Commit | Line | Data |
---|---|---|
f58851e6 | 1 | /* |
c56c65fb TT |
2 | * Copyright (c) 2003-2007 Network Appliance, Inc. 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 BSD-type | |
8 | * license below: | |
9 | * | |
10 | * Redistribution and use in source and binary forms, with or without | |
11 | * modification, are permitted provided that the following conditions | |
12 | * are met: | |
13 | * | |
14 | * Redistributions of source code must retain the above copyright | |
15 | * notice, this list of conditions and the following disclaimer. | |
16 | * | |
17 | * Redistributions in binary form must reproduce the above | |
18 | * copyright notice, this list of conditions and the following | |
19 | * disclaimer in the documentation and/or other materials provided | |
20 | * with the distribution. | |
21 | * | |
22 | * Neither the name of the Network Appliance, Inc. nor the names of | |
23 | * its contributors may be used to endorse or promote products | |
24 | * derived from this software without specific prior written | |
25 | * permission. | |
26 | * | |
27 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS | |
28 | * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT | |
29 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR | |
30 | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT | |
31 | * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, | |
32 | * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT | |
33 | * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, | |
34 | * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY | |
35 | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT | |
36 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE | |
37 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |
f58851e6 TT |
38 | */ |
39 | ||
c56c65fb TT |
40 | /* |
41 | * verbs.c | |
42 | * | |
43 | * Encapsulates the major functions managing: | |
44 | * o adapters | |
45 | * o endpoints | |
46 | * o connections | |
47 | * o buffer memory | |
48 | */ | |
49 | ||
a6b7a407 | 50 | #include <linux/interrupt.h> |
5a0e3ad6 | 51 | #include <linux/slab.h> |
eba8ff66 | 52 | #include <linux/prefetch.h> |
0dd39cae | 53 | #include <linux/sunrpc/addr.h> |
65866f82 | 54 | #include <asm/bitops.h> |
d0f36c46 | 55 | #include <linux/module.h> /* try_module_get()/module_put() */ |
c56c65fb | 56 | |
f58851e6 TT |
57 | #include "xprt_rdma.h" |
58 | ||
c56c65fb TT |
59 | /* |
60 | * Globals/Macros | |
61 | */ | |
62 | ||
f895b252 | 63 | #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) |
c56c65fb TT |
64 | # define RPCDBG_FACILITY RPCDBG_TRANS |
65 | #endif | |
66 | ||
67 | /* | |
68 | * internal functions | |
69 | */ | |
70 | ||
fe97b47c | 71 | static struct workqueue_struct *rpcrdma_receive_wq; |
c56c65fb | 72 | |
fe97b47c CL |
73 | int |
74 | rpcrdma_alloc_wq(void) | |
c56c65fb | 75 | { |
fe97b47c | 76 | struct workqueue_struct *recv_wq; |
c56c65fb | 77 | |
fe97b47c CL |
78 | recv_wq = alloc_workqueue("xprtrdma_receive", |
79 | WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_HIGHPRI, | |
80 | 0); | |
81 | if (!recv_wq) | |
82 | return -ENOMEM; | |
c56c65fb | 83 | |
fe97b47c CL |
84 | rpcrdma_receive_wq = recv_wq; |
85 | return 0; | |
c56c65fb TT |
86 | } |
87 | ||
fe97b47c CL |
88 | void |
89 | rpcrdma_destroy_wq(void) | |
f1a03b76 | 90 | { |
fe97b47c | 91 | struct workqueue_struct *wq; |
f1a03b76 | 92 | |
fe97b47c CL |
93 | if (rpcrdma_receive_wq) { |
94 | wq = rpcrdma_receive_wq; | |
95 | rpcrdma_receive_wq = NULL; | |
96 | destroy_workqueue(wq); | |
97 | } | |
f1a03b76 CL |
98 | } |
99 | ||
c56c65fb TT |
100 | static void |
101 | rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context) | |
102 | { | |
103 | struct rpcrdma_ep *ep = context; | |
104 | ||
7ff11de1 | 105 | pr_err("RPC: %s: %s on device %s ep %p\n", |
76357c71 | 106 | __func__, ib_event_msg(event->event), |
7ff11de1 | 107 | event->device->name, context); |
c56c65fb TT |
108 | if (ep->rep_connected == 1) { |
109 | ep->rep_connected = -EIO; | |
afadc468 | 110 | rpcrdma_conn_func(ep); |
c56c65fb TT |
111 | wake_up_all(&ep->rep_connect_wait); |
112 | } | |
113 | } | |
114 | ||
115 | static void | |
116 | rpcrdma_cq_async_error_upcall(struct ib_event *event, void *context) | |
117 | { | |
118 | struct rpcrdma_ep *ep = context; | |
119 | ||
7ff11de1 | 120 | pr_err("RPC: %s: %s on device %s ep %p\n", |
76357c71 | 121 | __func__, ib_event_msg(event->event), |
7ff11de1 | 122 | event->device->name, context); |
c56c65fb TT |
123 | if (ep->rep_connected == 1) { |
124 | ep->rep_connected = -EIO; | |
afadc468 | 125 | rpcrdma_conn_func(ep); |
c56c65fb TT |
126 | wake_up_all(&ep->rep_connect_wait); |
127 | } | |
128 | } | |
129 | ||
fc664485 CL |
130 | static void |
131 | rpcrdma_sendcq_process_wc(struct ib_wc *wc) | |
c56c65fb | 132 | { |
8502427c | 133 | /* WARNING: Only wr_id and status are reliable at this point */ |
e46ac34c CL |
134 | if (wc->wr_id == RPCRDMA_IGNORE_COMPLETION) { |
135 | if (wc->status != IB_WC_SUCCESS && | |
136 | wc->status != IB_WC_WR_FLUSH_ERR) | |
8502427c | 137 | pr_err("RPC: %s: SEND: %s\n", |
76357c71 | 138 | __func__, ib_wc_status_msg(wc->status)); |
8502427c CL |
139 | } else { |
140 | struct rpcrdma_mw *r; | |
141 | ||
142 | r = (struct rpcrdma_mw *)(unsigned long)wc->wr_id; | |
e46ac34c | 143 | r->mw_sendcompletion(wc); |
8502427c | 144 | } |
c56c65fb TT |
145 | } |
146 | ||
4220a072 CL |
147 | /* The common case is a single send completion is waiting. By |
148 | * passing two WC entries to ib_poll_cq, a return code of 1 | |
149 | * means there is exactly one WC waiting and no more. We don't | |
150 | * have to invoke ib_poll_cq again to know that the CQ has been | |
151 | * properly drained. | |
152 | */ | |
153 | static void | |
154 | rpcrdma_sendcq_poll(struct ib_cq *cq) | |
c56c65fb | 155 | { |
4220a072 CL |
156 | struct ib_wc *pos, wcs[2]; |
157 | int count, rc; | |
c56c65fb | 158 | |
1c00dd07 | 159 | do { |
4220a072 | 160 | pos = wcs; |
1c00dd07 | 161 | |
4220a072 CL |
162 | rc = ib_poll_cq(cq, ARRAY_SIZE(wcs), pos); |
163 | if (rc < 0) | |
164 | break; | |
1c00dd07 CL |
165 | |
166 | count = rc; | |
167 | while (count-- > 0) | |
4220a072 CL |
168 | rpcrdma_sendcq_process_wc(pos++); |
169 | } while (rc == ARRAY_SIZE(wcs)); | |
170 | return; | |
fc664485 | 171 | } |
c56c65fb | 172 | |
7b3d770c | 173 | /* Handle provider send completion upcalls. |
fc664485 CL |
174 | */ |
175 | static void | |
176 | rpcrdma_sendcq_upcall(struct ib_cq *cq, void *cq_context) | |
177 | { | |
7b3d770c | 178 | do { |
4220a072 | 179 | rpcrdma_sendcq_poll(cq); |
7b3d770c CL |
180 | } while (ib_req_notify_cq(cq, IB_CQ_NEXT_COMP | |
181 | IB_CQ_REPORT_MISSED_EVENTS) > 0); | |
fc664485 | 182 | } |
c56c65fb | 183 | |
fc664485 | 184 | static void |
fe97b47c CL |
185 | rpcrdma_receive_worker(struct work_struct *work) |
186 | { | |
187 | struct rpcrdma_rep *rep = | |
188 | container_of(work, struct rpcrdma_rep, rr_work); | |
fc664485 | 189 | |
fe97b47c | 190 | rpcrdma_reply_handler(rep); |
fc664485 CL |
191 | } |
192 | ||
193 | static void | |
fe97b47c | 194 | rpcrdma_recvcq_process_wc(struct ib_wc *wc) |
fc664485 CL |
195 | { |
196 | struct rpcrdma_rep *rep = | |
197 | (struct rpcrdma_rep *)(unsigned long)wc->wr_id; | |
198 | ||
8502427c CL |
199 | /* WARNING: Only wr_id and status are reliable at this point */ |
200 | if (wc->status != IB_WC_SUCCESS) | |
201 | goto out_fail; | |
fc664485 | 202 | |
8502427c | 203 | /* status == SUCCESS means all fields in wc are trustworthy */ |
fc664485 CL |
204 | if (wc->opcode != IB_WC_RECV) |
205 | return; | |
206 | ||
8502427c CL |
207 | dprintk("RPC: %s: rep %p opcode 'recv', length %u: success\n", |
208 | __func__, rep, wc->byte_len); | |
209 | ||
fc664485 | 210 | rep->rr_len = wc->byte_len; |
89e0d112 | 211 | ib_dma_sync_single_for_cpu(rep->rr_device, |
6b1184cd CL |
212 | rdmab_addr(rep->rr_rdmabuf), |
213 | rep->rr_len, DMA_FROM_DEVICE); | |
214 | prefetch(rdmab_to_msg(rep->rr_rdmabuf)); | |
fc664485 CL |
215 | |
216 | out_schedule: | |
fe97b47c | 217 | queue_work(rpcrdma_receive_wq, &rep->rr_work); |
8502427c | 218 | return; |
fe97b47c | 219 | |
8502427c CL |
220 | out_fail: |
221 | if (wc->status != IB_WC_WR_FLUSH_ERR) | |
222 | pr_err("RPC: %s: rep %p: %s\n", | |
76357c71 | 223 | __func__, rep, ib_wc_status_msg(wc->status)); |
b0e178a2 | 224 | rep->rr_len = RPCRDMA_BAD_LEN; |
8502427c | 225 | goto out_schedule; |
fc664485 CL |
226 | } |
227 | ||
4220a072 CL |
228 | /* The wc array is on stack: automatic memory is always CPU-local. |
229 | * | |
230 | * struct ib_wc is 64 bytes, making the poll array potentially | |
231 | * large. But this is at the bottom of the call chain. Further | |
232 | * substantial work is done in another thread. | |
233 | */ | |
234 | static void | |
235 | rpcrdma_recvcq_poll(struct ib_cq *cq) | |
fc664485 | 236 | { |
4220a072 | 237 | struct ib_wc *pos, wcs[4]; |
4220a072 | 238 | int count, rc; |
fc664485 | 239 | |
1c00dd07 | 240 | do { |
4220a072 | 241 | pos = wcs; |
1c00dd07 | 242 | |
4220a072 CL |
243 | rc = ib_poll_cq(cq, ARRAY_SIZE(wcs), pos); |
244 | if (rc < 0) | |
245 | break; | |
1c00dd07 CL |
246 | |
247 | count = rc; | |
248 | while (count-- > 0) | |
fe97b47c | 249 | rpcrdma_recvcq_process_wc(pos++); |
4220a072 | 250 | } while (rc == ARRAY_SIZE(wcs)); |
c56c65fb TT |
251 | } |
252 | ||
7b3d770c | 253 | /* Handle provider receive completion upcalls. |
c56c65fb TT |
254 | */ |
255 | static void | |
fc664485 | 256 | rpcrdma_recvcq_upcall(struct ib_cq *cq, void *cq_context) |
c56c65fb | 257 | { |
7b3d770c | 258 | do { |
4220a072 | 259 | rpcrdma_recvcq_poll(cq); |
7b3d770c CL |
260 | } while (ib_req_notify_cq(cq, IB_CQ_NEXT_COMP | |
261 | IB_CQ_REPORT_MISSED_EVENTS) > 0); | |
c56c65fb TT |
262 | } |
263 | ||
a7bc211a CL |
264 | static void |
265 | rpcrdma_flush_cqs(struct rpcrdma_ep *ep) | |
266 | { | |
5c166bef | 267 | struct ib_wc wc; |
5c166bef CL |
268 | |
269 | while (ib_poll_cq(ep->rep_attr.recv_cq, 1, &wc) > 0) | |
fe97b47c | 270 | rpcrdma_recvcq_process_wc(&wc); |
5c166bef CL |
271 | while (ib_poll_cq(ep->rep_attr.send_cq, 1, &wc) > 0) |
272 | rpcrdma_sendcq_process_wc(&wc); | |
a7bc211a CL |
273 | } |
274 | ||
c56c65fb TT |
275 | static int |
276 | rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event) | |
277 | { | |
278 | struct rpcrdma_xprt *xprt = id->context; | |
279 | struct rpcrdma_ia *ia = &xprt->rx_ia; | |
280 | struct rpcrdma_ep *ep = &xprt->rx_ep; | |
f895b252 | 281 | #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) |
0dd39cae | 282 | struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr; |
ff0db049 | 283 | #endif |
ce1ab9ab CL |
284 | struct ib_qp_attr *attr = &ia->ri_qp_attr; |
285 | struct ib_qp_init_attr *iattr = &ia->ri_qp_init_attr; | |
c56c65fb TT |
286 | int connstate = 0; |
287 | ||
288 | switch (event->event) { | |
289 | case RDMA_CM_EVENT_ADDR_RESOLVED: | |
290 | case RDMA_CM_EVENT_ROUTE_RESOLVED: | |
5675add3 | 291 | ia->ri_async_rc = 0; |
c56c65fb TT |
292 | complete(&ia->ri_done); |
293 | break; | |
294 | case RDMA_CM_EVENT_ADDR_ERROR: | |
295 | ia->ri_async_rc = -EHOSTUNREACH; | |
296 | dprintk("RPC: %s: CM address resolution error, ep 0x%p\n", | |
297 | __func__, ep); | |
298 | complete(&ia->ri_done); | |
299 | break; | |
300 | case RDMA_CM_EVENT_ROUTE_ERROR: | |
301 | ia->ri_async_rc = -ENETUNREACH; | |
302 | dprintk("RPC: %s: CM route resolution error, ep 0x%p\n", | |
303 | __func__, ep); | |
304 | complete(&ia->ri_done); | |
305 | break; | |
306 | case RDMA_CM_EVENT_ESTABLISHED: | |
307 | connstate = 1; | |
ce1ab9ab CL |
308 | ib_query_qp(ia->ri_id->qp, attr, |
309 | IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC, | |
310 | iattr); | |
c56c65fb TT |
311 | dprintk("RPC: %s: %d responder resources" |
312 | " (%d initiator)\n", | |
ce1ab9ab CL |
313 | __func__, attr->max_dest_rd_atomic, |
314 | attr->max_rd_atomic); | |
c56c65fb TT |
315 | goto connected; |
316 | case RDMA_CM_EVENT_CONNECT_ERROR: | |
317 | connstate = -ENOTCONN; | |
318 | goto connected; | |
319 | case RDMA_CM_EVENT_UNREACHABLE: | |
320 | connstate = -ENETDOWN; | |
321 | goto connected; | |
322 | case RDMA_CM_EVENT_REJECTED: | |
323 | connstate = -ECONNREFUSED; | |
324 | goto connected; | |
325 | case RDMA_CM_EVENT_DISCONNECTED: | |
326 | connstate = -ECONNABORTED; | |
327 | goto connected; | |
328 | case RDMA_CM_EVENT_DEVICE_REMOVAL: | |
329 | connstate = -ENODEV; | |
330 | connected: | |
c56c65fb TT |
331 | dprintk("RPC: %s: %sconnected\n", |
332 | __func__, connstate > 0 ? "" : "dis"); | |
333 | ep->rep_connected = connstate; | |
afadc468 | 334 | rpcrdma_conn_func(ep); |
c56c65fb | 335 | wake_up_all(&ep->rep_connect_wait); |
8079fb78 | 336 | /*FALLTHROUGH*/ |
c56c65fb | 337 | default: |
0dd39cae CL |
338 | dprintk("RPC: %s: %pIS:%u (ep 0x%p): %s\n", |
339 | __func__, sap, rpc_get_port(sap), ep, | |
76357c71 | 340 | rdma_event_msg(event->event)); |
c56c65fb TT |
341 | break; |
342 | } | |
343 | ||
f895b252 | 344 | #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) |
b3cd8d45 | 345 | if (connstate == 1) { |
ce1ab9ab | 346 | int ird = attr->max_dest_rd_atomic; |
b3cd8d45 | 347 | int tird = ep->rep_remote_cma.responder_resources; |
0dd39cae | 348 | |
a0ce85f5 | 349 | pr_info("rpcrdma: connection to %pIS:%u on %s, memreg '%s', %d credits, %d responders%s\n", |
0dd39cae | 350 | sap, rpc_get_port(sap), |
89e0d112 | 351 | ia->ri_device->name, |
a0ce85f5 | 352 | ia->ri_ops->ro_displayname, |
b3cd8d45 TT |
353 | xprt->rx_buf.rb_max_requests, |
354 | ird, ird < 4 && ird < tird / 2 ? " (low!)" : ""); | |
355 | } else if (connstate < 0) { | |
0dd39cae CL |
356 | pr_info("rpcrdma: connection to %pIS:%u closed (%d)\n", |
357 | sap, rpc_get_port(sap), connstate); | |
b3cd8d45 TT |
358 | } |
359 | #endif | |
360 | ||
c56c65fb TT |
361 | return 0; |
362 | } | |
363 | ||
d0f36c46 DS |
364 | static void rpcrdma_destroy_id(struct rdma_cm_id *id) |
365 | { | |
366 | if (id) { | |
367 | module_put(id->device->owner); | |
368 | rdma_destroy_id(id); | |
369 | } | |
370 | } | |
371 | ||
c56c65fb TT |
372 | static struct rdma_cm_id * |
373 | rpcrdma_create_id(struct rpcrdma_xprt *xprt, | |
374 | struct rpcrdma_ia *ia, struct sockaddr *addr) | |
375 | { | |
376 | struct rdma_cm_id *id; | |
377 | int rc; | |
378 | ||
1a954051 TT |
379 | init_completion(&ia->ri_done); |
380 | ||
fa20105e GS |
381 | id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP, |
382 | IB_QPT_RC); | |
c56c65fb TT |
383 | if (IS_ERR(id)) { |
384 | rc = PTR_ERR(id); | |
385 | dprintk("RPC: %s: rdma_create_id() failed %i\n", | |
386 | __func__, rc); | |
387 | return id; | |
388 | } | |
389 | ||
5675add3 | 390 | ia->ri_async_rc = -ETIMEDOUT; |
c56c65fb TT |
391 | rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT); |
392 | if (rc) { | |
393 | dprintk("RPC: %s: rdma_resolve_addr() failed %i\n", | |
394 | __func__, rc); | |
395 | goto out; | |
396 | } | |
5675add3 TT |
397 | wait_for_completion_interruptible_timeout(&ia->ri_done, |
398 | msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1); | |
d0f36c46 DS |
399 | |
400 | /* FIXME: | |
401 | * Until xprtrdma supports DEVICE_REMOVAL, the provider must | |
402 | * be pinned while there are active NFS/RDMA mounts to prevent | |
403 | * hangs and crashes at umount time. | |
404 | */ | |
405 | if (!ia->ri_async_rc && !try_module_get(id->device->owner)) { | |
406 | dprintk("RPC: %s: Failed to get device module\n", | |
407 | __func__); | |
408 | ia->ri_async_rc = -ENODEV; | |
409 | } | |
c56c65fb TT |
410 | rc = ia->ri_async_rc; |
411 | if (rc) | |
412 | goto out; | |
413 | ||
5675add3 | 414 | ia->ri_async_rc = -ETIMEDOUT; |
c56c65fb TT |
415 | rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT); |
416 | if (rc) { | |
417 | dprintk("RPC: %s: rdma_resolve_route() failed %i\n", | |
418 | __func__, rc); | |
d0f36c46 | 419 | goto put; |
c56c65fb | 420 | } |
5675add3 TT |
421 | wait_for_completion_interruptible_timeout(&ia->ri_done, |
422 | msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1); | |
c56c65fb TT |
423 | rc = ia->ri_async_rc; |
424 | if (rc) | |
d0f36c46 | 425 | goto put; |
c56c65fb TT |
426 | |
427 | return id; | |
d0f36c46 DS |
428 | put: |
429 | module_put(id->device->owner); | |
c56c65fb TT |
430 | out: |
431 | rdma_destroy_id(id); | |
432 | return ERR_PTR(rc); | |
433 | } | |
434 | ||
435 | /* | |
436 | * Drain any cq, prior to teardown. | |
437 | */ | |
438 | static void | |
439 | rpcrdma_clean_cq(struct ib_cq *cq) | |
440 | { | |
441 | struct ib_wc wc; | |
442 | int count = 0; | |
443 | ||
444 | while (1 == ib_poll_cq(cq, 1, &wc)) | |
445 | ++count; | |
446 | ||
447 | if (count) | |
448 | dprintk("RPC: %s: flushed %d events (last 0x%x)\n", | |
449 | __func__, count, wc.opcode); | |
450 | } | |
451 | ||
452 | /* | |
453 | * Exported functions. | |
454 | */ | |
455 | ||
456 | /* | |
457 | * Open and initialize an Interface Adapter. | |
458 | * o initializes fields of struct rpcrdma_ia, including | |
459 | * interface and provider attributes and protection zone. | |
460 | */ | |
461 | int | |
462 | rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg) | |
463 | { | |
c56c65fb | 464 | struct rpcrdma_ia *ia = &xprt->rx_ia; |
7bc7972c | 465 | struct ib_device_attr *devattr = &ia->ri_devattr; |
d1ed857e CL |
466 | int rc; |
467 | ||
468 | ia->ri_dma_mr = NULL; | |
c56c65fb | 469 | |
c56c65fb TT |
470 | ia->ri_id = rpcrdma_create_id(xprt, ia, addr); |
471 | if (IS_ERR(ia->ri_id)) { | |
472 | rc = PTR_ERR(ia->ri_id); | |
473 | goto out1; | |
474 | } | |
89e0d112 | 475 | ia->ri_device = ia->ri_id->device; |
c56c65fb | 476 | |
89e0d112 | 477 | ia->ri_pd = ib_alloc_pd(ia->ri_device); |
c56c65fb TT |
478 | if (IS_ERR(ia->ri_pd)) { |
479 | rc = PTR_ERR(ia->ri_pd); | |
480 | dprintk("RPC: %s: ib_alloc_pd() failed %i\n", | |
481 | __func__, rc); | |
482 | goto out2; | |
483 | } | |
484 | ||
89e0d112 | 485 | rc = ib_query_device(ia->ri_device, devattr); |
bd7ed1d1 TT |
486 | if (rc) { |
487 | dprintk("RPC: %s: ib_query_device failed %d\n", | |
488 | __func__, rc); | |
5ae711a2 | 489 | goto out3; |
bd7ed1d1 TT |
490 | } |
491 | ||
f10eafd3 | 492 | if (memreg == RPCRDMA_FRMR) { |
f022fa88 SG |
493 | if (!(devattr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) || |
494 | (devattr->max_fast_reg_page_list_len == 0)) { | |
3197d309 | 495 | dprintk("RPC: %s: FRMR registration " |
f10eafd3 CL |
496 | "not supported by HCA\n", __func__); |
497 | memreg = RPCRDMA_MTHCAFMR; | |
bd7ed1d1 | 498 | } |
f10eafd3 CL |
499 | } |
500 | if (memreg == RPCRDMA_MTHCAFMR) { | |
89e0d112 | 501 | if (!ia->ri_device->alloc_fmr) { |
f10eafd3 CL |
502 | dprintk("RPC: %s: MTHCAFMR registration " |
503 | "not supported by HCA\n", __func__); | |
f022fa88 | 504 | rc = -EINVAL; |
d2310930 | 505 | goto out3; |
f10eafd3 | 506 | } |
bd7ed1d1 TT |
507 | } |
508 | ||
bd7ed1d1 | 509 | switch (memreg) { |
3197d309 | 510 | case RPCRDMA_FRMR: |
a0ce85f5 | 511 | ia->ri_ops = &rpcrdma_frwr_memreg_ops; |
bd7ed1d1 | 512 | break; |
bd7ed1d1 | 513 | case RPCRDMA_ALLPHYSICAL: |
a0ce85f5 | 514 | ia->ri_ops = &rpcrdma_physical_memreg_ops; |
d1ed857e | 515 | break; |
bd7ed1d1 | 516 | case RPCRDMA_MTHCAFMR: |
a0ce85f5 | 517 | ia->ri_ops = &rpcrdma_fmr_memreg_ops; |
bd7ed1d1 TT |
518 | break; |
519 | default: | |
cdd9ade7 CL |
520 | printk(KERN_ERR "RPC: Unsupported memory " |
521 | "registration mode: %d\n", memreg); | |
522 | rc = -ENOMEM; | |
5ae711a2 | 523 | goto out3; |
c56c65fb | 524 | } |
a0ce85f5 CL |
525 | dprintk("RPC: %s: memory registration strategy is '%s'\n", |
526 | __func__, ia->ri_ops->ro_displayname); | |
c56c65fb | 527 | |
73806c88 | 528 | rwlock_init(&ia->ri_qplock); |
c56c65fb | 529 | return 0; |
5ae711a2 CL |
530 | |
531 | out3: | |
532 | ib_dealloc_pd(ia->ri_pd); | |
533 | ia->ri_pd = NULL; | |
c56c65fb | 534 | out2: |
d0f36c46 | 535 | rpcrdma_destroy_id(ia->ri_id); |
fee08caf | 536 | ia->ri_id = NULL; |
c56c65fb TT |
537 | out1: |
538 | return rc; | |
539 | } | |
540 | ||
541 | /* | |
542 | * Clean up/close an IA. | |
543 | * o if event handles and PD have been initialized, free them. | |
544 | * o close the IA | |
545 | */ | |
546 | void | |
547 | rpcrdma_ia_close(struct rpcrdma_ia *ia) | |
548 | { | |
c56c65fb | 549 | dprintk("RPC: %s: entering\n", __func__); |
fee08caf TT |
550 | if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) { |
551 | if (ia->ri_id->qp) | |
552 | rdma_destroy_qp(ia->ri_id); | |
d0f36c46 | 553 | rpcrdma_destroy_id(ia->ri_id); |
fee08caf TT |
554 | ia->ri_id = NULL; |
555 | } | |
6d44698d CL |
556 | |
557 | /* If the pd is still busy, xprtrdma missed freeing a resource */ | |
558 | if (ia->ri_pd && !IS_ERR(ia->ri_pd)) | |
7dd78647 | 559 | ib_dealloc_pd(ia->ri_pd); |
c56c65fb TT |
560 | } |
561 | ||
562 | /* | |
563 | * Create unconnected endpoint. | |
564 | */ | |
565 | int | |
566 | rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia, | |
567 | struct rpcrdma_create_data_internal *cdata) | |
568 | { | |
7bc7972c | 569 | struct ib_device_attr *devattr = &ia->ri_devattr; |
fc664485 | 570 | struct ib_cq *sendcq, *recvcq; |
8e37210b | 571 | struct ib_cq_init_attr cq_attr = {}; |
124fa17d | 572 | unsigned int max_qp_wr; |
5d40a8a5 | 573 | int rc, err; |
c56c65fb | 574 | |
b3221d6a CL |
575 | if (devattr->max_sge < RPCRDMA_MAX_IOVS) { |
576 | dprintk("RPC: %s: insufficient sge's available\n", | |
577 | __func__); | |
578 | return -ENOMEM; | |
579 | } | |
580 | ||
124fa17d CL |
581 | if (devattr->max_qp_wr <= RPCRDMA_BACKWARD_WRS) { |
582 | dprintk("RPC: %s: insufficient wqe's available\n", | |
583 | __func__); | |
584 | return -ENOMEM; | |
585 | } | |
586 | max_qp_wr = devattr->max_qp_wr - RPCRDMA_BACKWARD_WRS; | |
587 | ||
c56c65fb | 588 | /* check provider's send/recv wr limits */ |
124fa17d CL |
589 | if (cdata->max_requests > max_qp_wr) |
590 | cdata->max_requests = max_qp_wr; | |
c56c65fb TT |
591 | |
592 | ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall; | |
593 | ep->rep_attr.qp_context = ep; | |
c56c65fb TT |
594 | ep->rep_attr.srq = NULL; |
595 | ep->rep_attr.cap.max_send_wr = cdata->max_requests; | |
124fa17d | 596 | ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS; |
3968cb58 CL |
597 | rc = ia->ri_ops->ro_open(ia, ep, cdata); |
598 | if (rc) | |
599 | return rc; | |
c56c65fb | 600 | ep->rep_attr.cap.max_recv_wr = cdata->max_requests; |
124fa17d | 601 | ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS; |
b3221d6a | 602 | ep->rep_attr.cap.max_send_sge = RPCRDMA_MAX_IOVS; |
c56c65fb TT |
603 | ep->rep_attr.cap.max_recv_sge = 1; |
604 | ep->rep_attr.cap.max_inline_data = 0; | |
605 | ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR; | |
606 | ep->rep_attr.qp_type = IB_QPT_RC; | |
607 | ep->rep_attr.port_num = ~0; | |
608 | ||
609 | dprintk("RPC: %s: requested max: dtos: send %d recv %d; " | |
610 | "iovs: send %d recv %d\n", | |
611 | __func__, | |
612 | ep->rep_attr.cap.max_send_wr, | |
613 | ep->rep_attr.cap.max_recv_wr, | |
614 | ep->rep_attr.cap.max_send_sge, | |
615 | ep->rep_attr.cap.max_recv_sge); | |
616 | ||
617 | /* set trigger for requesting send completion */ | |
fc664485 | 618 | ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1; |
e7104a2a CL |
619 | if (ep->rep_cqinit > RPCRDMA_MAX_UNSIGNALED_SENDS) |
620 | ep->rep_cqinit = RPCRDMA_MAX_UNSIGNALED_SENDS; | |
621 | else if (ep->rep_cqinit <= 2) | |
c56c65fb TT |
622 | ep->rep_cqinit = 0; |
623 | INIT_CQCOUNT(ep); | |
c56c65fb | 624 | init_waitqueue_head(&ep->rep_connect_wait); |
254f91e2 | 625 | INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker); |
c56c65fb | 626 | |
8e37210b | 627 | cq_attr.cqe = ep->rep_attr.cap.max_send_wr + 1; |
89e0d112 | 628 | sendcq = ib_create_cq(ia->ri_device, rpcrdma_sendcq_upcall, |
4220a072 | 629 | rpcrdma_cq_async_error_upcall, NULL, &cq_attr); |
fc664485 CL |
630 | if (IS_ERR(sendcq)) { |
631 | rc = PTR_ERR(sendcq); | |
632 | dprintk("RPC: %s: failed to create send CQ: %i\n", | |
c56c65fb TT |
633 | __func__, rc); |
634 | goto out1; | |
635 | } | |
636 | ||
fc664485 | 637 | rc = ib_req_notify_cq(sendcq, IB_CQ_NEXT_COMP); |
c56c65fb TT |
638 | if (rc) { |
639 | dprintk("RPC: %s: ib_req_notify_cq failed: %i\n", | |
640 | __func__, rc); | |
641 | goto out2; | |
642 | } | |
643 | ||
8e37210b | 644 | cq_attr.cqe = ep->rep_attr.cap.max_recv_wr + 1; |
89e0d112 | 645 | recvcq = ib_create_cq(ia->ri_device, rpcrdma_recvcq_upcall, |
4220a072 | 646 | rpcrdma_cq_async_error_upcall, NULL, &cq_attr); |
fc664485 CL |
647 | if (IS_ERR(recvcq)) { |
648 | rc = PTR_ERR(recvcq); | |
649 | dprintk("RPC: %s: failed to create recv CQ: %i\n", | |
650 | __func__, rc); | |
651 | goto out2; | |
652 | } | |
653 | ||
654 | rc = ib_req_notify_cq(recvcq, IB_CQ_NEXT_COMP); | |
655 | if (rc) { | |
656 | dprintk("RPC: %s: ib_req_notify_cq failed: %i\n", | |
657 | __func__, rc); | |
658 | ib_destroy_cq(recvcq); | |
659 | goto out2; | |
660 | } | |
661 | ||
662 | ep->rep_attr.send_cq = sendcq; | |
663 | ep->rep_attr.recv_cq = recvcq; | |
c56c65fb TT |
664 | |
665 | /* Initialize cma parameters */ | |
666 | ||
667 | /* RPC/RDMA does not use private data */ | |
668 | ep->rep_remote_cma.private_data = NULL; | |
669 | ep->rep_remote_cma.private_data_len = 0; | |
670 | ||
671 | /* Client offers RDMA Read but does not initiate */ | |
b334eaab | 672 | ep->rep_remote_cma.initiator_depth = 0; |
7bc7972c | 673 | if (devattr->max_qp_rd_atom > 32) /* arbitrary but <= 255 */ |
b334eaab TT |
674 | ep->rep_remote_cma.responder_resources = 32; |
675 | else | |
7bc7972c CL |
676 | ep->rep_remote_cma.responder_resources = |
677 | devattr->max_qp_rd_atom; | |
c56c65fb TT |
678 | |
679 | ep->rep_remote_cma.retry_count = 7; | |
680 | ep->rep_remote_cma.flow_control = 0; | |
681 | ep->rep_remote_cma.rnr_retry_count = 0; | |
682 | ||
683 | return 0; | |
684 | ||
685 | out2: | |
fc664485 | 686 | err = ib_destroy_cq(sendcq); |
5d40a8a5 CL |
687 | if (err) |
688 | dprintk("RPC: %s: ib_destroy_cq returned %i\n", | |
689 | __func__, err); | |
c56c65fb | 690 | out1: |
d1ed857e CL |
691 | if (ia->ri_dma_mr) |
692 | ib_dereg_mr(ia->ri_dma_mr); | |
c56c65fb TT |
693 | return rc; |
694 | } | |
695 | ||
696 | /* | |
697 | * rpcrdma_ep_destroy | |
698 | * | |
699 | * Disconnect and destroy endpoint. After this, the only | |
700 | * valid operations on the ep are to free it (if dynamically | |
701 | * allocated) or re-create it. | |
c56c65fb | 702 | */ |
7f1d5419 | 703 | void |
c56c65fb TT |
704 | rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia) |
705 | { | |
706 | int rc; | |
707 | ||
708 | dprintk("RPC: %s: entering, connected is %d\n", | |
709 | __func__, ep->rep_connected); | |
710 | ||
254f91e2 CL |
711 | cancel_delayed_work_sync(&ep->rep_connect_worker); |
712 | ||
72c02173 | 713 | if (ia->ri_id->qp) |
282191cb | 714 | rpcrdma_ep_disconnect(ep, ia); |
72c02173 SW |
715 | |
716 | rpcrdma_clean_cq(ep->rep_attr.recv_cq); | |
717 | rpcrdma_clean_cq(ep->rep_attr.send_cq); | |
718 | ||
719 | if (ia->ri_id->qp) { | |
fee08caf TT |
720 | rdma_destroy_qp(ia->ri_id); |
721 | ia->ri_id->qp = NULL; | |
c56c65fb TT |
722 | } |
723 | ||
fc664485 CL |
724 | rc = ib_destroy_cq(ep->rep_attr.recv_cq); |
725 | if (rc) | |
726 | dprintk("RPC: %s: ib_destroy_cq returned %i\n", | |
727 | __func__, rc); | |
728 | ||
fc664485 | 729 | rc = ib_destroy_cq(ep->rep_attr.send_cq); |
c56c65fb TT |
730 | if (rc) |
731 | dprintk("RPC: %s: ib_destroy_cq returned %i\n", | |
732 | __func__, rc); | |
d1ed857e CL |
733 | |
734 | if (ia->ri_dma_mr) { | |
735 | rc = ib_dereg_mr(ia->ri_dma_mr); | |
736 | dprintk("RPC: %s: ib_dereg_mr returned %i\n", | |
737 | __func__, rc); | |
738 | } | |
c56c65fb TT |
739 | } |
740 | ||
741 | /* | |
742 | * Connect unconnected endpoint. | |
743 | */ | |
744 | int | |
745 | rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia) | |
746 | { | |
73806c88 | 747 | struct rdma_cm_id *id, *old; |
c56c65fb TT |
748 | int rc = 0; |
749 | int retry_count = 0; | |
c56c65fb | 750 | |
c055551e | 751 | if (ep->rep_connected != 0) { |
c56c65fb TT |
752 | struct rpcrdma_xprt *xprt; |
753 | retry: | |
ec62f40d | 754 | dprintk("RPC: %s: reconnecting...\n", __func__); |
282191cb CL |
755 | |
756 | rpcrdma_ep_disconnect(ep, ia); | |
a7bc211a | 757 | rpcrdma_flush_cqs(ep); |
c56c65fb TT |
758 | |
759 | xprt = container_of(ia, struct rpcrdma_xprt, rx_ia); | |
760 | id = rpcrdma_create_id(xprt, ia, | |
761 | (struct sockaddr *)&xprt->rx_data.addr); | |
762 | if (IS_ERR(id)) { | |
ec62f40d | 763 | rc = -EHOSTUNREACH; |
c56c65fb TT |
764 | goto out; |
765 | } | |
766 | /* TEMP TEMP TEMP - fail if new device: | |
767 | * Deregister/remarshal *all* requests! | |
768 | * Close and recreate adapter, pd, etc! | |
769 | * Re-determine all attributes still sane! | |
770 | * More stuff I haven't thought of! | |
771 | * Rrrgh! | |
772 | */ | |
89e0d112 | 773 | if (ia->ri_device != id->device) { |
c56c65fb TT |
774 | printk("RPC: %s: can't reconnect on " |
775 | "different device!\n", __func__); | |
d0f36c46 | 776 | rpcrdma_destroy_id(id); |
ec62f40d | 777 | rc = -ENETUNREACH; |
c56c65fb TT |
778 | goto out; |
779 | } | |
780 | /* END TEMP */ | |
ec62f40d CL |
781 | rc = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr); |
782 | if (rc) { | |
783 | dprintk("RPC: %s: rdma_create_qp failed %i\n", | |
784 | __func__, rc); | |
d0f36c46 | 785 | rpcrdma_destroy_id(id); |
ec62f40d CL |
786 | rc = -ENETUNREACH; |
787 | goto out; | |
788 | } | |
73806c88 CL |
789 | |
790 | write_lock(&ia->ri_qplock); | |
791 | old = ia->ri_id; | |
c56c65fb | 792 | ia->ri_id = id; |
73806c88 CL |
793 | write_unlock(&ia->ri_qplock); |
794 | ||
795 | rdma_destroy_qp(old); | |
d0f36c46 | 796 | rpcrdma_destroy_id(old); |
ec62f40d CL |
797 | } else { |
798 | dprintk("RPC: %s: connecting...\n", __func__); | |
799 | rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr); | |
800 | if (rc) { | |
801 | dprintk("RPC: %s: rdma_create_qp failed %i\n", | |
802 | __func__, rc); | |
803 | /* do not update ep->rep_connected */ | |
804 | return -ENETUNREACH; | |
805 | } | |
c56c65fb TT |
806 | } |
807 | ||
c56c65fb TT |
808 | ep->rep_connected = 0; |
809 | ||
810 | rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma); | |
811 | if (rc) { | |
812 | dprintk("RPC: %s: rdma_connect() failed with %i\n", | |
813 | __func__, rc); | |
814 | goto out; | |
815 | } | |
816 | ||
c56c65fb TT |
817 | wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0); |
818 | ||
819 | /* | |
820 | * Check state. A non-peer reject indicates no listener | |
821 | * (ECONNREFUSED), which may be a transient state. All | |
822 | * others indicate a transport condition which has already | |
823 | * undergone a best-effort. | |
824 | */ | |
f64f9e71 JP |
825 | if (ep->rep_connected == -ECONNREFUSED && |
826 | ++retry_count <= RDMA_CONNECT_RETRY_MAX) { | |
c56c65fb TT |
827 | dprintk("RPC: %s: non-peer_reject, retry\n", __func__); |
828 | goto retry; | |
829 | } | |
830 | if (ep->rep_connected <= 0) { | |
831 | /* Sometimes, the only way to reliably connect to remote | |
832 | * CMs is to use same nonzero values for ORD and IRD. */ | |
b334eaab TT |
833 | if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 && |
834 | (ep->rep_remote_cma.responder_resources == 0 || | |
835 | ep->rep_remote_cma.initiator_depth != | |
836 | ep->rep_remote_cma.responder_resources)) { | |
837 | if (ep->rep_remote_cma.responder_resources == 0) | |
838 | ep->rep_remote_cma.responder_resources = 1; | |
839 | ep->rep_remote_cma.initiator_depth = | |
840 | ep->rep_remote_cma.responder_resources; | |
c56c65fb | 841 | goto retry; |
b334eaab | 842 | } |
c56c65fb TT |
843 | rc = ep->rep_connected; |
844 | } else { | |
f531a5db CL |
845 | struct rpcrdma_xprt *r_xprt; |
846 | unsigned int extras; | |
847 | ||
c56c65fb | 848 | dprintk("RPC: %s: connected\n", __func__); |
f531a5db CL |
849 | |
850 | r_xprt = container_of(ia, struct rpcrdma_xprt, rx_ia); | |
851 | extras = r_xprt->rx_buf.rb_bc_srv_max_requests; | |
852 | ||
853 | if (extras) { | |
854 | rc = rpcrdma_ep_post_extra_recv(r_xprt, extras); | |
38b95bcf | 855 | if (rc) { |
f531a5db CL |
856 | pr_warn("%s: rpcrdma_ep_post_extra_recv: %i\n", |
857 | __func__, rc); | |
858 | rc = 0; | |
38b95bcf | 859 | } |
f531a5db | 860 | } |
c56c65fb TT |
861 | } |
862 | ||
863 | out: | |
864 | if (rc) | |
865 | ep->rep_connected = rc; | |
866 | return rc; | |
867 | } | |
868 | ||
869 | /* | |
870 | * rpcrdma_ep_disconnect | |
871 | * | |
872 | * This is separate from destroy to facilitate the ability | |
873 | * to reconnect without recreating the endpoint. | |
874 | * | |
875 | * This call is not reentrant, and must not be made in parallel | |
876 | * on the same endpoint. | |
877 | */ | |
282191cb | 878 | void |
c56c65fb TT |
879 | rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia) |
880 | { | |
881 | int rc; | |
882 | ||
a7bc211a | 883 | rpcrdma_flush_cqs(ep); |
c56c65fb TT |
884 | rc = rdma_disconnect(ia->ri_id); |
885 | if (!rc) { | |
886 | /* returns without wait if not connected */ | |
887 | wait_event_interruptible(ep->rep_connect_wait, | |
888 | ep->rep_connected != 1); | |
889 | dprintk("RPC: %s: after wait, %sconnected\n", __func__, | |
890 | (ep->rep_connected == 1) ? "still " : "dis"); | |
891 | } else { | |
892 | dprintk("RPC: %s: rdma_disconnect %i\n", __func__, rc); | |
893 | ep->rep_connected = rc; | |
894 | } | |
c56c65fb TT |
895 | } |
896 | ||
f531a5db | 897 | struct rpcrdma_req * |
1392402c CL |
898 | rpcrdma_create_req(struct rpcrdma_xprt *r_xprt) |
899 | { | |
f531a5db | 900 | struct rpcrdma_buffer *buffer = &r_xprt->rx_buf; |
1392402c | 901 | struct rpcrdma_req *req; |
1392402c | 902 | |
85275c87 | 903 | req = kzalloc(sizeof(*req), GFP_KERNEL); |
1392402c | 904 | if (req == NULL) |
85275c87 | 905 | return ERR_PTR(-ENOMEM); |
1392402c | 906 | |
f531a5db CL |
907 | INIT_LIST_HEAD(&req->rl_free); |
908 | spin_lock(&buffer->rb_reqslock); | |
909 | list_add(&req->rl_all, &buffer->rb_allreqs); | |
910 | spin_unlock(&buffer->rb_reqslock); | |
1392402c CL |
911 | req->rl_buffer = &r_xprt->rx_buf; |
912 | return req; | |
1392402c CL |
913 | } |
914 | ||
f531a5db | 915 | struct rpcrdma_rep * |
1392402c CL |
916 | rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt) |
917 | { | |
918 | struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data; | |
1392402c CL |
919 | struct rpcrdma_ia *ia = &r_xprt->rx_ia; |
920 | struct rpcrdma_rep *rep; | |
921 | int rc; | |
922 | ||
923 | rc = -ENOMEM; | |
6b1184cd | 924 | rep = kzalloc(sizeof(*rep), GFP_KERNEL); |
1392402c CL |
925 | if (rep == NULL) |
926 | goto out; | |
1392402c | 927 | |
6b1184cd CL |
928 | rep->rr_rdmabuf = rpcrdma_alloc_regbuf(ia, cdata->inline_rsize, |
929 | GFP_KERNEL); | |
930 | if (IS_ERR(rep->rr_rdmabuf)) { | |
931 | rc = PTR_ERR(rep->rr_rdmabuf); | |
1392402c | 932 | goto out_free; |
6b1184cd | 933 | } |
1392402c | 934 | |
89e0d112 | 935 | rep->rr_device = ia->ri_device; |
fed171b3 | 936 | rep->rr_rxprt = r_xprt; |
fe97b47c | 937 | INIT_WORK(&rep->rr_work, rpcrdma_receive_worker); |
1392402c CL |
938 | return rep; |
939 | ||
940 | out_free: | |
941 | kfree(rep); | |
942 | out: | |
943 | return ERR_PTR(rc); | |
944 | } | |
945 | ||
c56c65fb | 946 | int |
ac920d04 | 947 | rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt) |
c56c65fb | 948 | { |
ac920d04 CL |
949 | struct rpcrdma_buffer *buf = &r_xprt->rx_buf; |
950 | struct rpcrdma_ia *ia = &r_xprt->rx_ia; | |
c56c65fb TT |
951 | int i, rc; |
952 | ||
1e465fd4 | 953 | buf->rb_max_requests = r_xprt->rx_data.max_requests; |
f531a5db | 954 | buf->rb_bc_srv_max_requests = 0; |
c56c65fb | 955 | spin_lock_init(&buf->rb_lock); |
c56c65fb | 956 | |
91e70e70 CL |
957 | rc = ia->ri_ops->ro_init(r_xprt); |
958 | if (rc) | |
959 | goto out; | |
c56c65fb | 960 | |
1e465fd4 | 961 | INIT_LIST_HEAD(&buf->rb_send_bufs); |
f531a5db CL |
962 | INIT_LIST_HEAD(&buf->rb_allreqs); |
963 | spin_lock_init(&buf->rb_reqslock); | |
c56c65fb TT |
964 | for (i = 0; i < buf->rb_max_requests; i++) { |
965 | struct rpcrdma_req *req; | |
c56c65fb | 966 | |
1392402c CL |
967 | req = rpcrdma_create_req(r_xprt); |
968 | if (IS_ERR(req)) { | |
c56c65fb TT |
969 | dprintk("RPC: %s: request buffer %d alloc" |
970 | " failed\n", __func__, i); | |
1392402c | 971 | rc = PTR_ERR(req); |
c56c65fb TT |
972 | goto out; |
973 | } | |
f531a5db | 974 | req->rl_backchannel = false; |
1e465fd4 CL |
975 | list_add(&req->rl_free, &buf->rb_send_bufs); |
976 | } | |
977 | ||
978 | INIT_LIST_HEAD(&buf->rb_recv_bufs); | |
979 | for (i = 0; i < buf->rb_max_requests + 2; i++) { | |
980 | struct rpcrdma_rep *rep; | |
c56c65fb | 981 | |
1392402c CL |
982 | rep = rpcrdma_create_rep(r_xprt); |
983 | if (IS_ERR(rep)) { | |
c56c65fb TT |
984 | dprintk("RPC: %s: reply buffer %d alloc failed\n", |
985 | __func__, i); | |
1392402c | 986 | rc = PTR_ERR(rep); |
c56c65fb TT |
987 | goto out; |
988 | } | |
1e465fd4 | 989 | list_add(&rep->rr_list, &buf->rb_recv_bufs); |
c56c65fb | 990 | } |
1392402c | 991 | |
c56c65fb TT |
992 | return 0; |
993 | out: | |
994 | rpcrdma_buffer_destroy(buf); | |
995 | return rc; | |
996 | } | |
997 | ||
1e465fd4 CL |
998 | static struct rpcrdma_req * |
999 | rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf) | |
1000 | { | |
1001 | struct rpcrdma_req *req; | |
1002 | ||
1003 | req = list_first_entry(&buf->rb_send_bufs, | |
1004 | struct rpcrdma_req, rl_free); | |
1005 | list_del(&req->rl_free); | |
1006 | return req; | |
1007 | } | |
1008 | ||
1009 | static struct rpcrdma_rep * | |
1010 | rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf) | |
1011 | { | |
1012 | struct rpcrdma_rep *rep; | |
1013 | ||
1014 | rep = list_first_entry(&buf->rb_recv_bufs, | |
1015 | struct rpcrdma_rep, rr_list); | |
1016 | list_del(&rep->rr_list); | |
1017 | return rep; | |
1018 | } | |
1019 | ||
1392402c CL |
1020 | static void |
1021 | rpcrdma_destroy_rep(struct rpcrdma_ia *ia, struct rpcrdma_rep *rep) | |
1022 | { | |
6b1184cd | 1023 | rpcrdma_free_regbuf(ia, rep->rr_rdmabuf); |
1392402c CL |
1024 | kfree(rep); |
1025 | } | |
1026 | ||
f531a5db | 1027 | void |
1392402c CL |
1028 | rpcrdma_destroy_req(struct rpcrdma_ia *ia, struct rpcrdma_req *req) |
1029 | { | |
0ca77dc3 | 1030 | rpcrdma_free_regbuf(ia, req->rl_sendbuf); |
85275c87 | 1031 | rpcrdma_free_regbuf(ia, req->rl_rdmabuf); |
1392402c CL |
1032 | kfree(req); |
1033 | } | |
1034 | ||
c56c65fb TT |
1035 | void |
1036 | rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf) | |
1037 | { | |
c56c65fb TT |
1038 | struct rpcrdma_ia *ia = rdmab_to_ia(buf); |
1039 | ||
1e465fd4 CL |
1040 | while (!list_empty(&buf->rb_recv_bufs)) { |
1041 | struct rpcrdma_rep *rep; | |
c56c65fb | 1042 | |
1e465fd4 CL |
1043 | rep = rpcrdma_buffer_get_rep_locked(buf); |
1044 | rpcrdma_destroy_rep(ia, rep); | |
c56c65fb TT |
1045 | } |
1046 | ||
f531a5db CL |
1047 | spin_lock(&buf->rb_reqslock); |
1048 | while (!list_empty(&buf->rb_allreqs)) { | |
1e465fd4 | 1049 | struct rpcrdma_req *req; |
4034ba04 | 1050 | |
f531a5db CL |
1051 | req = list_first_entry(&buf->rb_allreqs, |
1052 | struct rpcrdma_req, rl_all); | |
1053 | list_del(&req->rl_all); | |
1054 | ||
1055 | spin_unlock(&buf->rb_reqslock); | |
1e465fd4 | 1056 | rpcrdma_destroy_req(ia, req); |
f531a5db | 1057 | spin_lock(&buf->rb_reqslock); |
1e465fd4 | 1058 | } |
f531a5db | 1059 | spin_unlock(&buf->rb_reqslock); |
4034ba04 | 1060 | |
1e465fd4 | 1061 | ia->ri_ops->ro_destroy(buf); |
c56c65fb TT |
1062 | } |
1063 | ||
346aa66b CL |
1064 | struct rpcrdma_mw * |
1065 | rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt) | |
c2922c02 | 1066 | { |
346aa66b CL |
1067 | struct rpcrdma_buffer *buf = &r_xprt->rx_buf; |
1068 | struct rpcrdma_mw *mw = NULL; | |
346aa66b | 1069 | |
58d1dcf5 | 1070 | spin_lock(&buf->rb_mwlock); |
346aa66b CL |
1071 | if (!list_empty(&buf->rb_mws)) { |
1072 | mw = list_first_entry(&buf->rb_mws, | |
1073 | struct rpcrdma_mw, mw_list); | |
1074 | list_del_init(&mw->mw_list); | |
c2922c02 | 1075 | } |
58d1dcf5 | 1076 | spin_unlock(&buf->rb_mwlock); |
346aa66b CL |
1077 | |
1078 | if (!mw) | |
1079 | pr_err("RPC: %s: no MWs available\n", __func__); | |
1080 | return mw; | |
c2922c02 CL |
1081 | } |
1082 | ||
346aa66b CL |
1083 | void |
1084 | rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw) | |
c2922c02 | 1085 | { |
346aa66b | 1086 | struct rpcrdma_buffer *buf = &r_xprt->rx_buf; |
c2922c02 | 1087 | |
58d1dcf5 | 1088 | spin_lock(&buf->rb_mwlock); |
346aa66b | 1089 | list_add_tail(&mw->mw_list, &buf->rb_mws); |
58d1dcf5 | 1090 | spin_unlock(&buf->rb_mwlock); |
c2922c02 CL |
1091 | } |
1092 | ||
c56c65fb TT |
1093 | /* |
1094 | * Get a set of request/reply buffers. | |
1095 | * | |
1e465fd4 | 1096 | * Reply buffer (if available) is attached to send buffer upon return. |
c56c65fb TT |
1097 | */ |
1098 | struct rpcrdma_req * | |
1099 | rpcrdma_buffer_get(struct rpcrdma_buffer *buffers) | |
1100 | { | |
1101 | struct rpcrdma_req *req; | |
c14d86e5 | 1102 | |
a5b027e1 | 1103 | spin_lock(&buffers->rb_lock); |
1e465fd4 CL |
1104 | if (list_empty(&buffers->rb_send_bufs)) |
1105 | goto out_reqbuf; | |
1106 | req = rpcrdma_buffer_get_req_locked(buffers); | |
1107 | if (list_empty(&buffers->rb_recv_bufs)) | |
1108 | goto out_repbuf; | |
1109 | req->rl_reply = rpcrdma_buffer_get_rep_locked(buffers); | |
a5b027e1 | 1110 | spin_unlock(&buffers->rb_lock); |
1e465fd4 | 1111 | return req; |
ddb6bebc | 1112 | |
1e465fd4 | 1113 | out_reqbuf: |
a5b027e1 | 1114 | spin_unlock(&buffers->rb_lock); |
1e465fd4 CL |
1115 | pr_warn("RPC: %s: out of request buffers\n", __func__); |
1116 | return NULL; | |
1117 | out_repbuf: | |
a5b027e1 | 1118 | spin_unlock(&buffers->rb_lock); |
1e465fd4 CL |
1119 | pr_warn("RPC: %s: out of reply buffers\n", __func__); |
1120 | req->rl_reply = NULL; | |
c56c65fb TT |
1121 | return req; |
1122 | } | |
1123 | ||
1124 | /* | |
1125 | * Put request/reply buffers back into pool. | |
1126 | * Pre-decrement counter/array index. | |
1127 | */ | |
1128 | void | |
1129 | rpcrdma_buffer_put(struct rpcrdma_req *req) | |
1130 | { | |
1131 | struct rpcrdma_buffer *buffers = req->rl_buffer; | |
1e465fd4 | 1132 | struct rpcrdma_rep *rep = req->rl_reply; |
c56c65fb | 1133 | |
1e465fd4 CL |
1134 | req->rl_niovs = 0; |
1135 | req->rl_reply = NULL; | |
1136 | ||
a5b027e1 | 1137 | spin_lock(&buffers->rb_lock); |
1e465fd4 CL |
1138 | list_add_tail(&req->rl_free, &buffers->rb_send_bufs); |
1139 | if (rep) | |
1140 | list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs); | |
a5b027e1 | 1141 | spin_unlock(&buffers->rb_lock); |
c56c65fb TT |
1142 | } |
1143 | ||
1144 | /* | |
1145 | * Recover reply buffers from pool. | |
1e465fd4 | 1146 | * This happens when recovering from disconnect. |
c56c65fb TT |
1147 | */ |
1148 | void | |
1149 | rpcrdma_recv_buffer_get(struct rpcrdma_req *req) | |
1150 | { | |
1151 | struct rpcrdma_buffer *buffers = req->rl_buffer; | |
c56c65fb | 1152 | |
a5b027e1 | 1153 | spin_lock(&buffers->rb_lock); |
1e465fd4 CL |
1154 | if (!list_empty(&buffers->rb_recv_bufs)) |
1155 | req->rl_reply = rpcrdma_buffer_get_rep_locked(buffers); | |
a5b027e1 | 1156 | spin_unlock(&buffers->rb_lock); |
c56c65fb TT |
1157 | } |
1158 | ||
1159 | /* | |
1160 | * Put reply buffers back into pool when not attached to | |
b45ccfd2 | 1161 | * request. This happens in error conditions. |
c56c65fb TT |
1162 | */ |
1163 | void | |
1164 | rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep) | |
1165 | { | |
fed171b3 | 1166 | struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf; |
c56c65fb | 1167 | |
a5b027e1 | 1168 | spin_lock(&buffers->rb_lock); |
1e465fd4 | 1169 | list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs); |
a5b027e1 | 1170 | spin_unlock(&buffers->rb_lock); |
c56c65fb TT |
1171 | } |
1172 | ||
1173 | /* | |
1174 | * Wrappers for internal-use kmalloc memory registration, used by buffer code. | |
1175 | */ | |
1176 | ||
d654788e CL |
1177 | void |
1178 | rpcrdma_mapping_error(struct rpcrdma_mr_seg *seg) | |
1179 | { | |
1180 | dprintk("RPC: map_one: offset %p iova %llx len %zu\n", | |
1181 | seg->mr_offset, | |
1182 | (unsigned long long)seg->mr_dma, seg->mr_dmalen); | |
1183 | } | |
1184 | ||
9128c3e7 CL |
1185 | /** |
1186 | * rpcrdma_alloc_regbuf - kmalloc and register memory for SEND/RECV buffers | |
1187 | * @ia: controlling rpcrdma_ia | |
1188 | * @size: size of buffer to be allocated, in bytes | |
1189 | * @flags: GFP flags | |
1190 | * | |
1191 | * Returns pointer to private header of an area of internally | |
1192 | * registered memory, or an ERR_PTR. The registered buffer follows | |
1193 | * the end of the private header. | |
1194 | * | |
1195 | * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for | |
1196 | * receiving the payload of RDMA RECV operations. regbufs are not | |
1197 | * used for RDMA READ/WRITE operations, thus are registered only for | |
1198 | * LOCAL access. | |
1199 | */ | |
1200 | struct rpcrdma_regbuf * | |
1201 | rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size, gfp_t flags) | |
1202 | { | |
1203 | struct rpcrdma_regbuf *rb; | |
e531dcab | 1204 | struct ib_sge *iov; |
9128c3e7 | 1205 | |
9128c3e7 CL |
1206 | rb = kmalloc(sizeof(*rb) + size, flags); |
1207 | if (rb == NULL) | |
1208 | goto out; | |
1209 | ||
e531dcab CL |
1210 | iov = &rb->rg_iov; |
1211 | iov->addr = ib_dma_map_single(ia->ri_device, | |
1212 | (void *)rb->rg_base, size, | |
1213 | DMA_BIDIRECTIONAL); | |
1214 | if (ib_dma_mapping_error(ia->ri_device, iov->addr)) | |
9128c3e7 CL |
1215 | goto out_free; |
1216 | ||
e531dcab | 1217 | iov->length = size; |
bb6c96d7 | 1218 | iov->lkey = ia->ri_pd->local_dma_lkey; |
e531dcab CL |
1219 | rb->rg_size = size; |
1220 | rb->rg_owner = NULL; | |
9128c3e7 CL |
1221 | return rb; |
1222 | ||
1223 | out_free: | |
1224 | kfree(rb); | |
1225 | out: | |
e531dcab | 1226 | return ERR_PTR(-ENOMEM); |
9128c3e7 CL |
1227 | } |
1228 | ||
1229 | /** | |
1230 | * rpcrdma_free_regbuf - deregister and free registered buffer | |
1231 | * @ia: controlling rpcrdma_ia | |
1232 | * @rb: regbuf to be deregistered and freed | |
1233 | */ | |
1234 | void | |
1235 | rpcrdma_free_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb) | |
1236 | { | |
e531dcab CL |
1237 | struct ib_sge *iov; |
1238 | ||
1239 | if (!rb) | |
1240 | return; | |
1241 | ||
1242 | iov = &rb->rg_iov; | |
1243 | ib_dma_unmap_single(ia->ri_device, | |
1244 | iov->addr, iov->length, DMA_BIDIRECTIONAL); | |
1245 | kfree(rb); | |
9128c3e7 CL |
1246 | } |
1247 | ||
c56c65fb TT |
1248 | /* |
1249 | * Prepost any receive buffer, then post send. | |
1250 | * | |
1251 | * Receive buffer is donated to hardware, reclaimed upon recv completion. | |
1252 | */ | |
1253 | int | |
1254 | rpcrdma_ep_post(struct rpcrdma_ia *ia, | |
1255 | struct rpcrdma_ep *ep, | |
1256 | struct rpcrdma_req *req) | |
1257 | { | |
b3221d6a | 1258 | struct ib_device *device = ia->ri_device; |
c56c65fb TT |
1259 | struct ib_send_wr send_wr, *send_wr_fail; |
1260 | struct rpcrdma_rep *rep = req->rl_reply; | |
b3221d6a CL |
1261 | struct ib_sge *iov = req->rl_send_iov; |
1262 | int i, rc; | |
c56c65fb TT |
1263 | |
1264 | if (rep) { | |
1265 | rc = rpcrdma_ep_post_recv(ia, ep, rep); | |
1266 | if (rc) | |
1267 | goto out; | |
1268 | req->rl_reply = NULL; | |
1269 | } | |
1270 | ||
1271 | send_wr.next = NULL; | |
e46ac34c | 1272 | send_wr.wr_id = RPCRDMA_IGNORE_COMPLETION; |
b3221d6a | 1273 | send_wr.sg_list = iov; |
c56c65fb TT |
1274 | send_wr.num_sge = req->rl_niovs; |
1275 | send_wr.opcode = IB_WR_SEND; | |
b3221d6a CL |
1276 | |
1277 | for (i = 0; i < send_wr.num_sge; i++) | |
1278 | ib_dma_sync_single_for_device(device, iov[i].addr, | |
1279 | iov[i].length, DMA_TO_DEVICE); | |
1280 | dprintk("RPC: %s: posting %d s/g entries\n", | |
1281 | __func__, send_wr.num_sge); | |
c56c65fb TT |
1282 | |
1283 | if (DECR_CQCOUNT(ep) > 0) | |
1284 | send_wr.send_flags = 0; | |
1285 | else { /* Provider must take a send completion every now and then */ | |
1286 | INIT_CQCOUNT(ep); | |
1287 | send_wr.send_flags = IB_SEND_SIGNALED; | |
1288 | } | |
1289 | ||
1290 | rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail); | |
1291 | if (rc) | |
1292 | dprintk("RPC: %s: ib_post_send returned %i\n", __func__, | |
1293 | rc); | |
1294 | out: | |
1295 | return rc; | |
1296 | } | |
1297 | ||
1298 | /* | |
1299 | * (Re)post a receive buffer. | |
1300 | */ | |
1301 | int | |
1302 | rpcrdma_ep_post_recv(struct rpcrdma_ia *ia, | |
1303 | struct rpcrdma_ep *ep, | |
1304 | struct rpcrdma_rep *rep) | |
1305 | { | |
1306 | struct ib_recv_wr recv_wr, *recv_wr_fail; | |
1307 | int rc; | |
1308 | ||
1309 | recv_wr.next = NULL; | |
1310 | recv_wr.wr_id = (u64) (unsigned long) rep; | |
6b1184cd | 1311 | recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov; |
c56c65fb TT |
1312 | recv_wr.num_sge = 1; |
1313 | ||
89e0d112 | 1314 | ib_dma_sync_single_for_cpu(ia->ri_device, |
6b1184cd CL |
1315 | rdmab_addr(rep->rr_rdmabuf), |
1316 | rdmab_length(rep->rr_rdmabuf), | |
1317 | DMA_BIDIRECTIONAL); | |
c56c65fb | 1318 | |
c56c65fb TT |
1319 | rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail); |
1320 | ||
1321 | if (rc) | |
1322 | dprintk("RPC: %s: ib_post_recv returned %i\n", __func__, | |
1323 | rc); | |
1324 | return rc; | |
1325 | } | |
43e95988 | 1326 | |
f531a5db CL |
1327 | /** |
1328 | * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests | |
1329 | * @r_xprt: transport associated with these backchannel resources | |
1330 | * @min_reqs: minimum number of incoming requests expected | |
1331 | * | |
1332 | * Returns zero if all requested buffers were posted, or a negative errno. | |
1333 | */ | |
1334 | int | |
1335 | rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count) | |
1336 | { | |
1337 | struct rpcrdma_buffer *buffers = &r_xprt->rx_buf; | |
1338 | struct rpcrdma_ia *ia = &r_xprt->rx_ia; | |
1339 | struct rpcrdma_ep *ep = &r_xprt->rx_ep; | |
1340 | struct rpcrdma_rep *rep; | |
1341 | unsigned long flags; | |
1342 | int rc; | |
1343 | ||
1344 | while (count--) { | |
1345 | spin_lock_irqsave(&buffers->rb_lock, flags); | |
1346 | if (list_empty(&buffers->rb_recv_bufs)) | |
1347 | goto out_reqbuf; | |
1348 | rep = rpcrdma_buffer_get_rep_locked(buffers); | |
1349 | spin_unlock_irqrestore(&buffers->rb_lock, flags); | |
1350 | ||
1351 | rc = rpcrdma_ep_post_recv(ia, ep, rep); | |
1352 | if (rc) | |
1353 | goto out_rc; | |
1354 | } | |
1355 | ||
1356 | return 0; | |
1357 | ||
1358 | out_reqbuf: | |
1359 | spin_unlock_irqrestore(&buffers->rb_lock, flags); | |
1360 | pr_warn("%s: no extra receive buffers\n", __func__); | |
1361 | return -ENOMEM; | |
1362 | ||
1363 | out_rc: | |
1364 | rpcrdma_recv_buffer_put(rep); | |
1365 | return rc; | |
1366 | } | |
1367 | ||
1c9351ee | 1368 | /* How many chunk list items fit within our inline buffers? |
43e95988 | 1369 | */ |
1c9351ee CL |
1370 | unsigned int |
1371 | rpcrdma_max_segments(struct rpcrdma_xprt *r_xprt) | |
43e95988 CL |
1372 | { |
1373 | struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data; | |
1c9351ee | 1374 | int bytes, segments; |
43e95988 | 1375 | |
1c9351ee CL |
1376 | bytes = min_t(unsigned int, cdata->inline_wsize, cdata->inline_rsize); |
1377 | bytes -= RPCRDMA_HDRLEN_MIN; | |
1378 | if (bytes < sizeof(struct rpcrdma_segment) * 2) { | |
1379 | pr_warn("RPC: %s: inline threshold too small\n", | |
1380 | __func__); | |
1381 | return 0; | |
43e95988 | 1382 | } |
1c9351ee CL |
1383 | |
1384 | segments = 1 << (fls(bytes / sizeof(struct rpcrdma_segment)) - 1); | |
1385 | dprintk("RPC: %s: max chunk list size = %d segments\n", | |
1386 | __func__, segments); | |
1387 | return segments; | |
43e95988 | 1388 | } |