 | #include <linux/vmalloc.h> |
cb0a6056 | 37 | #include <linux/ratelimit.h> |
eee2fa6a | 38 | #include <net/addrconf.h> |
ec16227e | 39 | |
0cb43965 | 40 | #include "rds_single_path.h" |
ec16227e AG |
41 | #include "rds.h" |
42 | #include "ib.h" | |
43 | ||
44 | /* | |
45 | * Set the selected protocol version | |
46 | */ | |
47 | static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version) | |
48 | { | |
49 | conn->c_version = version; | |
50 | } | |
51 | ||
52 | /* | |
53 | * Set up flow control | |
54 | */ | |
55 | static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits) | |
56 | { | |
57 | struct rds_ib_connection *ic = conn->c_transport_data; | |
58 | ||
59 | if (rds_ib_sysctl_flow_control && credits != 0) { | |
60 | /* We're doing flow control */ | |
61 | ic->i_flowctl = 1; | |
62 | rds_ib_send_add_credits(conn, credits); | |
63 | } else { | |
64 | ic->i_flowctl = 0; | |
65 | } | |
66 | } | |
67 | ||
68 | /* | |
69 | * Tune RNR behavior. Without flow control, we use a rather | |
70 | * low timeout, but not the absolute minimum - this should | |
71 | * be tunable. | |
72 | * | |
73 | * We already set the RNR retry count to 7 (which is the | |
74 | * smallest infinite number :-) above. | |
75 | * If flow control is off, we want to change this back to 0 | |
76 | * so that we learn quickly when our credit accounting is | |
77 | * buggy. | |
78 | * | |
79 | * Caller passes in a qp_attr pointer - don't waste stack spacv | |
80 | * by allocation this twice. | |
81 | */ | |
82 | static void | |
83 | rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr) | |
84 | { | |
85 | int ret; | |
86 | ||
87 | attr->min_rnr_timer = IB_RNR_TIMER_000_32; | |
88 | ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER); | |
89 | if (ret) | |
90 | printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret); | |
91 | } | |
92 | ||
93 | /* | |
94 | * Connection established. | |
95 | * We get here for both outgoing and incoming connection. | |
96 | */ | |
97 | void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event) | |
98 | { | |
ec16227e | 99 | struct rds_ib_connection *ic = conn->c_transport_data; |
eee2fa6a | 100 | const union rds_ib_conn_priv *dp = NULL; |
ec16227e | 101 | struct ib_qp_attr qp_attr; |
eee2fa6a KCP |
102 | __be64 ack_seq = 0; |
103 | __be32 credit = 0; | |
104 | u8 major = 0; | |
105 | u8 minor = 0; | |
ec16227e AG |
106 | int err; |
107 | ||
eee2fa6a KCP |
108 | dp = event->param.conn.private_data; |
109 | if (conn->c_isv6) { | |
110 | if (event->param.conn.private_data_len >= | |
111 | sizeof(struct rds6_ib_connect_private)) { | |
112 | major = dp->ricp_v6.dp_protocol_major; | |
113 | minor = dp->ricp_v6.dp_protocol_minor; | |
114 | credit = dp->ricp_v6.dp_credit; | |
115 | /* dp structure start is not guaranteed to be 8 bytes | |
116 | * aligned. Since dp_ack_seq is 64-bit extended load | |
117 | * operations can be used so go through get_unaligned | |
118 | * to avoid unaligned errors. | |
119 | */ | |
120 | ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq); | |
02a6a259 | 121 | } |
eee2fa6a KCP |
122 | } else if (event->param.conn.private_data_len >= |
123 | sizeof(struct rds_ib_connect_private)) { | |
124 | major = dp->ricp_v4.dp_protocol_major; | |
125 | minor = dp->ricp_v4.dp_protocol_minor; | |
126 | credit = dp->ricp_v4.dp_credit; | |
127 | ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq); | |
128 | } | |
129 | ||
130 | /* make sure it isn't empty data */ | |
131 | if (major) { | |
132 | rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor)); | |
133 | rds_ib_set_flow_control(conn, be32_to_cpu(credit)); | |
ec16227e AG |
134 | } |
135 | ||
cdc306a5 SS |
136 | if (conn->c_version < RDS_PROTOCOL_VERSION) { |
137 | if (conn->c_version != RDS_PROTOCOL_COMPAT_VERSION) { | |
138 | pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n", | |
139 | &conn->c_laddr, &conn->c_faddr, | |
140 | RDS_PROTOCOL_MAJOR(conn->c_version), | |
141 | RDS_PROTOCOL_MINOR(conn->c_version)); | |
142 | rds_conn_destroy(conn); | |
143 | return; | |
144 | } | |
f147dd9e | 145 | } |
ec16227e | 146 | |
fd261ce6 | 147 | pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c,%d> version %u.%u%s\n", |
cdc306a5 | 148 | ic->i_active_side ? "Active" : "Passive", |
fd261ce6 | 149 | &conn->c_laddr, &conn->c_faddr, conn->c_tos, |
cdc306a5 SS |
150 | RDS_PROTOCOL_MAJOR(conn->c_version), |
151 | RDS_PROTOCOL_MINOR(conn->c_version), | |
152 | ic->i_flowctl ? ", flow control" : ""); | |
153 | ||
cf657269 SS |
154 | atomic_set(&ic->i_cq_quiesce, 0); |
155 | ||
581d53c9 SS |
156 | /* Init rings and fill recv. this needs to wait until protocol |
157 | * negotiation is complete, since ring layout is different | |
158 | * from 3.1 to 4.1. | |
e11d912a AG |
159 | */ |
160 | rds_ib_send_init_ring(ic); | |
161 | rds_ib_recv_init_ring(ic); | |
162 | /* Post receive buffers - as a side effect, this will update | |
163 | * the posted credit count. */ | |
73ce4317 | 164 | rds_ib_recv_refill(conn, 1, GFP_KERNEL); |
e11d912a | 165 | |
ec16227e AG |
166 | /* Tune RNR behavior */ |
167 | rds_ib_tune_rnr(ic, &qp_attr); | |
168 | ||
169 | qp_attr.qp_state = IB_QPS_RTS; | |
170 | err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE); | |
171 | if (err) | |
172 | printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err); | |
173 | ||
3e0249f9 | 174 | /* update ib_device with this local ipaddr */ |
eee2fa6a | 175 | err = rds_ib_update_ipaddr(ic->rds_ibdev, &conn->c_laddr); |
ec16227e | 176 | if (err) |
3e0249f9 ZB |
177 | printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", |
178 | err); | |
ec16227e AG |
179 | |
180 | /* If the peer gave us the last packet it saw, process this as if | |
181 | * we had received a regular ACK. */ | |
c0adf54a | 182 | if (dp) { |
eee2fa6a KCP |
183 | if (ack_seq) |
184 | rds_send_drop_acked(conn, be64_to_cpu(ack_seq), | |
c0adf54a | 185 | NULL); |
186 | } | |
ec16227e | 187 | |
cdc306a5 | 188 | conn->c_proposed_version = conn->c_version; |
ec16227e AG |
189 | rds_connect_complete(conn); |
190 | } | |
191 | ||
192 | static void rds_ib_cm_fill_conn_param(struct rds_connection *conn, | |
eee2fa6a KCP |
193 | struct rdma_conn_param *conn_param, |
194 | union rds_ib_conn_priv *dp, | |
195 | u32 protocol_version, | |
196 | u32 max_responder_resources, | |
197 | u32 max_initiator_depth, | |
198 | bool isv6) | |
ec16227e | 199 | { |
40589e74 | 200 | struct rds_ib_connection *ic = conn->c_transport_data; |
3e0249f9 | 201 | struct rds_ib_device *rds_ibdev = ic->rds_ibdev; |
40589e74 | 202 | |
ec16227e | 203 | memset(conn_param, 0, sizeof(struct rdma_conn_param)); |
40589e74 | 204 | |
40589e74 AG |
205 | conn_param->responder_resources = |
206 | min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources); | |
207 | conn_param->initiator_depth = | |
208 | min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth); | |
3ba23ade | 209 | conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7); |
ec16227e AG |
210 | conn_param->rnr_retry_count = 7; |
211 | ||
212 | if (dp) { | |
ec16227e | 213 | memset(dp, 0, sizeof(*dp)); |
eee2fa6a KCP |
214 | if (isv6) { |
215 | dp->ricp_v6.dp_saddr = conn->c_laddr; | |
216 | dp->ricp_v6.dp_daddr = conn->c_faddr; | |
217 | dp->ricp_v6.dp_protocol_major = | |
218 | RDS_PROTOCOL_MAJOR(protocol_version); | |
219 | dp->ricp_v6.dp_protocol_minor = | |
220 | RDS_PROTOCOL_MINOR(protocol_version); | |
221 | dp->ricp_v6.dp_protocol_minor_mask = | |
222 | cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS); | |
223 | dp->ricp_v6.dp_ack_seq = | |
224 | cpu_to_be64(rds_ib_piggyb_ack(ic)); | |
fd261ce6 | 225 | dp->ricp_v6.dp_cmn.ricpc_dp_toss = conn->c_tos; |
eee2fa6a KCP |
226 | |
227 | conn_param->private_data = &dp->ricp_v6; | |
228 | conn_param->private_data_len = sizeof(dp->ricp_v6); | |
229 | } else { | |
230 | dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3]; | |
231 | dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3]; | |
232 | dp->ricp_v4.dp_protocol_major = | |
233 | RDS_PROTOCOL_MAJOR(protocol_version); | |
234 | dp->ricp_v4.dp_protocol_minor = | |
235 | RDS_PROTOCOL_MINOR(protocol_version); | |
236 | dp->ricp_v4.dp_protocol_minor_mask = | |
237 | cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS); | |
238 | dp->ricp_v4.dp_ack_seq = | |
239 | cpu_to_be64(rds_ib_piggyb_ack(ic)); | |
fd261ce6 | 240 | dp->ricp_v4.dp_cmn.ricpc_dp_toss = conn->c_tos; |
eee2fa6a KCP |
241 | |
242 | conn_param->private_data = &dp->ricp_v4; | |
243 | conn_param->private_data_len = sizeof(dp->ricp_v4); | |
244 | } | |
ec16227e AG |
245 | |
246 | /* Advertise flow control */ | |
247 | if (ic->i_flowctl) { | |
248 | unsigned int credits; | |
249 | ||
eee2fa6a KCP |
250 | credits = IB_GET_POST_CREDITS |
251 | (atomic_read(&ic->i_credits)); | |
252 | if (isv6) | |
253 | dp->ricp_v6.dp_credit = cpu_to_be32(credits); | |
254 | else | |
255 | dp->ricp_v4.dp_credit = cpu_to_be32(credits); | |
256 | atomic_sub(IB_SET_POST_CREDITS(credits), | |
257 | &ic->i_credits); | |
ec16227e | 258 | } |
ec16227e AG |
259 | } |
260 | } | |
261 | ||
262 | static void rds_ib_cq_event_handler(struct ib_event *event, void *data) | |
263 | { | |
1bde04a6 | 264 | rdsdebug("event %u (%s) data %p\n", |
3c88f3dc | 265 | event->event, ib_event_msg(event->event), data); |
ec16227e AG |
266 | } |
267 | ||
f4f943c9 SS |
268 | /* Plucking the oldest entry from the ring can be done concurrently with |
269 | * the thread refilling the ring. Each ring operation is protected by | |
270 | * spinlocks and the transient state of refilling doesn't change the | |
271 | * recording of which entry is oldest. | |
272 | * | |
273 | * This relies on IB only calling one cq comp_handler for each cq so that | |
274 | * there will only be one caller of rds_recv_incoming() per RDS connection. | |
275 | */ | |
276 | static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context) | |
277 | { | |
278 | struct rds_connection *conn = context; | |
279 | struct rds_ib_connection *ic = conn->c_transport_data; | |
280 | ||
281 | rdsdebug("conn %p cq %p\n", conn, cq); | |
282 | ||
283 | rds_ib_stats_inc(s_ib_evt_handler_call); | |
284 | ||
285 | tasklet_schedule(&ic->i_recv_tasklet); | |
286 | } | |
287 | ||
dcfd041c | 288 | static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq, |
289 | struct ib_wc *wcs) | |
f4f943c9 | 290 | { |
dcfd041c | 291 | int nr, i; |
f4f943c9 SS |
292 | struct ib_wc *wc; |
293 | ||
294 | while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) { | |
295 | for (i = 0; i < nr; i++) { | |
296 | wc = wcs + i; | |
297 | rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", | |
298 | (unsigned long long)wc->wr_id, wc->status, | |
299 | wc->byte_len, be32_to_cpu(wc->ex.imm_data)); | |
0c28c045 | 300 | |
1659185f AR |
301 | if (wc->wr_id <= ic->i_send_ring.w_nr || |
302 | wc->wr_id == RDS_IB_ACK_WR_ID) | |
303 | rds_ib_send_cqe_handler(ic, wc); | |
304 | else | |
305 | rds_ib_mr_cqe_handler(ic, wc); | |
306 | ||
f4f943c9 SS |
307 | } |
308 | } | |
309 | } | |
310 | ||
0c28c045 SS |
311 | static void rds_ib_tasklet_fn_send(unsigned long data) |
312 | { | |
313 | struct rds_ib_connection *ic = (struct rds_ib_connection *)data; | |
314 | struct rds_connection *conn = ic->conn; | |
0c28c045 SS |
315 | |
316 | rds_ib_stats_inc(s_ib_tasklet_call); | |
317 | ||
cf657269 SS |
318 | /* if cq has been already reaped, ignore incoming cq event */ |
319 | if (atomic_read(&ic->i_cq_quiesce)) | |
320 | return; | |
321 | ||
dcfd041c | 322 | poll_scq(ic, ic->i_send_cq, ic->i_send_wc); |
0c28c045 | 323 | ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); |
dcfd041c | 324 | poll_scq(ic, ic->i_send_cq, ic->i_send_wc); |
0c28c045 SS |
325 | |
326 | if (rds_conn_up(conn) && | |
327 | (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) || | |
328 | test_bit(0, &conn->c_map_queued))) | |
1f9ecd7e | 329 | rds_send_xmit(&ic->conn->c_path[0]); |
0c28c045 SS |
330 | } |
331 | ||
dcfd041c | 332 | static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq, |
333 | struct ib_wc *wcs, | |
334 | struct rds_ib_ack_state *ack_state) | |
335 | { | |
336 | int nr, i; | |
337 | struct ib_wc *wc; | |
338 | ||
339 | while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) { | |
340 | for (i = 0; i < nr; i++) { | |
341 | wc = wcs + i; | |
342 | rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", | |
343 | (unsigned long long)wc->wr_id, wc->status, | |
344 | wc->byte_len, be32_to_cpu(wc->ex.imm_data)); | |
345 | ||
346 | rds_ib_recv_cqe_handler(ic, wc, ack_state); | |
347 | } | |
348 | } | |
349 | } | |
350 | ||
f4f943c9 SS |
351 | static void rds_ib_tasklet_fn_recv(unsigned long data) |
352 | { | |
353 | struct rds_ib_connection *ic = (struct rds_ib_connection *)data; | |
354 | struct rds_connection *conn = ic->conn; | |
355 | struct rds_ib_device *rds_ibdev = ic->rds_ibdev; | |
356 | struct rds_ib_ack_state state; | |
357 | ||
9441c973 SS |
358 | if (!rds_ibdev) |
359 | rds_conn_drop(conn); | |
f4f943c9 SS |
360 | |
361 | rds_ib_stats_inc(s_ib_tasklet_call); | |
362 | ||
cf657269 SS |
363 | /* if cq has been already reaped, ignore incoming cq event */ |
364 | if (atomic_read(&ic->i_cq_quiesce)) | |
365 | return; | |
366 | ||
f4f943c9 | 367 | memset(&state, 0, sizeof(state)); |
dcfd041c | 368 | poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state); |
f4f943c9 | 369 | ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); |
dcfd041c | 370 | poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state); |
f4f943c9 SS |
371 | |
372 | if (state.ack_next_valid) | |
373 | rds_ib_set_ack(ic, state.ack_next, state.ack_required); | |
374 | if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) { | |
375 | rds_send_drop_acked(conn, state.ack_recv, NULL); | |
376 | ic->i_ack_recv = state.ack_recv; | |
377 | } | |
378 | ||
379 | if (rds_conn_up(conn)) | |
380 | rds_ib_attempt_ack(ic); | |
381 | } | |
382 | ||
ec16227e AG |
383 | static void rds_ib_qp_event_handler(struct ib_event *event, void *data) |
384 | { | |
385 | struct rds_connection *conn = data; | |
386 | struct rds_ib_connection *ic = conn->c_transport_data; | |
387 | ||
1bde04a6 | 388 | rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event, |
3c88f3dc | 389 | ib_event_msg(event->event)); |
ec16227e AG |
390 | |
391 | switch (event->event) { | |
392 | case IB_EVENT_COMM_EST: | |
393 | rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST); | |
394 | break; | |
395 | default: | |
fd261ce6 SS |
396 | rdsdebug("Fatal QP Event %u (%s) - connection %pI6c->%pI6c, reconnecting\n", |
397 | event->event, ib_event_msg(event->event), | |
398 | &conn->c_laddr, &conn->c_faddr); | |
97069788 | 399 | rds_conn_drop(conn); |
ec16227e AG |
400 | break; |
401 | } | |
402 | } | |
403 | ||
0c28c045 SS |
404 | static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context) |
405 | { | |
406 | struct rds_connection *conn = context; | |
407 | struct rds_ib_connection *ic = conn->c_transport_data; | |
408 | ||
409 | rdsdebug("conn %p cq %p\n", conn, cq); | |
410 | ||
411 | rds_ib_stats_inc(s_ib_evt_handler_call); | |
412 | ||
413 | tasklet_schedule(&ic->i_send_tasklet); | |
414 | } | |
415 | ||
be2f76ea SS |
416 | static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev) |
417 | { | |
418 | int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1]; | |
419 | int index = rds_ibdev->dev->num_comp_vectors - 1; | |
420 | int i; | |
421 | ||
422 | for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) { | |
423 | if (rds_ibdev->vector_load[i] < min) { | |
424 | index = i; | |
425 | min = rds_ibdev->vector_load[i]; | |
426 | } | |
427 | } | |
428 | ||
429 | rds_ibdev->vector_load[index]++; | |
430 | return index; | |
431 | } | |
432 | ||
433 | static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index) | |
434 | { | |
435 | rds_ibdev->vector_load[index]--; | |
436 | } | |
437 | ||
ec16227e AG |
438 | /* |
439 | * This needs to be very careful to not leave IS_ERR pointers around for | |
440 | * cleanup to trip over. | |
441 | */ | |
442 | static int rds_ib_setup_qp(struct rds_connection *conn) | |
443 | { | |
444 | struct rds_ib_connection *ic = conn->c_transport_data; | |
445 | struct ib_device *dev = ic->i_cm_id->device; | |
446 | struct ib_qp_init_attr attr; | |
8e37210b | 447 | struct ib_cq_init_attr cq_attr = {}; |
ec16227e | 448 | struct rds_ib_device *rds_ibdev; |
ad6832f9 | 449 | int ret, fr_queue_space; |
ec16227e | 450 | |
3e0249f9 ZB |
451 | /* |
452 | * It's normal to see a null device if an incoming connection races | |
453 | * with device removal, so we don't print a warning. | |
ec16227e | 454 | */ |
3e0249f9 ZB |
455 | rds_ibdev = rds_ib_get_client_data(dev); |
456 | if (!rds_ibdev) | |
ec16227e | 457 | return -EOPNOTSUPP; |
3e0249f9 | 458 | |
ad6832f9 | 459 | /* The fr_queue_space is currently set to 512, to add extra space on |
460 | * completion queue and send queue. This extra space is used for FRMR | |
461 | * registration and invalidation work requests | |
462 | */ | |
56012459 SS |
463 | fr_queue_space = rds_ibdev->use_fastreg ? |
464 | (RDS_IB_DEFAULT_FR_WR + 1) + | |
465 | (RDS_IB_DEFAULT_FR_INV_WR + 1) | |
466 | : 0; | |
ad6832f9 | 467 | |
3e0249f9 ZB |
468 | /* add the conn now so that connection establishment has the dev */ |
469 | rds_ib_add_conn(rds_ibdev, conn); | |
ec16227e AG |
470 | |
471 | if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1) | |
472 | rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1); | |
473 | if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1) | |
474 | rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1); | |
475 | ||
476 | /* Protection domain and memory range */ | |
477 | ic->i_pd = rds_ibdev->pd; | |
ec16227e | 478 | |
be2f76ea | 479 | ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev); |
ad6832f9 | 480 | cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1; |
be2f76ea | 481 | cq_attr.comp_vector = ic->i_scq_vector; |
0c28c045 | 482 | ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send, |
ec16227e | 483 | rds_ib_cq_event_handler, conn, |
8e37210b | 484 | &cq_attr); |
ec16227e AG |
485 | if (IS_ERR(ic->i_send_cq)) { |
486 | ret = PTR_ERR(ic->i_send_cq); | |
487 | ic->i_send_cq = NULL; | |
be2f76ea | 488 | ibdev_put_vector(rds_ibdev, ic->i_scq_vector); |
ec16227e | 489 | rdsdebug("ib_create_cq send failed: %d\n", ret); |
3b12f73a | 490 | goto rds_ibdev_out; |
ec16227e AG |
491 | } |
492 | ||
be2f76ea | 493 | ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev); |
8e37210b | 494 | cq_attr.cqe = ic->i_recv_ring.w_nr; |
be2f76ea | 495 | cq_attr.comp_vector = ic->i_rcq_vector; |
f4f943c9 | 496 | ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv, |
ec16227e | 497 | rds_ib_cq_event_handler, conn, |
8e37210b | 498 | &cq_attr); |
ec16227e AG |
499 | if (IS_ERR(ic->i_recv_cq)) { |
500 | ret = PTR_ERR(ic->i_recv_cq); | |
501 | ic->i_recv_cq = NULL; | |
be2f76ea | 502 | ibdev_put_vector(rds_ibdev, ic->i_rcq_vector); |
ec16227e | 503 | rdsdebug("ib_create_cq recv failed: %d\n", ret); |
3b12f73a | 504 | goto send_cq_out; |
ec16227e AG |
505 | } |
506 | ||
507 | ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); | |
508 | if (ret) { | |
509 | rdsdebug("ib_req_notify_cq send failed: %d\n", ret); | |
3b12f73a | 510 | goto recv_cq_out; |
ec16227e AG |
511 | } |
512 | ||
513 | ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); | |
514 | if (ret) { | |
515 | rdsdebug("ib_req_notify_cq recv failed: %d\n", ret); | |
3b12f73a | 516 | goto recv_cq_out; |
ec16227e AG |
517 | } |
518 | ||
519 | /* XXX negotiate max send/recv with remote? */ | |
520 | memset(&attr, 0, sizeof(attr)); | |
521 | attr.event_handler = rds_ib_qp_event_handler; | |
522 | attr.qp_context = conn; | |
523 | /* + 1 to allow for the single ack message */ | |
ad6832f9 | 524 | attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1; |
ec16227e AG |
525 | attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1; |
526 | attr.cap.max_send_sge = rds_ibdev->max_sge; | |
527 | attr.cap.max_recv_sge = RDS_IB_RECV_SGE; | |
528 | attr.sq_sig_type = IB_SIGNAL_REQ_WR; | |
529 | attr.qp_type = IB_QPT_RC; | |
530 | attr.send_cq = ic->i_send_cq; | |
531 | attr.recv_cq = ic->i_recv_cq; | |
ad6832f9 | 532 | atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR); |
56012459 | 533 | atomic_set(&ic->i_fastunreg_wrs, RDS_IB_DEFAULT_FR_INV_WR); |
ec16227e AG |
534 | |
535 | /* | |
536 | * XXX this can fail if max_*_wr is too large? Are we supposed | |
537 | * to back off until we get a value that the hardware can support? | |
538 | */ | |
539 | ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr); | |
540 | if (ret) { | |
541 | rdsdebug("rdma_create_qp failed: %d\n", ret); | |
3b12f73a | 542 | goto recv_cq_out; |
ec16227e AG |
543 | } |
544 | ||
545 | ic->i_send_hdrs = ib_dma_alloc_coherent(dev, | |
546 | ic->i_send_ring.w_nr * | |
547 | sizeof(struct rds_header), | |
548 | &ic->i_send_hdrs_dma, GFP_KERNEL); | |
8690bfa1 | 549 | if (!ic->i_send_hdrs) { |
ec16227e AG |
550 | ret = -ENOMEM; |
551 | rdsdebug("ib_dma_alloc_coherent send failed\n"); | |
3b12f73a | 552 | goto qp_out; |
ec16227e AG |
553 | } |
554 | ||
555 | ic->i_recv_hdrs = ib_dma_alloc_coherent(dev, | |
556 | ic->i_recv_ring.w_nr * | |
557 | sizeof(struct rds_header), | |
558 | &ic->i_recv_hdrs_dma, GFP_KERNEL); | |
8690bfa1 | 559 | if (!ic->i_recv_hdrs) { |
ec16227e AG |
560 | ret = -ENOMEM; |
561 | rdsdebug("ib_dma_alloc_coherent recv failed\n"); | |
3b12f73a | 562 | goto send_hdrs_dma_out; |
ec16227e AG |
563 | } |
564 | ||
565 | ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header), | |
566 | &ic->i_ack_dma, GFP_KERNEL); | |
8690bfa1 | 567 | if (!ic->i_ack) { |
ec16227e AG |
568 | ret = -ENOMEM; |
569 | rdsdebug("ib_dma_alloc_coherent ack failed\n"); | |
3b12f73a | 570 | goto recv_hdrs_dma_out; |
ec16227e AG |
571 | } |
572 | ||
fd7beced KC |
573 | ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work), |
574 | ic->i_send_ring.w_nr), | |
e4c52c98 | 575 | ibdev_to_node(dev)); |
8690bfa1 | 576 | if (!ic->i_sends) { |
ec16227e AG |
577 | ret = -ENOMEM; |
578 | rdsdebug("send allocation failed\n"); | |
3b12f73a | 579 | goto ack_dma_out; |
ec16227e | 580 | } |
ec16227e | 581 | |
fd7beced KC |
582 | ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work), |
583 | ic->i_recv_ring.w_nr), | |
e4c52c98 | 584 | ibdev_to_node(dev)); |
8690bfa1 | 585 | if (!ic->i_recvs) { |
ec16227e AG |
586 | ret = -ENOMEM; |
587 | rdsdebug("recv allocation failed\n"); | |
3b12f73a | 588 | goto sends_out; |
ec16227e AG |
589 | } |
590 | ||
ec16227e AG |
591 | rds_ib_recv_init_ack(ic); |
592 | ||
e5580242 | 593 | rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd, |
ec16227e AG |
594 | ic->i_send_cq, ic->i_recv_cq); |
595 | ||
91a82529 | 596 | goto out; |
3b12f73a ZY |
597 | |
598 | sends_out: | |
599 | vfree(ic->i_sends); | |
600 | ack_dma_out: | |
601 | ib_dma_free_coherent(dev, sizeof(struct rds_header), | |
602 | ic->i_ack, ic->i_ack_dma); | |
603 | recv_hdrs_dma_out: | |
604 | ib_dma_free_coherent(dev, ic->i_recv_ring.w_nr * | |
605 | sizeof(struct rds_header), | |
606 | ic->i_recv_hdrs, ic->i_recv_hdrs_dma); | |
607 | send_hdrs_dma_out: | |
608 | ib_dma_free_coherent(dev, ic->i_send_ring.w_nr * | |
609 | sizeof(struct rds_header), | |
610 | ic->i_send_hdrs, ic->i_send_hdrs_dma); | |
611 | qp_out: | |
612 | rdma_destroy_qp(ic->i_cm_id); | |
613 | recv_cq_out: | |
614 | if (!ib_destroy_cq(ic->i_recv_cq)) | |
615 | ic->i_recv_cq = NULL; | |
616 | send_cq_out: | |
617 | if (!ib_destroy_cq(ic->i_send_cq)) | |
618 | ic->i_send_cq = NULL; | |
619 | rds_ibdev_out: | |
620 | rds_ib_remove_conn(rds_ibdev, conn); | |
91a82529 | 621 | out: |
3e0249f9 | 622 | rds_ib_dev_put(rds_ibdev); |
3b12f73a | 623 | |
ec16227e AG |
624 | return ret; |
625 | } | |
626 | ||
eee2fa6a | 627 | static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6) |
ec16227e | 628 | { |
eee2fa6a KCP |
629 | const union rds_ib_conn_priv *dp = event->param.conn.private_data; |
630 | u8 data_len, major, minor; | |
ec16227e | 631 | u32 version = 0; |
eee2fa6a KCP |
632 | __be16 mask; |
633 | u16 common; | |
ec16227e | 634 | |
9ddbcfa0 AG |
635 | /* |
636 | * rdma_cm private data is odd - when there is any private data in the | |
ec16227e AG |
637 | * request, we will be given a pretty large buffer without telling us the |
638 | * original size. The only way to tell the difference is by looking at | |
639 | * the contents, which are initialized to zero. | |
640 | * If the protocol version fields aren't set, this is a connection attempt | |
641 | * from an older version. This could could be 3.0 or 2.0 - we can't tell. | |
9ddbcfa0 AG |
642 | * We really should have changed this for OFED 1.3 :-( |
643 | */ | |
644 | ||
645 | /* Be paranoid. RDS always has privdata */ | |
646 | if (!event->param.conn.private_data_len) { | |
647 | printk(KERN_NOTICE "RDS incoming connection has no private data, " | |
648 | "rejecting\n"); | |
649 | return 0; | |
650 | } | |
651 | ||
eee2fa6a KCP |
652 | if (isv6) { |
653 | data_len = sizeof(struct rds6_ib_connect_private); | |
654 | major = dp->ricp_v6.dp_protocol_major; | |
655 | minor = dp->ricp_v6.dp_protocol_minor; | |
656 | mask = dp->ricp_v6.dp_protocol_minor_mask; | |
657 | } else { | |
658 | data_len = sizeof(struct rds_ib_connect_private); | |
659 | major = dp->ricp_v4.dp_protocol_major; | |
660 | minor = dp->ricp_v4.dp_protocol_minor; | |
661 | mask = dp->ricp_v4.dp_protocol_minor_mask; | |
662 | } | |
663 | ||
9ddbcfa0 | 664 | /* Even if len is crap *now* I still want to check it. -ASG */ |
eee2fa6a | 665 | if (event->param.conn.private_data_len < data_len || major == 0) |
fd261ce6 | 666 | return RDS_PROTOCOL_4_0; |
ec16227e | 667 | |
eee2fa6a | 668 | common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS; |
fd261ce6 SS |
669 | if (major == 4 && common) { |
670 | version = RDS_PROTOCOL_4_0; | |
ec16227e AG |
671 | while ((common >>= 1) != 0) |
672 | version++; | |
cdc306a5 SS |
673 | } else if (RDS_PROTOCOL_COMPAT_VERSION == |
674 | RDS_PROTOCOL(major, minor)) { | |
675 | version = RDS_PROTOCOL_COMPAT_VERSION; | |
eee2fa6a KCP |
676 | } else { |
677 | if (isv6) | |
678 | printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n", | |
679 | &dp->ricp_v6.dp_saddr, major, minor); | |
680 | else | |
681 | printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n", | |
682 | &dp->ricp_v4.dp_saddr, major, minor); | |
683 | } | |
ec16227e AG |
684 | return version; |
685 | } | |
686 | ||
e65d4d96 | 687 | #if IS_ENABLED(CONFIG_IPV6) |
1e2b44e7 | 688 | /* Given an IPv6 address, find the net_device which hosts that address and |
eee2fa6a KCP |
689 | * return its index. This is used by the rds_ib_cm_handle_connect() code to |
690 | * find the interface index of where an incoming request comes from when | |
691 | * the request is using a link local address. | |
692 | * | |
693 | * Note one problem in this search. It is possible that two interfaces have | |
694 | * the same link local address. Unfortunately, this cannot be solved unless | |
695 | * the underlying layer gives us the interface which an incoming RDMA connect | |
696 | * request comes from. | |
697 | */ | |
698 | static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr) | |
699 | { | |
700 | struct net_device *dev; | |
701 | int idx = 0; | |
702 | ||
703 | rcu_read_lock(); | |
704 | for_each_netdev_rcu(net, dev) { | |
1e2b44e7 | 705 | if (ipv6_chk_addr(net, addr, dev, 1)) { |
eee2fa6a KCP |
706 | idx = dev->ifindex; |
707 | break; | |
708 | } | |
709 | } | |
710 | rcu_read_unlock(); | |
711 | ||
712 | return idx; | |
713 | } | |
e65d4d96 | 714 | #endif |
eee2fa6a | 715 | |
ec16227e | 716 | int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, |
eee2fa6a | 717 | struct rdma_cm_event *event, bool isv6) |
ec16227e AG |
718 | { |
719 | __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id; | |
720 | __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id; | |
eee2fa6a | 721 | const struct rds_ib_conn_priv_cmn *dp_cmn; |
ec16227e AG |
722 | struct rds_connection *conn = NULL; |
723 | struct rds_ib_connection *ic = NULL; | |
724 | struct rdma_conn_param conn_param; | |
eee2fa6a KCP |
725 | const union rds_ib_conn_priv *dp; |
726 | union rds_ib_conn_priv dp_rep; | |
727 | struct in6_addr s_mapped_addr; | |
728 | struct in6_addr d_mapped_addr; | |
729 | const struct in6_addr *saddr6; | |
730 | const struct in6_addr *daddr6; | |
731 | int destroy = 1; | |
732 | u32 ifindex = 0; | |
ec16227e | 733 | u32 version; |
eee2fa6a | 734 | int err = 1; |
ec16227e AG |
735 | |
736 | /* Check whether the remote protocol version matches ours. */ | |
eee2fa6a | 737 | version = rds_ib_protocol_compatible(event, isv6); |
d021fabf SS |
738 | if (!version) { |
739 | err = RDS_RDMA_REJ_INCOMPAT; | |
ec16227e | 740 | goto out; |
d021fabf | 741 | } |
ec16227e | 742 | |
eee2fa6a KCP |
743 | dp = event->param.conn.private_data; |
744 | if (isv6) { | |
e65d4d96 | 745 | #if IS_ENABLED(CONFIG_IPV6) |
eee2fa6a KCP |
746 | dp_cmn = &dp->ricp_v6.dp_cmn; |
747 | saddr6 = &dp->ricp_v6.dp_saddr; | |
748 | daddr6 = &dp->ricp_v6.dp_daddr; | |
1e2b44e7 | 749 | /* If either address is link local, need to find the |
eee2fa6a KCP |
750 | * interface index in order to create a proper RDS |
751 | * connection. | |
752 | */ | |
753 | if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) { | |
754 | /* Using init_net for now .. */ | |
755 | ifindex = __rds_find_ifindex(&init_net, daddr6); | |
756 | /* No index found... Need to bail out. */ | |
757 | if (ifindex == 0) { | |
758 | err = -EOPNOTSUPP; | |
759 | goto out; | |
760 | } | |
1e2b44e7 KCP |
761 | } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) { |
762 | /* Use our address to find the correct index. */ | |
763 | ifindex = __rds_find_ifindex(&init_net, daddr6); | |
764 | /* No index found... Need to bail out. */ | |
765 | if (ifindex == 0) { | |
766 | err = -EOPNOTSUPP; | |
767 | goto out; | |
768 | } | |
eee2fa6a | 769 | } |
e65d4d96 KCP |
770 | #else |
771 | err = -EOPNOTSUPP; | |
772 | goto out; | |
773 | #endif | |
eee2fa6a KCP |
774 | } else { |
775 | dp_cmn = &dp->ricp_v4.dp_cmn; | |
776 | ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr); | |
777 | ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr); | |
778 | saddr6 = &s_mapped_addr; | |
779 | daddr6 = &d_mapped_addr; | |
780 | } | |
781 | ||
fd261ce6 SS |
782 | rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid 0x%llx, tos:%d\n", |
783 | saddr6, daddr6, RDS_PROTOCOL_MAJOR(version), | |
784 | RDS_PROTOCOL_MINOR(version), | |
ec16227e | 785 | (unsigned long long)be64_to_cpu(lguid), |
fd261ce6 | 786 | (unsigned long long)be64_to_cpu(fguid), dp_cmn->ricpc_dp_toss); |
ec16227e | 787 | |
d5a8ac28 | 788 | /* RDS/IB is not currently netns aware, thus init_net */ |
eee2fa6a | 789 | conn = rds_conn_create(&init_net, daddr6, saddr6, |
fd261ce6 SS |
790 | &rds_ib_transport, dp_cmn->ricpc_dp_toss, |
791 | GFP_KERNEL, ifindex); | |
ec16227e AG |
792 | if (IS_ERR(conn)) { |
793 | rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn)); | |
794 | conn = NULL; | |
795 | goto out; | |
796 | } | |
797 | ||
798 | /* | |
799 | * The connection request may occur while the | |
800 | * previous connection exist, e.g. in case of failover. | |
801 | * But as connections may be initiated simultaneously | |
802 | * by both hosts, we have a random backoff mechanism - | |
803 | * see the comment above rds_queue_reconnect() | |
804 | */ | |
805 | mutex_lock(&conn->c_cm_lock); | |
806 | if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) { | |
807 | if (rds_conn_state(conn) == RDS_CONN_UP) { | |
808 | rdsdebug("incoming connect while connecting\n"); | |
809 | rds_conn_drop(conn); | |
810 | rds_ib_stats_inc(s_ib_listen_closed_stale); | |
811 | } else | |
812 | if (rds_conn_state(conn) == RDS_CONN_CONNECTING) { | |
813 | /* Wait and see - our connect may still be succeeding */ | |
814 | rds_ib_stats_inc(s_ib_connect_raced); | |
815 | } | |
ec16227e AG |
816 | goto out; |
817 | } | |
818 | ||
819 | ic = conn->c_transport_data; | |
820 | ||
821 | rds_ib_set_protocol(conn, version); | |
eee2fa6a | 822 | rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit)); |
ec16227e AG |
823 | |
824 | /* If the peer gave us the last packet it saw, process this as if | |
825 | * we had received a regular ACK. */ | |
eee2fa6a KCP |
826 | if (dp_cmn->ricpc_ack_seq) |
827 | rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq), | |
828 | NULL); | |
ec16227e AG |
829 | |
830 | BUG_ON(cm_id->context); | |
831 | BUG_ON(ic->i_cm_id); | |
832 | ||
833 | ic->i_cm_id = cm_id; | |
834 | cm_id->context = conn; | |
835 | ||
836 | /* We got halfway through setting up the ib_connection, if we | |
837 | * fail now, we have to take the long route out of this mess. */ | |
838 | destroy = 0; | |
839 | ||
840 | err = rds_ib_setup_qp(conn); | |
841 | if (err) { | |
842 | rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err); | |
843 | goto out; | |
844 | } | |
845 | ||
40589e74 | 846 | rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version, |
eee2fa6a KCP |
847 | event->param.conn.responder_resources, |
848 | event->param.conn.initiator_depth, isv6); | |
ec16227e AG |
849 | |
850 | /* rdma_accept() calls rdma_reject() internally if it fails */ | |
b418c527 ZY |
851 | if (rdma_accept(cm_id, &conn_param)) |
852 | rds_ib_conn_error(conn, "rdma_accept failed\n"); | |
ec16227e AG |
853 | |
854 | out: | |
a46ca94e ZB |
855 | if (conn) |
856 | mutex_unlock(&conn->c_cm_lock); | |
857 | if (err) | |
d021fabf | 858 | rdma_reject(cm_id, &err, sizeof(int)); |
ec16227e AG |
859 | return destroy; |
860 | } | |
861 | ||
862 | ||
eee2fa6a | 863 | int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6) |
ec16227e AG |
864 | { |
865 | struct rds_connection *conn = cm_id->context; | |
866 | struct rds_ib_connection *ic = conn->c_transport_data; | |
867 | struct rdma_conn_param conn_param; | |
eee2fa6a | 868 | union rds_ib_conn_priv dp; |
ec16227e AG |
869 | int ret; |
870 | ||
871 | /* If the peer doesn't do protocol negotiation, we must | |
872 | * default to RDSv3.0 */ | |
fd261ce6 | 873 | rds_ib_set_protocol(conn, RDS_PROTOCOL_4_1); |
ec16227e AG |
874 | ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */ |
875 | ||
876 | ret = rds_ib_setup_qp(conn); | |
877 | if (ret) { | |
878 | rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret); | |
879 | goto out; | |
880 | } | |
881 | ||
cdc306a5 SS |
882 | rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, |
883 | conn->c_proposed_version, | |
eee2fa6a | 884 | UINT_MAX, UINT_MAX, isv6); |
ec16227e AG |
885 | ret = rdma_connect(cm_id, &conn_param); |
886 | if (ret) | |
887 | rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret); | |
888 | ||
889 | out: | |
890 | /* Beware - returning non-zero tells the rdma_cm to destroy | |
891 | * the cm_id. We should certainly not do it as long as we still | |
892 | * "own" the cm_id. */ | |
893 | if (ret) { | |
894 | if (ic->i_cm_id == cm_id) | |
895 | ret = 0; | |
896 | } | |
581d53c9 | 897 | ic->i_active_side = true; |
ec16227e AG |
898 | return ret; |
899 | } | |
900 | ||
b04e8554 | 901 | int rds_ib_conn_path_connect(struct rds_conn_path *cp) |
ec16227e | 902 | { |
b04e8554 | 903 | struct rds_connection *conn = cp->cp_conn; |
eee2fa6a KCP |
904 | struct sockaddr_storage src, dest; |
905 | rdma_cm_event_handler handler; | |
906 | struct rds_ib_connection *ic; | |
ec16227e AG |
907 | int ret; |
908 | ||
eee2fa6a KCP |
909 | ic = conn->c_transport_data; |
910 | ||
ec16227e AG |
911 | /* XXX I wonder what affect the port space has */ |
912 | /* delegate cm event handler to rdma_transport */ | |
e65d4d96 | 913 | #if IS_ENABLED(CONFIG_IPV6) |
1e2b44e7 KCP |
914 | if (conn->c_isv6) |
915 | handler = rds6_rdma_cm_event_handler; | |
916 | else | |
e65d4d96 | 917 | #endif |
1e2b44e7 | 918 | handler = rds_rdma_cm_event_handler; |
eee2fa6a | 919 | ic->i_cm_id = rdma_create_id(&init_net, handler, conn, |
b26f9b99 | 920 | RDMA_PS_TCP, IB_QPT_RC); |
ec16227e AG |
921 | if (IS_ERR(ic->i_cm_id)) { |
922 | ret = PTR_ERR(ic->i_cm_id); | |
923 | ic->i_cm_id = NULL; | |
924 | rdsdebug("rdma_create_id() failed: %d\n", ret); | |
925 | goto out; | |
926 | } | |
927 | ||
928 | rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn); | |
929 | ||
eee2fa6a KCP |
930 | if (ipv6_addr_v4mapped(&conn->c_faddr)) { |
931 | struct sockaddr_in *sin; | |
932 | ||
933 | sin = (struct sockaddr_in *)&src; | |
934 | sin->sin_family = AF_INET; | |
935 | sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3]; | |
936 | sin->sin_port = 0; | |
ec16227e | 937 | |
eee2fa6a KCP |
938 | sin = (struct sockaddr_in *)&dest; |
939 | sin->sin_family = AF_INET; | |
940 | sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3]; | |
941 | sin->sin_port = htons(RDS_PORT); | |
942 | } else { | |
943 | struct sockaddr_in6 *sin6; | |
944 | ||
945 | sin6 = (struct sockaddr_in6 *)&src; | |
946 | sin6->sin6_family = AF_INET6; | |
947 | sin6->sin6_addr = conn->c_laddr; | |
948 | sin6->sin6_port = 0; | |
949 | sin6->sin6_scope_id = conn->c_dev_if; | |
950 | ||
951 | sin6 = (struct sockaddr_in6 *)&dest; | |
952 | sin6->sin6_family = AF_INET6; | |
953 | sin6->sin6_addr = conn->c_faddr; | |
954 | sin6->sin6_port = htons(RDS_CM_PORT); | |
955 | sin6->sin6_scope_id = conn->c_dev_if; | |
956 | } | |
ec16227e AG |
957 | |
958 | ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src, | |
959 | (struct sockaddr *)&dest, | |
960 | RDS_RDMA_RESOLVE_TIMEOUT_MS); | |
961 | if (ret) { | |
962 | rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id, | |
963 | ret); | |
964 | rdma_destroy_id(ic->i_cm_id); | |
965 | ic->i_cm_id = NULL; | |
966 | } | |
967 | ||
968 | out: | |
969 | return ret; | |
970 | } | |
971 | ||
972 | /* | |
973 | * This is so careful about only cleaning up resources that were built up | |
974 | * so that it can be called at any point during startup. In fact it | |
975 | * can be called multiple times for a given connection. | |
976 | */ | |
226f7a7d | 977 | void rds_ib_conn_path_shutdown(struct rds_conn_path *cp) |
ec16227e | 978 | { |
226f7a7d | 979 | struct rds_connection *conn = cp->cp_conn; |
ec16227e AG |
980 | struct rds_ib_connection *ic = conn->c_transport_data; |
981 | int err = 0; | |
982 | ||
983 | rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id, | |
984 | ic->i_pd, ic->i_send_cq, ic->i_recv_cq, | |
985 | ic->i_cm_id ? ic->i_cm_id->qp : NULL); | |
986 | ||
987 | if (ic->i_cm_id) { | |
988 | struct ib_device *dev = ic->i_cm_id->device; | |
989 | ||
990 | rdsdebug("disconnecting cm %p\n", ic->i_cm_id); | |
991 | err = rdma_disconnect(ic->i_cm_id); | |
992 | if (err) { | |
993 | /* Actually this may happen quite frequently, when | |
994 | * an outgoing connect raced with an incoming connect. | |
995 | */ | |
996 | rdsdebug("failed to disconnect, cm: %p err %d\n", | |
997 | ic->i_cm_id, err); | |
998 | } | |
999 | ||
e32b4a70 | 1000 | /* |
f046011c ZB |
1001 | * We want to wait for tx and rx completion to finish |
1002 | * before we tear down the connection, but we have to be | |
1003 | * careful not to get stuck waiting on a send ring that | |
1004 | * only has unsignaled sends in it. We've shutdown new | |
1005 | * sends before getting here so by waiting for signaled | |
1006 | * sends to complete we're ensured that there will be no | |
1007 | * more tx processing. | |
e32b4a70 | 1008 | */ |
ec16227e | 1009 | wait_event(rds_ib_ring_empty_wait, |
f046011c | 1010 | rds_ib_ring_empty(&ic->i_recv_ring) && |
ad6832f9 | 1011 | (atomic_read(&ic->i_signaled_sends) == 0) && |
56012459 SS |
1012 | (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR) && |
1013 | (atomic_read(&ic->i_fastunreg_wrs) == RDS_IB_DEFAULT_FR_INV_WR)); | |
0c28c045 | 1014 | tasklet_kill(&ic->i_send_tasklet); |
f046011c | 1015 | tasklet_kill(&ic->i_recv_tasklet); |
ec16227e | 1016 | |
cf657269 SS |
1017 | atomic_set(&ic->i_cq_quiesce, 1); |
1018 | ||
1bc7b863 | 1019 | /* first destroy the ib state that generates callbacks */ |
1020 | if (ic->i_cm_id->qp) | |
1021 | rdma_destroy_qp(ic->i_cm_id); | |
be2f76ea SS |
1022 | if (ic->i_send_cq) { |
1023 | if (ic->rds_ibdev) | |
1024 | ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector); | |
1bc7b863 | 1025 | ib_destroy_cq(ic->i_send_cq); |
be2f76ea SS |
1026 | } |
1027 | ||
1028 | if (ic->i_recv_cq) { | |
1029 | if (ic->rds_ibdev) | |
1030 | ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector); | |
1bc7b863 | 1031 | ib_destroy_cq(ic->i_recv_cq); |
be2f76ea | 1032 | } |
1bc7b863 | 1033 | |
1034 | /* then free the resources that ib callbacks use */ | |
ec16227e AG |
1035 | if (ic->i_send_hdrs) |
1036 | ib_dma_free_coherent(dev, | |
1037 | ic->i_send_ring.w_nr * | |
1038 | sizeof(struct rds_header), | |
1039 | ic->i_send_hdrs, | |
1040 | ic->i_send_hdrs_dma); | |
1041 | ||
1042 | if (ic->i_recv_hdrs) | |
1043 | ib_dma_free_coherent(dev, | |
1044 | ic->i_recv_ring.w_nr * | |
1045 | sizeof(struct rds_header), | |
1046 | ic->i_recv_hdrs, | |
1047 | ic->i_recv_hdrs_dma); | |
1048 | ||
1049 | if (ic->i_ack) | |
1050 | ib_dma_free_coherent(dev, sizeof(struct rds_header), | |
1051 | ic->i_ack, ic->i_ack_dma); | |
1052 | ||
1053 | if (ic->i_sends) | |
1054 | rds_ib_send_clear_ring(ic); | |
1055 | if (ic->i_recvs) | |
1056 | rds_ib_recv_clear_ring(ic); | |
1057 | ||
1c3be624 SS |
1058 | rdma_destroy_id(ic->i_cm_id); |
1059 | ||
ec16227e AG |
1060 | /* |
1061 | * Move connection back to the nodev list. | |
1062 | */ | |
745cbcca AG |
1063 | if (ic->rds_ibdev) |
1064 | rds_ib_remove_conn(ic->rds_ibdev, conn); | |
ec16227e AG |
1065 | |
1066 | ic->i_cm_id = NULL; | |
1067 | ic->i_pd = NULL; | |
ec16227e AG |
1068 | ic->i_send_cq = NULL; |
1069 | ic->i_recv_cq = NULL; | |
1070 | ic->i_send_hdrs = NULL; | |
1071 | ic->i_recv_hdrs = NULL; | |
1072 | ic->i_ack = NULL; | |
1073 | } | |
1074 | BUG_ON(ic->rds_ibdev); | |
1075 | ||
1076 | /* Clear pending transmit */ | |
ff3d7d36 AG |
1077 | if (ic->i_data_op) { |
1078 | struct rds_message *rm; | |
1079 | ||
1080 | rm = container_of(ic->i_data_op, struct rds_message, data); | |
1081 | rds_message_put(rm); | |
1082 | ic->i_data_op = NULL; | |
ec16227e AG |
1083 | } |
1084 | ||
1085 | /* Clear the ACK state */ | |
1086 | clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); | |
8cbd9606 AG |
1087 | #ifdef KERNEL_HAS_ATOMIC64 |
1088 | atomic64_set(&ic->i_ack_next, 0); | |
1089 | #else | |
1090 | ic->i_ack_next = 0; | |
1091 | #endif | |
ec16227e AG |
1092 | ic->i_ack_recv = 0; |
1093 | ||
1094 | /* Clear flow control state */ | |
1095 | ic->i_flowctl = 0; | |
1096 | atomic_set(&ic->i_credits, 0); | |
1097 | ||
1098 | rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); | |
1099 | rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); | |
1100 | ||
1101 | if (ic->i_ibinc) { | |
1102 | rds_inc_put(&ic->i_ibinc->ii_inc); | |
1103 | ic->i_ibinc = NULL; | |
1104 | } | |
1105 | ||
1106 | vfree(ic->i_sends); | |
1107 | ic->i_sends = NULL; | |
1108 | vfree(ic->i_recvs); | |
1109 | ic->i_recvs = NULL; | |
581d53c9 | 1110 | ic->i_active_side = false; |
ec16227e AG |
1111 | } |
1112 | ||
1113 | int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp) | |
1114 | { | |
1115 | struct rds_ib_connection *ic; | |
1116 | unsigned long flags; | |
33244125 | 1117 | int ret; |
ec16227e AG |
1118 | |
1119 | /* XXX too lazy? */ | |
f0229eaa | 1120 | ic = kzalloc(sizeof(struct rds_ib_connection), gfp); |
8690bfa1 | 1121 | if (!ic) |
ec16227e AG |
1122 | return -ENOMEM; |
1123 | ||
f394ad28 | 1124 | ret = rds_ib_recv_alloc_caches(ic, gfp); |
33244125 CM |
1125 | if (ret) { |
1126 | kfree(ic); | |
1127 | return ret; | |
1128 | } | |
1129 | ||
ec16227e | 1130 | INIT_LIST_HEAD(&ic->ib_node); |
0c28c045 SS |
1131 | tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send, |
1132 | (unsigned long)ic); | |
f4f943c9 | 1133 | tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv, |
0c28c045 | 1134 | (unsigned long)ic); |
ec16227e | 1135 | mutex_init(&ic->i_recv_mutex); |
8cbd9606 AG |
1136 | #ifndef KERNEL_HAS_ATOMIC64 |
1137 | spin_lock_init(&ic->i_ack_lock); | |
1138 | #endif | |
f046011c | 1139 | atomic_set(&ic->i_signaled_sends, 0); |
ec16227e AG |
1140 | |
1141 | /* | |
1142 | * rds_ib_conn_shutdown() waits for these to be emptied so they | |
1143 | * must be initialized before it can be called. | |
1144 | */ | |
1145 | rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); | |
1146 | rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); | |
1147 | ||
1148 | ic->conn = conn; | |
1149 | conn->c_transport_data = ic; | |
1150 | ||
1151 | spin_lock_irqsave(&ib_nodev_conns_lock, flags); | |
1152 | list_add_tail(&ic->ib_node, &ib_nodev_conns); | |
1153 | spin_unlock_irqrestore(&ib_nodev_conns_lock, flags); | |
1154 | ||
1155 | ||
1156 | rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data); | |
1157 | return 0; | |
1158 | } | |
1159 | ||
745cbcca AG |
1160 | /* |
1161 | * Free a connection. Connection must be shut down and not set for reconnect. | |
1162 | */ | |
ec16227e AG |
1163 | void rds_ib_conn_free(void *arg) |
1164 | { | |
1165 | struct rds_ib_connection *ic = arg; | |
745cbcca AG |
1166 | spinlock_t *lock_ptr; |
1167 | ||
ec16227e | 1168 | rdsdebug("ic %p\n", ic); |
745cbcca AG |
1169 | |
1170 | /* | |
1171 | * Conn is either on a dev's list or on the nodev list. | |
1172 | * A race with shutdown() or connect() would cause problems | |
1173 | * (since rds_ibdev would change) but that should never happen. | |
1174 | */ | |
1175 | lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock; | |
1176 | ||
1177 | spin_lock_irq(lock_ptr); | |
ec16227e | 1178 | list_del(&ic->ib_node); |
745cbcca AG |
1179 | spin_unlock_irq(lock_ptr); |
1180 | ||
33244125 CM |
1181 | rds_ib_recv_free_caches(ic); |
1182 | ||
ec16227e AG |
1183 | kfree(ic); |
1184 | } | |
1185 | ||
1186 | ||
1187 | /* | |
1188 | * An error occurred on the connection | |
1189 | */ | |
1190 | void | |
1191 | __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...) | |
1192 | { | |
1193 | va_list ap; | |
1194 | ||
1195 | rds_conn_drop(conn); | |
1196 | ||
1197 | va_start(ap, fmt); | |
1198 | vprintk(fmt, ap); | |
1199 | va_end(ap); | |
1200 | } |