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> |
05c97466 | 54 | #include <linux/sunrpc/svc_rdma.h> |
65866f82 | 55 | #include <asm/bitops.h> |
56a6bd15 | 56 | |
0a90487b | 57 | #include <rdma/ib_cm.h> |
c56c65fb | 58 | |
f58851e6 TT |
59 | #include "xprt_rdma.h" |
60 | ||
c56c65fb TT |
61 | /* |
62 | * Globals/Macros | |
63 | */ | |
64 | ||
f895b252 | 65 | #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) |
c56c65fb TT |
66 | # define RPCDBG_FACILITY RPCDBG_TRANS |
67 | #endif | |
68 | ||
69 | /* | |
70 | * internal functions | |
71 | */ | |
a9b0e381 | 72 | static void rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt); |
bebd0318 CL |
73 | static void rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf); |
74 | static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb); | |
c56c65fb | 75 | |
7d7fa9b5 | 76 | static struct workqueue_struct *rpcrdma_receive_wq __read_mostly; |
c56c65fb | 77 | |
fe97b47c CL |
78 | int |
79 | rpcrdma_alloc_wq(void) | |
c56c65fb | 80 | { |
fe97b47c | 81 | struct workqueue_struct *recv_wq; |
c56c65fb | 82 | |
fe97b47c CL |
83 | recv_wq = alloc_workqueue("xprtrdma_receive", |
84 | WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_HIGHPRI, | |
85 | 0); | |
86 | if (!recv_wq) | |
87 | return -ENOMEM; | |
c56c65fb | 88 | |
fe97b47c CL |
89 | rpcrdma_receive_wq = recv_wq; |
90 | return 0; | |
c56c65fb TT |
91 | } |
92 | ||
fe97b47c CL |
93 | void |
94 | rpcrdma_destroy_wq(void) | |
f1a03b76 | 95 | { |
fe97b47c | 96 | struct workqueue_struct *wq; |
f1a03b76 | 97 | |
fe97b47c CL |
98 | if (rpcrdma_receive_wq) { |
99 | wq = rpcrdma_receive_wq; | |
100 | rpcrdma_receive_wq = NULL; | |
101 | destroy_workqueue(wq); | |
102 | } | |
f1a03b76 CL |
103 | } |
104 | ||
c56c65fb TT |
105 | static void |
106 | rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context) | |
107 | { | |
108 | struct rpcrdma_ep *ep = context; | |
109 | ||
2f6922ca CL |
110 | pr_err("rpcrdma: %s on device %s ep %p\n", |
111 | ib_event_msg(event->event), event->device->name, context); | |
112 | ||
c56c65fb TT |
113 | if (ep->rep_connected == 1) { |
114 | ep->rep_connected = -EIO; | |
afadc468 | 115 | rpcrdma_conn_func(ep); |
c56c65fb TT |
116 | wake_up_all(&ep->rep_connect_wait); |
117 | } | |
118 | } | |
119 | ||
2fa8f88d CL |
120 | /** |
121 | * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC | |
122 | * @cq: completion queue (ignored) | |
123 | * @wc: completed WR | |
124 | * | |
fc664485 CL |
125 | */ |
126 | static void | |
2fa8f88d | 127 | rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc) |
fc664485 | 128 | { |
2fa8f88d CL |
129 | /* WARNING: Only wr_cqe and status are reliable at this point */ |
130 | if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR) | |
131 | pr_err("rpcrdma: Send: %s (%u/0x%x)\n", | |
132 | ib_wc_status_msg(wc->status), | |
133 | wc->status, wc->vendor_err); | |
fc664485 | 134 | } |
c56c65fb | 135 | |
23826c7a CL |
136 | /* Perform basic sanity checking to avoid using garbage |
137 | * to update the credit grant value. | |
138 | */ | |
139 | static void | |
140 | rpcrdma_update_granted_credits(struct rpcrdma_rep *rep) | |
141 | { | |
23826c7a | 142 | struct rpcrdma_buffer *buffer = &rep->rr_rxprt->rx_buf; |
c1bcb68e | 143 | __be32 *p = rep->rr_rdmabuf->rg_base; |
23826c7a CL |
144 | u32 credits; |
145 | ||
c1bcb68e | 146 | credits = be32_to_cpup(p + 2); |
23826c7a CL |
147 | if (credits == 0) |
148 | credits = 1; /* don't deadlock */ | |
149 | else if (credits > buffer->rb_max_requests) | |
150 | credits = buffer->rb_max_requests; | |
151 | ||
152 | atomic_set(&buffer->rb_credits, credits); | |
153 | } | |
154 | ||
552bf225 | 155 | /** |
1519e969 | 156 | * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC |
552bf225 CL |
157 | * @cq: completion queue (ignored) |
158 | * @wc: completed WR | |
159 | * | |
160 | */ | |
fc664485 | 161 | static void |
1519e969 | 162 | rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc) |
fc664485 | 163 | { |
552bf225 CL |
164 | struct ib_cqe *cqe = wc->wr_cqe; |
165 | struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep, | |
166 | rr_cqe); | |
fc664485 | 167 | |
8502427c CL |
168 | /* WARNING: Only wr_id and status are reliable at this point */ |
169 | if (wc->status != IB_WC_SUCCESS) | |
170 | goto out_fail; | |
fc664485 | 171 | |
8502427c | 172 | /* status == SUCCESS means all fields in wc are trustworthy */ |
8502427c CL |
173 | dprintk("RPC: %s: rep %p opcode 'recv', length %u: success\n", |
174 | __func__, rep, wc->byte_len); | |
175 | ||
96f8778f | 176 | rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len); |
c8b920bb CL |
177 | rep->rr_wc_flags = wc->wc_flags; |
178 | rep->rr_inv_rkey = wc->ex.invalidate_rkey; | |
179 | ||
91a10c52 | 180 | ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf), |
6b1184cd | 181 | rdmab_addr(rep->rr_rdmabuf), |
e2a67190 | 182 | wc->byte_len, DMA_FROM_DEVICE); |
23826c7a | 183 | |
e2a67190 CL |
184 | if (wc->byte_len >= RPCRDMA_HDRLEN_ERR) |
185 | rpcrdma_update_granted_credits(rep); | |
fc664485 CL |
186 | |
187 | out_schedule: | |
fe97b47c | 188 | queue_work(rpcrdma_receive_wq, &rep->rr_work); |
8502427c | 189 | return; |
fe97b47c | 190 | |
8502427c CL |
191 | out_fail: |
192 | if (wc->status != IB_WC_WR_FLUSH_ERR) | |
552bf225 CL |
193 | pr_err("rpcrdma: Recv: %s (%u/0x%x)\n", |
194 | ib_wc_status_msg(wc->status), | |
195 | wc->status, wc->vendor_err); | |
e2a67190 | 196 | rpcrdma_set_xdrlen(&rep->rr_hdrbuf, 0); |
8502427c | 197 | goto out_schedule; |
fc664485 CL |
198 | } |
199 | ||
87cfb9a0 CL |
200 | static void |
201 | rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt, | |
202 | struct rdma_conn_param *param) | |
203 | { | |
204 | struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data; | |
205 | const struct rpcrdma_connect_private *pmsg = param->private_data; | |
206 | unsigned int rsize, wsize; | |
207 | ||
c8b920bb CL |
208 | /* Default settings for RPC-over-RDMA Version One */ |
209 | r_xprt->rx_ia.ri_reminv_expected = false; | |
b5f0afbe | 210 | r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize; |
87cfb9a0 CL |
211 | rsize = RPCRDMA_V1_DEF_INLINE_SIZE; |
212 | wsize = RPCRDMA_V1_DEF_INLINE_SIZE; | |
213 | ||
214 | if (pmsg && | |
215 | pmsg->cp_magic == rpcrdma_cmp_magic && | |
216 | pmsg->cp_version == RPCRDMA_CMP_VERSION) { | |
c8b920bb | 217 | r_xprt->rx_ia.ri_reminv_expected = true; |
c95a3c6b | 218 | r_xprt->rx_ia.ri_implicit_roundup = true; |
87cfb9a0 CL |
219 | rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size); |
220 | wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size); | |
221 | } | |
222 | ||
223 | if (rsize < cdata->inline_rsize) | |
224 | cdata->inline_rsize = rsize; | |
225 | if (wsize < cdata->inline_wsize) | |
226 | cdata->inline_wsize = wsize; | |
6d6bf72d CL |
227 | dprintk("RPC: %s: max send %u, max recv %u\n", |
228 | __func__, cdata->inline_wsize, cdata->inline_rsize); | |
87cfb9a0 CL |
229 | rpcrdma_set_max_header_sizes(r_xprt); |
230 | } | |
231 | ||
c56c65fb TT |
232 | static int |
233 | rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event) | |
234 | { | |
235 | struct rpcrdma_xprt *xprt = id->context; | |
236 | struct rpcrdma_ia *ia = &xprt->rx_ia; | |
237 | struct rpcrdma_ep *ep = &xprt->rx_ep; | |
f895b252 | 238 | #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) |
0dd39cae | 239 | struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr; |
ff0db049 | 240 | #endif |
c56c65fb TT |
241 | int connstate = 0; |
242 | ||
243 | switch (event->event) { | |
244 | case RDMA_CM_EVENT_ADDR_RESOLVED: | |
245 | case RDMA_CM_EVENT_ROUTE_RESOLVED: | |
5675add3 | 246 | ia->ri_async_rc = 0; |
c56c65fb TT |
247 | complete(&ia->ri_done); |
248 | break; | |
249 | case RDMA_CM_EVENT_ADDR_ERROR: | |
250 | ia->ri_async_rc = -EHOSTUNREACH; | |
251 | dprintk("RPC: %s: CM address resolution error, ep 0x%p\n", | |
252 | __func__, ep); | |
253 | complete(&ia->ri_done); | |
254 | break; | |
255 | case RDMA_CM_EVENT_ROUTE_ERROR: | |
256 | ia->ri_async_rc = -ENETUNREACH; | |
257 | dprintk("RPC: %s: CM route resolution error, ep 0x%p\n", | |
258 | __func__, ep); | |
259 | complete(&ia->ri_done); | |
260 | break; | |
bebd0318 CL |
261 | case RDMA_CM_EVENT_DEVICE_REMOVAL: |
262 | #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) | |
173b8f49 CL |
263 | pr_info("rpcrdma: removing device %s for %pIS:%u\n", |
264 | ia->ri_device->name, | |
bebd0318 CL |
265 | sap, rpc_get_port(sap)); |
266 | #endif | |
267 | set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags); | |
268 | ep->rep_connected = -ENODEV; | |
269 | xprt_force_disconnect(&xprt->rx_xprt); | |
270 | wait_for_completion(&ia->ri_remove_done); | |
271 | ||
272 | ia->ri_id = NULL; | |
273 | ia->ri_pd = NULL; | |
274 | ia->ri_device = NULL; | |
275 | /* Return 1 to ensure the core destroys the id. */ | |
276 | return 1; | |
c56c65fb TT |
277 | case RDMA_CM_EVENT_ESTABLISHED: |
278 | connstate = 1; | |
87cfb9a0 | 279 | rpcrdma_update_connect_private(xprt, &event->param.conn); |
c56c65fb TT |
280 | goto connected; |
281 | case RDMA_CM_EVENT_CONNECT_ERROR: | |
282 | connstate = -ENOTCONN; | |
283 | goto connected; | |
284 | case RDMA_CM_EVENT_UNREACHABLE: | |
285 | connstate = -ENETDOWN; | |
286 | goto connected; | |
287 | case RDMA_CM_EVENT_REJECTED: | |
173b8f49 CL |
288 | dprintk("rpcrdma: connection to %pIS:%u rejected: %s\n", |
289 | sap, rpc_get_port(sap), | |
0a90487b | 290 | rdma_reject_msg(id, event->status)); |
c56c65fb | 291 | connstate = -ECONNREFUSED; |
0a90487b CL |
292 | if (event->status == IB_CM_REJ_STALE_CONN) |
293 | connstate = -EAGAIN; | |
c56c65fb TT |
294 | goto connected; |
295 | case RDMA_CM_EVENT_DISCONNECTED: | |
296 | connstate = -ECONNABORTED; | |
c56c65fb | 297 | connected: |
23826c7a | 298 | atomic_set(&xprt->rx_buf.rb_credits, 1); |
c56c65fb | 299 | ep->rep_connected = connstate; |
afadc468 | 300 | rpcrdma_conn_func(ep); |
c56c65fb | 301 | wake_up_all(&ep->rep_connect_wait); |
8079fb78 | 302 | /*FALLTHROUGH*/ |
c56c65fb | 303 | default: |
173b8f49 CL |
304 | dprintk("RPC: %s: %pIS:%u on %s/%s (ep 0x%p): %s\n", |
305 | __func__, sap, rpc_get_port(sap), | |
306 | ia->ri_device->name, ia->ri_ops->ro_displayname, | |
307 | ep, rdma_event_msg(event->event)); | |
c56c65fb TT |
308 | break; |
309 | } | |
310 | ||
311 | return 0; | |
312 | } | |
313 | ||
314 | static struct rdma_cm_id * | |
315 | rpcrdma_create_id(struct rpcrdma_xprt *xprt, | |
316 | struct rpcrdma_ia *ia, struct sockaddr *addr) | |
317 | { | |
109b88ab | 318 | unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1; |
c56c65fb TT |
319 | struct rdma_cm_id *id; |
320 | int rc; | |
321 | ||
1a954051 | 322 | init_completion(&ia->ri_done); |
bebd0318 | 323 | init_completion(&ia->ri_remove_done); |
1a954051 | 324 | |
fa20105e GS |
325 | id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP, |
326 | IB_QPT_RC); | |
c56c65fb TT |
327 | if (IS_ERR(id)) { |
328 | rc = PTR_ERR(id); | |
329 | dprintk("RPC: %s: rdma_create_id() failed %i\n", | |
330 | __func__, rc); | |
331 | return id; | |
332 | } | |
333 | ||
5675add3 | 334 | ia->ri_async_rc = -ETIMEDOUT; |
c56c65fb TT |
335 | rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT); |
336 | if (rc) { | |
337 | dprintk("RPC: %s: rdma_resolve_addr() failed %i\n", | |
338 | __func__, rc); | |
339 | goto out; | |
340 | } | |
109b88ab CL |
341 | rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout); |
342 | if (rc < 0) { | |
343 | dprintk("RPC: %s: wait() exited: %i\n", | |
344 | __func__, rc); | |
345 | goto out; | |
346 | } | |
d0f36c46 | 347 | |
c56c65fb TT |
348 | rc = ia->ri_async_rc; |
349 | if (rc) | |
350 | goto out; | |
351 | ||
5675add3 | 352 | ia->ri_async_rc = -ETIMEDOUT; |
c56c65fb TT |
353 | rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT); |
354 | if (rc) { | |
355 | dprintk("RPC: %s: rdma_resolve_route() failed %i\n", | |
356 | __func__, rc); | |
56a6bd15 | 357 | goto out; |
c56c65fb | 358 | } |
109b88ab CL |
359 | rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout); |
360 | if (rc < 0) { | |
361 | dprintk("RPC: %s: wait() exited: %i\n", | |
362 | __func__, rc); | |
56a6bd15 | 363 | goto out; |
109b88ab | 364 | } |
c56c65fb TT |
365 | rc = ia->ri_async_rc; |
366 | if (rc) | |
56a6bd15 | 367 | goto out; |
c56c65fb TT |
368 | |
369 | return id; | |
56a6bd15 | 370 | |
c56c65fb TT |
371 | out: |
372 | rdma_destroy_id(id); | |
373 | return ERR_PTR(rc); | |
374 | } | |
375 | ||
c56c65fb TT |
376 | /* |
377 | * Exported functions. | |
378 | */ | |
379 | ||
fff09594 CL |
380 | /** |
381 | * rpcrdma_ia_open - Open and initialize an Interface Adapter. | |
382 | * @xprt: controlling transport | |
383 | * @addr: IP address of remote peer | |
384 | * | |
385 | * Returns 0 on success, negative errno if an appropriate | |
386 | * Interface Adapter could not be found and opened. | |
c56c65fb TT |
387 | */ |
388 | int | |
fff09594 | 389 | rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr) |
c56c65fb | 390 | { |
c56c65fb | 391 | struct rpcrdma_ia *ia = &xprt->rx_ia; |
d1ed857e CL |
392 | int rc; |
393 | ||
c56c65fb TT |
394 | ia->ri_id = rpcrdma_create_id(xprt, ia, addr); |
395 | if (IS_ERR(ia->ri_id)) { | |
396 | rc = PTR_ERR(ia->ri_id); | |
fff09594 | 397 | goto out_err; |
c56c65fb | 398 | } |
89e0d112 | 399 | ia->ri_device = ia->ri_id->device; |
c56c65fb | 400 | |
ed082d36 | 401 | ia->ri_pd = ib_alloc_pd(ia->ri_device, 0); |
c56c65fb TT |
402 | if (IS_ERR(ia->ri_pd)) { |
403 | rc = PTR_ERR(ia->ri_pd); | |
b54054ca | 404 | pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc); |
fff09594 | 405 | goto out_err; |
c56c65fb TT |
406 | } |
407 | ||
fff09594 | 408 | switch (xprt_rdma_memreg_strategy) { |
3197d309 | 409 | case RPCRDMA_FRMR: |
b54054ca CL |
410 | if (frwr_is_supported(ia)) { |
411 | ia->ri_ops = &rpcrdma_frwr_memreg_ops; | |
412 | break; | |
413 | } | |
414 | /*FALLTHROUGH*/ | |
bd7ed1d1 | 415 | case RPCRDMA_MTHCAFMR: |
b54054ca CL |
416 | if (fmr_is_supported(ia)) { |
417 | ia->ri_ops = &rpcrdma_fmr_memreg_ops; | |
418 | break; | |
419 | } | |
420 | /*FALLTHROUGH*/ | |
bd7ed1d1 | 421 | default: |
fff09594 CL |
422 | pr_err("rpcrdma: Device %s does not support memreg mode %d\n", |
423 | ia->ri_device->name, xprt_rdma_memreg_strategy); | |
b54054ca | 424 | rc = -EINVAL; |
fff09594 | 425 | goto out_err; |
c56c65fb TT |
426 | } |
427 | ||
c56c65fb | 428 | return 0; |
5ae711a2 | 429 | |
fff09594 CL |
430 | out_err: |
431 | rpcrdma_ia_close(ia); | |
c56c65fb TT |
432 | return rc; |
433 | } | |
434 | ||
bebd0318 CL |
435 | /** |
436 | * rpcrdma_ia_remove - Handle device driver unload | |
437 | * @ia: interface adapter being removed | |
438 | * | |
439 | * Divest transport H/W resources associated with this adapter, | |
440 | * but allow it to be restored later. | |
441 | */ | |
442 | void | |
443 | rpcrdma_ia_remove(struct rpcrdma_ia *ia) | |
444 | { | |
445 | struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt, | |
446 | rx_ia); | |
447 | struct rpcrdma_ep *ep = &r_xprt->rx_ep; | |
448 | struct rpcrdma_buffer *buf = &r_xprt->rx_buf; | |
449 | struct rpcrdma_req *req; | |
450 | struct rpcrdma_rep *rep; | |
451 | ||
452 | cancel_delayed_work_sync(&buf->rb_refresh_worker); | |
453 | ||
454 | /* This is similar to rpcrdma_ep_destroy, but: | |
455 | * - Don't cancel the connect worker. | |
456 | * - Don't call rpcrdma_ep_disconnect, which waits | |
457 | * for another conn upcall, which will deadlock. | |
458 | * - rdma_disconnect is unneeded, the underlying | |
459 | * connection is already gone. | |
460 | */ | |
461 | if (ia->ri_id->qp) { | |
462 | ib_drain_qp(ia->ri_id->qp); | |
463 | rdma_destroy_qp(ia->ri_id); | |
464 | ia->ri_id->qp = NULL; | |
465 | } | |
466 | ib_free_cq(ep->rep_attr.recv_cq); | |
467 | ib_free_cq(ep->rep_attr.send_cq); | |
468 | ||
469 | /* The ULP is responsible for ensuring all DMA | |
470 | * mappings and MRs are gone. | |
471 | */ | |
472 | list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list) | |
473 | rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf); | |
474 | list_for_each_entry(req, &buf->rb_allreqs, rl_all) { | |
475 | rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf); | |
476 | rpcrdma_dma_unmap_regbuf(req->rl_sendbuf); | |
477 | rpcrdma_dma_unmap_regbuf(req->rl_recvbuf); | |
478 | } | |
479 | rpcrdma_destroy_mrs(buf); | |
480 | ||
481 | /* Allow waiters to continue */ | |
482 | complete(&ia->ri_remove_done); | |
483 | } | |
484 | ||
fff09594 CL |
485 | /** |
486 | * rpcrdma_ia_close - Clean up/close an IA. | |
487 | * @ia: interface adapter to close | |
488 | * | |
c56c65fb TT |
489 | */ |
490 | void | |
491 | rpcrdma_ia_close(struct rpcrdma_ia *ia) | |
492 | { | |
c56c65fb | 493 | dprintk("RPC: %s: entering\n", __func__); |
fee08caf TT |
494 | if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) { |
495 | if (ia->ri_id->qp) | |
496 | rdma_destroy_qp(ia->ri_id); | |
56a6bd15 | 497 | rdma_destroy_id(ia->ri_id); |
fee08caf | 498 | } |
fff09594 CL |
499 | ia->ri_id = NULL; |
500 | ia->ri_device = NULL; | |
6d44698d CL |
501 | |
502 | /* If the pd is still busy, xprtrdma missed freeing a resource */ | |
503 | if (ia->ri_pd && !IS_ERR(ia->ri_pd)) | |
7dd78647 | 504 | ib_dealloc_pd(ia->ri_pd); |
fff09594 | 505 | ia->ri_pd = NULL; |
c56c65fb TT |
506 | } |
507 | ||
508 | /* | |
509 | * Create unconnected endpoint. | |
510 | */ | |
511 | int | |
512 | rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia, | |
16f906d6 | 513 | struct rpcrdma_create_data_internal *cdata) |
c56c65fb | 514 | { |
87cfb9a0 | 515 | struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private; |
16f906d6 | 516 | unsigned int max_qp_wr, max_sge; |
fc664485 | 517 | struct ib_cq *sendcq, *recvcq; |
2fa8f88d | 518 | int rc; |
c56c65fb | 519 | |
eed50879 CL |
520 | max_sge = min_t(unsigned int, ia->ri_device->attrs.max_sge, |
521 | RPCRDMA_MAX_SEND_SGES); | |
16f906d6 CL |
522 | if (max_sge < RPCRDMA_MIN_SEND_SGES) { |
523 | pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge); | |
b3221d6a CL |
524 | return -ENOMEM; |
525 | } | |
16f906d6 | 526 | ia->ri_max_send_sges = max_sge - RPCRDMA_MIN_SEND_SGES; |
b3221d6a | 527 | |
e3e45b1b | 528 | if (ia->ri_device->attrs.max_qp_wr <= RPCRDMA_BACKWARD_WRS) { |
124fa17d CL |
529 | dprintk("RPC: %s: insufficient wqe's available\n", |
530 | __func__); | |
531 | return -ENOMEM; | |
532 | } | |
550d7502 | 533 | max_qp_wr = ia->ri_device->attrs.max_qp_wr - RPCRDMA_BACKWARD_WRS - 1; |
124fa17d | 534 | |
c56c65fb | 535 | /* check provider's send/recv wr limits */ |
124fa17d CL |
536 | if (cdata->max_requests > max_qp_wr) |
537 | cdata->max_requests = max_qp_wr; | |
c56c65fb TT |
538 | |
539 | ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall; | |
540 | ep->rep_attr.qp_context = ep; | |
c56c65fb TT |
541 | ep->rep_attr.srq = NULL; |
542 | ep->rep_attr.cap.max_send_wr = cdata->max_requests; | |
124fa17d | 543 | ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS; |
550d7502 | 544 | ep->rep_attr.cap.max_send_wr += 1; /* drain cqe */ |
3968cb58 CL |
545 | rc = ia->ri_ops->ro_open(ia, ep, cdata); |
546 | if (rc) | |
547 | return rc; | |
c56c65fb | 548 | ep->rep_attr.cap.max_recv_wr = cdata->max_requests; |
124fa17d | 549 | ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS; |
550d7502 | 550 | ep->rep_attr.cap.max_recv_wr += 1; /* drain cqe */ |
16f906d6 | 551 | ep->rep_attr.cap.max_send_sge = max_sge; |
c56c65fb TT |
552 | ep->rep_attr.cap.max_recv_sge = 1; |
553 | ep->rep_attr.cap.max_inline_data = 0; | |
554 | ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR; | |
555 | ep->rep_attr.qp_type = IB_QPT_RC; | |
556 | ep->rep_attr.port_num = ~0; | |
557 | ||
558 | dprintk("RPC: %s: requested max: dtos: send %d recv %d; " | |
559 | "iovs: send %d recv %d\n", | |
560 | __func__, | |
561 | ep->rep_attr.cap.max_send_wr, | |
562 | ep->rep_attr.cap.max_recv_wr, | |
563 | ep->rep_attr.cap.max_send_sge, | |
564 | ep->rep_attr.cap.max_recv_sge); | |
565 | ||
566 | /* set trigger for requesting send completion */ | |
fc664485 | 567 | ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1; |
26ae9d1c CL |
568 | if (ep->rep_cqinit <= 2) |
569 | ep->rep_cqinit = 0; /* always signal? */ | |
8d38de65 | 570 | rpcrdma_init_cqcount(ep, 0); |
c56c65fb | 571 | init_waitqueue_head(&ep->rep_connect_wait); |
254f91e2 | 572 | INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker); |
c56c65fb | 573 | |
2fa8f88d CL |
574 | sendcq = ib_alloc_cq(ia->ri_device, NULL, |
575 | ep->rep_attr.cap.max_send_wr + 1, | |
576 | 0, IB_POLL_SOFTIRQ); | |
fc664485 CL |
577 | if (IS_ERR(sendcq)) { |
578 | rc = PTR_ERR(sendcq); | |
579 | dprintk("RPC: %s: failed to create send CQ: %i\n", | |
c56c65fb TT |
580 | __func__, rc); |
581 | goto out1; | |
582 | } | |
583 | ||
552bf225 CL |
584 | recvcq = ib_alloc_cq(ia->ri_device, NULL, |
585 | ep->rep_attr.cap.max_recv_wr + 1, | |
586 | 0, IB_POLL_SOFTIRQ); | |
fc664485 CL |
587 | if (IS_ERR(recvcq)) { |
588 | rc = PTR_ERR(recvcq); | |
589 | dprintk("RPC: %s: failed to create recv CQ: %i\n", | |
590 | __func__, rc); | |
591 | goto out2; | |
592 | } | |
593 | ||
fc664485 CL |
594 | ep->rep_attr.send_cq = sendcq; |
595 | ep->rep_attr.recv_cq = recvcq; | |
c56c65fb TT |
596 | |
597 | /* Initialize cma parameters */ | |
b2dde94b | 598 | memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma)); |
c56c65fb | 599 | |
87cfb9a0 CL |
600 | /* Prepare RDMA-CM private message */ |
601 | pmsg->cp_magic = rpcrdma_cmp_magic; | |
602 | pmsg->cp_version = RPCRDMA_CMP_VERSION; | |
c8b920bb | 603 | pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok; |
87cfb9a0 CL |
604 | pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize); |
605 | pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize); | |
606 | ep->rep_remote_cma.private_data = pmsg; | |
607 | ep->rep_remote_cma.private_data_len = sizeof(*pmsg); | |
c56c65fb TT |
608 | |
609 | /* Client offers RDMA Read but does not initiate */ | |
b334eaab | 610 | ep->rep_remote_cma.initiator_depth = 0; |
e3e45b1b | 611 | if (ia->ri_device->attrs.max_qp_rd_atom > 32) /* arbitrary but <= 255 */ |
b334eaab TT |
612 | ep->rep_remote_cma.responder_resources = 32; |
613 | else | |
7bc7972c | 614 | ep->rep_remote_cma.responder_resources = |
e3e45b1b | 615 | ia->ri_device->attrs.max_qp_rd_atom; |
c56c65fb | 616 | |
b2dde94b CL |
617 | /* Limit transport retries so client can detect server |
618 | * GID changes quickly. RPC layer handles re-establishing | |
619 | * transport connection and retransmission. | |
620 | */ | |
621 | ep->rep_remote_cma.retry_count = 6; | |
622 | ||
623 | /* RPC-over-RDMA handles its own flow control. In addition, | |
624 | * make all RNR NAKs visible so we know that RPC-over-RDMA | |
625 | * flow control is working correctly (no NAKs should be seen). | |
626 | */ | |
c56c65fb TT |
627 | ep->rep_remote_cma.flow_control = 0; |
628 | ep->rep_remote_cma.rnr_retry_count = 0; | |
629 | ||
630 | return 0; | |
631 | ||
632 | out2: | |
2fa8f88d | 633 | ib_free_cq(sendcq); |
c56c65fb TT |
634 | out1: |
635 | return rc; | |
636 | } | |
637 | ||
638 | /* | |
639 | * rpcrdma_ep_destroy | |
640 | * | |
641 | * Disconnect and destroy endpoint. After this, the only | |
642 | * valid operations on the ep are to free it (if dynamically | |
643 | * allocated) or re-create it. | |
c56c65fb | 644 | */ |
7f1d5419 | 645 | void |
c56c65fb TT |
646 | rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia) |
647 | { | |
c56c65fb TT |
648 | dprintk("RPC: %s: entering, connected is %d\n", |
649 | __func__, ep->rep_connected); | |
650 | ||
254f91e2 CL |
651 | cancel_delayed_work_sync(&ep->rep_connect_worker); |
652 | ||
72c02173 | 653 | if (ia->ri_id->qp) { |
550d7502 | 654 | rpcrdma_ep_disconnect(ep, ia); |
fee08caf TT |
655 | rdma_destroy_qp(ia->ri_id); |
656 | ia->ri_id->qp = NULL; | |
c56c65fb TT |
657 | } |
658 | ||
552bf225 | 659 | ib_free_cq(ep->rep_attr.recv_cq); |
2fa8f88d | 660 | ib_free_cq(ep->rep_attr.send_cq); |
c56c65fb TT |
661 | } |
662 | ||
a9b0e381 CL |
663 | /* Re-establish a connection after a device removal event. |
664 | * Unlike a normal reconnection, a fresh PD and a new set | |
665 | * of MRs and buffers is needed. | |
666 | */ | |
667 | static int | |
668 | rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt, | |
669 | struct rpcrdma_ep *ep, struct rpcrdma_ia *ia) | |
670 | { | |
671 | struct sockaddr *sap = (struct sockaddr *)&r_xprt->rx_data.addr; | |
672 | int rc, err; | |
673 | ||
674 | pr_info("%s: r_xprt = %p\n", __func__, r_xprt); | |
675 | ||
676 | rc = -EHOSTUNREACH; | |
677 | if (rpcrdma_ia_open(r_xprt, sap)) | |
678 | goto out1; | |
679 | ||
680 | rc = -ENOMEM; | |
681 | err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data); | |
682 | if (err) { | |
683 | pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err); | |
684 | goto out2; | |
685 | } | |
686 | ||
687 | rc = -ENETUNREACH; | |
688 | err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr); | |
689 | if (err) { | |
690 | pr_err("rpcrdma: rdma_create_qp returned %d\n", err); | |
691 | goto out3; | |
692 | } | |
693 | ||
694 | rpcrdma_create_mrs(r_xprt); | |
695 | return 0; | |
696 | ||
697 | out3: | |
698 | rpcrdma_ep_destroy(ep, ia); | |
699 | out2: | |
700 | rpcrdma_ia_close(ia); | |
701 | out1: | |
702 | return rc; | |
703 | } | |
704 | ||
1890896b CL |
705 | static int |
706 | rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep, | |
707 | struct rpcrdma_ia *ia) | |
708 | { | |
709 | struct sockaddr *sap = (struct sockaddr *)&r_xprt->rx_data.addr; | |
710 | struct rdma_cm_id *id, *old; | |
711 | int err, rc; | |
712 | ||
713 | dprintk("RPC: %s: reconnecting...\n", __func__); | |
714 | ||
715 | rpcrdma_ep_disconnect(ep, ia); | |
716 | ||
717 | rc = -EHOSTUNREACH; | |
718 | id = rpcrdma_create_id(r_xprt, ia, sap); | |
719 | if (IS_ERR(id)) | |
720 | goto out; | |
721 | ||
722 | /* As long as the new ID points to the same device as the | |
723 | * old ID, we can reuse the transport's existing PD and all | |
724 | * previously allocated MRs. Also, the same device means | |
725 | * the transport's previous DMA mappings are still valid. | |
726 | * | |
727 | * This is a sanity check only. There should be no way these | |
728 | * point to two different devices here. | |
729 | */ | |
730 | old = id; | |
731 | rc = -ENETUNREACH; | |
732 | if (ia->ri_device != id->device) { | |
733 | pr_err("rpcrdma: can't reconnect on different device!\n"); | |
734 | goto out_destroy; | |
735 | } | |
736 | ||
737 | err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr); | |
738 | if (err) { | |
739 | dprintk("RPC: %s: rdma_create_qp returned %d\n", | |
740 | __func__, err); | |
741 | goto out_destroy; | |
742 | } | |
743 | ||
744 | /* Atomically replace the transport's ID and QP. */ | |
745 | rc = 0; | |
746 | old = ia->ri_id; | |
747 | ia->ri_id = id; | |
748 | rdma_destroy_qp(old); | |
749 | ||
750 | out_destroy: | |
56a6bd15 | 751 | rdma_destroy_id(old); |
1890896b CL |
752 | out: |
753 | return rc; | |
754 | } | |
755 | ||
c56c65fb TT |
756 | /* |
757 | * Connect unconnected endpoint. | |
758 | */ | |
759 | int | |
760 | rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia) | |
761 | { | |
0a90487b CL |
762 | struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt, |
763 | rx_ia); | |
0a90487b | 764 | unsigned int extras; |
1890896b | 765 | int rc; |
c56c65fb | 766 | |
c56c65fb | 767 | retry: |
1890896b CL |
768 | switch (ep->rep_connected) { |
769 | case 0: | |
ec62f40d CL |
770 | dprintk("RPC: %s: connecting...\n", __func__); |
771 | rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr); | |
772 | if (rc) { | |
773 | dprintk("RPC: %s: rdma_create_qp failed %i\n", | |
774 | __func__, rc); | |
1890896b CL |
775 | rc = -ENETUNREACH; |
776 | goto out_noupdate; | |
ec62f40d | 777 | } |
1890896b | 778 | break; |
a9b0e381 CL |
779 | case -ENODEV: |
780 | rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia); | |
781 | if (rc) | |
782 | goto out_noupdate; | |
783 | break; | |
1890896b CL |
784 | default: |
785 | rc = rpcrdma_ep_reconnect(r_xprt, ep, ia); | |
786 | if (rc) | |
787 | goto out; | |
c56c65fb TT |
788 | } |
789 | ||
c56c65fb TT |
790 | ep->rep_connected = 0; |
791 | ||
792 | rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma); | |
793 | if (rc) { | |
794 | dprintk("RPC: %s: rdma_connect() failed with %i\n", | |
795 | __func__, rc); | |
796 | goto out; | |
797 | } | |
798 | ||
c56c65fb | 799 | wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0); |
c56c65fb | 800 | if (ep->rep_connected <= 0) { |
0a90487b | 801 | if (ep->rep_connected == -EAGAIN) |
c56c65fb TT |
802 | goto retry; |
803 | rc = ep->rep_connected; | |
0a90487b | 804 | goto out; |
c56c65fb TT |
805 | } |
806 | ||
0a90487b CL |
807 | dprintk("RPC: %s: connected\n", __func__); |
808 | extras = r_xprt->rx_buf.rb_bc_srv_max_requests; | |
809 | if (extras) | |
810 | rpcrdma_ep_post_extra_recv(r_xprt, extras); | |
811 | ||
c56c65fb TT |
812 | out: |
813 | if (rc) | |
814 | ep->rep_connected = rc; | |
1890896b CL |
815 | |
816 | out_noupdate: | |
c56c65fb TT |
817 | return rc; |
818 | } | |
819 | ||
820 | /* | |
821 | * rpcrdma_ep_disconnect | |
822 | * | |
823 | * This is separate from destroy to facilitate the ability | |
824 | * to reconnect without recreating the endpoint. | |
825 | * | |
826 | * This call is not reentrant, and must not be made in parallel | |
827 | * on the same endpoint. | |
828 | */ | |
282191cb | 829 | void |
c56c65fb TT |
830 | rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia) |
831 | { | |
832 | int rc; | |
833 | ||
c56c65fb TT |
834 | rc = rdma_disconnect(ia->ri_id); |
835 | if (!rc) { | |
836 | /* returns without wait if not connected */ | |
837 | wait_event_interruptible(ep->rep_connect_wait, | |
838 | ep->rep_connected != 1); | |
839 | dprintk("RPC: %s: after wait, %sconnected\n", __func__, | |
840 | (ep->rep_connected == 1) ? "still " : "dis"); | |
841 | } else { | |
842 | dprintk("RPC: %s: rdma_disconnect %i\n", __func__, rc); | |
843 | ep->rep_connected = rc; | |
844 | } | |
550d7502 CL |
845 | |
846 | ib_drain_qp(ia->ri_id->qp); | |
c56c65fb TT |
847 | } |
848 | ||
505bbe64 CL |
849 | static void |
850 | rpcrdma_mr_recovery_worker(struct work_struct *work) | |
851 | { | |
852 | struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer, | |
853 | rb_recovery_worker.work); | |
854 | struct rpcrdma_mw *mw; | |
855 | ||
856 | spin_lock(&buf->rb_recovery_lock); | |
857 | while (!list_empty(&buf->rb_stale_mrs)) { | |
9a5c63e9 | 858 | mw = rpcrdma_pop_mw(&buf->rb_stale_mrs); |
505bbe64 CL |
859 | spin_unlock(&buf->rb_recovery_lock); |
860 | ||
861 | dprintk("RPC: %s: recovering MR %p\n", __func__, mw); | |
862 | mw->mw_xprt->rx_ia.ri_ops->ro_recover_mr(mw); | |
863 | ||
864 | spin_lock(&buf->rb_recovery_lock); | |
53d78523 | 865 | } |
505bbe64 CL |
866 | spin_unlock(&buf->rb_recovery_lock); |
867 | } | |
868 | ||
869 | void | |
870 | rpcrdma_defer_mr_recovery(struct rpcrdma_mw *mw) | |
871 | { | |
872 | struct rpcrdma_xprt *r_xprt = mw->mw_xprt; | |
873 | struct rpcrdma_buffer *buf = &r_xprt->rx_buf; | |
874 | ||
875 | spin_lock(&buf->rb_recovery_lock); | |
9a5c63e9 | 876 | rpcrdma_push_mw(mw, &buf->rb_stale_mrs); |
505bbe64 CL |
877 | spin_unlock(&buf->rb_recovery_lock); |
878 | ||
879 | schedule_delayed_work(&buf->rb_recovery_worker, 0); | |
880 | } | |
881 | ||
e2ac236c CL |
882 | static void |
883 | rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt) | |
884 | { | |
885 | struct rpcrdma_buffer *buf = &r_xprt->rx_buf; | |
886 | struct rpcrdma_ia *ia = &r_xprt->rx_ia; | |
887 | unsigned int count; | |
888 | LIST_HEAD(free); | |
889 | LIST_HEAD(all); | |
890 | ||
891 | for (count = 0; count < 32; count++) { | |
892 | struct rpcrdma_mw *mw; | |
893 | int rc; | |
894 | ||
895 | mw = kzalloc(sizeof(*mw), GFP_KERNEL); | |
896 | if (!mw) | |
897 | break; | |
898 | ||
899 | rc = ia->ri_ops->ro_init_mr(ia, mw); | |
900 | if (rc) { | |
901 | kfree(mw); | |
902 | break; | |
903 | } | |
904 | ||
905 | mw->mw_xprt = r_xprt; | |
906 | ||
907 | list_add(&mw->mw_list, &free); | |
908 | list_add(&mw->mw_all, &all); | |
909 | } | |
910 | ||
911 | spin_lock(&buf->rb_mwlock); | |
912 | list_splice(&free, &buf->rb_mws); | |
913 | list_splice(&all, &buf->rb_all); | |
914 | r_xprt->rx_stats.mrs_allocated += count; | |
915 | spin_unlock(&buf->rb_mwlock); | |
916 | ||
917 | dprintk("RPC: %s: created %u MRs\n", __func__, count); | |
918 | } | |
919 | ||
920 | static void | |
921 | rpcrdma_mr_refresh_worker(struct work_struct *work) | |
922 | { | |
923 | struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer, | |
924 | rb_refresh_worker.work); | |
925 | struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt, | |
926 | rx_buf); | |
927 | ||
928 | rpcrdma_create_mrs(r_xprt); | |
929 | } | |
930 | ||
f531a5db | 931 | struct rpcrdma_req * |
1392402c CL |
932 | rpcrdma_create_req(struct rpcrdma_xprt *r_xprt) |
933 | { | |
f531a5db | 934 | struct rpcrdma_buffer *buffer = &r_xprt->rx_buf; |
1392402c | 935 | struct rpcrdma_req *req; |
1392402c | 936 | |
85275c87 | 937 | req = kzalloc(sizeof(*req), GFP_KERNEL); |
1392402c | 938 | if (req == NULL) |
85275c87 | 939 | return ERR_PTR(-ENOMEM); |
1392402c | 940 | |
f531a5db CL |
941 | spin_lock(&buffer->rb_reqslock); |
942 | list_add(&req->rl_all, &buffer->rb_allreqs); | |
943 | spin_unlock(&buffer->rb_reqslock); | |
2fa8f88d | 944 | req->rl_cqe.done = rpcrdma_wc_send; |
1392402c | 945 | req->rl_buffer = &r_xprt->rx_buf; |
9d6b0409 | 946 | INIT_LIST_HEAD(&req->rl_registered); |
90aab602 CL |
947 | req->rl_send_wr.next = NULL; |
948 | req->rl_send_wr.wr_cqe = &req->rl_cqe; | |
655fec69 | 949 | req->rl_send_wr.sg_list = req->rl_send_sge; |
90aab602 | 950 | req->rl_send_wr.opcode = IB_WR_SEND; |
1392402c | 951 | return req; |
1392402c CL |
952 | } |
953 | ||
f531a5db | 954 | struct rpcrdma_rep * |
1392402c CL |
955 | rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt) |
956 | { | |
957 | struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data; | |
1392402c CL |
958 | struct rpcrdma_rep *rep; |
959 | int rc; | |
960 | ||
961 | rc = -ENOMEM; | |
6b1184cd | 962 | rep = kzalloc(sizeof(*rep), GFP_KERNEL); |
1392402c CL |
963 | if (rep == NULL) |
964 | goto out; | |
1392402c | 965 | |
13650c23 | 966 | rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize, |
99ef4db3 | 967 | DMA_FROM_DEVICE, GFP_KERNEL); |
6b1184cd CL |
968 | if (IS_ERR(rep->rr_rdmabuf)) { |
969 | rc = PTR_ERR(rep->rr_rdmabuf); | |
1392402c | 970 | goto out_free; |
6b1184cd | 971 | } |
96f8778f CL |
972 | xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base, |
973 | rdmab_length(rep->rr_rdmabuf)); | |
1392402c | 974 | |
1519e969 | 975 | rep->rr_cqe.done = rpcrdma_wc_receive; |
fed171b3 | 976 | rep->rr_rxprt = r_xprt; |
496b77a5 | 977 | INIT_WORK(&rep->rr_work, rpcrdma_reply_handler); |
6ea8e711 CL |
978 | rep->rr_recv_wr.next = NULL; |
979 | rep->rr_recv_wr.wr_cqe = &rep->rr_cqe; | |
980 | rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov; | |
981 | rep->rr_recv_wr.num_sge = 1; | |
1392402c CL |
982 | return rep; |
983 | ||
984 | out_free: | |
985 | kfree(rep); | |
986 | out: | |
987 | return ERR_PTR(rc); | |
988 | } | |
989 | ||
c56c65fb | 990 | int |
ac920d04 | 991 | rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt) |
c56c65fb | 992 | { |
ac920d04 | 993 | struct rpcrdma_buffer *buf = &r_xprt->rx_buf; |
c56c65fb TT |
994 | int i, rc; |
995 | ||
1e465fd4 | 996 | buf->rb_max_requests = r_xprt->rx_data.max_requests; |
f531a5db | 997 | buf->rb_bc_srv_max_requests = 0; |
23826c7a | 998 | atomic_set(&buf->rb_credits, 1); |
e2ac236c | 999 | spin_lock_init(&buf->rb_mwlock); |
505bbe64 CL |
1000 | spin_lock_init(&buf->rb_lock); |
1001 | spin_lock_init(&buf->rb_recovery_lock); | |
e2ac236c CL |
1002 | INIT_LIST_HEAD(&buf->rb_mws); |
1003 | INIT_LIST_HEAD(&buf->rb_all); | |
431af645 | 1004 | INIT_LIST_HEAD(&buf->rb_pending); |
505bbe64 | 1005 | INIT_LIST_HEAD(&buf->rb_stale_mrs); |
e2ac236c CL |
1006 | INIT_DELAYED_WORK(&buf->rb_refresh_worker, |
1007 | rpcrdma_mr_refresh_worker); | |
505bbe64 CL |
1008 | INIT_DELAYED_WORK(&buf->rb_recovery_worker, |
1009 | rpcrdma_mr_recovery_worker); | |
c56c65fb | 1010 | |
e2ac236c | 1011 | rpcrdma_create_mrs(r_xprt); |
c56c65fb | 1012 | |
1e465fd4 | 1013 | INIT_LIST_HEAD(&buf->rb_send_bufs); |
f531a5db CL |
1014 | INIT_LIST_HEAD(&buf->rb_allreqs); |
1015 | spin_lock_init(&buf->rb_reqslock); | |
c56c65fb TT |
1016 | for (i = 0; i < buf->rb_max_requests; i++) { |
1017 | struct rpcrdma_req *req; | |
c56c65fb | 1018 | |
1392402c CL |
1019 | req = rpcrdma_create_req(r_xprt); |
1020 | if (IS_ERR(req)) { | |
c56c65fb TT |
1021 | dprintk("RPC: %s: request buffer %d alloc" |
1022 | " failed\n", __func__, i); | |
1392402c | 1023 | rc = PTR_ERR(req); |
c56c65fb TT |
1024 | goto out; |
1025 | } | |
f531a5db | 1026 | req->rl_backchannel = false; |
a80d66c9 | 1027 | list_add(&req->rl_list, &buf->rb_send_bufs); |
1e465fd4 CL |
1028 | } |
1029 | ||
1030 | INIT_LIST_HEAD(&buf->rb_recv_bufs); | |
05c97466 | 1031 | for (i = 0; i < buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS; i++) { |
1e465fd4 | 1032 | struct rpcrdma_rep *rep; |
c56c65fb | 1033 | |
1392402c CL |
1034 | rep = rpcrdma_create_rep(r_xprt); |
1035 | if (IS_ERR(rep)) { | |
c56c65fb TT |
1036 | dprintk("RPC: %s: reply buffer %d alloc failed\n", |
1037 | __func__, i); | |
1392402c | 1038 | rc = PTR_ERR(rep); |
c56c65fb TT |
1039 | goto out; |
1040 | } | |
1e465fd4 | 1041 | list_add(&rep->rr_list, &buf->rb_recv_bufs); |
c56c65fb | 1042 | } |
1392402c | 1043 | |
c56c65fb TT |
1044 | return 0; |
1045 | out: | |
1046 | rpcrdma_buffer_destroy(buf); | |
1047 | return rc; | |
1048 | } | |
1049 | ||
1e465fd4 CL |
1050 | static struct rpcrdma_req * |
1051 | rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf) | |
1052 | { | |
1053 | struct rpcrdma_req *req; | |
1054 | ||
1055 | req = list_first_entry(&buf->rb_send_bufs, | |
a80d66c9 | 1056 | struct rpcrdma_req, rl_list); |
431af645 | 1057 | list_del_init(&req->rl_list); |
1e465fd4 CL |
1058 | return req; |
1059 | } | |
1060 | ||
1061 | static struct rpcrdma_rep * | |
1062 | rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf) | |
1063 | { | |
1064 | struct rpcrdma_rep *rep; | |
1065 | ||
1066 | rep = list_first_entry(&buf->rb_recv_bufs, | |
1067 | struct rpcrdma_rep, rr_list); | |
1068 | list_del(&rep->rr_list); | |
1069 | return rep; | |
1070 | } | |
1071 | ||
1392402c | 1072 | static void |
13650c23 | 1073 | rpcrdma_destroy_rep(struct rpcrdma_rep *rep) |
1392402c | 1074 | { |
13650c23 | 1075 | rpcrdma_free_regbuf(rep->rr_rdmabuf); |
1392402c CL |
1076 | kfree(rep); |
1077 | } | |
1078 | ||
f531a5db | 1079 | void |
13650c23 | 1080 | rpcrdma_destroy_req(struct rpcrdma_req *req) |
1392402c | 1081 | { |
13650c23 CL |
1082 | rpcrdma_free_regbuf(req->rl_recvbuf); |
1083 | rpcrdma_free_regbuf(req->rl_sendbuf); | |
1084 | rpcrdma_free_regbuf(req->rl_rdmabuf); | |
1392402c CL |
1085 | kfree(req); |
1086 | } | |
1087 | ||
e2ac236c CL |
1088 | static void |
1089 | rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf) | |
1090 | { | |
1091 | struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt, | |
1092 | rx_buf); | |
1093 | struct rpcrdma_ia *ia = rdmab_to_ia(buf); | |
1094 | struct rpcrdma_mw *mw; | |
1095 | unsigned int count; | |
1096 | ||
1097 | count = 0; | |
1098 | spin_lock(&buf->rb_mwlock); | |
1099 | while (!list_empty(&buf->rb_all)) { | |
1100 | mw = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all); | |
1101 | list_del(&mw->mw_all); | |
1102 | ||
1103 | spin_unlock(&buf->rb_mwlock); | |
1104 | ia->ri_ops->ro_release_mr(mw); | |
1105 | count++; | |
1106 | spin_lock(&buf->rb_mwlock); | |
1107 | } | |
1108 | spin_unlock(&buf->rb_mwlock); | |
1109 | r_xprt->rx_stats.mrs_allocated = 0; | |
1110 | ||
1111 | dprintk("RPC: %s: released %u MRs\n", __func__, count); | |
1112 | } | |
1113 | ||
c56c65fb TT |
1114 | void |
1115 | rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf) | |
1116 | { | |
505bbe64 | 1117 | cancel_delayed_work_sync(&buf->rb_recovery_worker); |
9378b274 | 1118 | cancel_delayed_work_sync(&buf->rb_refresh_worker); |
505bbe64 | 1119 | |
1e465fd4 CL |
1120 | while (!list_empty(&buf->rb_recv_bufs)) { |
1121 | struct rpcrdma_rep *rep; | |
c56c65fb | 1122 | |
1e465fd4 | 1123 | rep = rpcrdma_buffer_get_rep_locked(buf); |
13650c23 | 1124 | rpcrdma_destroy_rep(rep); |
c56c65fb | 1125 | } |
05c97466 | 1126 | buf->rb_send_count = 0; |
c56c65fb | 1127 | |
f531a5db CL |
1128 | spin_lock(&buf->rb_reqslock); |
1129 | while (!list_empty(&buf->rb_allreqs)) { | |
1e465fd4 | 1130 | struct rpcrdma_req *req; |
4034ba04 | 1131 | |
f531a5db CL |
1132 | req = list_first_entry(&buf->rb_allreqs, |
1133 | struct rpcrdma_req, rl_all); | |
1134 | list_del(&req->rl_all); | |
1135 | ||
1136 | spin_unlock(&buf->rb_reqslock); | |
13650c23 | 1137 | rpcrdma_destroy_req(req); |
f531a5db | 1138 | spin_lock(&buf->rb_reqslock); |
1e465fd4 | 1139 | } |
f531a5db | 1140 | spin_unlock(&buf->rb_reqslock); |
05c97466 | 1141 | buf->rb_recv_count = 0; |
4034ba04 | 1142 | |
e2ac236c | 1143 | rpcrdma_destroy_mrs(buf); |
c56c65fb TT |
1144 | } |
1145 | ||
346aa66b CL |
1146 | struct rpcrdma_mw * |
1147 | rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt) | |
c2922c02 | 1148 | { |
346aa66b CL |
1149 | struct rpcrdma_buffer *buf = &r_xprt->rx_buf; |
1150 | struct rpcrdma_mw *mw = NULL; | |
346aa66b | 1151 | |
58d1dcf5 | 1152 | spin_lock(&buf->rb_mwlock); |
9a5c63e9 CL |
1153 | if (!list_empty(&buf->rb_mws)) |
1154 | mw = rpcrdma_pop_mw(&buf->rb_mws); | |
58d1dcf5 | 1155 | spin_unlock(&buf->rb_mwlock); |
346aa66b CL |
1156 | |
1157 | if (!mw) | |
e2ac236c | 1158 | goto out_nomws; |
4b196dc6 | 1159 | mw->mw_flags = 0; |
346aa66b | 1160 | return mw; |
e2ac236c CL |
1161 | |
1162 | out_nomws: | |
1163 | dprintk("RPC: %s: no MWs available\n", __func__); | |
bebd0318 CL |
1164 | if (r_xprt->rx_ep.rep_connected != -ENODEV) |
1165 | schedule_delayed_work(&buf->rb_refresh_worker, 0); | |
e2ac236c CL |
1166 | |
1167 | /* Allow the reply handler and refresh worker to run */ | |
1168 | cond_resched(); | |
1169 | ||
1170 | return NULL; | |
c2922c02 CL |
1171 | } |
1172 | ||
346aa66b CL |
1173 | void |
1174 | rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw) | |
c2922c02 | 1175 | { |
346aa66b | 1176 | struct rpcrdma_buffer *buf = &r_xprt->rx_buf; |
c2922c02 | 1177 | |
58d1dcf5 | 1178 | spin_lock(&buf->rb_mwlock); |
9a5c63e9 | 1179 | rpcrdma_push_mw(mw, &buf->rb_mws); |
58d1dcf5 | 1180 | spin_unlock(&buf->rb_mwlock); |
c2922c02 CL |
1181 | } |
1182 | ||
05c97466 CL |
1183 | static struct rpcrdma_rep * |
1184 | rpcrdma_buffer_get_rep(struct rpcrdma_buffer *buffers) | |
1185 | { | |
1186 | /* If an RPC previously completed without a reply (say, a | |
1187 | * credential problem or a soft timeout occurs) then hold off | |
1188 | * on supplying more Receive buffers until the number of new | |
1189 | * pending RPCs catches up to the number of posted Receives. | |
1190 | */ | |
1191 | if (unlikely(buffers->rb_send_count < buffers->rb_recv_count)) | |
1192 | return NULL; | |
1193 | ||
1194 | if (unlikely(list_empty(&buffers->rb_recv_bufs))) | |
1195 | return NULL; | |
1196 | buffers->rb_recv_count++; | |
1197 | return rpcrdma_buffer_get_rep_locked(buffers); | |
1198 | } | |
1199 | ||
c56c65fb TT |
1200 | /* |
1201 | * Get a set of request/reply buffers. | |
78d506e1 CL |
1202 | * |
1203 | * Reply buffer (if available) is attached to send buffer upon return. | |
c56c65fb TT |
1204 | */ |
1205 | struct rpcrdma_req * | |
1206 | rpcrdma_buffer_get(struct rpcrdma_buffer *buffers) | |
1207 | { | |
1208 | struct rpcrdma_req *req; | |
c14d86e5 | 1209 | |
a5b027e1 | 1210 | spin_lock(&buffers->rb_lock); |
1e465fd4 CL |
1211 | if (list_empty(&buffers->rb_send_bufs)) |
1212 | goto out_reqbuf; | |
05c97466 | 1213 | buffers->rb_send_count++; |
1e465fd4 | 1214 | req = rpcrdma_buffer_get_req_locked(buffers); |
05c97466 | 1215 | req->rl_reply = rpcrdma_buffer_get_rep(buffers); |
a5b027e1 | 1216 | spin_unlock(&buffers->rb_lock); |
1e465fd4 | 1217 | return req; |
ddb6bebc | 1218 | |
1e465fd4 | 1219 | out_reqbuf: |
a5b027e1 | 1220 | spin_unlock(&buffers->rb_lock); |
78d506e1 | 1221 | pr_warn("RPC: %s: out of request buffers\n", __func__); |
1e465fd4 | 1222 | return NULL; |
c56c65fb TT |
1223 | } |
1224 | ||
1225 | /* | |
1226 | * Put request/reply buffers back into pool. | |
1227 | * Pre-decrement counter/array index. | |
1228 | */ | |
1229 | void | |
1230 | rpcrdma_buffer_put(struct rpcrdma_req *req) | |
1231 | { | |
1232 | struct rpcrdma_buffer *buffers = req->rl_buffer; | |
1e465fd4 | 1233 | struct rpcrdma_rep *rep = req->rl_reply; |
c56c65fb | 1234 | |
90aab602 | 1235 | req->rl_send_wr.num_sge = 0; |
1e465fd4 CL |
1236 | req->rl_reply = NULL; |
1237 | ||
a5b027e1 | 1238 | spin_lock(&buffers->rb_lock); |
05c97466 | 1239 | buffers->rb_send_count--; |
a80d66c9 | 1240 | list_add_tail(&req->rl_list, &buffers->rb_send_bufs); |
05c97466 CL |
1241 | if (rep) { |
1242 | buffers->rb_recv_count--; | |
1e465fd4 | 1243 | list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs); |
05c97466 | 1244 | } |
a5b027e1 | 1245 | spin_unlock(&buffers->rb_lock); |
c56c65fb TT |
1246 | } |
1247 | ||
1248 | /* | |
1249 | * Recover reply buffers from pool. | |
1e465fd4 | 1250 | * This happens when recovering from disconnect. |
c56c65fb TT |
1251 | */ |
1252 | void | |
1253 | rpcrdma_recv_buffer_get(struct rpcrdma_req *req) | |
1254 | { | |
1255 | struct rpcrdma_buffer *buffers = req->rl_buffer; | |
c56c65fb | 1256 | |
a5b027e1 | 1257 | spin_lock(&buffers->rb_lock); |
05c97466 | 1258 | req->rl_reply = rpcrdma_buffer_get_rep(buffers); |
a5b027e1 | 1259 | spin_unlock(&buffers->rb_lock); |
c56c65fb TT |
1260 | } |
1261 | ||
1262 | /* | |
1263 | * Put reply buffers back into pool when not attached to | |
b45ccfd2 | 1264 | * request. This happens in error conditions. |
c56c65fb TT |
1265 | */ |
1266 | void | |
1267 | rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep) | |
1268 | { | |
fed171b3 | 1269 | struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf; |
c56c65fb | 1270 | |
a5b027e1 | 1271 | spin_lock(&buffers->rb_lock); |
05c97466 | 1272 | buffers->rb_recv_count--; |
1e465fd4 | 1273 | list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs); |
a5b027e1 | 1274 | spin_unlock(&buffers->rb_lock); |
c56c65fb TT |
1275 | } |
1276 | ||
9128c3e7 | 1277 | /** |
99ef4db3 | 1278 | * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers |
9128c3e7 | 1279 | * @size: size of buffer to be allocated, in bytes |
99ef4db3 | 1280 | * @direction: direction of data movement |
9128c3e7 CL |
1281 | * @flags: GFP flags |
1282 | * | |
54cbd6b0 CL |
1283 | * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that |
1284 | * can be persistently DMA-mapped for I/O. | |
9128c3e7 CL |
1285 | * |
1286 | * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for | |
99ef4db3 CL |
1287 | * receiving the payload of RDMA RECV operations. During Long Calls |
1288 | * or Replies they may be registered externally via ro_map. | |
9128c3e7 CL |
1289 | */ |
1290 | struct rpcrdma_regbuf * | |
13650c23 CL |
1291 | rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction, |
1292 | gfp_t flags) | |
9128c3e7 CL |
1293 | { |
1294 | struct rpcrdma_regbuf *rb; | |
9128c3e7 | 1295 | |
9128c3e7 CL |
1296 | rb = kmalloc(sizeof(*rb) + size, flags); |
1297 | if (rb == NULL) | |
54cbd6b0 | 1298 | return ERR_PTR(-ENOMEM); |
9128c3e7 | 1299 | |
54cbd6b0 | 1300 | rb->rg_device = NULL; |
99ef4db3 | 1301 | rb->rg_direction = direction; |
54cbd6b0 | 1302 | rb->rg_iov.length = size; |
9128c3e7 CL |
1303 | |
1304 | return rb; | |
54cbd6b0 | 1305 | } |
9128c3e7 | 1306 | |
54cbd6b0 CL |
1307 | /** |
1308 | * __rpcrdma_map_regbuf - DMA-map a regbuf | |
1309 | * @ia: controlling rpcrdma_ia | |
1310 | * @rb: regbuf to be mapped | |
1311 | */ | |
1312 | bool | |
1313 | __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb) | |
1314 | { | |
91a10c52 CL |
1315 | struct ib_device *device = ia->ri_device; |
1316 | ||
54cbd6b0 CL |
1317 | if (rb->rg_direction == DMA_NONE) |
1318 | return false; | |
1319 | ||
91a10c52 | 1320 | rb->rg_iov.addr = ib_dma_map_single(device, |
54cbd6b0 CL |
1321 | (void *)rb->rg_base, |
1322 | rdmab_length(rb), | |
1323 | rb->rg_direction); | |
91a10c52 | 1324 | if (ib_dma_mapping_error(device, rdmab_addr(rb))) |
54cbd6b0 CL |
1325 | return false; |
1326 | ||
91a10c52 | 1327 | rb->rg_device = device; |
54cbd6b0 CL |
1328 | rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey; |
1329 | return true; | |
1330 | } | |
1331 | ||
1332 | static void | |
1333 | rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb) | |
1334 | { | |
1335 | if (!rpcrdma_regbuf_is_mapped(rb)) | |
1336 | return; | |
1337 | ||
1338 | ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb), | |
1339 | rdmab_length(rb), rb->rg_direction); | |
1340 | rb->rg_device = NULL; | |
9128c3e7 CL |
1341 | } |
1342 | ||
1343 | /** | |
1344 | * rpcrdma_free_regbuf - deregister and free registered buffer | |
9128c3e7 CL |
1345 | * @rb: regbuf to be deregistered and freed |
1346 | */ | |
1347 | void | |
13650c23 | 1348 | rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb) |
9128c3e7 | 1349 | { |
e531dcab CL |
1350 | if (!rb) |
1351 | return; | |
1352 | ||
54cbd6b0 | 1353 | rpcrdma_dma_unmap_regbuf(rb); |
e531dcab | 1354 | kfree(rb); |
9128c3e7 CL |
1355 | } |
1356 | ||
c56c65fb TT |
1357 | /* |
1358 | * Prepost any receive buffer, then post send. | |
1359 | * | |
1360 | * Receive buffer is donated to hardware, reclaimed upon recv completion. | |
1361 | */ | |
1362 | int | |
1363 | rpcrdma_ep_post(struct rpcrdma_ia *ia, | |
1364 | struct rpcrdma_ep *ep, | |
1365 | struct rpcrdma_req *req) | |
1366 | { | |
90aab602 CL |
1367 | struct ib_send_wr *send_wr = &req->rl_send_wr; |
1368 | struct ib_send_wr *send_wr_fail; | |
655fec69 | 1369 | int rc; |
c56c65fb | 1370 | |
90aab602 CL |
1371 | if (req->rl_reply) { |
1372 | rc = rpcrdma_ep_post_recv(ia, req->rl_reply); | |
c56c65fb | 1373 | if (rc) |
7a89f9c6 | 1374 | return rc; |
c56c65fb TT |
1375 | req->rl_reply = NULL; |
1376 | } | |
1377 | ||
b3221d6a | 1378 | dprintk("RPC: %s: posting %d s/g entries\n", |
90aab602 | 1379 | __func__, send_wr->num_sge); |
c56c65fb | 1380 | |
8d38de65 | 1381 | rpcrdma_set_signaled(ep, send_wr); |
90aab602 | 1382 | rc = ib_post_send(ia->ri_id->qp, send_wr, &send_wr_fail); |
c56c65fb | 1383 | if (rc) |
7a89f9c6 CL |
1384 | goto out_postsend_err; |
1385 | return 0; | |
1386 | ||
1387 | out_postsend_err: | |
1388 | pr_err("rpcrdma: RDMA Send ib_post_send returned %i\n", rc); | |
1389 | return -ENOTCONN; | |
c56c65fb TT |
1390 | } |
1391 | ||
c56c65fb TT |
1392 | int |
1393 | rpcrdma_ep_post_recv(struct rpcrdma_ia *ia, | |
c56c65fb TT |
1394 | struct rpcrdma_rep *rep) |
1395 | { | |
6ea8e711 | 1396 | struct ib_recv_wr *recv_wr_fail; |
c56c65fb TT |
1397 | int rc; |
1398 | ||
54cbd6b0 CL |
1399 | if (!rpcrdma_dma_map_regbuf(ia, rep->rr_rdmabuf)) |
1400 | goto out_map; | |
6ea8e711 | 1401 | rc = ib_post_recv(ia->ri_id->qp, &rep->rr_recv_wr, &recv_wr_fail); |
c56c65fb | 1402 | if (rc) |
7a89f9c6 CL |
1403 | goto out_postrecv; |
1404 | return 0; | |
1405 | ||
54cbd6b0 CL |
1406 | out_map: |
1407 | pr_err("rpcrdma: failed to DMA map the Receive buffer\n"); | |
1408 | return -EIO; | |
1409 | ||
7a89f9c6 CL |
1410 | out_postrecv: |
1411 | pr_err("rpcrdma: ib_post_recv returned %i\n", rc); | |
1412 | return -ENOTCONN; | |
c56c65fb | 1413 | } |
43e95988 | 1414 | |
f531a5db CL |
1415 | /** |
1416 | * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests | |
1417 | * @r_xprt: transport associated with these backchannel resources | |
1418 | * @min_reqs: minimum number of incoming requests expected | |
1419 | * | |
1420 | * Returns zero if all requested buffers were posted, or a negative errno. | |
1421 | */ | |
1422 | int | |
1423 | rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count) | |
1424 | { | |
1425 | struct rpcrdma_buffer *buffers = &r_xprt->rx_buf; | |
1426 | struct rpcrdma_ia *ia = &r_xprt->rx_ia; | |
f531a5db | 1427 | struct rpcrdma_rep *rep; |
f531a5db CL |
1428 | int rc; |
1429 | ||
1430 | while (count--) { | |
9b06688b | 1431 | spin_lock(&buffers->rb_lock); |
f531a5db CL |
1432 | if (list_empty(&buffers->rb_recv_bufs)) |
1433 | goto out_reqbuf; | |
1434 | rep = rpcrdma_buffer_get_rep_locked(buffers); | |
9b06688b | 1435 | spin_unlock(&buffers->rb_lock); |
f531a5db | 1436 | |
b157380a | 1437 | rc = rpcrdma_ep_post_recv(ia, rep); |
f531a5db CL |
1438 | if (rc) |
1439 | goto out_rc; | |
1440 | } | |
1441 | ||
1442 | return 0; | |
1443 | ||
1444 | out_reqbuf: | |
9b06688b | 1445 | spin_unlock(&buffers->rb_lock); |
f531a5db CL |
1446 | pr_warn("%s: no extra receive buffers\n", __func__); |
1447 | return -ENOMEM; | |
1448 | ||
1449 | out_rc: | |
1450 | rpcrdma_recv_buffer_put(rep); | |
1451 | return rc; | |
1452 | } |