Commit | Line | Data |
---|---|---|
1f327613 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
fc79d4b1 | 2 | /* |
fc79d4b1 TT |
3 | * RDMA transport layer based on the trans_fd.c implementation. |
4 | * | |
5 | * Copyright (C) 2008 by Tom Tucker <tom@opengridcomputing.com> | |
6 | * Copyright (C) 2006 by Russ Cox <rsc@swtch.com> | |
7 | * Copyright (C) 2004-2005 by Latchesar Ionkov <lucho@ionkov.net> | |
8 | * Copyright (C) 2004-2008 by Eric Van Hensbergen <ericvh@gmail.com> | |
9 | * Copyright (C) 1997-2002 by Ron Minnich <rminnich@sarnoff.com> | |
fc79d4b1 TT |
10 | */ |
11 | ||
5d385153 JP |
12 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
13 | ||
fc79d4b1 TT |
14 | #include <linux/in.h> |
15 | #include <linux/module.h> | |
16 | #include <linux/net.h> | |
17 | #include <linux/ipv6.h> | |
18 | #include <linux/kthread.h> | |
19 | #include <linux/errno.h> | |
20 | #include <linux/kernel.h> | |
21 | #include <linux/un.h> | |
22 | #include <linux/uaccess.h> | |
23 | #include <linux/inet.h> | |
24 | #include <linux/idr.h> | |
25 | #include <linux/file.h> | |
26 | #include <linux/parser.h> | |
27 | #include <linux/semaphore.h> | |
5a0e3ad6 | 28 | #include <linux/slab.h> |
c4fac910 | 29 | #include <linux/seq_file.h> |
fc79d4b1 TT |
30 | #include <net/9p/9p.h> |
31 | #include <net/9p/client.h> | |
32 | #include <net/9p/transport.h> | |
33 | #include <rdma/ib_verbs.h> | |
34 | #include <rdma/rdma_cm.h> | |
fc79d4b1 TT |
35 | |
36 | #define P9_PORT 5640 | |
37 | #define P9_RDMA_SQ_DEPTH 32 | |
38 | #define P9_RDMA_RQ_DEPTH 32 | |
39 | #define P9_RDMA_SEND_SGE 4 | |
40 | #define P9_RDMA_RECV_SGE 4 | |
41 | #define P9_RDMA_IRD 0 | |
42 | #define P9_RDMA_ORD 0 | |
43 | #define P9_RDMA_TIMEOUT 30000 /* 30 seconds */ | |
3fcc62f4 | 44 | #define P9_RDMA_MAXSIZE (1024*1024) /* 1MB */ |
fc79d4b1 | 45 | |
fc79d4b1 TT |
46 | /** |
47 | * struct p9_trans_rdma - RDMA transport instance | |
48 | * | |
49 | * @state: tracks the transport state machine for connection setup and tear down | |
50 | * @cm_id: The RDMA CM ID | |
51 | * @pd: Protection Domain pointer | |
52 | * @qp: Queue Pair pointer | |
53 | * @cq: Completion Queue pointer | |
fc79d4b1 | 54 | * @timeout: Number of uSecs to wait for connection management events |
c4fac910 DH |
55 | * @privport: Whether a privileged port may be used |
56 | * @port: The port to use | |
fc79d4b1 TT |
57 | * @sq_depth: The depth of the Send Queue |
58 | * @sq_sem: Semaphore for the SQ | |
59 | * @rq_depth: The depth of the Receive Queue. | |
fd453d0e | 60 | * @rq_sem: Semaphore for the RQ |
1cff3306 SD |
61 | * @excess_rc : Amount of posted Receive Contexts without a pending request. |
62 | * See rdma_request() | |
fc79d4b1 TT |
63 | * @addr: The remote peer's address |
64 | * @req_lock: Protects the active request list | |
fc79d4b1 TT |
65 | * @cm_done: Completion event for connection management tracking |
66 | */ | |
67 | struct p9_trans_rdma { | |
68 | enum { | |
69 | P9_RDMA_INIT, | |
70 | P9_RDMA_ADDR_RESOLVED, | |
71 | P9_RDMA_ROUTE_RESOLVED, | |
72 | P9_RDMA_CONNECTED, | |
73 | P9_RDMA_FLUSHING, | |
74 | P9_RDMA_CLOSING, | |
75 | P9_RDMA_CLOSED, | |
76 | } state; | |
77 | struct rdma_cm_id *cm_id; | |
78 | struct ib_pd *pd; | |
79 | struct ib_qp *qp; | |
80 | struct ib_cq *cq; | |
fc79d4b1 | 81 | long timeout; |
c4fac910 DH |
82 | bool privport; |
83 | u16 port; | |
fc79d4b1 TT |
84 | int sq_depth; |
85 | struct semaphore sq_sem; | |
86 | int rq_depth; | |
fd453d0e | 87 | struct semaphore rq_sem; |
1cff3306 | 88 | atomic_t excess_rc; |
fc79d4b1 TT |
89 | struct sockaddr_in addr; |
90 | spinlock_t req_lock; | |
91 | ||
92 | struct completion cm_done; | |
93 | }; | |
94 | ||
15e522a7 AL |
95 | struct p9_rdma_req; |
96 | ||
fc79d4b1 | 97 | /** |
15e522a7 | 98 | * struct p9_rdma_context - Keeps track of in-process WR |
fc79d4b1 | 99 | * |
760b3d61 | 100 | * @cqe: completion queue entry |
fc79d4b1 TT |
101 | * @busa: Bus address to unmap when the WR completes |
102 | * @req: Keeps track of requests (send) | |
103 | * @rc: Keepts track of replies (receive) | |
104 | */ | |
fc79d4b1 | 105 | struct p9_rdma_context { |
7cf20fc6 | 106 | struct ib_cqe cqe; |
fc79d4b1 TT |
107 | dma_addr_t busa; |
108 | union { | |
109 | struct p9_req_t *req; | |
523adb6c | 110 | struct p9_fcall rc; |
fc79d4b1 TT |
111 | }; |
112 | }; | |
113 | ||
114 | /** | |
15e522a7 | 115 | * struct p9_rdma_opts - Collection of mount options |
fc79d4b1 | 116 | * @port: port of connection |
760b3d61 | 117 | * @privport: Whether a privileged port may be used |
fc79d4b1 TT |
118 | * @sq_depth: The requested depth of the SQ. This really doesn't need |
119 | * to be any deeper than the number of threads used in the client | |
120 | * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth | |
121 | * @timeout: Time to wait in msecs for CM events | |
122 | */ | |
123 | struct p9_rdma_opts { | |
124 | short port; | |
c4fac910 | 125 | bool privport; |
fc79d4b1 TT |
126 | int sq_depth; |
127 | int rq_depth; | |
128 | long timeout; | |
129 | }; | |
130 | ||
131 | /* | |
132 | * Option Parsing (code inspired by NFS code) | |
133 | */ | |
134 | enum { | |
135 | /* Options that take integer arguments */ | |
f569d3ef DM |
136 | Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout, |
137 | /* Options that take no argument */ | |
138 | Opt_privport, | |
139 | Opt_err, | |
fc79d4b1 TT |
140 | }; |
141 | ||
142 | static match_table_t tokens = { | |
143 | {Opt_port, "port=%u"}, | |
144 | {Opt_sq_depth, "sq=%u"}, | |
145 | {Opt_rq_depth, "rq=%u"}, | |
146 | {Opt_timeout, "timeout=%u"}, | |
f569d3ef | 147 | {Opt_privport, "privport"}, |
fc79d4b1 TT |
148 | {Opt_err, NULL}, |
149 | }; | |
150 | ||
c4fac910 DH |
151 | static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt) |
152 | { | |
153 | struct p9_trans_rdma *rdma = clnt->trans; | |
154 | ||
155 | if (rdma->port != P9_PORT) | |
156 | seq_printf(m, ",port=%u", rdma->port); | |
157 | if (rdma->sq_depth != P9_RDMA_SQ_DEPTH) | |
158 | seq_printf(m, ",sq=%u", rdma->sq_depth); | |
159 | if (rdma->rq_depth != P9_RDMA_RQ_DEPTH) | |
160 | seq_printf(m, ",rq=%u", rdma->rq_depth); | |
161 | if (rdma->timeout != P9_RDMA_TIMEOUT) | |
162 | seq_printf(m, ",timeout=%lu", rdma->timeout); | |
163 | if (rdma->privport) | |
164 | seq_puts(m, ",privport"); | |
165 | return 0; | |
166 | } | |
167 | ||
fc79d4b1 | 168 | /** |
0e15597e AK |
169 | * parse_opts - parse mount options into rdma options structure |
170 | * @params: options string passed from mount | |
171 | * @opts: rdma transport-specific structure to parse options into | |
fc79d4b1 TT |
172 | * |
173 | * Returns 0 upon success, -ERRNO upon failure | |
174 | */ | |
175 | static int parse_opts(char *params, struct p9_rdma_opts *opts) | |
176 | { | |
177 | char *p; | |
178 | substring_t args[MAX_OPT_ARGS]; | |
179 | int option; | |
d8c8a9e3 | 180 | char *options, *tmp_options; |
fc79d4b1 TT |
181 | |
182 | opts->port = P9_PORT; | |
183 | opts->sq_depth = P9_RDMA_SQ_DEPTH; | |
184 | opts->rq_depth = P9_RDMA_RQ_DEPTH; | |
185 | opts->timeout = P9_RDMA_TIMEOUT; | |
c4fac910 | 186 | opts->privport = false; |
fc79d4b1 TT |
187 | |
188 | if (!params) | |
189 | return 0; | |
190 | ||
d8c8a9e3 EVH |
191 | tmp_options = kstrdup(params, GFP_KERNEL); |
192 | if (!tmp_options) { | |
5d385153 JP |
193 | p9_debug(P9_DEBUG_ERROR, |
194 | "failed to allocate copy of option string\n"); | |
fc79d4b1 TT |
195 | return -ENOMEM; |
196 | } | |
d8c8a9e3 | 197 | options = tmp_options; |
fc79d4b1 TT |
198 | |
199 | while ((p = strsep(&options, ",")) != NULL) { | |
200 | int token; | |
201 | int r; | |
202 | if (!*p) | |
203 | continue; | |
204 | token = match_token(p, tokens, args); | |
f569d3ef DM |
205 | if ((token != Opt_err) && (token != Opt_privport)) { |
206 | r = match_int(&args[0], &option); | |
207 | if (r < 0) { | |
208 | p9_debug(P9_DEBUG_ERROR, | |
209 | "integer field, but no integer?\n"); | |
210 | continue; | |
211 | } | |
fc79d4b1 TT |
212 | } |
213 | switch (token) { | |
214 | case Opt_port: | |
215 | opts->port = option; | |
216 | break; | |
217 | case Opt_sq_depth: | |
218 | opts->sq_depth = option; | |
219 | break; | |
220 | case Opt_rq_depth: | |
221 | opts->rq_depth = option; | |
222 | break; | |
223 | case Opt_timeout: | |
224 | opts->timeout = option; | |
225 | break; | |
f569d3ef | 226 | case Opt_privport: |
c4fac910 | 227 | opts->privport = true; |
f569d3ef | 228 | break; |
fc79d4b1 TT |
229 | default: |
230 | continue; | |
231 | } | |
232 | } | |
233 | /* RQ must be at least as large as the SQ */ | |
234 | opts->rq_depth = max(opts->rq_depth, opts->sq_depth); | |
d8c8a9e3 | 235 | kfree(tmp_options); |
fc79d4b1 TT |
236 | return 0; |
237 | } | |
238 | ||
239 | static int | |
240 | p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event) | |
241 | { | |
242 | struct p9_client *c = id->context; | |
243 | struct p9_trans_rdma *rdma = c->trans; | |
244 | switch (event->event) { | |
245 | case RDMA_CM_EVENT_ADDR_RESOLVED: | |
246 | BUG_ON(rdma->state != P9_RDMA_INIT); | |
247 | rdma->state = P9_RDMA_ADDR_RESOLVED; | |
248 | break; | |
249 | ||
250 | case RDMA_CM_EVENT_ROUTE_RESOLVED: | |
251 | BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED); | |
252 | rdma->state = P9_RDMA_ROUTE_RESOLVED; | |
253 | break; | |
254 | ||
255 | case RDMA_CM_EVENT_ESTABLISHED: | |
256 | BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED); | |
257 | rdma->state = P9_RDMA_CONNECTED; | |
258 | break; | |
259 | ||
260 | case RDMA_CM_EVENT_DISCONNECTED: | |
261 | if (rdma) | |
262 | rdma->state = P9_RDMA_CLOSED; | |
473c7dd1 | 263 | c->status = Disconnected; |
fc79d4b1 TT |
264 | break; |
265 | ||
266 | case RDMA_CM_EVENT_TIMEWAIT_EXIT: | |
267 | break; | |
268 | ||
269 | case RDMA_CM_EVENT_ADDR_CHANGE: | |
270 | case RDMA_CM_EVENT_ROUTE_ERROR: | |
271 | case RDMA_CM_EVENT_DEVICE_REMOVAL: | |
272 | case RDMA_CM_EVENT_MULTICAST_JOIN: | |
273 | case RDMA_CM_EVENT_MULTICAST_ERROR: | |
274 | case RDMA_CM_EVENT_REJECTED: | |
275 | case RDMA_CM_EVENT_CONNECT_REQUEST: | |
276 | case RDMA_CM_EVENT_CONNECT_RESPONSE: | |
277 | case RDMA_CM_EVENT_CONNECT_ERROR: | |
278 | case RDMA_CM_EVENT_ADDR_ERROR: | |
279 | case RDMA_CM_EVENT_UNREACHABLE: | |
280 | c->status = Disconnected; | |
281 | rdma_disconnect(rdma->cm_id); | |
282 | break; | |
283 | default: | |
284 | BUG(); | |
285 | } | |
286 | complete(&rdma->cm_done); | |
287 | return 0; | |
288 | } | |
289 | ||
290 | static void | |
7cf20fc6 | 291 | recv_done(struct ib_cq *cq, struct ib_wc *wc) |
fc79d4b1 | 292 | { |
7cf20fc6 CH |
293 | struct p9_client *client = cq->cq_context; |
294 | struct p9_trans_rdma *rdma = client->trans; | |
295 | struct p9_rdma_context *c = | |
296 | container_of(wc->wr_cqe, struct p9_rdma_context, cqe); | |
fc79d4b1 TT |
297 | struct p9_req_t *req; |
298 | int err = 0; | |
299 | int16_t tag; | |
300 | ||
301 | req = NULL; | |
302 | ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize, | |
303 | DMA_FROM_DEVICE); | |
304 | ||
7cf20fc6 | 305 | if (wc->status != IB_WC_SUCCESS) |
fc79d4b1 TT |
306 | goto err_out; |
307 | ||
523adb6c DM |
308 | c->rc.size = wc->byte_len; |
309 | err = p9_parse_header(&c->rc, NULL, NULL, &tag, 1); | |
fc79d4b1 TT |
310 | if (err) |
311 | goto err_out; | |
312 | ||
313 | req = p9_tag_lookup(client, tag); | |
314 | if (!req) | |
315 | goto err_out; | |
316 | ||
47229ff8 SD |
317 | /* Check that we have not yet received a reply for this request. |
318 | */ | |
523adb6c | 319 | if (unlikely(req->rc.sdata)) { |
47229ff8 SD |
320 | pr_err("Duplicate reply for request %d", tag); |
321 | goto err_out; | |
322 | } | |
323 | ||
523adb6c DM |
324 | req->rc.size = c->rc.size; |
325 | req->rc.sdata = c->rc.sdata; | |
2b6e72ed | 326 | p9_client_cb(client, req, REQ_STATUS_RCVD); |
fc79d4b1 | 327 | |
7cf20fc6 CH |
328 | out: |
329 | up(&rdma->rq_sem); | |
330 | kfree(c); | |
fc79d4b1 TT |
331 | return; |
332 | ||
333 | err_out: | |
7cf20fc6 CH |
334 | p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", |
335 | req, err, wc->status); | |
fc79d4b1 TT |
336 | rdma->state = P9_RDMA_FLUSHING; |
337 | client->status = Disconnected; | |
7cf20fc6 | 338 | goto out; |
fc79d4b1 TT |
339 | } |
340 | ||
341 | static void | |
7cf20fc6 | 342 | send_done(struct ib_cq *cq, struct ib_wc *wc) |
fc79d4b1 | 343 | { |
7cf20fc6 CH |
344 | struct p9_client *client = cq->cq_context; |
345 | struct p9_trans_rdma *rdma = client->trans; | |
346 | struct p9_rdma_context *c = | |
347 | container_of(wc->wr_cqe, struct p9_rdma_context, cqe); | |
348 | ||
fc79d4b1 | 349 | ib_dma_unmap_single(rdma->cm_id->device, |
523adb6c | 350 | c->busa, c->req->tc.size, |
fc79d4b1 | 351 | DMA_TO_DEVICE); |
7cf20fc6 | 352 | up(&rdma->sq_sem); |
8b11ff09 | 353 | p9_req_put(client, c->req); |
7cf20fc6 | 354 | kfree(c); |
fc79d4b1 TT |
355 | } |
356 | ||
357 | static void qp_event_handler(struct ib_event *event, void *context) | |
358 | { | |
5d385153 JP |
359 | p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n", |
360 | event->event, context); | |
fc79d4b1 TT |
361 | } |
362 | ||
fc79d4b1 TT |
363 | static void rdma_destroy_trans(struct p9_trans_rdma *rdma) |
364 | { | |
365 | if (!rdma) | |
366 | return; | |
367 | ||
fc79d4b1 TT |
368 | if (rdma->qp && !IS_ERR(rdma->qp)) |
369 | ib_destroy_qp(rdma->qp); | |
370 | ||
371 | if (rdma->pd && !IS_ERR(rdma->pd)) | |
372 | ib_dealloc_pd(rdma->pd); | |
373 | ||
374 | if (rdma->cq && !IS_ERR(rdma->cq)) | |
7cf20fc6 | 375 | ib_free_cq(rdma->cq); |
fc79d4b1 TT |
376 | |
377 | if (rdma->cm_id && !IS_ERR(rdma->cm_id)) | |
378 | rdma_destroy_id(rdma->cm_id); | |
379 | ||
380 | kfree(rdma); | |
381 | } | |
382 | ||
383 | static int | |
384 | post_recv(struct p9_client *client, struct p9_rdma_context *c) | |
385 | { | |
386 | struct p9_trans_rdma *rdma = client->trans; | |
72bc4d37 | 387 | struct ib_recv_wr wr; |
fc79d4b1 TT |
388 | struct ib_sge sge; |
389 | ||
390 | c->busa = ib_dma_map_single(rdma->cm_id->device, | |
523adb6c | 391 | c->rc.sdata, client->msize, |
fc79d4b1 TT |
392 | DMA_FROM_DEVICE); |
393 | if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) | |
394 | goto error; | |
395 | ||
7cf20fc6 CH |
396 | c->cqe.done = recv_done; |
397 | ||
fc79d4b1 TT |
398 | sge.addr = c->busa; |
399 | sge.length = client->msize; | |
2f31fa88 | 400 | sge.lkey = rdma->pd->local_dma_lkey; |
fc79d4b1 TT |
401 | |
402 | wr.next = NULL; | |
7cf20fc6 | 403 | wr.wr_cqe = &c->cqe; |
fc79d4b1 TT |
404 | wr.sg_list = &sge; |
405 | wr.num_sge = 1; | |
72bc4d37 | 406 | return ib_post_recv(rdma->qp, &wr, NULL); |
fc79d4b1 TT |
407 | |
408 | error: | |
5d385153 | 409 | p9_debug(P9_DEBUG_ERROR, "EIO\n"); |
fc79d4b1 TT |
410 | return -EIO; |
411 | } | |
412 | ||
413 | static int rdma_request(struct p9_client *client, struct p9_req_t *req) | |
414 | { | |
415 | struct p9_trans_rdma *rdma = client->trans; | |
72bc4d37 | 416 | struct ib_send_wr wr; |
fc79d4b1 TT |
417 | struct ib_sge sge; |
418 | int err = 0; | |
419 | unsigned long flags; | |
420 | struct p9_rdma_context *c = NULL; | |
421 | struct p9_rdma_context *rpl_context = NULL; | |
422 | ||
1cff3306 SD |
423 | /* When an error occurs between posting the recv and the send, |
424 | * there will be a receive context posted without a pending request. | |
425 | * Since there is no way to "un-post" it, we remember it and skip | |
426 | * post_recv() for the next request. | |
427 | * So here, | |
428 | * see if we are this `next request' and need to absorb an excess rc. | |
429 | * If yes, then drop and free our own, and do not recv_post(). | |
430 | **/ | |
431 | if (unlikely(atomic_read(&rdma->excess_rc) > 0)) { | |
432 | if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) { | |
523adb6c DM |
433 | /* Got one! */ |
434 | p9_fcall_fini(&req->rc); | |
435 | req->rc.sdata = NULL; | |
1cff3306 SD |
436 | goto dont_need_post_recv; |
437 | } else { | |
438 | /* We raced and lost. */ | |
439 | atomic_inc(&rdma->excess_rc); | |
440 | } | |
441 | } | |
442 | ||
fc79d4b1 | 443 | /* Allocate an fcall for the reply */ |
eeff66ef | 444 | rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS); |
1d6400c7 DB |
445 | if (!rpl_context) { |
446 | err = -ENOMEM; | |
2f52d07c | 447 | goto recv_error; |
1d6400c7 | 448 | } |
523adb6c | 449 | rpl_context->rc.sdata = req->rc.sdata; |
fc79d4b1 TT |
450 | |
451 | /* | |
452 | * Post a receive buffer for this request. We need to ensure | |
453 | * there is a reply buffer available for every outstanding | |
454 | * request. A flushed request can result in no reply for an | |
455 | * outstanding request, so we must keep a count to avoid | |
456 | * overflowing the RQ. | |
457 | */ | |
2f52d07c SD |
458 | if (down_interruptible(&rdma->rq_sem)) { |
459 | err = -EINTR; | |
460 | goto recv_error; | |
461 | } | |
fd453d0e SD |
462 | |
463 | err = post_recv(client, rpl_context); | |
464 | if (err) { | |
8b894adb | 465 | p9_debug(P9_DEBUG_ERROR, "POST RECV failed: %d\n", err); |
2f52d07c | 466 | goto recv_error; |
fd453d0e | 467 | } |
fc79d4b1 | 468 | /* remove posted receive buffer from request structure */ |
523adb6c | 469 | req->rc.sdata = NULL; |
fc79d4b1 | 470 | |
1cff3306 | 471 | dont_need_post_recv: |
fc79d4b1 | 472 | /* Post the request */ |
eeff66ef | 473 | c = kmalloc(sizeof *c, GFP_NOFS); |
1d6400c7 DB |
474 | if (!c) { |
475 | err = -ENOMEM; | |
2f52d07c | 476 | goto send_error; |
1d6400c7 | 477 | } |
fc79d4b1 TT |
478 | c->req = req; |
479 | ||
480 | c->busa = ib_dma_map_single(rdma->cm_id->device, | |
523adb6c | 481 | c->req->tc.sdata, c->req->tc.size, |
fc79d4b1 | 482 | DMA_TO_DEVICE); |
2f52d07c SD |
483 | if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) { |
484 | err = -EIO; | |
485 | goto send_error; | |
486 | } | |
fc79d4b1 | 487 | |
7cf20fc6 CH |
488 | c->cqe.done = send_done; |
489 | ||
fc79d4b1 | 490 | sge.addr = c->busa; |
523adb6c | 491 | sge.length = c->req->tc.size; |
2f31fa88 | 492 | sge.lkey = rdma->pd->local_dma_lkey; |
fc79d4b1 TT |
493 | |
494 | wr.next = NULL; | |
7cf20fc6 | 495 | wr.wr_cqe = &c->cqe; |
fc79d4b1 TT |
496 | wr.opcode = IB_WR_SEND; |
497 | wr.send_flags = IB_SEND_SIGNALED; | |
498 | wr.sg_list = &sge; | |
499 | wr.num_sge = 1; | |
500 | ||
2f52d07c SD |
501 | if (down_interruptible(&rdma->sq_sem)) { |
502 | err = -EINTR; | |
503 | goto send_error; | |
504 | } | |
505 | ||
3f9d5b8d SD |
506 | /* Mark request as `sent' *before* we actually send it, |
507 | * because doing if after could erase the REQ_STATUS_RCVD | |
508 | * status in case of a very fast reply. | |
509 | */ | |
510 | req->status = REQ_STATUS_SENT; | |
72bc4d37 | 511 | err = ib_post_send(rdma->qp, &wr, NULL); |
2f52d07c SD |
512 | if (err) |
513 | goto send_error; | |
fc79d4b1 | 514 | |
2f52d07c SD |
515 | /* Success */ |
516 | return 0; | |
fc79d4b1 | 517 | |
2f52d07c SD |
518 | /* Handle errors that happened during or while preparing the send: */ |
519 | send_error: | |
3f9d5b8d | 520 | req->status = REQ_STATUS_ERROR; |
1d6400c7 | 521 | kfree(c); |
2f52d07c | 522 | p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err); |
1cff3306 SD |
523 | |
524 | /* Ach. | |
525 | * We did recv_post(), but not send. We have one recv_post in excess. | |
526 | */ | |
527 | atomic_inc(&rdma->excess_rc); | |
2f52d07c SD |
528 | return err; |
529 | ||
530 | /* Handle errors that happened during or while preparing post_recv(): */ | |
531 | recv_error: | |
1d6400c7 | 532 | kfree(rpl_context); |
fc79d4b1 | 533 | spin_lock_irqsave(&rdma->req_lock, flags); |
8b894adb | 534 | if (err != -EINTR && rdma->state < P9_RDMA_CLOSING) { |
fc79d4b1 TT |
535 | rdma->state = P9_RDMA_CLOSING; |
536 | spin_unlock_irqrestore(&rdma->req_lock, flags); | |
537 | rdma_disconnect(rdma->cm_id); | |
538 | } else | |
539 | spin_unlock_irqrestore(&rdma->req_lock, flags); | |
540 | return err; | |
541 | } | |
542 | ||
543 | static void rdma_close(struct p9_client *client) | |
544 | { | |
545 | struct p9_trans_rdma *rdma; | |
546 | ||
547 | if (!client) | |
548 | return; | |
549 | ||
550 | rdma = client->trans; | |
551 | if (!rdma) | |
552 | return; | |
553 | ||
554 | client->status = Disconnected; | |
555 | rdma_disconnect(rdma->cm_id); | |
556 | rdma_destroy_trans(rdma); | |
557 | } | |
558 | ||
559 | /** | |
560 | * alloc_rdma - Allocate and initialize the rdma transport structure | |
fc79d4b1 TT |
561 | * @opts: Mount options structure |
562 | */ | |
563 | static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts) | |
564 | { | |
565 | struct p9_trans_rdma *rdma; | |
566 | ||
567 | rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL); | |
568 | if (!rdma) | |
569 | return NULL; | |
570 | ||
c4fac910 DH |
571 | rdma->port = opts->port; |
572 | rdma->privport = opts->privport; | |
fc79d4b1 TT |
573 | rdma->sq_depth = opts->sq_depth; |
574 | rdma->rq_depth = opts->rq_depth; | |
575 | rdma->timeout = opts->timeout; | |
576 | spin_lock_init(&rdma->req_lock); | |
577 | init_completion(&rdma->cm_done); | |
578 | sema_init(&rdma->sq_sem, rdma->sq_depth); | |
fd453d0e | 579 | sema_init(&rdma->rq_sem, rdma->rq_depth); |
1cff3306 | 580 | atomic_set(&rdma->excess_rc, 0); |
fc79d4b1 TT |
581 | |
582 | return rdma; | |
583 | } | |
584 | ||
fc79d4b1 TT |
585 | static int rdma_cancel(struct p9_client *client, struct p9_req_t *req) |
586 | { | |
931700d2 SD |
587 | /* Nothing to do here. |
588 | * We will take care of it (if we have to) in rdma_cancelled() | |
589 | */ | |
fc79d4b1 TT |
590 | return 1; |
591 | } | |
592 | ||
931700d2 SD |
593 | /* A request has been fully flushed without a reply. |
594 | * That means we have posted one buffer in excess. | |
595 | */ | |
596 | static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req) | |
597 | { | |
598 | struct p9_trans_rdma *rdma = client->trans; | |
599 | atomic_inc(&rdma->excess_rc); | |
600 | return 0; | |
601 | } | |
602 | ||
f569d3ef DM |
603 | static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma) |
604 | { | |
605 | struct sockaddr_in cl = { | |
606 | .sin_family = AF_INET, | |
607 | .sin_addr.s_addr = htonl(INADDR_ANY), | |
608 | }; | |
609 | int port, err = -EINVAL; | |
610 | ||
611 | for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) { | |
612 | cl.sin_port = htons((ushort)port); | |
613 | err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl); | |
614 | if (err != -EADDRINUSE) | |
615 | break; | |
616 | } | |
617 | return err; | |
618 | } | |
619 | ||
fc79d4b1 | 620 | /** |
4a026da9 | 621 | * rdma_create_trans - Transport method for creating a transport instance |
fc79d4b1 TT |
622 | * @client: client instance |
623 | * @addr: IP address string | |
624 | * @args: Mount options string | |
625 | */ | |
626 | static int | |
627 | rdma_create_trans(struct p9_client *client, const char *addr, char *args) | |
628 | { | |
629 | int err; | |
630 | struct p9_rdma_opts opts; | |
631 | struct p9_trans_rdma *rdma; | |
632 | struct rdma_conn_param conn_param; | |
633 | struct ib_qp_init_attr qp_attr; | |
fc79d4b1 | 634 | |
10aa1452 TB |
635 | if (addr == NULL) |
636 | return -EINVAL; | |
637 | ||
fc79d4b1 TT |
638 | /* Parse the transport specific mount options */ |
639 | err = parse_opts(args, &opts); | |
640 | if (err < 0) | |
641 | return err; | |
642 | ||
643 | /* Create and initialize the RDMA transport structure */ | |
644 | rdma = alloc_rdma(&opts); | |
645 | if (!rdma) | |
646 | return -ENOMEM; | |
647 | ||
648 | /* Create the RDMA CM ID */ | |
fa20105e GS |
649 | rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client, |
650 | RDMA_PS_TCP, IB_QPT_RC); | |
fc79d4b1 TT |
651 | if (IS_ERR(rdma->cm_id)) |
652 | goto error; | |
653 | ||
517ac45a TT |
654 | /* Associate the client with the transport */ |
655 | client->trans = rdma; | |
656 | ||
f569d3ef DM |
657 | /* Bind to a privileged port if we need to */ |
658 | if (opts.privport) { | |
659 | err = p9_rdma_bind_privport(rdma); | |
660 | if (err < 0) { | |
661 | pr_err("%s (%d): problem binding to privport: %d\n", | |
662 | __func__, task_pid_nr(current), -err); | |
663 | goto error; | |
664 | } | |
665 | } | |
666 | ||
fc79d4b1 TT |
667 | /* Resolve the server's address */ |
668 | rdma->addr.sin_family = AF_INET; | |
669 | rdma->addr.sin_addr.s_addr = in_aton(addr); | |
670 | rdma->addr.sin_port = htons(opts.port); | |
671 | err = rdma_resolve_addr(rdma->cm_id, NULL, | |
672 | (struct sockaddr *)&rdma->addr, | |
673 | rdma->timeout); | |
674 | if (err) | |
675 | goto error; | |
676 | err = wait_for_completion_interruptible(&rdma->cm_done); | |
677 | if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED)) | |
678 | goto error; | |
679 | ||
680 | /* Resolve the route to the server */ | |
681 | err = rdma_resolve_route(rdma->cm_id, rdma->timeout); | |
682 | if (err) | |
683 | goto error; | |
684 | err = wait_for_completion_interruptible(&rdma->cm_done); | |
685 | if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED)) | |
686 | goto error; | |
687 | ||
fc79d4b1 | 688 | /* Create the Completion Queue */ |
20cf4e02 CL |
689 | rdma->cq = ib_alloc_cq_any(rdma->cm_id->device, client, |
690 | opts.sq_depth + opts.rq_depth + 1, | |
691 | IB_POLL_SOFTIRQ); | |
fc79d4b1 TT |
692 | if (IS_ERR(rdma->cq)) |
693 | goto error; | |
fc79d4b1 TT |
694 | |
695 | /* Create the Protection Domain */ | |
ed082d36 | 696 | rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0); |
fc79d4b1 TT |
697 | if (IS_ERR(rdma->pd)) |
698 | goto error; | |
699 | ||
fc79d4b1 TT |
700 | /* Create the Queue Pair */ |
701 | memset(&qp_attr, 0, sizeof qp_attr); | |
702 | qp_attr.event_handler = qp_event_handler; | |
703 | qp_attr.qp_context = client; | |
704 | qp_attr.cap.max_send_wr = opts.sq_depth; | |
705 | qp_attr.cap.max_recv_wr = opts.rq_depth; | |
706 | qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE; | |
707 | qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE; | |
708 | qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR; | |
709 | qp_attr.qp_type = IB_QPT_RC; | |
710 | qp_attr.send_cq = rdma->cq; | |
711 | qp_attr.recv_cq = rdma->cq; | |
712 | err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr); | |
713 | if (err) | |
714 | goto error; | |
715 | rdma->qp = rdma->cm_id->qp; | |
716 | ||
717 | /* Request a connection */ | |
718 | memset(&conn_param, 0, sizeof(conn_param)); | |
719 | conn_param.private_data = NULL; | |
720 | conn_param.private_data_len = 0; | |
721 | conn_param.responder_resources = P9_RDMA_IRD; | |
722 | conn_param.initiator_depth = P9_RDMA_ORD; | |
723 | err = rdma_connect(rdma->cm_id, &conn_param); | |
724 | if (err) | |
725 | goto error; | |
726 | err = wait_for_completion_interruptible(&rdma->cm_done); | |
727 | if (err || (rdma->state != P9_RDMA_CONNECTED)) | |
728 | goto error; | |
729 | ||
fc79d4b1 TT |
730 | client->status = Connected; |
731 | ||
732 | return 0; | |
733 | ||
734 | error: | |
735 | rdma_destroy_trans(rdma); | |
736 | return -ENOTCONN; | |
737 | } | |
738 | ||
739 | static struct p9_trans_module p9_rdma_trans = { | |
740 | .name = "rdma", | |
741 | .maxsize = P9_RDMA_MAXSIZE, | |
01d205d9 | 742 | .pooled_rbuffers = true, |
fc79d4b1 TT |
743 | .def = 0, |
744 | .owner = THIS_MODULE, | |
745 | .create = rdma_create_trans, | |
746 | .close = rdma_close, | |
747 | .request = rdma_request, | |
748 | .cancel = rdma_cancel, | |
931700d2 | 749 | .cancelled = rdma_cancelled, |
c4fac910 | 750 | .show_options = p9_rdma_show_options, |
fc79d4b1 TT |
751 | }; |
752 | ||
753 | /** | |
754 | * p9_trans_rdma_init - Register the 9P RDMA transport driver | |
755 | */ | |
756 | static int __init p9_trans_rdma_init(void) | |
757 | { | |
758 | v9fs_register_trans(&p9_rdma_trans); | |
759 | return 0; | |
760 | } | |
761 | ||
762 | static void __exit p9_trans_rdma_exit(void) | |
763 | { | |
764 | v9fs_unregister_trans(&p9_rdma_trans); | |
765 | } | |
766 | ||
767 | module_init(p9_trans_rdma_init); | |
768 | module_exit(p9_trans_rdma_exit); | |
4cd82a5b | 769 | MODULE_ALIAS_9P("rdma"); |
fc79d4b1 TT |
770 | |
771 | MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>"); | |
772 | MODULE_DESCRIPTION("RDMA Transport for 9P"); | |
773 | MODULE_LICENSE("Dual BSD/GPL"); |