Commit | Line | Data |
---|---|---|
377f9b2f | 1 | /* |
0bf48289 | 2 | * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved. |
377f9b2f TT |
3 | * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved. |
4 | * | |
5 | * This software is available to you under a choice of one of two | |
6 | * licenses. You may choose to be licensed under the terms of the GNU | |
7 | * General Public License (GPL) Version 2, available from the file | |
8 | * COPYING in the main directory of this source tree, or the BSD-type | |
9 | * license below: | |
10 | * | |
11 | * Redistribution and use in source and binary forms, with or without | |
12 | * modification, are permitted provided that the following conditions | |
13 | * are met: | |
14 | * | |
15 | * Redistributions of source code must retain the above copyright | |
16 | * notice, this list of conditions and the following disclaimer. | |
17 | * | |
18 | * Redistributions in binary form must reproduce the above | |
19 | * copyright notice, this list of conditions and the following | |
20 | * disclaimer in the documentation and/or other materials provided | |
21 | * with the distribution. | |
22 | * | |
23 | * Neither the name of the Network Appliance, Inc. nor the names of | |
24 | * its contributors may be used to endorse or promote products | |
25 | * derived from this software without specific prior written | |
26 | * permission. | |
27 | * | |
28 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS | |
29 | * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT | |
30 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR | |
31 | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT | |
32 | * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, | |
33 | * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT | |
34 | * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, | |
35 | * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY | |
36 | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT | |
37 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE | |
38 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |
39 | * | |
40 | * Author: Tom Tucker <tom@opengridcomputing.com> | |
41 | */ | |
42 | ||
43 | #include <linux/sunrpc/svc_xprt.h> | |
44 | #include <linux/sunrpc/debug.h> | |
45 | #include <linux/sunrpc/rpc_rdma.h> | |
a6b7a407 | 46 | #include <linux/interrupt.h> |
d43c36dc | 47 | #include <linux/sched.h> |
5a0e3ad6 | 48 | #include <linux/slab.h> |
377f9b2f | 49 | #include <linux/spinlock.h> |
a25e758c | 50 | #include <linux/workqueue.h> |
377f9b2f TT |
51 | #include <rdma/ib_verbs.h> |
52 | #include <rdma/rdma_cm.h> | |
53 | #include <linux/sunrpc/svc_rdma.h> | |
bc3b2d7f | 54 | #include <linux/export.h> |
cec56c8f | 55 | #include "xprt_rdma.h" |
377f9b2f TT |
56 | |
57 | #define RPCDBG_FACILITY RPCDBG_SVCXPRT | |
58 | ||
59 | static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, | |
62832c03 | 60 | struct net *net, |
377f9b2f TT |
61 | struct sockaddr *sa, int salen, |
62 | int flags); | |
63 | static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt); | |
64 | static void svc_rdma_release_rqst(struct svc_rqst *); | |
377f9b2f TT |
65 | static void dto_tasklet_func(unsigned long data); |
66 | static void svc_rdma_detach(struct svc_xprt *xprt); | |
67 | static void svc_rdma_free(struct svc_xprt *xprt); | |
68 | static int svc_rdma_has_wspace(struct svc_xprt *xprt); | |
16e4d93f | 69 | static int svc_rdma_secure_port(struct svc_rqst *); |
377f9b2f TT |
70 | static void rq_cq_reap(struct svcxprt_rdma *xprt); |
71 | static void sq_cq_reap(struct svcxprt_rdma *xprt); | |
72 | ||
5eaa65b2 | 73 | static DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL); |
377f9b2f TT |
74 | static DEFINE_SPINLOCK(dto_lock); |
75 | static LIST_HEAD(dto_xprt_q); | |
76 | ||
77 | static struct svc_xprt_ops svc_rdma_ops = { | |
78 | .xpo_create = svc_rdma_create, | |
79 | .xpo_recvfrom = svc_rdma_recvfrom, | |
80 | .xpo_sendto = svc_rdma_sendto, | |
81 | .xpo_release_rqst = svc_rdma_release_rqst, | |
82 | .xpo_detach = svc_rdma_detach, | |
83 | .xpo_free = svc_rdma_free, | |
84 | .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr, | |
85 | .xpo_has_wspace = svc_rdma_has_wspace, | |
86 | .xpo_accept = svc_rdma_accept, | |
16e4d93f | 87 | .xpo_secure_port = svc_rdma_secure_port, |
377f9b2f TT |
88 | }; |
89 | ||
90 | struct svc_xprt_class svc_rdma_class = { | |
91 | .xcl_name = "rdma", | |
92 | .xcl_owner = THIS_MODULE, | |
93 | .xcl_ops = &svc_rdma_ops, | |
7e5be288 | 94 | .xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA, |
3c45ddf8 | 95 | .xcl_ident = XPRT_TRANSPORT_RDMA, |
377f9b2f TT |
96 | }; |
97 | ||
377f9b2f TT |
98 | struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt) |
99 | { | |
100 | struct svc_rdma_op_ctxt *ctxt; | |
101 | ||
102 | while (1) { | |
8948896c TT |
103 | ctxt = kmem_cache_alloc(svc_rdma_ctxt_cachep, GFP_KERNEL); |
104 | if (ctxt) | |
105 | break; | |
106 | schedule_timeout_uninterruptible(msecs_to_jiffies(500)); | |
377f9b2f | 107 | } |
8948896c TT |
108 | ctxt->xprt = xprt; |
109 | INIT_LIST_HEAD(&ctxt->dto_q); | |
110 | ctxt->count = 0; | |
64be8608 | 111 | ctxt->frmr = NULL; |
8948896c | 112 | atomic_inc(&xprt->sc_ctxt_used); |
377f9b2f TT |
113 | return ctxt; |
114 | } | |
115 | ||
146b6df6 | 116 | void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt) |
e6ab9143 TT |
117 | { |
118 | struct svcxprt_rdma *xprt = ctxt->xprt; | |
119 | int i; | |
120 | for (i = 0; i < ctxt->count && ctxt->sge[i].length; i++) { | |
64be8608 TT |
121 | /* |
122 | * Unmap the DMA addr in the SGE if the lkey matches | |
123 | * the sc_dma_lkey, otherwise, ignore it since it is | |
124 | * an FRMR lkey and will be unmapped later when the | |
125 | * last WR that uses it completes. | |
126 | */ | |
127 | if (ctxt->sge[i].lkey == xprt->sc_dma_lkey) { | |
128 | atomic_dec(&xprt->sc_dma_used); | |
b432e6b3 | 129 | ib_dma_unmap_page(xprt->sc_cm_id->device, |
64be8608 TT |
130 | ctxt->sge[i].addr, |
131 | ctxt->sge[i].length, | |
132 | ctxt->direction); | |
133 | } | |
e6ab9143 TT |
134 | } |
135 | } | |
136 | ||
377f9b2f TT |
137 | void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages) |
138 | { | |
139 | struct svcxprt_rdma *xprt; | |
140 | int i; | |
141 | ||
377f9b2f TT |
142 | xprt = ctxt->xprt; |
143 | if (free_pages) | |
144 | for (i = 0; i < ctxt->count; i++) | |
145 | put_page(ctxt->pages[i]); | |
146 | ||
8948896c | 147 | kmem_cache_free(svc_rdma_ctxt_cachep, ctxt); |
87407673 | 148 | atomic_dec(&xprt->sc_ctxt_used); |
377f9b2f TT |
149 | } |
150 | ||
ab96dddb TT |
151 | /* |
152 | * Temporary NFS req mappings are shared across all transport | |
153 | * instances. These are short lived and should be bounded by the number | |
154 | * of concurrent server threads * depth of the SQ. | |
155 | */ | |
156 | struct svc_rdma_req_map *svc_rdma_get_req_map(void) | |
157 | { | |
158 | struct svc_rdma_req_map *map; | |
159 | while (1) { | |
160 | map = kmem_cache_alloc(svc_rdma_map_cachep, GFP_KERNEL); | |
161 | if (map) | |
162 | break; | |
163 | schedule_timeout_uninterruptible(msecs_to_jiffies(500)); | |
164 | } | |
165 | map->count = 0; | |
166 | return map; | |
167 | } | |
168 | ||
169 | void svc_rdma_put_req_map(struct svc_rdma_req_map *map) | |
170 | { | |
171 | kmem_cache_free(svc_rdma_map_cachep, map); | |
172 | } | |
173 | ||
377f9b2f TT |
174 | /* ib_cq event handler */ |
175 | static void cq_event_handler(struct ib_event *event, void *context) | |
176 | { | |
177 | struct svc_xprt *xprt = context; | |
76357c71 SG |
178 | dprintk("svcrdma: received CQ event %s (%d), context=%p\n", |
179 | ib_event_msg(event->event), event->event, context); | |
377f9b2f TT |
180 | set_bit(XPT_CLOSE, &xprt->xpt_flags); |
181 | } | |
182 | ||
183 | /* QP event handler */ | |
184 | static void qp_event_handler(struct ib_event *event, void *context) | |
185 | { | |
186 | struct svc_xprt *xprt = context; | |
187 | ||
188 | switch (event->event) { | |
189 | /* These are considered benign events */ | |
190 | case IB_EVENT_PATH_MIG: | |
191 | case IB_EVENT_COMM_EST: | |
192 | case IB_EVENT_SQ_DRAINED: | |
193 | case IB_EVENT_QP_LAST_WQE_REACHED: | |
76357c71 SG |
194 | dprintk("svcrdma: QP event %s (%d) received for QP=%p\n", |
195 | ib_event_msg(event->event), event->event, | |
196 | event->element.qp); | |
377f9b2f TT |
197 | break; |
198 | /* These are considered fatal events */ | |
199 | case IB_EVENT_PATH_MIG_ERR: | |
200 | case IB_EVENT_QP_FATAL: | |
201 | case IB_EVENT_QP_REQ_ERR: | |
202 | case IB_EVENT_QP_ACCESS_ERR: | |
203 | case IB_EVENT_DEVICE_FATAL: | |
204 | default: | |
76357c71 | 205 | dprintk("svcrdma: QP ERROR event %s (%d) received for QP=%p, " |
377f9b2f | 206 | "closing transport\n", |
76357c71 SG |
207 | ib_event_msg(event->event), event->event, |
208 | event->element.qp); | |
377f9b2f TT |
209 | set_bit(XPT_CLOSE, &xprt->xpt_flags); |
210 | break; | |
211 | } | |
212 | } | |
213 | ||
214 | /* | |
215 | * Data Transfer Operation Tasklet | |
216 | * | |
217 | * Walks a list of transports with I/O pending, removing entries as | |
218 | * they are added to the server's I/O pending list. Two bits indicate | |
219 | * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave | |
220 | * spinlock that serializes access to the transport list with the RQ | |
221 | * and SQ interrupt handlers. | |
222 | */ | |
223 | static void dto_tasklet_func(unsigned long data) | |
224 | { | |
225 | struct svcxprt_rdma *xprt; | |
226 | unsigned long flags; | |
227 | ||
228 | spin_lock_irqsave(&dto_lock, flags); | |
229 | while (!list_empty(&dto_xprt_q)) { | |
230 | xprt = list_entry(dto_xprt_q.next, | |
231 | struct svcxprt_rdma, sc_dto_q); | |
232 | list_del_init(&xprt->sc_dto_q); | |
233 | spin_unlock_irqrestore(&dto_lock, flags); | |
234 | ||
dbcd00eb TT |
235 | rq_cq_reap(xprt); |
236 | sq_cq_reap(xprt); | |
377f9b2f | 237 | |
c48cbb40 | 238 | svc_xprt_put(&xprt->sc_xprt); |
377f9b2f TT |
239 | spin_lock_irqsave(&dto_lock, flags); |
240 | } | |
241 | spin_unlock_irqrestore(&dto_lock, flags); | |
242 | } | |
243 | ||
244 | /* | |
245 | * Receive Queue Completion Handler | |
246 | * | |
247 | * Since an RQ completion handler is called on interrupt context, we | |
248 | * need to defer the handling of the I/O to a tasklet | |
249 | */ | |
250 | static void rq_comp_handler(struct ib_cq *cq, void *cq_context) | |
251 | { | |
252 | struct svcxprt_rdma *xprt = cq_context; | |
253 | unsigned long flags; | |
254 | ||
1711386c TT |
255 | /* Guard against unconditional flush call for destroyed QP */ |
256 | if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0) | |
257 | return; | |
258 | ||
377f9b2f TT |
259 | /* |
260 | * Set the bit regardless of whether or not it's on the list | |
261 | * because it may be on the list already due to an SQ | |
262 | * completion. | |
1711386c | 263 | */ |
377f9b2f TT |
264 | set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags); |
265 | ||
266 | /* | |
267 | * If this transport is not already on the DTO transport queue, | |
268 | * add it | |
269 | */ | |
270 | spin_lock_irqsave(&dto_lock, flags); | |
c48cbb40 TT |
271 | if (list_empty(&xprt->sc_dto_q)) { |
272 | svc_xprt_get(&xprt->sc_xprt); | |
377f9b2f | 273 | list_add_tail(&xprt->sc_dto_q, &dto_xprt_q); |
c48cbb40 | 274 | } |
377f9b2f TT |
275 | spin_unlock_irqrestore(&dto_lock, flags); |
276 | ||
277 | /* Tasklet does all the work to avoid irqsave locks. */ | |
278 | tasklet_schedule(&dto_tasklet); | |
279 | } | |
280 | ||
281 | /* | |
282 | * rq_cq_reap - Process the RQ CQ. | |
283 | * | |
284 | * Take all completing WC off the CQE and enqueue the associated DTO | |
285 | * context on the dto_q for the transport. | |
0905c0f0 TT |
286 | * |
287 | * Note that caller must hold a transport reference. | |
377f9b2f TT |
288 | */ |
289 | static void rq_cq_reap(struct svcxprt_rdma *xprt) | |
290 | { | |
291 | int ret; | |
292 | struct ib_wc wc; | |
293 | struct svc_rdma_op_ctxt *ctxt = NULL; | |
294 | ||
dbcd00eb TT |
295 | if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags)) |
296 | return; | |
297 | ||
298 | ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP); | |
377f9b2f TT |
299 | atomic_inc(&rdma_stat_rq_poll); |
300 | ||
377f9b2f TT |
301 | while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) { |
302 | ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id; | |
303 | ctxt->wc_status = wc.status; | |
304 | ctxt->byte_len = wc.byte_len; | |
e6ab9143 | 305 | svc_rdma_unmap_dma(ctxt); |
377f9b2f TT |
306 | if (wc.status != IB_WC_SUCCESS) { |
307 | /* Close the transport */ | |
0905c0f0 | 308 | dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt); |
377f9b2f TT |
309 | set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); |
310 | svc_rdma_put_context(ctxt, 1); | |
0905c0f0 | 311 | svc_xprt_put(&xprt->sc_xprt); |
377f9b2f TT |
312 | continue; |
313 | } | |
47698e08 | 314 | spin_lock_bh(&xprt->sc_rq_dto_lock); |
377f9b2f | 315 | list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q); |
47698e08 | 316 | spin_unlock_bh(&xprt->sc_rq_dto_lock); |
0905c0f0 | 317 | svc_xprt_put(&xprt->sc_xprt); |
377f9b2f | 318 | } |
377f9b2f TT |
319 | |
320 | if (ctxt) | |
321 | atomic_inc(&rdma_stat_rq_prod); | |
dbcd00eb TT |
322 | |
323 | set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags); | |
324 | /* | |
325 | * If data arrived before established event, | |
326 | * don't enqueue. This defers RPC I/O until the | |
327 | * RDMA connection is complete. | |
328 | */ | |
329 | if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags)) | |
330 | svc_xprt_enqueue(&xprt->sc_xprt); | |
377f9b2f TT |
331 | } |
332 | ||
e1183210 | 333 | /* |
70f23fd6 | 334 | * Process a completion context |
e1183210 TT |
335 | */ |
336 | static void process_context(struct svcxprt_rdma *xprt, | |
337 | struct svc_rdma_op_ctxt *ctxt) | |
338 | { | |
339 | svc_rdma_unmap_dma(ctxt); | |
340 | ||
341 | switch (ctxt->wr_op) { | |
342 | case IB_WR_SEND: | |
3fe04ee9 CL |
343 | if (ctxt->frmr) |
344 | pr_err("svcrdma: SEND: ctxt->frmr != NULL\n"); | |
e1183210 TT |
345 | svc_rdma_put_context(ctxt, 1); |
346 | break; | |
347 | ||
348 | case IB_WR_RDMA_WRITE: | |
3fe04ee9 CL |
349 | if (ctxt->frmr) |
350 | pr_err("svcrdma: WRITE: ctxt->frmr != NULL\n"); | |
e1183210 TT |
351 | svc_rdma_put_context(ctxt, 0); |
352 | break; | |
353 | ||
354 | case IB_WR_RDMA_READ: | |
146b6df6 | 355 | case IB_WR_RDMA_READ_WITH_INV: |
0bf48289 | 356 | svc_rdma_put_frmr(xprt, ctxt->frmr); |
e1183210 TT |
357 | if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) { |
358 | struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr; | |
3fe04ee9 CL |
359 | if (read_hdr) { |
360 | spin_lock_bh(&xprt->sc_rq_dto_lock); | |
361 | set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags); | |
362 | list_add_tail(&read_hdr->dto_q, | |
363 | &xprt->sc_read_complete_q); | |
364 | spin_unlock_bh(&xprt->sc_rq_dto_lock); | |
365 | } else { | |
366 | pr_err("svcrdma: ctxt->read_hdr == NULL\n"); | |
367 | } | |
e1183210 TT |
368 | svc_xprt_enqueue(&xprt->sc_xprt); |
369 | } | |
370 | svc_rdma_put_context(ctxt, 0); | |
371 | break; | |
372 | ||
373 | default: | |
374 | printk(KERN_ERR "svcrdma: unexpected completion type, " | |
375 | "opcode=%d\n", | |
376 | ctxt->wr_op); | |
377 | break; | |
378 | } | |
379 | } | |
380 | ||
377f9b2f TT |
381 | /* |
382 | * Send Queue Completion Handler - potentially called on interrupt context. | |
0905c0f0 TT |
383 | * |
384 | * Note that caller must hold a transport reference. | |
377f9b2f TT |
385 | */ |
386 | static void sq_cq_reap(struct svcxprt_rdma *xprt) | |
387 | { | |
388 | struct svc_rdma_op_ctxt *ctxt = NULL; | |
0bf48289 SW |
389 | struct ib_wc wc_a[6]; |
390 | struct ib_wc *wc; | |
377f9b2f TT |
391 | struct ib_cq *cq = xprt->sc_sq_cq; |
392 | int ret; | |
393 | ||
0bf48289 SW |
394 | memset(wc_a, 0, sizeof(wc_a)); |
395 | ||
dbcd00eb TT |
396 | if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags)) |
397 | return; | |
398 | ||
399 | ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP); | |
377f9b2f | 400 | atomic_inc(&rdma_stat_sq_poll); |
0bf48289 SW |
401 | while ((ret = ib_poll_cq(cq, ARRAY_SIZE(wc_a), wc_a)) > 0) { |
402 | int i; | |
377f9b2f | 403 | |
0bf48289 SW |
404 | for (i = 0; i < ret; i++) { |
405 | wc = &wc_a[i]; | |
406 | if (wc->status != IB_WC_SUCCESS) { | |
76357c71 SG |
407 | dprintk("svcrdma: sq wc err status %s (%d)\n", |
408 | ib_wc_status_msg(wc->status), | |
0bf48289 | 409 | wc->status); |
377f9b2f | 410 | |
0bf48289 SW |
411 | /* Close the transport */ |
412 | set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); | |
413 | } | |
377f9b2f | 414 | |
0bf48289 SW |
415 | /* Decrement used SQ WR count */ |
416 | atomic_dec(&xprt->sc_sq_count); | |
417 | wake_up(&xprt->sc_send_wait); | |
418 | ||
419 | ctxt = (struct svc_rdma_op_ctxt *) | |
420 | (unsigned long)wc->wr_id; | |
421 | if (ctxt) | |
422 | process_context(xprt, ctxt); | |
423 | ||
424 | svc_xprt_put(&xprt->sc_xprt); | |
425 | } | |
377f9b2f TT |
426 | } |
427 | ||
428 | if (ctxt) | |
429 | atomic_inc(&rdma_stat_sq_prod); | |
430 | } | |
431 | ||
432 | static void sq_comp_handler(struct ib_cq *cq, void *cq_context) | |
433 | { | |
434 | struct svcxprt_rdma *xprt = cq_context; | |
435 | unsigned long flags; | |
436 | ||
1711386c TT |
437 | /* Guard against unconditional flush call for destroyed QP */ |
438 | if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0) | |
439 | return; | |
440 | ||
377f9b2f TT |
441 | /* |
442 | * Set the bit regardless of whether or not it's on the list | |
443 | * because it may be on the list already due to an RQ | |
444 | * completion. | |
1711386c | 445 | */ |
377f9b2f TT |
446 | set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags); |
447 | ||
448 | /* | |
449 | * If this transport is not already on the DTO transport queue, | |
450 | * add it | |
451 | */ | |
452 | spin_lock_irqsave(&dto_lock, flags); | |
c48cbb40 TT |
453 | if (list_empty(&xprt->sc_dto_q)) { |
454 | svc_xprt_get(&xprt->sc_xprt); | |
377f9b2f | 455 | list_add_tail(&xprt->sc_dto_q, &dto_xprt_q); |
c48cbb40 | 456 | } |
377f9b2f TT |
457 | spin_unlock_irqrestore(&dto_lock, flags); |
458 | ||
459 | /* Tasklet does all the work to avoid irqsave locks. */ | |
460 | tasklet_schedule(&dto_tasklet); | |
461 | } | |
462 | ||
377f9b2f TT |
463 | static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv, |
464 | int listener) | |
465 | { | |
466 | struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL); | |
467 | ||
468 | if (!cma_xprt) | |
469 | return NULL; | |
bd4620dd | 470 | svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv); |
377f9b2f TT |
471 | INIT_LIST_HEAD(&cma_xprt->sc_accept_q); |
472 | INIT_LIST_HEAD(&cma_xprt->sc_dto_q); | |
473 | INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q); | |
474 | INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q); | |
64be8608 | 475 | INIT_LIST_HEAD(&cma_xprt->sc_frmr_q); |
377f9b2f TT |
476 | init_waitqueue_head(&cma_xprt->sc_send_wait); |
477 | ||
478 | spin_lock_init(&cma_xprt->sc_lock); | |
377f9b2f | 479 | spin_lock_init(&cma_xprt->sc_rq_dto_lock); |
64be8608 | 480 | spin_lock_init(&cma_xprt->sc_frmr_q_lock); |
377f9b2f TT |
481 | |
482 | cma_xprt->sc_ord = svcrdma_ord; | |
483 | ||
484 | cma_xprt->sc_max_req_size = svcrdma_max_req_size; | |
485 | cma_xprt->sc_max_requests = svcrdma_max_requests; | |
486 | cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT; | |
487 | atomic_set(&cma_xprt->sc_sq_count, 0); | |
87295b6c | 488 | atomic_set(&cma_xprt->sc_ctxt_used, 0); |
377f9b2f | 489 | |
8948896c | 490 | if (listener) |
377f9b2f TT |
491 | set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags); |
492 | ||
493 | return cma_xprt; | |
494 | } | |
495 | ||
496 | struct page *svc_rdma_get_page(void) | |
497 | { | |
498 | struct page *page; | |
499 | ||
500 | while ((page = alloc_page(GFP_KERNEL)) == NULL) { | |
501 | /* If we can't get memory, wait a bit and try again */ | |
c42a01ee | 502 | printk(KERN_INFO "svcrdma: out of memory...retrying in 1s\n"); |
377f9b2f TT |
503 | schedule_timeout_uninterruptible(msecs_to_jiffies(1000)); |
504 | } | |
505 | return page; | |
506 | } | |
507 | ||
508 | int svc_rdma_post_recv(struct svcxprt_rdma *xprt) | |
509 | { | |
510 | struct ib_recv_wr recv_wr, *bad_recv_wr; | |
511 | struct svc_rdma_op_ctxt *ctxt; | |
512 | struct page *page; | |
a5abf4e8 | 513 | dma_addr_t pa; |
377f9b2f TT |
514 | int sge_no; |
515 | int buflen; | |
516 | int ret; | |
517 | ||
518 | ctxt = svc_rdma_get_context(xprt); | |
519 | buflen = 0; | |
520 | ctxt->direction = DMA_FROM_DEVICE; | |
521 | for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) { | |
3fe04ee9 CL |
522 | if (sge_no >= xprt->sc_max_sge) { |
523 | pr_err("svcrdma: Too many sges (%d)\n", sge_no); | |
524 | goto err_put_ctxt; | |
525 | } | |
377f9b2f TT |
526 | page = svc_rdma_get_page(); |
527 | ctxt->pages[sge_no] = page; | |
b432e6b3 TT |
528 | pa = ib_dma_map_page(xprt->sc_cm_id->device, |
529 | page, 0, PAGE_SIZE, | |
377f9b2f | 530 | DMA_FROM_DEVICE); |
a5abf4e8 TT |
531 | if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa)) |
532 | goto err_put_ctxt; | |
533 | atomic_inc(&xprt->sc_dma_used); | |
377f9b2f TT |
534 | ctxt->sge[sge_no].addr = pa; |
535 | ctxt->sge[sge_no].length = PAGE_SIZE; | |
a5abf4e8 | 536 | ctxt->sge[sge_no].lkey = xprt->sc_dma_lkey; |
4a84386f | 537 | ctxt->count = sge_no + 1; |
377f9b2f TT |
538 | buflen += PAGE_SIZE; |
539 | } | |
377f9b2f TT |
540 | recv_wr.next = NULL; |
541 | recv_wr.sg_list = &ctxt->sge[0]; | |
542 | recv_wr.num_sge = ctxt->count; | |
543 | recv_wr.wr_id = (u64)(unsigned long)ctxt; | |
544 | ||
0905c0f0 | 545 | svc_xprt_get(&xprt->sc_xprt); |
377f9b2f | 546 | ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr); |
0905c0f0 | 547 | if (ret) { |
21515e46 | 548 | svc_rdma_unmap_dma(ctxt); |
05a0826a | 549 | svc_rdma_put_context(ctxt, 1); |
21515e46 | 550 | svc_xprt_put(&xprt->sc_xprt); |
0905c0f0 | 551 | } |
377f9b2f | 552 | return ret; |
a5abf4e8 TT |
553 | |
554 | err_put_ctxt: | |
4a84386f | 555 | svc_rdma_unmap_dma(ctxt); |
a5abf4e8 TT |
556 | svc_rdma_put_context(ctxt, 1); |
557 | return -ENOMEM; | |
377f9b2f TT |
558 | } |
559 | ||
560 | /* | |
561 | * This function handles the CONNECT_REQUEST event on a listening | |
562 | * endpoint. It is passed the cma_id for the _new_ connection. The context in | |
563 | * this cma_id is inherited from the listening cma_id and is the svc_xprt | |
564 | * structure for the listening endpoint. | |
565 | * | |
566 | * This function creates a new xprt for the new connection and enqueues it on | |
567 | * the accept queue for the listent xprt. When the listen thread is kicked, it | |
568 | * will call the recvfrom method on the listen xprt which will accept the new | |
569 | * connection. | |
570 | */ | |
36ef25e4 | 571 | static void handle_connect_req(struct rdma_cm_id *new_cma_id, size_t client_ird) |
377f9b2f TT |
572 | { |
573 | struct svcxprt_rdma *listen_xprt = new_cma_id->context; | |
574 | struct svcxprt_rdma *newxprt; | |
af261af4 | 575 | struct sockaddr *sa; |
377f9b2f TT |
576 | |
577 | /* Create a new transport */ | |
578 | newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0); | |
579 | if (!newxprt) { | |
580 | dprintk("svcrdma: failed to create new transport\n"); | |
581 | return; | |
582 | } | |
583 | newxprt->sc_cm_id = new_cma_id; | |
584 | new_cma_id->context = newxprt; | |
585 | dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n", | |
586 | newxprt, newxprt->sc_cm_id, listen_xprt); | |
587 | ||
36ef25e4 TT |
588 | /* Save client advertised inbound read limit for use later in accept. */ |
589 | newxprt->sc_ord = client_ird; | |
590 | ||
af261af4 TT |
591 | /* Set the local and remote addresses in the transport */ |
592 | sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr; | |
593 | svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa)); | |
594 | sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr; | |
595 | svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa)); | |
596 | ||
377f9b2f TT |
597 | /* |
598 | * Enqueue the new transport on the accept queue of the listening | |
599 | * transport | |
600 | */ | |
601 | spin_lock_bh(&listen_xprt->sc_lock); | |
602 | list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q); | |
603 | spin_unlock_bh(&listen_xprt->sc_lock); | |
604 | ||
377f9b2f TT |
605 | set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags); |
606 | svc_xprt_enqueue(&listen_xprt->sc_xprt); | |
607 | } | |
608 | ||
609 | /* | |
610 | * Handles events generated on the listening endpoint. These events will be | |
611 | * either be incoming connect requests or adapter removal events. | |
612 | */ | |
613 | static int rdma_listen_handler(struct rdma_cm_id *cma_id, | |
614 | struct rdma_cm_event *event) | |
615 | { | |
616 | struct svcxprt_rdma *xprt = cma_id->context; | |
617 | int ret = 0; | |
618 | ||
619 | switch (event->event) { | |
620 | case RDMA_CM_EVENT_CONNECT_REQUEST: | |
621 | dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, " | |
76357c71 SG |
622 | "event = %s (%d)\n", cma_id, cma_id->context, |
623 | rdma_event_msg(event->event), event->event); | |
36ef25e4 | 624 | handle_connect_req(cma_id, |
67080c82 | 625 | event->param.conn.initiator_depth); |
377f9b2f TT |
626 | break; |
627 | ||
628 | case RDMA_CM_EVENT_ESTABLISHED: | |
629 | /* Accept complete */ | |
630 | dprintk("svcrdma: Connection completed on LISTEN xprt=%p, " | |
631 | "cm_id=%p\n", xprt, cma_id); | |
632 | break; | |
633 | ||
634 | case RDMA_CM_EVENT_DEVICE_REMOVAL: | |
635 | dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n", | |
636 | xprt, cma_id); | |
637 | if (xprt) | |
638 | set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); | |
639 | break; | |
640 | ||
641 | default: | |
642 | dprintk("svcrdma: Unexpected event on listening endpoint %p, " | |
76357c71 SG |
643 | "event = %s (%d)\n", cma_id, |
644 | rdma_event_msg(event->event), event->event); | |
377f9b2f TT |
645 | break; |
646 | } | |
647 | ||
648 | return ret; | |
649 | } | |
650 | ||
651 | static int rdma_cma_handler(struct rdma_cm_id *cma_id, | |
652 | struct rdma_cm_event *event) | |
653 | { | |
654 | struct svc_xprt *xprt = cma_id->context; | |
655 | struct svcxprt_rdma *rdma = | |
656 | container_of(xprt, struct svcxprt_rdma, sc_xprt); | |
657 | switch (event->event) { | |
658 | case RDMA_CM_EVENT_ESTABLISHED: | |
659 | /* Accept complete */ | |
c48cbb40 | 660 | svc_xprt_get(xprt); |
377f9b2f TT |
661 | dprintk("svcrdma: Connection completed on DTO xprt=%p, " |
662 | "cm_id=%p\n", xprt, cma_id); | |
663 | clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags); | |
664 | svc_xprt_enqueue(xprt); | |
665 | break; | |
666 | case RDMA_CM_EVENT_DISCONNECTED: | |
667 | dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n", | |
668 | xprt, cma_id); | |
669 | if (xprt) { | |
670 | set_bit(XPT_CLOSE, &xprt->xpt_flags); | |
671 | svc_xprt_enqueue(xprt); | |
120693d1 | 672 | svc_xprt_put(xprt); |
377f9b2f TT |
673 | } |
674 | break; | |
675 | case RDMA_CM_EVENT_DEVICE_REMOVAL: | |
676 | dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, " | |
76357c71 SG |
677 | "event = %s (%d)\n", cma_id, xprt, |
678 | rdma_event_msg(event->event), event->event); | |
377f9b2f TT |
679 | if (xprt) { |
680 | set_bit(XPT_CLOSE, &xprt->xpt_flags); | |
681 | svc_xprt_enqueue(xprt); | |
682 | } | |
683 | break; | |
684 | default: | |
685 | dprintk("svcrdma: Unexpected event on DTO endpoint %p, " | |
76357c71 SG |
686 | "event = %s (%d)\n", cma_id, |
687 | rdma_event_msg(event->event), event->event); | |
377f9b2f TT |
688 | break; |
689 | } | |
690 | return 0; | |
691 | } | |
692 | ||
693 | /* | |
694 | * Create a listening RDMA service endpoint. | |
695 | */ | |
696 | static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, | |
62832c03 | 697 | struct net *net, |
377f9b2f TT |
698 | struct sockaddr *sa, int salen, |
699 | int flags) | |
700 | { | |
701 | struct rdma_cm_id *listen_id; | |
702 | struct svcxprt_rdma *cma_xprt; | |
377f9b2f TT |
703 | int ret; |
704 | ||
705 | dprintk("svcrdma: Creating RDMA socket\n"); | |
bade732a TT |
706 | if (sa->sa_family != AF_INET) { |
707 | dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family); | |
708 | return ERR_PTR(-EAFNOSUPPORT); | |
709 | } | |
377f9b2f TT |
710 | cma_xprt = rdma_create_xprt(serv, 1); |
711 | if (!cma_xprt) | |
58e8f621 | 712 | return ERR_PTR(-ENOMEM); |
377f9b2f | 713 | |
b26f9b99 SH |
714 | listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP, |
715 | IB_QPT_RC); | |
377f9b2f | 716 | if (IS_ERR(listen_id)) { |
58e8f621 TT |
717 | ret = PTR_ERR(listen_id); |
718 | dprintk("svcrdma: rdma_create_id failed = %d\n", ret); | |
719 | goto err0; | |
377f9b2f | 720 | } |
58e8f621 | 721 | |
377f9b2f TT |
722 | ret = rdma_bind_addr(listen_id, sa); |
723 | if (ret) { | |
377f9b2f | 724 | dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret); |
58e8f621 | 725 | goto err1; |
377f9b2f TT |
726 | } |
727 | cma_xprt->sc_cm_id = listen_id; | |
728 | ||
729 | ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG); | |
730 | if (ret) { | |
377f9b2f | 731 | dprintk("svcrdma: rdma_listen failed = %d\n", ret); |
58e8f621 | 732 | goto err1; |
377f9b2f TT |
733 | } |
734 | ||
735 | /* | |
736 | * We need to use the address from the cm_id in case the | |
737 | * caller specified 0 for the port number. | |
738 | */ | |
739 | sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr; | |
740 | svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen); | |
741 | ||
742 | return &cma_xprt->sc_xprt; | |
58e8f621 TT |
743 | |
744 | err1: | |
745 | rdma_destroy_id(listen_id); | |
746 | err0: | |
747 | kfree(cma_xprt); | |
748 | return ERR_PTR(ret); | |
377f9b2f TT |
749 | } |
750 | ||
64be8608 TT |
751 | static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt) |
752 | { | |
753 | struct ib_mr *mr; | |
754 | struct ib_fast_reg_page_list *pl; | |
755 | struct svc_rdma_fastreg_mr *frmr; | |
756 | ||
757 | frmr = kmalloc(sizeof(*frmr), GFP_KERNEL); | |
758 | if (!frmr) | |
759 | goto err; | |
760 | ||
761 | mr = ib_alloc_fast_reg_mr(xprt->sc_pd, RPCSVC_MAXPAGES); | |
846d8e7c | 762 | if (IS_ERR(mr)) |
64be8608 TT |
763 | goto err_free_frmr; |
764 | ||
765 | pl = ib_alloc_fast_reg_page_list(xprt->sc_cm_id->device, | |
766 | RPCSVC_MAXPAGES); | |
846d8e7c | 767 | if (IS_ERR(pl)) |
64be8608 TT |
768 | goto err_free_mr; |
769 | ||
770 | frmr->mr = mr; | |
771 | frmr->page_list = pl; | |
772 | INIT_LIST_HEAD(&frmr->frmr_list); | |
773 | return frmr; | |
774 | ||
775 | err_free_mr: | |
776 | ib_dereg_mr(mr); | |
777 | err_free_frmr: | |
778 | kfree(frmr); | |
779 | err: | |
780 | return ERR_PTR(-ENOMEM); | |
781 | } | |
782 | ||
783 | static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt) | |
784 | { | |
785 | struct svc_rdma_fastreg_mr *frmr; | |
786 | ||
787 | while (!list_empty(&xprt->sc_frmr_q)) { | |
788 | frmr = list_entry(xprt->sc_frmr_q.next, | |
789 | struct svc_rdma_fastreg_mr, frmr_list); | |
790 | list_del_init(&frmr->frmr_list); | |
791 | ib_dereg_mr(frmr->mr); | |
792 | ib_free_fast_reg_page_list(frmr->page_list); | |
793 | kfree(frmr); | |
794 | } | |
795 | } | |
796 | ||
797 | struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma) | |
798 | { | |
799 | struct svc_rdma_fastreg_mr *frmr = NULL; | |
800 | ||
801 | spin_lock_bh(&rdma->sc_frmr_q_lock); | |
802 | if (!list_empty(&rdma->sc_frmr_q)) { | |
803 | frmr = list_entry(rdma->sc_frmr_q.next, | |
804 | struct svc_rdma_fastreg_mr, frmr_list); | |
805 | list_del_init(&frmr->frmr_list); | |
806 | frmr->map_len = 0; | |
807 | frmr->page_list_len = 0; | |
808 | } | |
809 | spin_unlock_bh(&rdma->sc_frmr_q_lock); | |
810 | if (frmr) | |
811 | return frmr; | |
812 | ||
813 | return rdma_alloc_frmr(rdma); | |
814 | } | |
815 | ||
816 | static void frmr_unmap_dma(struct svcxprt_rdma *xprt, | |
817 | struct svc_rdma_fastreg_mr *frmr) | |
818 | { | |
819 | int page_no; | |
820 | for (page_no = 0; page_no < frmr->page_list_len; page_no++) { | |
821 | dma_addr_t addr = frmr->page_list->page_list[page_no]; | |
822 | if (ib_dma_mapping_error(frmr->mr->device, addr)) | |
823 | continue; | |
824 | atomic_dec(&xprt->sc_dma_used); | |
b432e6b3 TT |
825 | ib_dma_unmap_page(frmr->mr->device, addr, PAGE_SIZE, |
826 | frmr->direction); | |
64be8608 TT |
827 | } |
828 | } | |
829 | ||
830 | void svc_rdma_put_frmr(struct svcxprt_rdma *rdma, | |
831 | struct svc_rdma_fastreg_mr *frmr) | |
832 | { | |
833 | if (frmr) { | |
834 | frmr_unmap_dma(rdma, frmr); | |
835 | spin_lock_bh(&rdma->sc_frmr_q_lock); | |
3fe04ee9 | 836 | WARN_ON_ONCE(!list_empty(&frmr->frmr_list)); |
64be8608 TT |
837 | list_add(&frmr->frmr_list, &rdma->sc_frmr_q); |
838 | spin_unlock_bh(&rdma->sc_frmr_q_lock); | |
839 | } | |
840 | } | |
841 | ||
377f9b2f TT |
842 | /* |
843 | * This is the xpo_recvfrom function for listening endpoints. Its | |
844 | * purpose is to accept incoming connections. The CMA callback handler | |
845 | * has already created a new transport and attached it to the new CMA | |
846 | * ID. | |
847 | * | |
848 | * There is a queue of pending connections hung on the listening | |
849 | * transport. This queue contains the new svc_xprt structure. This | |
850 | * function takes svc_xprt structures off the accept_q and completes | |
851 | * the connection. | |
852 | */ | |
853 | static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) | |
854 | { | |
855 | struct svcxprt_rdma *listen_rdma; | |
856 | struct svcxprt_rdma *newxprt = NULL; | |
857 | struct rdma_conn_param conn_param; | |
8e37210b | 858 | struct ib_cq_init_attr cq_attr = {}; |
377f9b2f TT |
859 | struct ib_qp_init_attr qp_attr; |
860 | struct ib_device_attr devattr; | |
ed72b9c6 | 861 | int uninitialized_var(dma_mr_acc); |
3de2c31c | 862 | int need_dma_mr = 0; |
377f9b2f TT |
863 | int ret; |
864 | int i; | |
865 | ||
866 | listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt); | |
867 | clear_bit(XPT_CONN, &xprt->xpt_flags); | |
868 | /* Get the next entry off the accept list */ | |
869 | spin_lock_bh(&listen_rdma->sc_lock); | |
870 | if (!list_empty(&listen_rdma->sc_accept_q)) { | |
871 | newxprt = list_entry(listen_rdma->sc_accept_q.next, | |
872 | struct svcxprt_rdma, sc_accept_q); | |
873 | list_del_init(&newxprt->sc_accept_q); | |
874 | } | |
875 | if (!list_empty(&listen_rdma->sc_accept_q)) | |
876 | set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags); | |
877 | spin_unlock_bh(&listen_rdma->sc_lock); | |
878 | if (!newxprt) | |
879 | return NULL; | |
880 | ||
881 | dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n", | |
882 | newxprt, newxprt->sc_cm_id); | |
883 | ||
884 | ret = ib_query_device(newxprt->sc_cm_id->device, &devattr); | |
885 | if (ret) { | |
886 | dprintk("svcrdma: could not query device attributes on " | |
887 | "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret); | |
888 | goto errout; | |
889 | } | |
890 | ||
891 | /* Qualify the transport resource defaults with the | |
892 | * capabilities of this particular device */ | |
893 | newxprt->sc_max_sge = min((size_t)devattr.max_sge, | |
894 | (size_t)RPCSVC_MAXPAGES); | |
895 | newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr, | |
896 | (size_t)svcrdma_max_requests); | |
897 | newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests; | |
898 | ||
36ef25e4 TT |
899 | /* |
900 | * Limit ORD based on client limit, local device limit, and | |
901 | * configured svcrdma limit. | |
902 | */ | |
903 | newxprt->sc_ord = min_t(size_t, devattr.max_qp_rd_atom, newxprt->sc_ord); | |
904 | newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord); | |
377f9b2f TT |
905 | |
906 | newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device); | |
907 | if (IS_ERR(newxprt->sc_pd)) { | |
908 | dprintk("svcrdma: error creating PD for connect request\n"); | |
909 | goto errout; | |
910 | } | |
8e37210b | 911 | cq_attr.cqe = newxprt->sc_sq_depth; |
377f9b2f TT |
912 | newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device, |
913 | sq_comp_handler, | |
914 | cq_event_handler, | |
915 | newxprt, | |
8e37210b | 916 | &cq_attr); |
377f9b2f TT |
917 | if (IS_ERR(newxprt->sc_sq_cq)) { |
918 | dprintk("svcrdma: error creating SQ CQ for connect request\n"); | |
919 | goto errout; | |
920 | } | |
8e37210b | 921 | cq_attr.cqe = newxprt->sc_max_requests; |
377f9b2f TT |
922 | newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device, |
923 | rq_comp_handler, | |
924 | cq_event_handler, | |
925 | newxprt, | |
8e37210b | 926 | &cq_attr); |
377f9b2f TT |
927 | if (IS_ERR(newxprt->sc_rq_cq)) { |
928 | dprintk("svcrdma: error creating RQ CQ for connect request\n"); | |
929 | goto errout; | |
930 | } | |
931 | ||
932 | memset(&qp_attr, 0, sizeof qp_attr); | |
933 | qp_attr.event_handler = qp_event_handler; | |
934 | qp_attr.qp_context = &newxprt->sc_xprt; | |
935 | qp_attr.cap.max_send_wr = newxprt->sc_sq_depth; | |
936 | qp_attr.cap.max_recv_wr = newxprt->sc_max_requests; | |
937 | qp_attr.cap.max_send_sge = newxprt->sc_max_sge; | |
938 | qp_attr.cap.max_recv_sge = newxprt->sc_max_sge; | |
939 | qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR; | |
940 | qp_attr.qp_type = IB_QPT_RC; | |
941 | qp_attr.send_cq = newxprt->sc_sq_cq; | |
942 | qp_attr.recv_cq = newxprt->sc_rq_cq; | |
943 | dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n" | |
944 | " cm_id->device=%p, sc_pd->device=%p\n" | |
945 | " cap.max_send_wr = %d\n" | |
946 | " cap.max_recv_wr = %d\n" | |
947 | " cap.max_send_sge = %d\n" | |
948 | " cap.max_recv_sge = %d\n", | |
949 | newxprt->sc_cm_id, newxprt->sc_pd, | |
950 | newxprt->sc_cm_id->device, newxprt->sc_pd->device, | |
951 | qp_attr.cap.max_send_wr, | |
952 | qp_attr.cap.max_recv_wr, | |
953 | qp_attr.cap.max_send_sge, | |
954 | qp_attr.cap.max_recv_sge); | |
955 | ||
956 | ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr); | |
957 | if (ret) { | |
d1e458fe SW |
958 | dprintk("svcrdma: failed to create QP, ret=%d\n", ret); |
959 | goto errout; | |
377f9b2f TT |
960 | } |
961 | newxprt->sc_qp = newxprt->sc_cm_id->qp; | |
962 | ||
3a5c6380 TT |
963 | /* |
964 | * Use the most secure set of MR resources based on the | |
965 | * transport type and available memory management features in | |
966 | * the device. Here's the table implemented below: | |
967 | * | |
968 | * Fast Global DMA Remote WR | |
969 | * Reg LKEY MR Access | |
970 | * Sup'd Sup'd Needed Needed | |
971 | * | |
972 | * IWARP N N Y Y | |
973 | * N Y Y Y | |
974 | * Y N Y N | |
975 | * Y Y N - | |
976 | * | |
977 | * IB N N Y N | |
978 | * N Y N - | |
979 | * Y N Y N | |
980 | * Y Y N - | |
981 | * | |
982 | * NB: iWARP requires remote write access for the data sink | |
983 | * of an RDMA_READ. IB does not. | |
984 | */ | |
e5452411 | 985 | newxprt->sc_reader = rdma_read_chunk_lcl; |
3a5c6380 TT |
986 | if (devattr.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) { |
987 | newxprt->sc_frmr_pg_list_len = | |
988 | devattr.max_fast_reg_page_list_len; | |
989 | newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG; | |
e5452411 | 990 | newxprt->sc_reader = rdma_read_chunk_frmr; |
3a5c6380 TT |
991 | } |
992 | ||
993 | /* | |
994 | * Determine if a DMA MR is required and if so, what privs are required | |
995 | */ | |
3de2c31c MW |
996 | if (!rdma_protocol_iwarp(newxprt->sc_cm_id->device, |
997 | newxprt->sc_cm_id->port_num) && | |
5d9fb044 | 998 | !rdma_ib_or_roce(newxprt->sc_cm_id->device, |
3de2c31c | 999 | newxprt->sc_cm_id->port_num)) |
377f9b2f | 1000 | goto errout; |
3de2c31c MW |
1001 | |
1002 | if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG) || | |
1003 | !(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) { | |
1004 | need_dma_mr = 1; | |
1005 | dma_mr_acc = IB_ACCESS_LOCAL_WRITE; | |
1006 | if (rdma_protocol_iwarp(newxprt->sc_cm_id->device, | |
1007 | newxprt->sc_cm_id->port_num) && | |
1008 | !(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) | |
1009 | dma_mr_acc |= IB_ACCESS_REMOTE_WRITE; | |
377f9b2f TT |
1010 | } |
1011 | ||
3de2c31c MW |
1012 | if (rdma_protocol_iwarp(newxprt->sc_cm_id->device, |
1013 | newxprt->sc_cm_id->port_num)) | |
1014 | newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV; | |
1015 | ||
3a5c6380 TT |
1016 | /* Create the DMA MR if needed, otherwise, use the DMA LKEY */ |
1017 | if (need_dma_mr) { | |
1018 | /* Register all of physical memory */ | |
1019 | newxprt->sc_phys_mr = | |
1020 | ib_get_dma_mr(newxprt->sc_pd, dma_mr_acc); | |
1021 | if (IS_ERR(newxprt->sc_phys_mr)) { | |
1022 | dprintk("svcrdma: Failed to create DMA MR ret=%d\n", | |
1023 | ret); | |
1024 | goto errout; | |
1025 | } | |
1026 | newxprt->sc_dma_lkey = newxprt->sc_phys_mr->lkey; | |
1027 | } else | |
1028 | newxprt->sc_dma_lkey = | |
1029 | newxprt->sc_cm_id->device->local_dma_lkey; | |
1030 | ||
377f9b2f TT |
1031 | /* Post receive buffers */ |
1032 | for (i = 0; i < newxprt->sc_max_requests; i++) { | |
1033 | ret = svc_rdma_post_recv(newxprt); | |
1034 | if (ret) { | |
1035 | dprintk("svcrdma: failure posting receive buffers\n"); | |
1036 | goto errout; | |
1037 | } | |
1038 | } | |
1039 | ||
1040 | /* Swap out the handler */ | |
1041 | newxprt->sc_cm_id->event_handler = rdma_cma_handler; | |
1042 | ||
af261af4 TT |
1043 | /* |
1044 | * Arm the CQs for the SQ and RQ before accepting so we can't | |
1045 | * miss the first message | |
1046 | */ | |
1047 | ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP); | |
1048 | ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP); | |
1049 | ||
377f9b2f TT |
1050 | /* Accept Connection */ |
1051 | set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags); | |
1052 | memset(&conn_param, 0, sizeof conn_param); | |
1053 | conn_param.responder_resources = 0; | |
1054 | conn_param.initiator_depth = newxprt->sc_ord; | |
1055 | ret = rdma_accept(newxprt->sc_cm_id, &conn_param); | |
1056 | if (ret) { | |
1057 | dprintk("svcrdma: failed to accept new connection, ret=%d\n", | |
1058 | ret); | |
1059 | goto errout; | |
1060 | } | |
1061 | ||
1062 | dprintk("svcrdma: new connection %p accepted with the following " | |
1063 | "attributes:\n" | |
21454aaa | 1064 | " local_ip : %pI4\n" |
377f9b2f | 1065 | " local_port : %d\n" |
21454aaa | 1066 | " remote_ip : %pI4\n" |
377f9b2f TT |
1067 | " remote_port : %d\n" |
1068 | " max_sge : %d\n" | |
1069 | " sq_depth : %d\n" | |
1070 | " max_requests : %d\n" | |
1071 | " ord : %d\n", | |
1072 | newxprt, | |
21454aaa HH |
1073 | &((struct sockaddr_in *)&newxprt->sc_cm_id-> |
1074 | route.addr.src_addr)->sin_addr.s_addr, | |
377f9b2f TT |
1075 | ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> |
1076 | route.addr.src_addr)->sin_port), | |
21454aaa HH |
1077 | &((struct sockaddr_in *)&newxprt->sc_cm_id-> |
1078 | route.addr.dst_addr)->sin_addr.s_addr, | |
377f9b2f TT |
1079 | ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> |
1080 | route.addr.dst_addr)->sin_port), | |
1081 | newxprt->sc_max_sge, | |
1082 | newxprt->sc_sq_depth, | |
1083 | newxprt->sc_max_requests, | |
1084 | newxprt->sc_ord); | |
1085 | ||
377f9b2f TT |
1086 | return &newxprt->sc_xprt; |
1087 | ||
1088 | errout: | |
1089 | dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret); | |
c48cbb40 TT |
1090 | /* Take a reference in case the DTO handler runs */ |
1091 | svc_xprt_get(&newxprt->sc_xprt); | |
1711386c | 1092 | if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) |
c48cbb40 | 1093 | ib_destroy_qp(newxprt->sc_qp); |
377f9b2f | 1094 | rdma_destroy_id(newxprt->sc_cm_id); |
c48cbb40 TT |
1095 | /* This call to put will destroy the transport */ |
1096 | svc_xprt_put(&newxprt->sc_xprt); | |
377f9b2f TT |
1097 | return NULL; |
1098 | } | |
1099 | ||
377f9b2f TT |
1100 | static void svc_rdma_release_rqst(struct svc_rqst *rqstp) |
1101 | { | |
377f9b2f TT |
1102 | } |
1103 | ||
c48cbb40 | 1104 | /* |
1711386c | 1105 | * When connected, an svc_xprt has at least two references: |
c48cbb40 TT |
1106 | * |
1107 | * - A reference held by the cm_id between the ESTABLISHED and | |
1108 | * DISCONNECTED events. If the remote peer disconnected first, this | |
1109 | * reference could be gone. | |
1110 | * | |
1111 | * - A reference held by the svc_recv code that called this function | |
1112 | * as part of close processing. | |
1113 | * | |
1711386c | 1114 | * At a minimum one references should still be held. |
c48cbb40 | 1115 | */ |
377f9b2f TT |
1116 | static void svc_rdma_detach(struct svc_xprt *xprt) |
1117 | { | |
1118 | struct svcxprt_rdma *rdma = | |
1119 | container_of(xprt, struct svcxprt_rdma, sc_xprt); | |
377f9b2f | 1120 | dprintk("svc: svc_rdma_detach(%p)\n", xprt); |
c48cbb40 TT |
1121 | |
1122 | /* Disconnect and flush posted WQE */ | |
377f9b2f | 1123 | rdma_disconnect(rdma->sc_cm_id); |
377f9b2f TT |
1124 | } |
1125 | ||
8da91ea8 | 1126 | static void __svc_rdma_free(struct work_struct *work) |
377f9b2f | 1127 | { |
8da91ea8 TT |
1128 | struct svcxprt_rdma *rdma = |
1129 | container_of(work, struct svcxprt_rdma, sc_work); | |
377f9b2f | 1130 | dprintk("svcrdma: svc_rdma_free(%p)\n", rdma); |
8da91ea8 | 1131 | |
c48cbb40 | 1132 | /* We should only be called from kref_put */ |
3fe04ee9 CL |
1133 | if (atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0) |
1134 | pr_err("svcrdma: sc_xprt still in use? (%d)\n", | |
1135 | atomic_read(&rdma->sc_xprt.xpt_ref.refcount)); | |
8da91ea8 | 1136 | |
356d0a15 TT |
1137 | /* |
1138 | * Destroy queued, but not processed read completions. Note | |
1139 | * that this cleanup has to be done before destroying the | |
1140 | * cm_id because the device ptr is needed to unmap the dma in | |
1141 | * svc_rdma_put_context. | |
1142 | */ | |
356d0a15 TT |
1143 | while (!list_empty(&rdma->sc_read_complete_q)) { |
1144 | struct svc_rdma_op_ctxt *ctxt; | |
1145 | ctxt = list_entry(rdma->sc_read_complete_q.next, | |
1146 | struct svc_rdma_op_ctxt, | |
1147 | dto_q); | |
1148 | list_del_init(&ctxt->dto_q); | |
1149 | svc_rdma_put_context(ctxt, 1); | |
1150 | } | |
356d0a15 TT |
1151 | |
1152 | /* Destroy queued, but not processed recv completions */ | |
356d0a15 TT |
1153 | while (!list_empty(&rdma->sc_rq_dto_q)) { |
1154 | struct svc_rdma_op_ctxt *ctxt; | |
1155 | ctxt = list_entry(rdma->sc_rq_dto_q.next, | |
1156 | struct svc_rdma_op_ctxt, | |
1157 | dto_q); | |
1158 | list_del_init(&ctxt->dto_q); | |
1159 | svc_rdma_put_context(ctxt, 1); | |
1160 | } | |
356d0a15 TT |
1161 | |
1162 | /* Warn if we leaked a resource or under-referenced */ | |
3fe04ee9 CL |
1163 | if (atomic_read(&rdma->sc_ctxt_used) != 0) |
1164 | pr_err("svcrdma: ctxt still in use? (%d)\n", | |
1165 | atomic_read(&rdma->sc_ctxt_used)); | |
1166 | if (atomic_read(&rdma->sc_dma_used) != 0) | |
1167 | pr_err("svcrdma: dma still in use? (%d)\n", | |
1168 | atomic_read(&rdma->sc_dma_used)); | |
356d0a15 | 1169 | |
64be8608 TT |
1170 | /* De-allocate fastreg mr */ |
1171 | rdma_dealloc_frmr_q(rdma); | |
1172 | ||
1711386c TT |
1173 | /* Destroy the QP if present (not a listener) */ |
1174 | if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) | |
1175 | ib_destroy_qp(rdma->sc_qp); | |
1176 | ||
c48cbb40 TT |
1177 | if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq)) |
1178 | ib_destroy_cq(rdma->sc_sq_cq); | |
377f9b2f | 1179 | |
c48cbb40 TT |
1180 | if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq)) |
1181 | ib_destroy_cq(rdma->sc_rq_cq); | |
377f9b2f | 1182 | |
c48cbb40 TT |
1183 | if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr)) |
1184 | ib_dereg_mr(rdma->sc_phys_mr); | |
377f9b2f | 1185 | |
c48cbb40 TT |
1186 | if (rdma->sc_pd && !IS_ERR(rdma->sc_pd)) |
1187 | ib_dealloc_pd(rdma->sc_pd); | |
377f9b2f | 1188 | |
356d0a15 TT |
1189 | /* Destroy the CM ID */ |
1190 | rdma_destroy_id(rdma->sc_cm_id); | |
1191 | ||
c48cbb40 | 1192 | kfree(rdma); |
377f9b2f TT |
1193 | } |
1194 | ||
8da91ea8 TT |
1195 | static void svc_rdma_free(struct svc_xprt *xprt) |
1196 | { | |
1197 | struct svcxprt_rdma *rdma = | |
1198 | container_of(xprt, struct svcxprt_rdma, sc_xprt); | |
1199 | INIT_WORK(&rdma->sc_work, __svc_rdma_free); | |
a25e758c | 1200 | queue_work(svc_rdma_wq, &rdma->sc_work); |
8da91ea8 TT |
1201 | } |
1202 | ||
377f9b2f TT |
1203 | static int svc_rdma_has_wspace(struct svc_xprt *xprt) |
1204 | { | |
1205 | struct svcxprt_rdma *rdma = | |
1206 | container_of(xprt, struct svcxprt_rdma, sc_xprt); | |
1207 | ||
1208 | /* | |
0bf48289 | 1209 | * If there are already waiters on the SQ, |
377f9b2f TT |
1210 | * return false. |
1211 | */ | |
1212 | if (waitqueue_active(&rdma->sc_send_wait)) | |
1213 | return 0; | |
1214 | ||
1215 | /* Otherwise return true. */ | |
1216 | return 1; | |
1217 | } | |
1218 | ||
16e4d93f CL |
1219 | static int svc_rdma_secure_port(struct svc_rqst *rqstp) |
1220 | { | |
1221 | return 1; | |
1222 | } | |
1223 | ||
e1183210 TT |
1224 | /* |
1225 | * Attempt to register the kvec representing the RPC memory with the | |
1226 | * device. | |
1227 | * | |
1228 | * Returns: | |
1229 | * NULL : The device does not support fastreg or there were no more | |
1230 | * fastreg mr. | |
1231 | * frmr : The kvec register request was successfully posted. | |
1232 | * <0 : An error was encountered attempting to register the kvec. | |
1233 | */ | |
1234 | int svc_rdma_fastreg(struct svcxprt_rdma *xprt, | |
1235 | struct svc_rdma_fastreg_mr *frmr) | |
1236 | { | |
1237 | struct ib_send_wr fastreg_wr; | |
1238 | u8 key; | |
1239 | ||
1240 | /* Bump the key */ | |
1241 | key = (u8)(frmr->mr->lkey & 0x000000FF); | |
1242 | ib_update_fast_reg_key(frmr->mr, ++key); | |
1243 | ||
1244 | /* Prepare FASTREG WR */ | |
1245 | memset(&fastreg_wr, 0, sizeof fastreg_wr); | |
1246 | fastreg_wr.opcode = IB_WR_FAST_REG_MR; | |
1247 | fastreg_wr.send_flags = IB_SEND_SIGNALED; | |
1248 | fastreg_wr.wr.fast_reg.iova_start = (unsigned long)frmr->kva; | |
1249 | fastreg_wr.wr.fast_reg.page_list = frmr->page_list; | |
1250 | fastreg_wr.wr.fast_reg.page_list_len = frmr->page_list_len; | |
1251 | fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT; | |
1252 | fastreg_wr.wr.fast_reg.length = frmr->map_len; | |
1253 | fastreg_wr.wr.fast_reg.access_flags = frmr->access_flags; | |
1254 | fastreg_wr.wr.fast_reg.rkey = frmr->mr->lkey; | |
1255 | return svc_rdma_send(xprt, &fastreg_wr); | |
1256 | } | |
1257 | ||
377f9b2f TT |
1258 | int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr) |
1259 | { | |
5b180a9a TT |
1260 | struct ib_send_wr *bad_wr, *n_wr; |
1261 | int wr_count; | |
1262 | int i; | |
377f9b2f TT |
1263 | int ret; |
1264 | ||
1265 | if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) | |
9d6347ac | 1266 | return -ENOTCONN; |
377f9b2f | 1267 | |
5b180a9a TT |
1268 | wr_count = 1; |
1269 | for (n_wr = wr->next; n_wr; n_wr = n_wr->next) | |
1270 | wr_count++; | |
1271 | ||
377f9b2f TT |
1272 | /* If the SQ is full, wait until an SQ entry is available */ |
1273 | while (1) { | |
1274 | spin_lock_bh(&xprt->sc_lock); | |
5b180a9a | 1275 | if (xprt->sc_sq_depth < atomic_read(&xprt->sc_sq_count) + wr_count) { |
377f9b2f TT |
1276 | spin_unlock_bh(&xprt->sc_lock); |
1277 | atomic_inc(&rdma_stat_sq_starve); | |
dbcd00eb TT |
1278 | |
1279 | /* See if we can opportunistically reap SQ WR to make room */ | |
377f9b2f TT |
1280 | sq_cq_reap(xprt); |
1281 | ||
1282 | /* Wait until SQ WR available if SQ still full */ | |
1283 | wait_event(xprt->sc_send_wait, | |
1284 | atomic_read(&xprt->sc_sq_count) < | |
1285 | xprt->sc_sq_depth); | |
830bb59b | 1286 | if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) |
b432e6b3 | 1287 | return -ENOTCONN; |
377f9b2f TT |
1288 | continue; |
1289 | } | |
5b180a9a TT |
1290 | /* Take a transport ref for each WR posted */ |
1291 | for (i = 0; i < wr_count; i++) | |
1292 | svc_xprt_get(&xprt->sc_xprt); | |
1293 | ||
1294 | /* Bump used SQ WR count and post */ | |
1295 | atomic_add(wr_count, &xprt->sc_sq_count); | |
377f9b2f | 1296 | ret = ib_post_send(xprt->sc_qp, wr, &bad_wr); |
5b180a9a TT |
1297 | if (ret) { |
1298 | set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); | |
1299 | atomic_sub(wr_count, &xprt->sc_sq_count); | |
1300 | for (i = 0; i < wr_count; i ++) | |
1301 | svc_xprt_put(&xprt->sc_xprt); | |
377f9b2f TT |
1302 | dprintk("svcrdma: failed to post SQ WR rc=%d, " |
1303 | "sc_sq_count=%d, sc_sq_depth=%d\n", | |
1304 | ret, atomic_read(&xprt->sc_sq_count), | |
1305 | xprt->sc_sq_depth); | |
0905c0f0 | 1306 | } |
377f9b2f | 1307 | spin_unlock_bh(&xprt->sc_lock); |
5b180a9a TT |
1308 | if (ret) |
1309 | wake_up(&xprt->sc_send_wait); | |
377f9b2f TT |
1310 | break; |
1311 | } | |
1312 | return ret; | |
1313 | } | |
1314 | ||
008fdbc5 TT |
1315 | void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp, |
1316 | enum rpcrdma_errcode err) | |
377f9b2f TT |
1317 | { |
1318 | struct ib_send_wr err_wr; | |
377f9b2f TT |
1319 | struct page *p; |
1320 | struct svc_rdma_op_ctxt *ctxt; | |
1321 | u32 *va; | |
1322 | int length; | |
1323 | int ret; | |
1324 | ||
1325 | p = svc_rdma_get_page(); | |
1326 | va = page_address(p); | |
1327 | ||
1328 | /* XDR encode error */ | |
1329 | length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va); | |
1330 | ||
4a84386f TT |
1331 | ctxt = svc_rdma_get_context(xprt); |
1332 | ctxt->direction = DMA_FROM_DEVICE; | |
1333 | ctxt->count = 1; | |
1334 | ctxt->pages[0] = p; | |
1335 | ||
377f9b2f | 1336 | /* Prepare SGE for local address */ |
4a84386f TT |
1337 | ctxt->sge[0].addr = ib_dma_map_page(xprt->sc_cm_id->device, |
1338 | p, 0, length, DMA_FROM_DEVICE); | |
1339 | if (ib_dma_mapping_error(xprt->sc_cm_id->device, ctxt->sge[0].addr)) { | |
04911b53 | 1340 | put_page(p); |
a5e50268 | 1341 | svc_rdma_put_context(ctxt, 1); |
04911b53 TT |
1342 | return; |
1343 | } | |
1344 | atomic_inc(&xprt->sc_dma_used); | |
4a84386f TT |
1345 | ctxt->sge[0].lkey = xprt->sc_dma_lkey; |
1346 | ctxt->sge[0].length = length; | |
377f9b2f TT |
1347 | |
1348 | /* Prepare SEND WR */ | |
1349 | memset(&err_wr, 0, sizeof err_wr); | |
1350 | ctxt->wr_op = IB_WR_SEND; | |
1351 | err_wr.wr_id = (unsigned long)ctxt; | |
4a84386f | 1352 | err_wr.sg_list = ctxt->sge; |
377f9b2f TT |
1353 | err_wr.num_sge = 1; |
1354 | err_wr.opcode = IB_WR_SEND; | |
1355 | err_wr.send_flags = IB_SEND_SIGNALED; | |
1356 | ||
1357 | /* Post It */ | |
1358 | ret = svc_rdma_send(xprt, &err_wr); | |
1359 | if (ret) { | |
008fdbc5 TT |
1360 | dprintk("svcrdma: Error %d posting send for protocol error\n", |
1361 | ret); | |
4a84386f | 1362 | svc_rdma_unmap_dma(ctxt); |
377f9b2f TT |
1363 | svc_rdma_put_context(ctxt, 1); |
1364 | } | |
377f9b2f | 1365 | } |