2 * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved.
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
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
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.
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
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.
39 * Author: Tom Tucker <tom@opengridcomputing.com>
42 #include <linux/sunrpc/svc_xprt.h>
43 #include <linux/sunrpc/debug.h>
44 #include <linux/sunrpc/rpc_rdma.h>
45 #include <linux/spinlock.h>
46 #include <rdma/ib_verbs.h>
47 #include <rdma/rdma_cm.h>
48 #include <linux/sunrpc/svc_rdma.h>
50 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
52 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
53 struct sockaddr *sa, int salen,
55 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
56 static void svc_rdma_release_rqst(struct svc_rqst *);
57 static void dto_tasklet_func(unsigned long data);
58 static void svc_rdma_detach(struct svc_xprt *xprt);
59 static void svc_rdma_free(struct svc_xprt *xprt);
60 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
61 static void rq_cq_reap(struct svcxprt_rdma *xprt);
62 static void sq_cq_reap(struct svcxprt_rdma *xprt);
64 DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL);
65 static DEFINE_SPINLOCK(dto_lock);
66 static LIST_HEAD(dto_xprt_q);
68 static struct svc_xprt_ops svc_rdma_ops = {
69 .xpo_create = svc_rdma_create,
70 .xpo_recvfrom = svc_rdma_recvfrom,
71 .xpo_sendto = svc_rdma_sendto,
72 .xpo_release_rqst = svc_rdma_release_rqst,
73 .xpo_detach = svc_rdma_detach,
74 .xpo_free = svc_rdma_free,
75 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
76 .xpo_has_wspace = svc_rdma_has_wspace,
77 .xpo_accept = svc_rdma_accept,
80 struct svc_xprt_class svc_rdma_class = {
82 .xcl_owner = THIS_MODULE,
83 .xcl_ops = &svc_rdma_ops,
84 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
87 static int rdma_bump_context_cache(struct svcxprt_rdma *xprt)
91 struct svc_rdma_op_ctxt *ctxt;
93 target = min(xprt->sc_ctxt_cnt + xprt->sc_ctxt_bump,
96 spin_lock_bh(&xprt->sc_ctxt_lock);
97 while (xprt->sc_ctxt_cnt < target) {
99 spin_unlock_bh(&xprt->sc_ctxt_lock);
101 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
103 spin_lock_bh(&xprt->sc_ctxt_lock);
106 INIT_LIST_HEAD(&ctxt->free_list);
107 list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
109 /* kmalloc failed...give up for now */
114 spin_unlock_bh(&xprt->sc_ctxt_lock);
115 dprintk("svcrdma: sc_ctxt_max=%d, sc_ctxt_cnt=%d\n",
116 xprt->sc_ctxt_max, xprt->sc_ctxt_cnt);
120 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
122 struct svc_rdma_op_ctxt *ctxt;
125 spin_lock_bh(&xprt->sc_ctxt_lock);
126 if (unlikely(list_empty(&xprt->sc_ctxt_free))) {
127 /* Try to bump my cache. */
128 spin_unlock_bh(&xprt->sc_ctxt_lock);
130 if (rdma_bump_context_cache(xprt))
133 printk(KERN_INFO "svcrdma: sleeping waiting for "
134 "context memory on xprt=%p\n",
136 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
139 ctxt = list_entry(xprt->sc_ctxt_free.next,
140 struct svc_rdma_op_ctxt,
142 list_del_init(&ctxt->free_list);
143 spin_unlock_bh(&xprt->sc_ctxt_lock);
145 INIT_LIST_HEAD(&ctxt->dto_q);
147 atomic_inc(&xprt->sc_ctxt_used);
153 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
155 struct svcxprt_rdma *xprt;
161 for (i = 0; i < ctxt->count; i++)
162 put_page(ctxt->pages[i]);
164 for (i = 0; i < ctxt->count; i++)
165 dma_unmap_single(xprt->sc_cm_id->device->dma_device,
170 spin_lock_bh(&xprt->sc_ctxt_lock);
171 list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
172 spin_unlock_bh(&xprt->sc_ctxt_lock);
173 atomic_dec(&xprt->sc_ctxt_used);
176 /* ib_cq event handler */
177 static void cq_event_handler(struct ib_event *event, void *context)
179 struct svc_xprt *xprt = context;
180 dprintk("svcrdma: received CQ event id=%d, context=%p\n",
181 event->event, context);
182 set_bit(XPT_CLOSE, &xprt->xpt_flags);
185 /* QP event handler */
186 static void qp_event_handler(struct ib_event *event, void *context)
188 struct svc_xprt *xprt = context;
190 switch (event->event) {
191 /* These are considered benign events */
192 case IB_EVENT_PATH_MIG:
193 case IB_EVENT_COMM_EST:
194 case IB_EVENT_SQ_DRAINED:
195 case IB_EVENT_QP_LAST_WQE_REACHED:
196 dprintk("svcrdma: QP event %d received for QP=%p\n",
197 event->event, event->element.qp);
199 /* These are considered fatal events */
200 case IB_EVENT_PATH_MIG_ERR:
201 case IB_EVENT_QP_FATAL:
202 case IB_EVENT_QP_REQ_ERR:
203 case IB_EVENT_QP_ACCESS_ERR:
204 case IB_EVENT_DEVICE_FATAL:
206 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
207 "closing transport\n",
208 event->event, event->element.qp);
209 set_bit(XPT_CLOSE, &xprt->xpt_flags);
215 * Data Transfer Operation Tasklet
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.
223 static void dto_tasklet_func(unsigned long data)
225 struct svcxprt_rdma *xprt;
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);
238 svc_xprt_put(&xprt->sc_xprt);
239 spin_lock_irqsave(&dto_lock, flags);
241 spin_unlock_irqrestore(&dto_lock, flags);
245 * Receive Queue Completion Handler
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
250 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
252 struct svcxprt_rdma *xprt = cq_context;
256 * Set the bit regardless of whether or not it's on the list
257 * because it may be on the list already due to an SQ
260 set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
263 * If this transport is not already on the DTO transport queue,
266 spin_lock_irqsave(&dto_lock, flags);
267 if (list_empty(&xprt->sc_dto_q)) {
268 svc_xprt_get(&xprt->sc_xprt);
269 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
271 spin_unlock_irqrestore(&dto_lock, flags);
273 /* Tasklet does all the work to avoid irqsave locks. */
274 tasklet_schedule(&dto_tasklet);
278 * rq_cq_reap - Process the RQ CQ.
280 * Take all completing WC off the CQE and enqueue the associated DTO
281 * context on the dto_q for the transport.
283 static void rq_cq_reap(struct svcxprt_rdma *xprt)
287 struct svc_rdma_op_ctxt *ctxt = NULL;
289 if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
292 ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
293 atomic_inc(&rdma_stat_rq_poll);
295 spin_lock_bh(&xprt->sc_rq_dto_lock);
296 while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
297 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
298 ctxt->wc_status = wc.status;
299 ctxt->byte_len = wc.byte_len;
300 if (wc.status != IB_WC_SUCCESS) {
301 /* Close the transport */
302 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
303 svc_rdma_put_context(ctxt, 1);
306 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
308 spin_unlock_bh(&xprt->sc_rq_dto_lock);
311 atomic_inc(&rdma_stat_rq_prod);
313 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
315 * If data arrived before established event,
316 * don't enqueue. This defers RPC I/O until the
317 * RDMA connection is complete.
319 if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
320 svc_xprt_enqueue(&xprt->sc_xprt);
324 * Send Queue Completion Handler - potentially called on interrupt context.
326 static void sq_cq_reap(struct svcxprt_rdma *xprt)
328 struct svc_rdma_op_ctxt *ctxt = NULL;
330 struct ib_cq *cq = xprt->sc_sq_cq;
334 if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
337 ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
338 atomic_inc(&rdma_stat_sq_poll);
339 while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
340 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
343 if (wc.status != IB_WC_SUCCESS)
344 /* Close the transport */
345 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
347 /* Decrement used SQ WR count */
348 atomic_dec(&xprt->sc_sq_count);
349 wake_up(&xprt->sc_send_wait);
351 switch (ctxt->wr_op) {
353 case IB_WR_RDMA_WRITE:
354 svc_rdma_put_context(ctxt, 1);
357 case IB_WR_RDMA_READ:
358 if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
359 struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
361 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
362 spin_lock_bh(&xprt->sc_read_complete_lock);
363 list_add_tail(&read_hdr->dto_q,
364 &xprt->sc_read_complete_q);
365 spin_unlock_bh(&xprt->sc_read_complete_lock);
366 svc_xprt_enqueue(&xprt->sc_xprt);
368 svc_rdma_put_context(ctxt, 0);
372 printk(KERN_ERR "svcrdma: unexpected completion type, "
373 "opcode=%d, status=%d\n",
374 wc.opcode, wc.status);
380 atomic_inc(&rdma_stat_sq_prod);
383 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
385 struct svcxprt_rdma *xprt = cq_context;
389 * Set the bit regardless of whether or not it's on the list
390 * because it may be on the list already due to an RQ
393 set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
396 * If this transport is not already on the DTO transport queue,
399 spin_lock_irqsave(&dto_lock, flags);
400 if (list_empty(&xprt->sc_dto_q)) {
401 svc_xprt_get(&xprt->sc_xprt);
402 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
404 spin_unlock_irqrestore(&dto_lock, flags);
406 /* Tasklet does all the work to avoid irqsave locks. */
407 tasklet_schedule(&dto_tasklet);
410 static void create_context_cache(struct svcxprt_rdma *xprt,
411 int ctxt_count, int ctxt_bump, int ctxt_max)
413 struct svc_rdma_op_ctxt *ctxt;
416 xprt->sc_ctxt_max = ctxt_max;
417 xprt->sc_ctxt_bump = ctxt_bump;
418 xprt->sc_ctxt_cnt = 0;
419 atomic_set(&xprt->sc_ctxt_used, 0);
421 INIT_LIST_HEAD(&xprt->sc_ctxt_free);
422 for (i = 0; i < ctxt_count; i++) {
423 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
425 INIT_LIST_HEAD(&ctxt->free_list);
426 list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
432 static void destroy_context_cache(struct svcxprt_rdma *xprt)
434 while (!list_empty(&xprt->sc_ctxt_free)) {
435 struct svc_rdma_op_ctxt *ctxt;
436 ctxt = list_entry(xprt->sc_ctxt_free.next,
437 struct svc_rdma_op_ctxt,
439 list_del_init(&ctxt->free_list);
444 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
447 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
451 svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
452 INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
453 INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
454 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
455 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
456 init_waitqueue_head(&cma_xprt->sc_send_wait);
458 spin_lock_init(&cma_xprt->sc_lock);
459 spin_lock_init(&cma_xprt->sc_read_complete_lock);
460 spin_lock_init(&cma_xprt->sc_ctxt_lock);
461 spin_lock_init(&cma_xprt->sc_rq_dto_lock);
463 cma_xprt->sc_ord = svcrdma_ord;
465 cma_xprt->sc_max_req_size = svcrdma_max_req_size;
466 cma_xprt->sc_max_requests = svcrdma_max_requests;
467 cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
468 atomic_set(&cma_xprt->sc_sq_count, 0);
471 int reqs = cma_xprt->sc_max_requests;
472 create_context_cache(cma_xprt,
473 reqs << 1, /* starting size */
474 reqs, /* bump amount */
476 cma_xprt->sc_sq_depth +
477 RPCRDMA_MAX_THREADS + 1); /* max */
478 if (list_empty(&cma_xprt->sc_ctxt_free)) {
482 clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
484 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
489 struct page *svc_rdma_get_page(void)
493 while ((page = alloc_page(GFP_KERNEL)) == NULL) {
494 /* If we can't get memory, wait a bit and try again */
495 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
497 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
502 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
504 struct ib_recv_wr recv_wr, *bad_recv_wr;
505 struct svc_rdma_op_ctxt *ctxt;
512 ctxt = svc_rdma_get_context(xprt);
514 ctxt->direction = DMA_FROM_DEVICE;
515 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
516 BUG_ON(sge_no >= xprt->sc_max_sge);
517 page = svc_rdma_get_page();
518 ctxt->pages[sge_no] = page;
519 pa = ib_dma_map_page(xprt->sc_cm_id->device,
522 ctxt->sge[sge_no].addr = pa;
523 ctxt->sge[sge_no].length = PAGE_SIZE;
524 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
527 ctxt->count = sge_no;
529 recv_wr.sg_list = &ctxt->sge[0];
530 recv_wr.num_sge = ctxt->count;
531 recv_wr.wr_id = (u64)(unsigned long)ctxt;
533 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
535 svc_rdma_put_context(ctxt, 1);
540 * This function handles the CONNECT_REQUEST event on a listening
541 * endpoint. It is passed the cma_id for the _new_ connection. The context in
542 * this cma_id is inherited from the listening cma_id and is the svc_xprt
543 * structure for the listening endpoint.
545 * This function creates a new xprt for the new connection and enqueues it on
546 * the accept queue for the listent xprt. When the listen thread is kicked, it
547 * will call the recvfrom method on the listen xprt which will accept the new
550 static void handle_connect_req(struct rdma_cm_id *new_cma_id)
552 struct svcxprt_rdma *listen_xprt = new_cma_id->context;
553 struct svcxprt_rdma *newxprt;
555 /* Create a new transport */
556 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
558 dprintk("svcrdma: failed to create new transport\n");
561 newxprt->sc_cm_id = new_cma_id;
562 new_cma_id->context = newxprt;
563 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
564 newxprt, newxprt->sc_cm_id, listen_xprt);
567 * Enqueue the new transport on the accept queue of the listening
570 spin_lock_bh(&listen_xprt->sc_lock);
571 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
572 spin_unlock_bh(&listen_xprt->sc_lock);
575 * Can't use svc_xprt_received here because we are not on a
578 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
579 svc_xprt_enqueue(&listen_xprt->sc_xprt);
583 * Handles events generated on the listening endpoint. These events will be
584 * either be incoming connect requests or adapter removal events.
586 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
587 struct rdma_cm_event *event)
589 struct svcxprt_rdma *xprt = cma_id->context;
592 switch (event->event) {
593 case RDMA_CM_EVENT_CONNECT_REQUEST:
594 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
595 "event=%d\n", cma_id, cma_id->context, event->event);
596 handle_connect_req(cma_id);
599 case RDMA_CM_EVENT_ESTABLISHED:
600 /* Accept complete */
601 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
602 "cm_id=%p\n", xprt, cma_id);
605 case RDMA_CM_EVENT_DEVICE_REMOVAL:
606 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
609 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
613 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
614 "event=%d\n", cma_id, event->event);
621 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
622 struct rdma_cm_event *event)
624 struct svc_xprt *xprt = cma_id->context;
625 struct svcxprt_rdma *rdma =
626 container_of(xprt, struct svcxprt_rdma, sc_xprt);
627 switch (event->event) {
628 case RDMA_CM_EVENT_ESTABLISHED:
629 /* Accept complete */
631 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
632 "cm_id=%p\n", xprt, cma_id);
633 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
634 svc_xprt_enqueue(xprt);
636 case RDMA_CM_EVENT_DISCONNECTED:
637 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
640 set_bit(XPT_CLOSE, &xprt->xpt_flags);
641 svc_xprt_enqueue(xprt);
645 case RDMA_CM_EVENT_DEVICE_REMOVAL:
646 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
647 "event=%d\n", cma_id, xprt, event->event);
649 set_bit(XPT_CLOSE, &xprt->xpt_flags);
650 svc_xprt_enqueue(xprt);
654 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
655 "event=%d\n", cma_id, event->event);
662 * Create a listening RDMA service endpoint.
664 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
665 struct sockaddr *sa, int salen,
668 struct rdma_cm_id *listen_id;
669 struct svcxprt_rdma *cma_xprt;
670 struct svc_xprt *xprt;
673 dprintk("svcrdma: Creating RDMA socket\n");
675 cma_xprt = rdma_create_xprt(serv, 1);
677 return ERR_PTR(-ENOMEM);
678 xprt = &cma_xprt->sc_xprt;
680 listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
681 if (IS_ERR(listen_id)) {
682 ret = PTR_ERR(listen_id);
683 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
687 ret = rdma_bind_addr(listen_id, sa);
689 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
692 cma_xprt->sc_cm_id = listen_id;
694 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
696 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
701 * We need to use the address from the cm_id in case the
702 * caller specified 0 for the port number.
704 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
705 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
707 return &cma_xprt->sc_xprt;
710 rdma_destroy_id(listen_id);
717 * This is the xpo_recvfrom function for listening endpoints. Its
718 * purpose is to accept incoming connections. The CMA callback handler
719 * has already created a new transport and attached it to the new CMA
722 * There is a queue of pending connections hung on the listening
723 * transport. This queue contains the new svc_xprt structure. This
724 * function takes svc_xprt structures off the accept_q and completes
727 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
729 struct svcxprt_rdma *listen_rdma;
730 struct svcxprt_rdma *newxprt = NULL;
731 struct rdma_conn_param conn_param;
732 struct ib_qp_init_attr qp_attr;
733 struct ib_device_attr devattr;
738 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
739 clear_bit(XPT_CONN, &xprt->xpt_flags);
740 /* Get the next entry off the accept list */
741 spin_lock_bh(&listen_rdma->sc_lock);
742 if (!list_empty(&listen_rdma->sc_accept_q)) {
743 newxprt = list_entry(listen_rdma->sc_accept_q.next,
744 struct svcxprt_rdma, sc_accept_q);
745 list_del_init(&newxprt->sc_accept_q);
747 if (!list_empty(&listen_rdma->sc_accept_q))
748 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
749 spin_unlock_bh(&listen_rdma->sc_lock);
753 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
754 newxprt, newxprt->sc_cm_id);
756 ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
758 dprintk("svcrdma: could not query device attributes on "
759 "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
763 /* Qualify the transport resource defaults with the
764 * capabilities of this particular device */
765 newxprt->sc_max_sge = min((size_t)devattr.max_sge,
766 (size_t)RPCSVC_MAXPAGES);
767 newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
768 (size_t)svcrdma_max_requests);
769 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
771 newxprt->sc_ord = min((size_t)devattr.max_qp_rd_atom,
772 (size_t)svcrdma_ord);
774 newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
775 if (IS_ERR(newxprt->sc_pd)) {
776 dprintk("svcrdma: error creating PD for connect request\n");
779 newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
783 newxprt->sc_sq_depth,
785 if (IS_ERR(newxprt->sc_sq_cq)) {
786 dprintk("svcrdma: error creating SQ CQ for connect request\n");
789 newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
793 newxprt->sc_max_requests,
795 if (IS_ERR(newxprt->sc_rq_cq)) {
796 dprintk("svcrdma: error creating RQ CQ for connect request\n");
800 memset(&qp_attr, 0, sizeof qp_attr);
801 qp_attr.event_handler = qp_event_handler;
802 qp_attr.qp_context = &newxprt->sc_xprt;
803 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
804 qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
805 qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
806 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
807 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
808 qp_attr.qp_type = IB_QPT_RC;
809 qp_attr.send_cq = newxprt->sc_sq_cq;
810 qp_attr.recv_cq = newxprt->sc_rq_cq;
811 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
812 " cm_id->device=%p, sc_pd->device=%p\n"
813 " cap.max_send_wr = %d\n"
814 " cap.max_recv_wr = %d\n"
815 " cap.max_send_sge = %d\n"
816 " cap.max_recv_sge = %d\n",
817 newxprt->sc_cm_id, newxprt->sc_pd,
818 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
819 qp_attr.cap.max_send_wr,
820 qp_attr.cap.max_recv_wr,
821 qp_attr.cap.max_send_sge,
822 qp_attr.cap.max_recv_sge);
824 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
827 * XXX: This is a hack. We need a xx_request_qp interface
828 * that will adjust the qp_attr's with a best-effort
831 qp_attr.cap.max_send_sge -= 2;
832 qp_attr.cap.max_recv_sge -= 2;
833 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
836 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
839 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
840 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
841 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
842 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
844 svc_xprt_get(&newxprt->sc_xprt);
845 newxprt->sc_qp = newxprt->sc_cm_id->qp;
847 /* Register all of physical memory */
848 newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
849 IB_ACCESS_LOCAL_WRITE |
850 IB_ACCESS_REMOTE_WRITE);
851 if (IS_ERR(newxprt->sc_phys_mr)) {
852 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
856 /* Post receive buffers */
857 for (i = 0; i < newxprt->sc_max_requests; i++) {
858 ret = svc_rdma_post_recv(newxprt);
860 dprintk("svcrdma: failure posting receive buffers\n");
865 /* Swap out the handler */
866 newxprt->sc_cm_id->event_handler = rdma_cma_handler;
868 /* Accept Connection */
869 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
870 memset(&conn_param, 0, sizeof conn_param);
871 conn_param.responder_resources = 0;
872 conn_param.initiator_depth = newxprt->sc_ord;
873 ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
875 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
880 dprintk("svcrdma: new connection %p accepted with the following "
882 " local_ip : %d.%d.%d.%d\n"
884 " remote_ip : %d.%d.%d.%d\n"
885 " remote_port : %d\n"
888 " max_requests : %d\n"
891 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
892 route.addr.src_addr)->sin_addr.s_addr),
893 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
894 route.addr.src_addr)->sin_port),
895 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
896 route.addr.dst_addr)->sin_addr.s_addr),
897 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
898 route.addr.dst_addr)->sin_port),
900 newxprt->sc_sq_depth,
901 newxprt->sc_max_requests,
904 /* Set the local and remote addresses in the transport */
905 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
906 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
907 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
908 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
910 ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
911 ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
912 return &newxprt->sc_xprt;
915 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
916 /* Take a reference in case the DTO handler runs */
917 svc_xprt_get(&newxprt->sc_xprt);
918 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) {
919 ib_destroy_qp(newxprt->sc_qp);
920 svc_xprt_put(&newxprt->sc_xprt);
922 rdma_destroy_id(newxprt->sc_cm_id);
923 /* This call to put will destroy the transport */
924 svc_xprt_put(&newxprt->sc_xprt);
928 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
933 * When connected, an svc_xprt has at least three references:
935 * - A reference held by the QP. We still hold that here because this
936 * code deletes the QP and puts the reference.
938 * - A reference held by the cm_id between the ESTABLISHED and
939 * DISCONNECTED events. If the remote peer disconnected first, this
940 * reference could be gone.
942 * - A reference held by the svc_recv code that called this function
943 * as part of close processing.
945 * At a minimum two references should still be held.
947 static void svc_rdma_detach(struct svc_xprt *xprt)
949 struct svcxprt_rdma *rdma =
950 container_of(xprt, struct svcxprt_rdma, sc_xprt);
951 dprintk("svc: svc_rdma_detach(%p)\n", xprt);
953 /* Disconnect and flush posted WQE */
954 rdma_disconnect(rdma->sc_cm_id);
956 /* Destroy the QP if present (not a listener) */
957 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) {
958 ib_destroy_qp(rdma->sc_qp);
962 /* Destroy the CM ID */
963 rdma_destroy_id(rdma->sc_cm_id);
966 static void svc_rdma_free(struct svc_xprt *xprt)
968 struct svcxprt_rdma *rdma = (struct svcxprt_rdma *)xprt;
969 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
970 /* We should only be called from kref_put */
971 BUG_ON(atomic_read(&xprt->xpt_ref.refcount) != 0);
972 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
973 ib_destroy_cq(rdma->sc_sq_cq);
975 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
976 ib_destroy_cq(rdma->sc_rq_cq);
978 if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
979 ib_dereg_mr(rdma->sc_phys_mr);
981 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
982 ib_dealloc_pd(rdma->sc_pd);
984 destroy_context_cache(rdma);
988 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
990 struct svcxprt_rdma *rdma =
991 container_of(xprt, struct svcxprt_rdma, sc_xprt);
994 * If there are fewer SQ WR available than required to send a
995 * simple response, return false.
997 if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
1001 * ...or there are already waiters on the SQ,
1004 if (waitqueue_active(&rdma->sc_send_wait))
1007 /* Otherwise return true. */
1011 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1013 struct ib_send_wr *bad_wr;
1016 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1019 BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1020 BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1022 /* If the SQ is full, wait until an SQ entry is available */
1024 spin_lock_bh(&xprt->sc_lock);
1025 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1026 spin_unlock_bh(&xprt->sc_lock);
1027 atomic_inc(&rdma_stat_sq_starve);
1029 /* See if we can opportunistically reap SQ WR to make room */
1032 /* Wait until SQ WR available if SQ still full */
1033 wait_event(xprt->sc_send_wait,
1034 atomic_read(&xprt->sc_sq_count) <
1036 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1040 /* Bumped used SQ WR count and post */
1041 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1043 atomic_inc(&xprt->sc_sq_count);
1045 dprintk("svcrdma: failed to post SQ WR rc=%d, "
1046 "sc_sq_count=%d, sc_sq_depth=%d\n",
1047 ret, atomic_read(&xprt->sc_sq_count),
1049 spin_unlock_bh(&xprt->sc_lock);
1055 int svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1056 enum rpcrdma_errcode err)
1058 struct ib_send_wr err_wr;
1061 struct svc_rdma_op_ctxt *ctxt;
1066 p = svc_rdma_get_page();
1067 va = page_address(p);
1069 /* XDR encode error */
1070 length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1072 /* Prepare SGE for local address */
1073 sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1074 p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1075 sge.lkey = xprt->sc_phys_mr->lkey;
1076 sge.length = length;
1078 ctxt = svc_rdma_get_context(xprt);
1082 /* Prepare SEND WR */
1083 memset(&err_wr, 0, sizeof err_wr);
1084 ctxt->wr_op = IB_WR_SEND;
1085 err_wr.wr_id = (unsigned long)ctxt;
1086 err_wr.sg_list = &sge;
1088 err_wr.opcode = IB_WR_SEND;
1089 err_wr.send_flags = IB_SEND_SIGNALED;
1092 ret = svc_rdma_send(xprt, &err_wr);
1094 dprintk("svcrdma: Error posting send = %d\n", ret);
1095 svc_rdma_put_context(ctxt, 1);