]>
Commit | Line | Data |
---|---|---|
377f9b2f TT |
1 | /* |
2 | * Copyright (c) 2005-2007 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. | |
38 | * | |
39 | * Author: Tom Tucker <tom@opengridcomputing.com> | |
40 | */ | |
41 | ||
42 | #include <linux/sunrpc/svc_xprt.h> | |
43 | #include <linux/sunrpc/debug.h> | |
44 | #include <linux/sunrpc/rpc_rdma.h> | |
a6b7a407 | 45 | #include <linux/interrupt.h> |
d43c36dc | 46 | #include <linux/sched.h> |
5a0e3ad6 | 47 | #include <linux/slab.h> |
377f9b2f | 48 | #include <linux/spinlock.h> |
a25e758c | 49 | #include <linux/workqueue.h> |
377f9b2f TT |
50 | #include <rdma/ib_verbs.h> |
51 | #include <rdma/rdma_cm.h> | |
52 | #include <linux/sunrpc/svc_rdma.h> | |
53 | ||
54 | #define RPCDBG_FACILITY RPCDBG_SVCXPRT | |
55 | ||
56 | static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, | |
62832c03 | 57 | struct net *net, |
377f9b2f TT |
58 | struct sockaddr *sa, int salen, |
59 | int flags); | |
60 | static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt); | |
61 | static void svc_rdma_release_rqst(struct svc_rqst *); | |
377f9b2f TT |
62 | static void dto_tasklet_func(unsigned long data); |
63 | static void svc_rdma_detach(struct svc_xprt *xprt); | |
64 | static void svc_rdma_free(struct svc_xprt *xprt); | |
65 | static int svc_rdma_has_wspace(struct svc_xprt *xprt); | |
66 | static void rq_cq_reap(struct svcxprt_rdma *xprt); | |
67 | static void sq_cq_reap(struct svcxprt_rdma *xprt); | |
68 | ||
5eaa65b2 | 69 | static DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL); |
377f9b2f TT |
70 | static DEFINE_SPINLOCK(dto_lock); |
71 | static LIST_HEAD(dto_xprt_q); | |
72 | ||
73 | static struct svc_xprt_ops svc_rdma_ops = { | |
74 | .xpo_create = svc_rdma_create, | |
75 | .xpo_recvfrom = svc_rdma_recvfrom, | |
76 | .xpo_sendto = svc_rdma_sendto, | |
77 | .xpo_release_rqst = svc_rdma_release_rqst, | |
78 | .xpo_detach = svc_rdma_detach, | |
79 | .xpo_free = svc_rdma_free, | |
80 | .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr, | |
81 | .xpo_has_wspace = svc_rdma_has_wspace, | |
82 | .xpo_accept = svc_rdma_accept, | |
83 | }; | |
84 | ||
85 | struct svc_xprt_class svc_rdma_class = { | |
86 | .xcl_name = "rdma", | |
87 | .xcl_owner = THIS_MODULE, | |
88 | .xcl_ops = &svc_rdma_ops, | |
89 | .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP, | |
90 | }; | |
91 | ||
8948896c TT |
92 | /* WR context cache. Created in svc_rdma.c */ |
93 | extern struct kmem_cache *svc_rdma_ctxt_cachep; | |
377f9b2f | 94 | |
a25e758c TH |
95 | /* Workqueue created in svc_rdma.c */ |
96 | extern struct workqueue_struct *svc_rdma_wq; | |
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 | ||
142 | BUG_ON(!ctxt); | |
143 | xprt = ctxt->xprt; | |
144 | if (free_pages) | |
145 | for (i = 0; i < ctxt->count; i++) | |
146 | put_page(ctxt->pages[i]); | |
147 | ||
8948896c | 148 | kmem_cache_free(svc_rdma_ctxt_cachep, ctxt); |
87407673 | 149 | atomic_dec(&xprt->sc_ctxt_used); |
377f9b2f TT |
150 | } |
151 | ||
ab96dddb TT |
152 | /* Temporary NFS request map cache. Created in svc_rdma.c */ |
153 | extern struct kmem_cache *svc_rdma_map_cachep; | |
154 | ||
155 | /* | |
156 | * Temporary NFS req mappings are shared across all transport | |
157 | * instances. These are short lived and should be bounded by the number | |
158 | * of concurrent server threads * depth of the SQ. | |
159 | */ | |
160 | struct svc_rdma_req_map *svc_rdma_get_req_map(void) | |
161 | { | |
162 | struct svc_rdma_req_map *map; | |
163 | while (1) { | |
164 | map = kmem_cache_alloc(svc_rdma_map_cachep, GFP_KERNEL); | |
165 | if (map) | |
166 | break; | |
167 | schedule_timeout_uninterruptible(msecs_to_jiffies(500)); | |
168 | } | |
169 | map->count = 0; | |
64be8608 | 170 | map->frmr = NULL; |
ab96dddb TT |
171 | return map; |
172 | } | |
173 | ||
174 | void svc_rdma_put_req_map(struct svc_rdma_req_map *map) | |
175 | { | |
176 | kmem_cache_free(svc_rdma_map_cachep, map); | |
177 | } | |
178 | ||
377f9b2f TT |
179 | /* ib_cq event handler */ |
180 | static void cq_event_handler(struct ib_event *event, void *context) | |
181 | { | |
182 | struct svc_xprt *xprt = context; | |
183 | dprintk("svcrdma: received CQ event id=%d, context=%p\n", | |
184 | event->event, context); | |
185 | set_bit(XPT_CLOSE, &xprt->xpt_flags); | |
186 | } | |
187 | ||
188 | /* QP event handler */ | |
189 | static void qp_event_handler(struct ib_event *event, void *context) | |
190 | { | |
191 | struct svc_xprt *xprt = context; | |
192 | ||
193 | switch (event->event) { | |
194 | /* These are considered benign events */ | |
195 | case IB_EVENT_PATH_MIG: | |
196 | case IB_EVENT_COMM_EST: | |
197 | case IB_EVENT_SQ_DRAINED: | |
198 | case IB_EVENT_QP_LAST_WQE_REACHED: | |
199 | dprintk("svcrdma: QP event %d received for QP=%p\n", | |
200 | event->event, event->element.qp); | |
201 | break; | |
202 | /* These are considered fatal events */ | |
203 | case IB_EVENT_PATH_MIG_ERR: | |
204 | case IB_EVENT_QP_FATAL: | |
205 | case IB_EVENT_QP_REQ_ERR: | |
206 | case IB_EVENT_QP_ACCESS_ERR: | |
207 | case IB_EVENT_DEVICE_FATAL: | |
208 | default: | |
209 | dprintk("svcrdma: QP ERROR event %d received for QP=%p, " | |
210 | "closing transport\n", | |
211 | event->event, event->element.qp); | |
212 | set_bit(XPT_CLOSE, &xprt->xpt_flags); | |
213 | break; | |
214 | } | |
215 | } | |
216 | ||
217 | /* | |
218 | * Data Transfer Operation Tasklet | |
219 | * | |
220 | * Walks a list of transports with I/O pending, removing entries as | |
221 | * they are added to the server's I/O pending list. Two bits indicate | |
222 | * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave | |
223 | * spinlock that serializes access to the transport list with the RQ | |
224 | * and SQ interrupt handlers. | |
225 | */ | |
226 | static void dto_tasklet_func(unsigned long data) | |
227 | { | |
228 | struct svcxprt_rdma *xprt; | |
229 | unsigned long flags; | |
230 | ||
231 | spin_lock_irqsave(&dto_lock, flags); | |
232 | while (!list_empty(&dto_xprt_q)) { | |
233 | xprt = list_entry(dto_xprt_q.next, | |
234 | struct svcxprt_rdma, sc_dto_q); | |
235 | list_del_init(&xprt->sc_dto_q); | |
236 | spin_unlock_irqrestore(&dto_lock, flags); | |
237 | ||
dbcd00eb TT |
238 | rq_cq_reap(xprt); |
239 | sq_cq_reap(xprt); | |
377f9b2f | 240 | |
c48cbb40 | 241 | svc_xprt_put(&xprt->sc_xprt); |
377f9b2f TT |
242 | spin_lock_irqsave(&dto_lock, flags); |
243 | } | |
244 | spin_unlock_irqrestore(&dto_lock, flags); | |
245 | } | |
246 | ||
247 | /* | |
248 | * Receive Queue Completion Handler | |
249 | * | |
250 | * Since an RQ completion handler is called on interrupt context, we | |
251 | * need to defer the handling of the I/O to a tasklet | |
252 | */ | |
253 | static void rq_comp_handler(struct ib_cq *cq, void *cq_context) | |
254 | { | |
255 | struct svcxprt_rdma *xprt = cq_context; | |
256 | unsigned long flags; | |
257 | ||
1711386c TT |
258 | /* Guard against unconditional flush call for destroyed QP */ |
259 | if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0) | |
260 | return; | |
261 | ||
377f9b2f TT |
262 | /* |
263 | * Set the bit regardless of whether or not it's on the list | |
264 | * because it may be on the list already due to an SQ | |
265 | * completion. | |
1711386c | 266 | */ |
377f9b2f TT |
267 | set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags); |
268 | ||
269 | /* | |
270 | * If this transport is not already on the DTO transport queue, | |
271 | * add it | |
272 | */ | |
273 | spin_lock_irqsave(&dto_lock, flags); | |
c48cbb40 TT |
274 | if (list_empty(&xprt->sc_dto_q)) { |
275 | svc_xprt_get(&xprt->sc_xprt); | |
377f9b2f | 276 | list_add_tail(&xprt->sc_dto_q, &dto_xprt_q); |
c48cbb40 | 277 | } |
377f9b2f TT |
278 | spin_unlock_irqrestore(&dto_lock, flags); |
279 | ||
280 | /* Tasklet does all the work to avoid irqsave locks. */ | |
281 | tasklet_schedule(&dto_tasklet); | |
282 | } | |
283 | ||
284 | /* | |
285 | * rq_cq_reap - Process the RQ CQ. | |
286 | * | |
287 | * Take all completing WC off the CQE and enqueue the associated DTO | |
288 | * context on the dto_q for the transport. | |
0905c0f0 TT |
289 | * |
290 | * Note that caller must hold a transport reference. | |
377f9b2f TT |
291 | */ |
292 | static void rq_cq_reap(struct svcxprt_rdma *xprt) | |
293 | { | |
294 | int ret; | |
295 | struct ib_wc wc; | |
296 | struct svc_rdma_op_ctxt *ctxt = NULL; | |
297 | ||
dbcd00eb TT |
298 | if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags)) |
299 | return; | |
300 | ||
301 | ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP); | |
377f9b2f TT |
302 | atomic_inc(&rdma_stat_rq_poll); |
303 | ||
377f9b2f TT |
304 | while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) { |
305 | ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id; | |
306 | ctxt->wc_status = wc.status; | |
307 | ctxt->byte_len = wc.byte_len; | |
e6ab9143 | 308 | svc_rdma_unmap_dma(ctxt); |
377f9b2f TT |
309 | if (wc.status != IB_WC_SUCCESS) { |
310 | /* Close the transport */ | |
0905c0f0 | 311 | dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt); |
377f9b2f TT |
312 | set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); |
313 | svc_rdma_put_context(ctxt, 1); | |
0905c0f0 | 314 | svc_xprt_put(&xprt->sc_xprt); |
377f9b2f TT |
315 | continue; |
316 | } | |
47698e08 | 317 | spin_lock_bh(&xprt->sc_rq_dto_lock); |
377f9b2f | 318 | list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q); |
47698e08 | 319 | spin_unlock_bh(&xprt->sc_rq_dto_lock); |
0905c0f0 | 320 | svc_xprt_put(&xprt->sc_xprt); |
377f9b2f | 321 | } |
377f9b2f TT |
322 | |
323 | if (ctxt) | |
324 | atomic_inc(&rdma_stat_rq_prod); | |
dbcd00eb TT |
325 | |
326 | set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags); | |
327 | /* | |
328 | * If data arrived before established event, | |
329 | * don't enqueue. This defers RPC I/O until the | |
330 | * RDMA connection is complete. | |
331 | */ | |
332 | if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags)) | |
333 | svc_xprt_enqueue(&xprt->sc_xprt); | |
377f9b2f TT |
334 | } |
335 | ||
e1183210 | 336 | /* |
70f23fd6 | 337 | * Process a completion context |
e1183210 TT |
338 | */ |
339 | static void process_context(struct svcxprt_rdma *xprt, | |
340 | struct svc_rdma_op_ctxt *ctxt) | |
341 | { | |
342 | svc_rdma_unmap_dma(ctxt); | |
343 | ||
344 | switch (ctxt->wr_op) { | |
345 | case IB_WR_SEND: | |
afd566ea TT |
346 | if (test_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags)) |
347 | svc_rdma_put_frmr(xprt, ctxt->frmr); | |
e1183210 TT |
348 | svc_rdma_put_context(ctxt, 1); |
349 | break; | |
350 | ||
351 | case IB_WR_RDMA_WRITE: | |
352 | svc_rdma_put_context(ctxt, 0); | |
353 | break; | |
354 | ||
355 | case IB_WR_RDMA_READ: | |
146b6df6 | 356 | case IB_WR_RDMA_READ_WITH_INV: |
e1183210 TT |
357 | if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) { |
358 | struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr; | |
359 | BUG_ON(!read_hdr); | |
146b6df6 TT |
360 | if (test_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags)) |
361 | svc_rdma_put_frmr(xprt, ctxt->frmr); | |
e1183210 TT |
362 | spin_lock_bh(&xprt->sc_rq_dto_lock); |
363 | set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags); | |
364 | list_add_tail(&read_hdr->dto_q, | |
365 | &xprt->sc_read_complete_q); | |
366 | spin_unlock_bh(&xprt->sc_rq_dto_lock); | |
367 | svc_xprt_enqueue(&xprt->sc_xprt); | |
368 | } | |
369 | svc_rdma_put_context(ctxt, 0); | |
370 | break; | |
371 | ||
372 | default: | |
373 | printk(KERN_ERR "svcrdma: unexpected completion type, " | |
374 | "opcode=%d\n", | |
375 | ctxt->wr_op); | |
376 | break; | |
377 | } | |
378 | } | |
379 | ||
377f9b2f TT |
380 | /* |
381 | * Send Queue Completion Handler - potentially called on interrupt context. | |
0905c0f0 TT |
382 | * |
383 | * Note that caller must hold a transport reference. | |
377f9b2f TT |
384 | */ |
385 | static void sq_cq_reap(struct svcxprt_rdma *xprt) | |
386 | { | |
387 | struct svc_rdma_op_ctxt *ctxt = NULL; | |
388 | struct ib_wc wc; | |
389 | struct ib_cq *cq = xprt->sc_sq_cq; | |
390 | int ret; | |
391 | ||
dbcd00eb TT |
392 | if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags)) |
393 | return; | |
394 | ||
395 | ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP); | |
377f9b2f TT |
396 | atomic_inc(&rdma_stat_sq_poll); |
397 | while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) { | |
377f9b2f TT |
398 | if (wc.status != IB_WC_SUCCESS) |
399 | /* Close the transport */ | |
400 | set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); | |
401 | ||
402 | /* Decrement used SQ WR count */ | |
403 | atomic_dec(&xprt->sc_sq_count); | |
404 | wake_up(&xprt->sc_send_wait); | |
405 | ||
e1183210 TT |
406 | ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id; |
407 | if (ctxt) | |
408 | process_context(xprt, ctxt); | |
377f9b2f | 409 | |
0905c0f0 | 410 | svc_xprt_put(&xprt->sc_xprt); |
377f9b2f TT |
411 | } |
412 | ||
413 | if (ctxt) | |
414 | atomic_inc(&rdma_stat_sq_prod); | |
415 | } | |
416 | ||
417 | static void sq_comp_handler(struct ib_cq *cq, void *cq_context) | |
418 | { | |
419 | struct svcxprt_rdma *xprt = cq_context; | |
420 | unsigned long flags; | |
421 | ||
1711386c TT |
422 | /* Guard against unconditional flush call for destroyed QP */ |
423 | if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0) | |
424 | return; | |
425 | ||
377f9b2f TT |
426 | /* |
427 | * Set the bit regardless of whether or not it's on the list | |
428 | * because it may be on the list already due to an RQ | |
429 | * completion. | |
1711386c | 430 | */ |
377f9b2f TT |
431 | set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags); |
432 | ||
433 | /* | |
434 | * If this transport is not already on the DTO transport queue, | |
435 | * add it | |
436 | */ | |
437 | spin_lock_irqsave(&dto_lock, flags); | |
c48cbb40 TT |
438 | if (list_empty(&xprt->sc_dto_q)) { |
439 | svc_xprt_get(&xprt->sc_xprt); | |
377f9b2f | 440 | list_add_tail(&xprt->sc_dto_q, &dto_xprt_q); |
c48cbb40 | 441 | } |
377f9b2f TT |
442 | spin_unlock_irqrestore(&dto_lock, flags); |
443 | ||
444 | /* Tasklet does all the work to avoid irqsave locks. */ | |
445 | tasklet_schedule(&dto_tasklet); | |
446 | } | |
447 | ||
377f9b2f TT |
448 | static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv, |
449 | int listener) | |
450 | { | |
451 | struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL); | |
452 | ||
453 | if (!cma_xprt) | |
454 | return NULL; | |
455 | svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv); | |
456 | INIT_LIST_HEAD(&cma_xprt->sc_accept_q); | |
457 | INIT_LIST_HEAD(&cma_xprt->sc_dto_q); | |
458 | INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q); | |
459 | INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q); | |
64be8608 | 460 | INIT_LIST_HEAD(&cma_xprt->sc_frmr_q); |
377f9b2f TT |
461 | init_waitqueue_head(&cma_xprt->sc_send_wait); |
462 | ||
463 | spin_lock_init(&cma_xprt->sc_lock); | |
377f9b2f | 464 | spin_lock_init(&cma_xprt->sc_rq_dto_lock); |
64be8608 | 465 | spin_lock_init(&cma_xprt->sc_frmr_q_lock); |
377f9b2f TT |
466 | |
467 | cma_xprt->sc_ord = svcrdma_ord; | |
468 | ||
469 | cma_xprt->sc_max_req_size = svcrdma_max_req_size; | |
470 | cma_xprt->sc_max_requests = svcrdma_max_requests; | |
471 | cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT; | |
472 | atomic_set(&cma_xprt->sc_sq_count, 0); | |
87295b6c | 473 | atomic_set(&cma_xprt->sc_ctxt_used, 0); |
377f9b2f | 474 | |
8948896c | 475 | if (listener) |
377f9b2f TT |
476 | set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags); |
477 | ||
478 | return cma_xprt; | |
479 | } | |
480 | ||
481 | struct page *svc_rdma_get_page(void) | |
482 | { | |
483 | struct page *page; | |
484 | ||
485 | while ((page = alloc_page(GFP_KERNEL)) == NULL) { | |
486 | /* If we can't get memory, wait a bit and try again */ | |
487 | printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 " | |
488 | "jiffies.\n"); | |
489 | schedule_timeout_uninterruptible(msecs_to_jiffies(1000)); | |
490 | } | |
491 | return page; | |
492 | } | |
493 | ||
494 | int svc_rdma_post_recv(struct svcxprt_rdma *xprt) | |
495 | { | |
496 | struct ib_recv_wr recv_wr, *bad_recv_wr; | |
497 | struct svc_rdma_op_ctxt *ctxt; | |
498 | struct page *page; | |
a5abf4e8 | 499 | dma_addr_t pa; |
377f9b2f TT |
500 | int sge_no; |
501 | int buflen; | |
502 | int ret; | |
503 | ||
504 | ctxt = svc_rdma_get_context(xprt); | |
505 | buflen = 0; | |
506 | ctxt->direction = DMA_FROM_DEVICE; | |
507 | for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) { | |
508 | BUG_ON(sge_no >= xprt->sc_max_sge); | |
509 | page = svc_rdma_get_page(); | |
510 | ctxt->pages[sge_no] = page; | |
b432e6b3 TT |
511 | pa = ib_dma_map_page(xprt->sc_cm_id->device, |
512 | page, 0, PAGE_SIZE, | |
377f9b2f | 513 | DMA_FROM_DEVICE); |
a5abf4e8 TT |
514 | if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa)) |
515 | goto err_put_ctxt; | |
516 | atomic_inc(&xprt->sc_dma_used); | |
377f9b2f TT |
517 | ctxt->sge[sge_no].addr = pa; |
518 | ctxt->sge[sge_no].length = PAGE_SIZE; | |
a5abf4e8 | 519 | ctxt->sge[sge_no].lkey = xprt->sc_dma_lkey; |
4a84386f | 520 | ctxt->count = sge_no + 1; |
377f9b2f TT |
521 | buflen += PAGE_SIZE; |
522 | } | |
377f9b2f TT |
523 | recv_wr.next = NULL; |
524 | recv_wr.sg_list = &ctxt->sge[0]; | |
525 | recv_wr.num_sge = ctxt->count; | |
526 | recv_wr.wr_id = (u64)(unsigned long)ctxt; | |
527 | ||
0905c0f0 | 528 | svc_xprt_get(&xprt->sc_xprt); |
377f9b2f | 529 | ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr); |
0905c0f0 | 530 | if (ret) { |
21515e46 | 531 | svc_rdma_unmap_dma(ctxt); |
05a0826a | 532 | svc_rdma_put_context(ctxt, 1); |
21515e46 | 533 | svc_xprt_put(&xprt->sc_xprt); |
0905c0f0 | 534 | } |
377f9b2f | 535 | return ret; |
a5abf4e8 TT |
536 | |
537 | err_put_ctxt: | |
4a84386f | 538 | svc_rdma_unmap_dma(ctxt); |
a5abf4e8 TT |
539 | svc_rdma_put_context(ctxt, 1); |
540 | return -ENOMEM; | |
377f9b2f TT |
541 | } |
542 | ||
543 | /* | |
544 | * This function handles the CONNECT_REQUEST event on a listening | |
545 | * endpoint. It is passed the cma_id for the _new_ connection. The context in | |
546 | * this cma_id is inherited from the listening cma_id and is the svc_xprt | |
547 | * structure for the listening endpoint. | |
548 | * | |
549 | * This function creates a new xprt for the new connection and enqueues it on | |
550 | * the accept queue for the listent xprt. When the listen thread is kicked, it | |
551 | * will call the recvfrom method on the listen xprt which will accept the new | |
552 | * connection. | |
553 | */ | |
36ef25e4 | 554 | static void handle_connect_req(struct rdma_cm_id *new_cma_id, size_t client_ird) |
377f9b2f TT |
555 | { |
556 | struct svcxprt_rdma *listen_xprt = new_cma_id->context; | |
557 | struct svcxprt_rdma *newxprt; | |
af261af4 | 558 | struct sockaddr *sa; |
377f9b2f TT |
559 | |
560 | /* Create a new transport */ | |
561 | newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0); | |
562 | if (!newxprt) { | |
563 | dprintk("svcrdma: failed to create new transport\n"); | |
564 | return; | |
565 | } | |
566 | newxprt->sc_cm_id = new_cma_id; | |
567 | new_cma_id->context = newxprt; | |
568 | dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n", | |
569 | newxprt, newxprt->sc_cm_id, listen_xprt); | |
570 | ||
36ef25e4 TT |
571 | /* Save client advertised inbound read limit for use later in accept. */ |
572 | newxprt->sc_ord = client_ird; | |
573 | ||
af261af4 TT |
574 | /* Set the local and remote addresses in the transport */ |
575 | sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr; | |
576 | svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa)); | |
577 | sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr; | |
578 | svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa)); | |
579 | ||
377f9b2f TT |
580 | /* |
581 | * Enqueue the new transport on the accept queue of the listening | |
582 | * transport | |
583 | */ | |
584 | spin_lock_bh(&listen_xprt->sc_lock); | |
585 | list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q); | |
586 | spin_unlock_bh(&listen_xprt->sc_lock); | |
587 | ||
588 | /* | |
589 | * Can't use svc_xprt_received here because we are not on a | |
590 | * rqstp thread | |
591 | */ | |
592 | set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags); | |
593 | svc_xprt_enqueue(&listen_xprt->sc_xprt); | |
594 | } | |
595 | ||
596 | /* | |
597 | * Handles events generated on the listening endpoint. These events will be | |
598 | * either be incoming connect requests or adapter removal events. | |
599 | */ | |
600 | static int rdma_listen_handler(struct rdma_cm_id *cma_id, | |
601 | struct rdma_cm_event *event) | |
602 | { | |
603 | struct svcxprt_rdma *xprt = cma_id->context; | |
604 | int ret = 0; | |
605 | ||
606 | switch (event->event) { | |
607 | case RDMA_CM_EVENT_CONNECT_REQUEST: | |
608 | dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, " | |
609 | "event=%d\n", cma_id, cma_id->context, event->event); | |
36ef25e4 | 610 | handle_connect_req(cma_id, |
67080c82 | 611 | event->param.conn.initiator_depth); |
377f9b2f TT |
612 | break; |
613 | ||
614 | case RDMA_CM_EVENT_ESTABLISHED: | |
615 | /* Accept complete */ | |
616 | dprintk("svcrdma: Connection completed on LISTEN xprt=%p, " | |
617 | "cm_id=%p\n", xprt, cma_id); | |
618 | break; | |
619 | ||
620 | case RDMA_CM_EVENT_DEVICE_REMOVAL: | |
621 | dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n", | |
622 | xprt, cma_id); | |
623 | if (xprt) | |
624 | set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); | |
625 | break; | |
626 | ||
627 | default: | |
628 | dprintk("svcrdma: Unexpected event on listening endpoint %p, " | |
629 | "event=%d\n", cma_id, event->event); | |
630 | break; | |
631 | } | |
632 | ||
633 | return ret; | |
634 | } | |
635 | ||
636 | static int rdma_cma_handler(struct rdma_cm_id *cma_id, | |
637 | struct rdma_cm_event *event) | |
638 | { | |
639 | struct svc_xprt *xprt = cma_id->context; | |
640 | struct svcxprt_rdma *rdma = | |
641 | container_of(xprt, struct svcxprt_rdma, sc_xprt); | |
642 | switch (event->event) { | |
643 | case RDMA_CM_EVENT_ESTABLISHED: | |
644 | /* Accept complete */ | |
c48cbb40 | 645 | svc_xprt_get(xprt); |
377f9b2f TT |
646 | dprintk("svcrdma: Connection completed on DTO xprt=%p, " |
647 | "cm_id=%p\n", xprt, cma_id); | |
648 | clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags); | |
649 | svc_xprt_enqueue(xprt); | |
650 | break; | |
651 | case RDMA_CM_EVENT_DISCONNECTED: | |
652 | dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n", | |
653 | xprt, cma_id); | |
654 | if (xprt) { | |
655 | set_bit(XPT_CLOSE, &xprt->xpt_flags); | |
656 | svc_xprt_enqueue(xprt); | |
120693d1 | 657 | svc_xprt_put(xprt); |
377f9b2f TT |
658 | } |
659 | break; | |
660 | case RDMA_CM_EVENT_DEVICE_REMOVAL: | |
661 | dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, " | |
662 | "event=%d\n", cma_id, xprt, event->event); | |
663 | if (xprt) { | |
664 | set_bit(XPT_CLOSE, &xprt->xpt_flags); | |
665 | svc_xprt_enqueue(xprt); | |
666 | } | |
667 | break; | |
668 | default: | |
669 | dprintk("svcrdma: Unexpected event on DTO endpoint %p, " | |
670 | "event=%d\n", cma_id, event->event); | |
671 | break; | |
672 | } | |
673 | return 0; | |
674 | } | |
675 | ||
676 | /* | |
677 | * Create a listening RDMA service endpoint. | |
678 | */ | |
679 | static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, | |
62832c03 | 680 | struct net *net, |
377f9b2f TT |
681 | struct sockaddr *sa, int salen, |
682 | int flags) | |
683 | { | |
684 | struct rdma_cm_id *listen_id; | |
685 | struct svcxprt_rdma *cma_xprt; | |
686 | struct svc_xprt *xprt; | |
687 | int ret; | |
688 | ||
689 | dprintk("svcrdma: Creating RDMA socket\n"); | |
bade732a TT |
690 | if (sa->sa_family != AF_INET) { |
691 | dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family); | |
692 | return ERR_PTR(-EAFNOSUPPORT); | |
693 | } | |
377f9b2f TT |
694 | cma_xprt = rdma_create_xprt(serv, 1); |
695 | if (!cma_xprt) | |
58e8f621 | 696 | return ERR_PTR(-ENOMEM); |
377f9b2f TT |
697 | xprt = &cma_xprt->sc_xprt; |
698 | ||
b26f9b99 SH |
699 | listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP, |
700 | IB_QPT_RC); | |
377f9b2f | 701 | if (IS_ERR(listen_id)) { |
58e8f621 TT |
702 | ret = PTR_ERR(listen_id); |
703 | dprintk("svcrdma: rdma_create_id failed = %d\n", ret); | |
704 | goto err0; | |
377f9b2f | 705 | } |
58e8f621 | 706 | |
377f9b2f TT |
707 | ret = rdma_bind_addr(listen_id, sa); |
708 | if (ret) { | |
377f9b2f | 709 | dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret); |
58e8f621 | 710 | goto err1; |
377f9b2f TT |
711 | } |
712 | cma_xprt->sc_cm_id = listen_id; | |
713 | ||
714 | ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG); | |
715 | if (ret) { | |
377f9b2f | 716 | dprintk("svcrdma: rdma_listen failed = %d\n", ret); |
58e8f621 | 717 | goto err1; |
377f9b2f TT |
718 | } |
719 | ||
720 | /* | |
721 | * We need to use the address from the cm_id in case the | |
722 | * caller specified 0 for the port number. | |
723 | */ | |
724 | sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr; | |
725 | svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen); | |
726 | ||
727 | return &cma_xprt->sc_xprt; | |
58e8f621 TT |
728 | |
729 | err1: | |
730 | rdma_destroy_id(listen_id); | |
731 | err0: | |
732 | kfree(cma_xprt); | |
733 | return ERR_PTR(ret); | |
377f9b2f TT |
734 | } |
735 | ||
64be8608 TT |
736 | static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt) |
737 | { | |
738 | struct ib_mr *mr; | |
739 | struct ib_fast_reg_page_list *pl; | |
740 | struct svc_rdma_fastreg_mr *frmr; | |
741 | ||
742 | frmr = kmalloc(sizeof(*frmr), GFP_KERNEL); | |
743 | if (!frmr) | |
744 | goto err; | |
745 | ||
746 | mr = ib_alloc_fast_reg_mr(xprt->sc_pd, RPCSVC_MAXPAGES); | |
846d8e7c | 747 | if (IS_ERR(mr)) |
64be8608 TT |
748 | goto err_free_frmr; |
749 | ||
750 | pl = ib_alloc_fast_reg_page_list(xprt->sc_cm_id->device, | |
751 | RPCSVC_MAXPAGES); | |
846d8e7c | 752 | if (IS_ERR(pl)) |
64be8608 TT |
753 | goto err_free_mr; |
754 | ||
755 | frmr->mr = mr; | |
756 | frmr->page_list = pl; | |
757 | INIT_LIST_HEAD(&frmr->frmr_list); | |
758 | return frmr; | |
759 | ||
760 | err_free_mr: | |
761 | ib_dereg_mr(mr); | |
762 | err_free_frmr: | |
763 | kfree(frmr); | |
764 | err: | |
765 | return ERR_PTR(-ENOMEM); | |
766 | } | |
767 | ||
768 | static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt) | |
769 | { | |
770 | struct svc_rdma_fastreg_mr *frmr; | |
771 | ||
772 | while (!list_empty(&xprt->sc_frmr_q)) { | |
773 | frmr = list_entry(xprt->sc_frmr_q.next, | |
774 | struct svc_rdma_fastreg_mr, frmr_list); | |
775 | list_del_init(&frmr->frmr_list); | |
776 | ib_dereg_mr(frmr->mr); | |
777 | ib_free_fast_reg_page_list(frmr->page_list); | |
778 | kfree(frmr); | |
779 | } | |
780 | } | |
781 | ||
782 | struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma) | |
783 | { | |
784 | struct svc_rdma_fastreg_mr *frmr = NULL; | |
785 | ||
786 | spin_lock_bh(&rdma->sc_frmr_q_lock); | |
787 | if (!list_empty(&rdma->sc_frmr_q)) { | |
788 | frmr = list_entry(rdma->sc_frmr_q.next, | |
789 | struct svc_rdma_fastreg_mr, frmr_list); | |
790 | list_del_init(&frmr->frmr_list); | |
791 | frmr->map_len = 0; | |
792 | frmr->page_list_len = 0; | |
793 | } | |
794 | spin_unlock_bh(&rdma->sc_frmr_q_lock); | |
795 | if (frmr) | |
796 | return frmr; | |
797 | ||
798 | return rdma_alloc_frmr(rdma); | |
799 | } | |
800 | ||
801 | static void frmr_unmap_dma(struct svcxprt_rdma *xprt, | |
802 | struct svc_rdma_fastreg_mr *frmr) | |
803 | { | |
804 | int page_no; | |
805 | for (page_no = 0; page_no < frmr->page_list_len; page_no++) { | |
806 | dma_addr_t addr = frmr->page_list->page_list[page_no]; | |
807 | if (ib_dma_mapping_error(frmr->mr->device, addr)) | |
808 | continue; | |
809 | atomic_dec(&xprt->sc_dma_used); | |
b432e6b3 TT |
810 | ib_dma_unmap_page(frmr->mr->device, addr, PAGE_SIZE, |
811 | frmr->direction); | |
64be8608 TT |
812 | } |
813 | } | |
814 | ||
815 | void svc_rdma_put_frmr(struct svcxprt_rdma *rdma, | |
816 | struct svc_rdma_fastreg_mr *frmr) | |
817 | { | |
818 | if (frmr) { | |
819 | frmr_unmap_dma(rdma, frmr); | |
820 | spin_lock_bh(&rdma->sc_frmr_q_lock); | |
821 | BUG_ON(!list_empty(&frmr->frmr_list)); | |
822 | list_add(&frmr->frmr_list, &rdma->sc_frmr_q); | |
823 | spin_unlock_bh(&rdma->sc_frmr_q_lock); | |
824 | } | |
825 | } | |
826 | ||
377f9b2f TT |
827 | /* |
828 | * This is the xpo_recvfrom function for listening endpoints. Its | |
829 | * purpose is to accept incoming connections. The CMA callback handler | |
830 | * has already created a new transport and attached it to the new CMA | |
831 | * ID. | |
832 | * | |
833 | * There is a queue of pending connections hung on the listening | |
834 | * transport. This queue contains the new svc_xprt structure. This | |
835 | * function takes svc_xprt structures off the accept_q and completes | |
836 | * the connection. | |
837 | */ | |
838 | static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) | |
839 | { | |
840 | struct svcxprt_rdma *listen_rdma; | |
841 | struct svcxprt_rdma *newxprt = NULL; | |
842 | struct rdma_conn_param conn_param; | |
843 | struct ib_qp_init_attr qp_attr; | |
844 | struct ib_device_attr devattr; | |
ed72b9c6 | 845 | int uninitialized_var(dma_mr_acc); |
3a5c6380 | 846 | int need_dma_mr; |
377f9b2f TT |
847 | int ret; |
848 | int i; | |
849 | ||
850 | listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt); | |
851 | clear_bit(XPT_CONN, &xprt->xpt_flags); | |
852 | /* Get the next entry off the accept list */ | |
853 | spin_lock_bh(&listen_rdma->sc_lock); | |
854 | if (!list_empty(&listen_rdma->sc_accept_q)) { | |
855 | newxprt = list_entry(listen_rdma->sc_accept_q.next, | |
856 | struct svcxprt_rdma, sc_accept_q); | |
857 | list_del_init(&newxprt->sc_accept_q); | |
858 | } | |
859 | if (!list_empty(&listen_rdma->sc_accept_q)) | |
860 | set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags); | |
861 | spin_unlock_bh(&listen_rdma->sc_lock); | |
862 | if (!newxprt) | |
863 | return NULL; | |
864 | ||
865 | dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n", | |
866 | newxprt, newxprt->sc_cm_id); | |
867 | ||
868 | ret = ib_query_device(newxprt->sc_cm_id->device, &devattr); | |
869 | if (ret) { | |
870 | dprintk("svcrdma: could not query device attributes on " | |
871 | "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret); | |
872 | goto errout; | |
873 | } | |
874 | ||
875 | /* Qualify the transport resource defaults with the | |
876 | * capabilities of this particular device */ | |
877 | newxprt->sc_max_sge = min((size_t)devattr.max_sge, | |
878 | (size_t)RPCSVC_MAXPAGES); | |
879 | newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr, | |
880 | (size_t)svcrdma_max_requests); | |
881 | newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests; | |
882 | ||
36ef25e4 TT |
883 | /* |
884 | * Limit ORD based on client limit, local device limit, and | |
885 | * configured svcrdma limit. | |
886 | */ | |
887 | newxprt->sc_ord = min_t(size_t, devattr.max_qp_rd_atom, newxprt->sc_ord); | |
888 | newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord); | |
377f9b2f TT |
889 | |
890 | newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device); | |
891 | if (IS_ERR(newxprt->sc_pd)) { | |
892 | dprintk("svcrdma: error creating PD for connect request\n"); | |
893 | goto errout; | |
894 | } | |
895 | newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device, | |
896 | sq_comp_handler, | |
897 | cq_event_handler, | |
898 | newxprt, | |
899 | newxprt->sc_sq_depth, | |
900 | 0); | |
901 | if (IS_ERR(newxprt->sc_sq_cq)) { | |
902 | dprintk("svcrdma: error creating SQ CQ for connect request\n"); | |
903 | goto errout; | |
904 | } | |
905 | newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device, | |
906 | rq_comp_handler, | |
907 | cq_event_handler, | |
908 | newxprt, | |
909 | newxprt->sc_max_requests, | |
910 | 0); | |
911 | if (IS_ERR(newxprt->sc_rq_cq)) { | |
912 | dprintk("svcrdma: error creating RQ CQ for connect request\n"); | |
913 | goto errout; | |
914 | } | |
915 | ||
916 | memset(&qp_attr, 0, sizeof qp_attr); | |
917 | qp_attr.event_handler = qp_event_handler; | |
918 | qp_attr.qp_context = &newxprt->sc_xprt; | |
919 | qp_attr.cap.max_send_wr = newxprt->sc_sq_depth; | |
920 | qp_attr.cap.max_recv_wr = newxprt->sc_max_requests; | |
921 | qp_attr.cap.max_send_sge = newxprt->sc_max_sge; | |
922 | qp_attr.cap.max_recv_sge = newxprt->sc_max_sge; | |
923 | qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR; | |
924 | qp_attr.qp_type = IB_QPT_RC; | |
925 | qp_attr.send_cq = newxprt->sc_sq_cq; | |
926 | qp_attr.recv_cq = newxprt->sc_rq_cq; | |
927 | dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n" | |
928 | " cm_id->device=%p, sc_pd->device=%p\n" | |
929 | " cap.max_send_wr = %d\n" | |
930 | " cap.max_recv_wr = %d\n" | |
931 | " cap.max_send_sge = %d\n" | |
932 | " cap.max_recv_sge = %d\n", | |
933 | newxprt->sc_cm_id, newxprt->sc_pd, | |
934 | newxprt->sc_cm_id->device, newxprt->sc_pd->device, | |
935 | qp_attr.cap.max_send_wr, | |
936 | qp_attr.cap.max_recv_wr, | |
937 | qp_attr.cap.max_send_sge, | |
938 | qp_attr.cap.max_recv_sge); | |
939 | ||
940 | ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr); | |
941 | if (ret) { | |
942 | /* | |
943 | * XXX: This is a hack. We need a xx_request_qp interface | |
944 | * that will adjust the qp_attr's with a best-effort | |
945 | * number | |
946 | */ | |
947 | qp_attr.cap.max_send_sge -= 2; | |
948 | qp_attr.cap.max_recv_sge -= 2; | |
949 | ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, | |
950 | &qp_attr); | |
951 | if (ret) { | |
952 | dprintk("svcrdma: failed to create QP, ret=%d\n", ret); | |
953 | goto errout; | |
954 | } | |
955 | newxprt->sc_max_sge = qp_attr.cap.max_send_sge; | |
956 | newxprt->sc_max_sge = qp_attr.cap.max_recv_sge; | |
957 | newxprt->sc_sq_depth = qp_attr.cap.max_send_wr; | |
958 | newxprt->sc_max_requests = qp_attr.cap.max_recv_wr; | |
959 | } | |
960 | newxprt->sc_qp = newxprt->sc_cm_id->qp; | |
961 | ||
3a5c6380 TT |
962 | /* |
963 | * Use the most secure set of MR resources based on the | |
964 | * transport type and available memory management features in | |
965 | * the device. Here's the table implemented below: | |
966 | * | |
967 | * Fast Global DMA Remote WR | |
968 | * Reg LKEY MR Access | |
969 | * Sup'd Sup'd Needed Needed | |
970 | * | |
971 | * IWARP N N Y Y | |
972 | * N Y Y Y | |
973 | * Y N Y N | |
974 | * Y Y N - | |
975 | * | |
976 | * IB N N Y N | |
977 | * N Y N - | |
978 | * Y N Y N | |
979 | * Y Y N - | |
980 | * | |
981 | * NB: iWARP requires remote write access for the data sink | |
982 | * of an RDMA_READ. IB does not. | |
983 | */ | |
984 | if (devattr.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) { | |
985 | newxprt->sc_frmr_pg_list_len = | |
986 | devattr.max_fast_reg_page_list_len; | |
987 | newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG; | |
988 | } | |
989 | ||
990 | /* | |
991 | * Determine if a DMA MR is required and if so, what privs are required | |
992 | */ | |
993 | switch (rdma_node_get_transport(newxprt->sc_cm_id->device->node_type)) { | |
994 | case RDMA_TRANSPORT_IWARP: | |
995 | newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV; | |
996 | if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { | |
997 | need_dma_mr = 1; | |
998 | dma_mr_acc = | |
999 | (IB_ACCESS_LOCAL_WRITE | | |
1000 | IB_ACCESS_REMOTE_WRITE); | |
1001 | } else if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) { | |
1002 | need_dma_mr = 1; | |
1003 | dma_mr_acc = IB_ACCESS_LOCAL_WRITE; | |
1004 | } else | |
1005 | need_dma_mr = 0; | |
1006 | break; | |
1007 | case RDMA_TRANSPORT_IB: | |
1008 | if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) { | |
1009 | need_dma_mr = 1; | |
1010 | dma_mr_acc = IB_ACCESS_LOCAL_WRITE; | |
1011 | } else | |
1012 | need_dma_mr = 0; | |
1013 | break; | |
1014 | default: | |
377f9b2f TT |
1015 | goto errout; |
1016 | } | |
1017 | ||
3a5c6380 TT |
1018 | /* Create the DMA MR if needed, otherwise, use the DMA LKEY */ |
1019 | if (need_dma_mr) { | |
1020 | /* Register all of physical memory */ | |
1021 | newxprt->sc_phys_mr = | |
1022 | ib_get_dma_mr(newxprt->sc_pd, dma_mr_acc); | |
1023 | if (IS_ERR(newxprt->sc_phys_mr)) { | |
1024 | dprintk("svcrdma: Failed to create DMA MR ret=%d\n", | |
1025 | ret); | |
1026 | goto errout; | |
1027 | } | |
1028 | newxprt->sc_dma_lkey = newxprt->sc_phys_mr->lkey; | |
1029 | } else | |
1030 | newxprt->sc_dma_lkey = | |
1031 | newxprt->sc_cm_id->device->local_dma_lkey; | |
1032 | ||
377f9b2f TT |
1033 | /* Post receive buffers */ |
1034 | for (i = 0; i < newxprt->sc_max_requests; i++) { | |
1035 | ret = svc_rdma_post_recv(newxprt); | |
1036 | if (ret) { | |
1037 | dprintk("svcrdma: failure posting receive buffers\n"); | |
1038 | goto errout; | |
1039 | } | |
1040 | } | |
1041 | ||
1042 | /* Swap out the handler */ | |
1043 | newxprt->sc_cm_id->event_handler = rdma_cma_handler; | |
1044 | ||
af261af4 TT |
1045 | /* |
1046 | * Arm the CQs for the SQ and RQ before accepting so we can't | |
1047 | * miss the first message | |
1048 | */ | |
1049 | ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP); | |
1050 | ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP); | |
1051 | ||
377f9b2f TT |
1052 | /* Accept Connection */ |
1053 | set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags); | |
1054 | memset(&conn_param, 0, sizeof conn_param); | |
1055 | conn_param.responder_resources = 0; | |
1056 | conn_param.initiator_depth = newxprt->sc_ord; | |
1057 | ret = rdma_accept(newxprt->sc_cm_id, &conn_param); | |
1058 | if (ret) { | |
1059 | dprintk("svcrdma: failed to accept new connection, ret=%d\n", | |
1060 | ret); | |
1061 | goto errout; | |
1062 | } | |
1063 | ||
1064 | dprintk("svcrdma: new connection %p accepted with the following " | |
1065 | "attributes:\n" | |
21454aaa | 1066 | " local_ip : %pI4\n" |
377f9b2f | 1067 | " local_port : %d\n" |
21454aaa | 1068 | " remote_ip : %pI4\n" |
377f9b2f TT |
1069 | " remote_port : %d\n" |
1070 | " max_sge : %d\n" | |
1071 | " sq_depth : %d\n" | |
1072 | " max_requests : %d\n" | |
1073 | " ord : %d\n", | |
1074 | newxprt, | |
21454aaa HH |
1075 | &((struct sockaddr_in *)&newxprt->sc_cm_id-> |
1076 | route.addr.src_addr)->sin_addr.s_addr, | |
377f9b2f TT |
1077 | ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> |
1078 | route.addr.src_addr)->sin_port), | |
21454aaa HH |
1079 | &((struct sockaddr_in *)&newxprt->sc_cm_id-> |
1080 | route.addr.dst_addr)->sin_addr.s_addr, | |
377f9b2f TT |
1081 | ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> |
1082 | route.addr.dst_addr)->sin_port), | |
1083 | newxprt->sc_max_sge, | |
1084 | newxprt->sc_sq_depth, | |
1085 | newxprt->sc_max_requests, | |
1086 | newxprt->sc_ord); | |
1087 | ||
377f9b2f TT |
1088 | return &newxprt->sc_xprt; |
1089 | ||
1090 | errout: | |
1091 | dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret); | |
c48cbb40 TT |
1092 | /* Take a reference in case the DTO handler runs */ |
1093 | svc_xprt_get(&newxprt->sc_xprt); | |
1711386c | 1094 | if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) |
c48cbb40 | 1095 | ib_destroy_qp(newxprt->sc_qp); |
377f9b2f | 1096 | rdma_destroy_id(newxprt->sc_cm_id); |
c48cbb40 TT |
1097 | /* This call to put will destroy the transport */ |
1098 | svc_xprt_put(&newxprt->sc_xprt); | |
377f9b2f TT |
1099 | return NULL; |
1100 | } | |
1101 | ||
377f9b2f TT |
1102 | static void svc_rdma_release_rqst(struct svc_rqst *rqstp) |
1103 | { | |
377f9b2f TT |
1104 | } |
1105 | ||
c48cbb40 | 1106 | /* |
1711386c | 1107 | * When connected, an svc_xprt has at least two references: |
c48cbb40 TT |
1108 | * |
1109 | * - A reference held by the cm_id between the ESTABLISHED and | |
1110 | * DISCONNECTED events. If the remote peer disconnected first, this | |
1111 | * reference could be gone. | |
1112 | * | |
1113 | * - A reference held by the svc_recv code that called this function | |
1114 | * as part of close processing. | |
1115 | * | |
1711386c | 1116 | * At a minimum one references should still be held. |
c48cbb40 | 1117 | */ |
377f9b2f TT |
1118 | static void svc_rdma_detach(struct svc_xprt *xprt) |
1119 | { | |
1120 | struct svcxprt_rdma *rdma = | |
1121 | container_of(xprt, struct svcxprt_rdma, sc_xprt); | |
377f9b2f | 1122 | dprintk("svc: svc_rdma_detach(%p)\n", xprt); |
c48cbb40 TT |
1123 | |
1124 | /* Disconnect and flush posted WQE */ | |
377f9b2f | 1125 | rdma_disconnect(rdma->sc_cm_id); |
377f9b2f TT |
1126 | } |
1127 | ||
8da91ea8 | 1128 | static void __svc_rdma_free(struct work_struct *work) |
377f9b2f | 1129 | { |
8da91ea8 TT |
1130 | struct svcxprt_rdma *rdma = |
1131 | container_of(work, struct svcxprt_rdma, sc_work); | |
377f9b2f | 1132 | dprintk("svcrdma: svc_rdma_free(%p)\n", rdma); |
8da91ea8 | 1133 | |
c48cbb40 | 1134 | /* We should only be called from kref_put */ |
8da91ea8 TT |
1135 | BUG_ON(atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0); |
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 */ | |
1163 | WARN_ON(atomic_read(&rdma->sc_ctxt_used) != 0); | |
87295b6c | 1164 | WARN_ON(atomic_read(&rdma->sc_dma_used) != 0); |
356d0a15 | 1165 | |
64be8608 TT |
1166 | /* De-allocate fastreg mr */ |
1167 | rdma_dealloc_frmr_q(rdma); | |
1168 | ||
1711386c TT |
1169 | /* Destroy the QP if present (not a listener) */ |
1170 | if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) | |
1171 | ib_destroy_qp(rdma->sc_qp); | |
1172 | ||
c48cbb40 TT |
1173 | if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq)) |
1174 | ib_destroy_cq(rdma->sc_sq_cq); | |
377f9b2f | 1175 | |
c48cbb40 TT |
1176 | if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq)) |
1177 | ib_destroy_cq(rdma->sc_rq_cq); | |
377f9b2f | 1178 | |
c48cbb40 TT |
1179 | if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr)) |
1180 | ib_dereg_mr(rdma->sc_phys_mr); | |
377f9b2f | 1181 | |
c48cbb40 TT |
1182 | if (rdma->sc_pd && !IS_ERR(rdma->sc_pd)) |
1183 | ib_dealloc_pd(rdma->sc_pd); | |
377f9b2f | 1184 | |
356d0a15 TT |
1185 | /* Destroy the CM ID */ |
1186 | rdma_destroy_id(rdma->sc_cm_id); | |
1187 | ||
c48cbb40 | 1188 | kfree(rdma); |
377f9b2f TT |
1189 | } |
1190 | ||
8da91ea8 TT |
1191 | static void svc_rdma_free(struct svc_xprt *xprt) |
1192 | { | |
1193 | struct svcxprt_rdma *rdma = | |
1194 | container_of(xprt, struct svcxprt_rdma, sc_xprt); | |
1195 | INIT_WORK(&rdma->sc_work, __svc_rdma_free); | |
a25e758c | 1196 | queue_work(svc_rdma_wq, &rdma->sc_work); |
8da91ea8 TT |
1197 | } |
1198 | ||
377f9b2f TT |
1199 | static int svc_rdma_has_wspace(struct svc_xprt *xprt) |
1200 | { | |
1201 | struct svcxprt_rdma *rdma = | |
1202 | container_of(xprt, struct svcxprt_rdma, sc_xprt); | |
1203 | ||
1204 | /* | |
1205 | * If there are fewer SQ WR available than required to send a | |
1206 | * simple response, return false. | |
1207 | */ | |
1208 | if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3)) | |
1209 | return 0; | |
1210 | ||
1211 | /* | |
1212 | * ...or there are already waiters on the SQ, | |
1213 | * return false. | |
1214 | */ | |
1215 | if (waitqueue_active(&rdma->sc_send_wait)) | |
1216 | return 0; | |
1217 | ||
1218 | /* Otherwise return true. */ | |
1219 | return 1; | |
1220 | } | |
1221 | ||
e1183210 TT |
1222 | /* |
1223 | * Attempt to register the kvec representing the RPC memory with the | |
1224 | * device. | |
1225 | * | |
1226 | * Returns: | |
1227 | * NULL : The device does not support fastreg or there were no more | |
1228 | * fastreg mr. | |
1229 | * frmr : The kvec register request was successfully posted. | |
1230 | * <0 : An error was encountered attempting to register the kvec. | |
1231 | */ | |
1232 | int svc_rdma_fastreg(struct svcxprt_rdma *xprt, | |
1233 | struct svc_rdma_fastreg_mr *frmr) | |
1234 | { | |
1235 | struct ib_send_wr fastreg_wr; | |
1236 | u8 key; | |
1237 | ||
1238 | /* Bump the key */ | |
1239 | key = (u8)(frmr->mr->lkey & 0x000000FF); | |
1240 | ib_update_fast_reg_key(frmr->mr, ++key); | |
1241 | ||
1242 | /* Prepare FASTREG WR */ | |
1243 | memset(&fastreg_wr, 0, sizeof fastreg_wr); | |
1244 | fastreg_wr.opcode = IB_WR_FAST_REG_MR; | |
1245 | fastreg_wr.send_flags = IB_SEND_SIGNALED; | |
1246 | fastreg_wr.wr.fast_reg.iova_start = (unsigned long)frmr->kva; | |
1247 | fastreg_wr.wr.fast_reg.page_list = frmr->page_list; | |
1248 | fastreg_wr.wr.fast_reg.page_list_len = frmr->page_list_len; | |
1249 | fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT; | |
1250 | fastreg_wr.wr.fast_reg.length = frmr->map_len; | |
1251 | fastreg_wr.wr.fast_reg.access_flags = frmr->access_flags; | |
1252 | fastreg_wr.wr.fast_reg.rkey = frmr->mr->lkey; | |
1253 | return svc_rdma_send(xprt, &fastreg_wr); | |
1254 | } | |
1255 | ||
377f9b2f TT |
1256 | int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr) |
1257 | { | |
5b180a9a TT |
1258 | struct ib_send_wr *bad_wr, *n_wr; |
1259 | int wr_count; | |
1260 | int i; | |
377f9b2f TT |
1261 | int ret; |
1262 | ||
1263 | if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) | |
9d6347ac | 1264 | return -ENOTCONN; |
377f9b2f TT |
1265 | |
1266 | BUG_ON(wr->send_flags != IB_SEND_SIGNALED); | |
5b180a9a TT |
1267 | wr_count = 1; |
1268 | for (n_wr = wr->next; n_wr; n_wr = n_wr->next) | |
1269 | wr_count++; | |
1270 | ||
377f9b2f TT |
1271 | /* If the SQ is full, wait until an SQ entry is available */ |
1272 | while (1) { | |
1273 | spin_lock_bh(&xprt->sc_lock); | |
5b180a9a | 1274 | if (xprt->sc_sq_depth < atomic_read(&xprt->sc_sq_count) + wr_count) { |
377f9b2f TT |
1275 | spin_unlock_bh(&xprt->sc_lock); |
1276 | atomic_inc(&rdma_stat_sq_starve); | |
dbcd00eb TT |
1277 | |
1278 | /* See if we can opportunistically reap SQ WR to make room */ | |
377f9b2f TT |
1279 | sq_cq_reap(xprt); |
1280 | ||
1281 | /* Wait until SQ WR available if SQ still full */ | |
1282 | wait_event(xprt->sc_send_wait, | |
1283 | atomic_read(&xprt->sc_sq_count) < | |
1284 | xprt->sc_sq_depth); | |
830bb59b | 1285 | if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) |
b432e6b3 | 1286 | return -ENOTCONN; |
377f9b2f TT |
1287 | continue; |
1288 | } | |
5b180a9a TT |
1289 | /* Take a transport ref for each WR posted */ |
1290 | for (i = 0; i < wr_count; i++) | |
1291 | svc_xprt_get(&xprt->sc_xprt); | |
1292 | ||
1293 | /* Bump used SQ WR count and post */ | |
1294 | atomic_add(wr_count, &xprt->sc_sq_count); | |
377f9b2f | 1295 | ret = ib_post_send(xprt->sc_qp, wr, &bad_wr); |
5b180a9a TT |
1296 | if (ret) { |
1297 | set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); | |
1298 | atomic_sub(wr_count, &xprt->sc_sq_count); | |
1299 | for (i = 0; i < wr_count; i ++) | |
1300 | svc_xprt_put(&xprt->sc_xprt); | |
377f9b2f TT |
1301 | dprintk("svcrdma: failed to post SQ WR rc=%d, " |
1302 | "sc_sq_count=%d, sc_sq_depth=%d\n", | |
1303 | ret, atomic_read(&xprt->sc_sq_count), | |
1304 | xprt->sc_sq_depth); | |
0905c0f0 | 1305 | } |
377f9b2f | 1306 | spin_unlock_bh(&xprt->sc_lock); |
5b180a9a TT |
1307 | if (ret) |
1308 | wake_up(&xprt->sc_send_wait); | |
377f9b2f TT |
1309 | break; |
1310 | } | |
1311 | return ret; | |
1312 | } | |
1313 | ||
008fdbc5 TT |
1314 | void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp, |
1315 | enum rpcrdma_errcode err) | |
377f9b2f TT |
1316 | { |
1317 | struct ib_send_wr err_wr; | |
377f9b2f TT |
1318 | struct page *p; |
1319 | struct svc_rdma_op_ctxt *ctxt; | |
1320 | u32 *va; | |
1321 | int length; | |
1322 | int ret; | |
1323 | ||
1324 | p = svc_rdma_get_page(); | |
1325 | va = page_address(p); | |
1326 | ||
1327 | /* XDR encode error */ | |
1328 | length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va); | |
1329 | ||
4a84386f TT |
1330 | ctxt = svc_rdma_get_context(xprt); |
1331 | ctxt->direction = DMA_FROM_DEVICE; | |
1332 | ctxt->count = 1; | |
1333 | ctxt->pages[0] = p; | |
1334 | ||
377f9b2f | 1335 | /* Prepare SGE for local address */ |
4a84386f TT |
1336 | ctxt->sge[0].addr = ib_dma_map_page(xprt->sc_cm_id->device, |
1337 | p, 0, length, DMA_FROM_DEVICE); | |
1338 | if (ib_dma_mapping_error(xprt->sc_cm_id->device, ctxt->sge[0].addr)) { | |
04911b53 | 1339 | put_page(p); |
a5e50268 | 1340 | svc_rdma_put_context(ctxt, 1); |
04911b53 TT |
1341 | return; |
1342 | } | |
1343 | atomic_inc(&xprt->sc_dma_used); | |
4a84386f TT |
1344 | ctxt->sge[0].lkey = xprt->sc_dma_lkey; |
1345 | ctxt->sge[0].length = length; | |
377f9b2f TT |
1346 | |
1347 | /* Prepare SEND WR */ | |
1348 | memset(&err_wr, 0, sizeof err_wr); | |
1349 | ctxt->wr_op = IB_WR_SEND; | |
1350 | err_wr.wr_id = (unsigned long)ctxt; | |
4a84386f | 1351 | err_wr.sg_list = ctxt->sge; |
377f9b2f TT |
1352 | err_wr.num_sge = 1; |
1353 | err_wr.opcode = IB_WR_SEND; | |
1354 | err_wr.send_flags = IB_SEND_SIGNALED; | |
1355 | ||
1356 | /* Post It */ | |
1357 | ret = svc_rdma_send(xprt, &err_wr); | |
1358 | if (ret) { | |
008fdbc5 TT |
1359 | dprintk("svcrdma: Error %d posting send for protocol error\n", |
1360 | ret); | |
4a84386f | 1361 | svc_rdma_unmap_dma(ctxt); |
377f9b2f TT |
1362 | svc_rdma_put_context(ctxt, 1); |
1363 | } | |
377f9b2f | 1364 | } |