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 ib_dma_unmap_single(xprt->sc_cm_id->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 /* Temporary NFS request map cache. Created in svc_rdma.c */
177 extern struct kmem_cache *svc_rdma_map_cachep;
180 * Temporary NFS req mappings are shared across all transport
181 * instances. These are short lived and should be bounded by the number
182 * of concurrent server threads * depth of the SQ.
184 struct svc_rdma_req_map *svc_rdma_get_req_map(void)
186 struct svc_rdma_req_map *map;
188 map = kmem_cache_alloc(svc_rdma_map_cachep, GFP_KERNEL);
191 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
197 void svc_rdma_put_req_map(struct svc_rdma_req_map *map)
199 kmem_cache_free(svc_rdma_map_cachep, map);
202 /* ib_cq event handler */
203 static void cq_event_handler(struct ib_event *event, void *context)
205 struct svc_xprt *xprt = context;
206 dprintk("svcrdma: received CQ event id=%d, context=%p\n",
207 event->event, context);
208 set_bit(XPT_CLOSE, &xprt->xpt_flags);
211 /* QP event handler */
212 static void qp_event_handler(struct ib_event *event, void *context)
214 struct svc_xprt *xprt = context;
216 switch (event->event) {
217 /* These are considered benign events */
218 case IB_EVENT_PATH_MIG:
219 case IB_EVENT_COMM_EST:
220 case IB_EVENT_SQ_DRAINED:
221 case IB_EVENT_QP_LAST_WQE_REACHED:
222 dprintk("svcrdma: QP event %d received for QP=%p\n",
223 event->event, event->element.qp);
225 /* These are considered fatal events */
226 case IB_EVENT_PATH_MIG_ERR:
227 case IB_EVENT_QP_FATAL:
228 case IB_EVENT_QP_REQ_ERR:
229 case IB_EVENT_QP_ACCESS_ERR:
230 case IB_EVENT_DEVICE_FATAL:
232 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
233 "closing transport\n",
234 event->event, event->element.qp);
235 set_bit(XPT_CLOSE, &xprt->xpt_flags);
241 * Data Transfer Operation Tasklet
243 * Walks a list of transports with I/O pending, removing entries as
244 * they are added to the server's I/O pending list. Two bits indicate
245 * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
246 * spinlock that serializes access to the transport list with the RQ
247 * and SQ interrupt handlers.
249 static void dto_tasklet_func(unsigned long data)
251 struct svcxprt_rdma *xprt;
254 spin_lock_irqsave(&dto_lock, flags);
255 while (!list_empty(&dto_xprt_q)) {
256 xprt = list_entry(dto_xprt_q.next,
257 struct svcxprt_rdma, sc_dto_q);
258 list_del_init(&xprt->sc_dto_q);
259 spin_unlock_irqrestore(&dto_lock, flags);
264 svc_xprt_put(&xprt->sc_xprt);
265 spin_lock_irqsave(&dto_lock, flags);
267 spin_unlock_irqrestore(&dto_lock, flags);
271 * Receive Queue Completion Handler
273 * Since an RQ completion handler is called on interrupt context, we
274 * need to defer the handling of the I/O to a tasklet
276 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
278 struct svcxprt_rdma *xprt = cq_context;
281 /* Guard against unconditional flush call for destroyed QP */
282 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
286 * Set the bit regardless of whether or not it's on the list
287 * because it may be on the list already due to an SQ
290 set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
293 * If this transport is not already on the DTO transport queue,
296 spin_lock_irqsave(&dto_lock, flags);
297 if (list_empty(&xprt->sc_dto_q)) {
298 svc_xprt_get(&xprt->sc_xprt);
299 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
301 spin_unlock_irqrestore(&dto_lock, flags);
303 /* Tasklet does all the work to avoid irqsave locks. */
304 tasklet_schedule(&dto_tasklet);
308 * rq_cq_reap - Process the RQ CQ.
310 * Take all completing WC off the CQE and enqueue the associated DTO
311 * context on the dto_q for the transport.
313 * Note that caller must hold a transport reference.
315 static void rq_cq_reap(struct svcxprt_rdma *xprt)
319 struct svc_rdma_op_ctxt *ctxt = NULL;
321 if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
324 ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
325 atomic_inc(&rdma_stat_rq_poll);
327 while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
328 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
329 ctxt->wc_status = wc.status;
330 ctxt->byte_len = wc.byte_len;
331 if (wc.status != IB_WC_SUCCESS) {
332 /* Close the transport */
333 dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt);
334 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
335 svc_rdma_put_context(ctxt, 1);
336 svc_xprt_put(&xprt->sc_xprt);
339 spin_lock_bh(&xprt->sc_rq_dto_lock);
340 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
341 spin_unlock_bh(&xprt->sc_rq_dto_lock);
342 svc_xprt_put(&xprt->sc_xprt);
346 atomic_inc(&rdma_stat_rq_prod);
348 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
350 * If data arrived before established event,
351 * don't enqueue. This defers RPC I/O until the
352 * RDMA connection is complete.
354 if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
355 svc_xprt_enqueue(&xprt->sc_xprt);
359 * Send Queue Completion Handler - potentially called on interrupt context.
361 * Note that caller must hold a transport reference.
363 static void sq_cq_reap(struct svcxprt_rdma *xprt)
365 struct svc_rdma_op_ctxt *ctxt = NULL;
367 struct ib_cq *cq = xprt->sc_sq_cq;
371 if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
374 ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
375 atomic_inc(&rdma_stat_sq_poll);
376 while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
377 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
380 if (wc.status != IB_WC_SUCCESS)
381 /* Close the transport */
382 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
384 /* Decrement used SQ WR count */
385 atomic_dec(&xprt->sc_sq_count);
386 wake_up(&xprt->sc_send_wait);
388 switch (ctxt->wr_op) {
390 svc_rdma_put_context(ctxt, 1);
393 case IB_WR_RDMA_WRITE:
394 svc_rdma_put_context(ctxt, 0);
397 case IB_WR_RDMA_READ:
398 if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
399 struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
401 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
402 spin_lock_bh(&xprt->sc_read_complete_lock);
403 list_add_tail(&read_hdr->dto_q,
404 &xprt->sc_read_complete_q);
405 spin_unlock_bh(&xprt->sc_read_complete_lock);
406 svc_xprt_enqueue(&xprt->sc_xprt);
408 svc_rdma_put_context(ctxt, 0);
412 printk(KERN_ERR "svcrdma: unexpected completion type, "
413 "opcode=%d, status=%d\n",
414 wc.opcode, wc.status);
417 svc_xprt_put(&xprt->sc_xprt);
421 atomic_inc(&rdma_stat_sq_prod);
424 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
426 struct svcxprt_rdma *xprt = cq_context;
429 /* Guard against unconditional flush call for destroyed QP */
430 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
434 * Set the bit regardless of whether or not it's on the list
435 * because it may be on the list already due to an RQ
438 set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
441 * If this transport is not already on the DTO transport queue,
444 spin_lock_irqsave(&dto_lock, flags);
445 if (list_empty(&xprt->sc_dto_q)) {
446 svc_xprt_get(&xprt->sc_xprt);
447 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
449 spin_unlock_irqrestore(&dto_lock, flags);
451 /* Tasklet does all the work to avoid irqsave locks. */
452 tasklet_schedule(&dto_tasklet);
455 static void create_context_cache(struct svcxprt_rdma *xprt,
456 int ctxt_count, int ctxt_bump, int ctxt_max)
458 struct svc_rdma_op_ctxt *ctxt;
461 xprt->sc_ctxt_max = ctxt_max;
462 xprt->sc_ctxt_bump = ctxt_bump;
463 xprt->sc_ctxt_cnt = 0;
464 atomic_set(&xprt->sc_ctxt_used, 0);
466 INIT_LIST_HEAD(&xprt->sc_ctxt_free);
467 for (i = 0; i < ctxt_count; i++) {
468 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
470 INIT_LIST_HEAD(&ctxt->free_list);
471 list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
477 static void destroy_context_cache(struct svcxprt_rdma *xprt)
479 while (!list_empty(&xprt->sc_ctxt_free)) {
480 struct svc_rdma_op_ctxt *ctxt;
481 ctxt = list_entry(xprt->sc_ctxt_free.next,
482 struct svc_rdma_op_ctxt,
484 list_del_init(&ctxt->free_list);
489 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
492 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
496 svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
497 INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
498 INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
499 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
500 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
501 init_waitqueue_head(&cma_xprt->sc_send_wait);
503 spin_lock_init(&cma_xprt->sc_lock);
504 spin_lock_init(&cma_xprt->sc_read_complete_lock);
505 spin_lock_init(&cma_xprt->sc_ctxt_lock);
506 spin_lock_init(&cma_xprt->sc_rq_dto_lock);
508 cma_xprt->sc_ord = svcrdma_ord;
510 cma_xprt->sc_max_req_size = svcrdma_max_req_size;
511 cma_xprt->sc_max_requests = svcrdma_max_requests;
512 cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
513 atomic_set(&cma_xprt->sc_sq_count, 0);
516 int reqs = cma_xprt->sc_max_requests;
517 create_context_cache(cma_xprt,
518 reqs << 1, /* starting size */
519 reqs, /* bump amount */
521 cma_xprt->sc_sq_depth +
522 RPCRDMA_MAX_THREADS + 1); /* max */
523 if (list_empty(&cma_xprt->sc_ctxt_free)) {
527 clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
529 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
534 struct page *svc_rdma_get_page(void)
538 while ((page = alloc_page(GFP_KERNEL)) == NULL) {
539 /* If we can't get memory, wait a bit and try again */
540 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
542 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
547 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
549 struct ib_recv_wr recv_wr, *bad_recv_wr;
550 struct svc_rdma_op_ctxt *ctxt;
557 ctxt = svc_rdma_get_context(xprt);
559 ctxt->direction = DMA_FROM_DEVICE;
560 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
561 BUG_ON(sge_no >= xprt->sc_max_sge);
562 page = svc_rdma_get_page();
563 ctxt->pages[sge_no] = page;
564 pa = ib_dma_map_page(xprt->sc_cm_id->device,
567 ctxt->sge[sge_no].addr = pa;
568 ctxt->sge[sge_no].length = PAGE_SIZE;
569 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
572 ctxt->count = sge_no;
574 recv_wr.sg_list = &ctxt->sge[0];
575 recv_wr.num_sge = ctxt->count;
576 recv_wr.wr_id = (u64)(unsigned long)ctxt;
578 svc_xprt_get(&xprt->sc_xprt);
579 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
581 svc_xprt_put(&xprt->sc_xprt);
582 svc_rdma_put_context(ctxt, 1);
588 * This function handles the CONNECT_REQUEST event on a listening
589 * endpoint. It is passed the cma_id for the _new_ connection. The context in
590 * this cma_id is inherited from the listening cma_id and is the svc_xprt
591 * structure for the listening endpoint.
593 * This function creates a new xprt for the new connection and enqueues it on
594 * the accept queue for the listent xprt. When the listen thread is kicked, it
595 * will call the recvfrom method on the listen xprt which will accept the new
598 static void handle_connect_req(struct rdma_cm_id *new_cma_id)
600 struct svcxprt_rdma *listen_xprt = new_cma_id->context;
601 struct svcxprt_rdma *newxprt;
604 /* Create a new transport */
605 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
607 dprintk("svcrdma: failed to create new transport\n");
610 newxprt->sc_cm_id = new_cma_id;
611 new_cma_id->context = newxprt;
612 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
613 newxprt, newxprt->sc_cm_id, listen_xprt);
615 /* Set the local and remote addresses in the transport */
616 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
617 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
618 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
619 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
622 * Enqueue the new transport on the accept queue of the listening
625 spin_lock_bh(&listen_xprt->sc_lock);
626 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
627 spin_unlock_bh(&listen_xprt->sc_lock);
630 * Can't use svc_xprt_received here because we are not on a
633 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
634 svc_xprt_enqueue(&listen_xprt->sc_xprt);
638 * Handles events generated on the listening endpoint. These events will be
639 * either be incoming connect requests or adapter removal events.
641 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
642 struct rdma_cm_event *event)
644 struct svcxprt_rdma *xprt = cma_id->context;
647 switch (event->event) {
648 case RDMA_CM_EVENT_CONNECT_REQUEST:
649 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
650 "event=%d\n", cma_id, cma_id->context, event->event);
651 handle_connect_req(cma_id);
654 case RDMA_CM_EVENT_ESTABLISHED:
655 /* Accept complete */
656 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
657 "cm_id=%p\n", xprt, cma_id);
660 case RDMA_CM_EVENT_DEVICE_REMOVAL:
661 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
664 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
668 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
669 "event=%d\n", cma_id, event->event);
676 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
677 struct rdma_cm_event *event)
679 struct svc_xprt *xprt = cma_id->context;
680 struct svcxprt_rdma *rdma =
681 container_of(xprt, struct svcxprt_rdma, sc_xprt);
682 switch (event->event) {
683 case RDMA_CM_EVENT_ESTABLISHED:
684 /* Accept complete */
686 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
687 "cm_id=%p\n", xprt, cma_id);
688 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
689 svc_xprt_enqueue(xprt);
691 case RDMA_CM_EVENT_DISCONNECTED:
692 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
695 set_bit(XPT_CLOSE, &xprt->xpt_flags);
696 svc_xprt_enqueue(xprt);
700 case RDMA_CM_EVENT_DEVICE_REMOVAL:
701 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
702 "event=%d\n", cma_id, xprt, event->event);
704 set_bit(XPT_CLOSE, &xprt->xpt_flags);
705 svc_xprt_enqueue(xprt);
709 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
710 "event=%d\n", cma_id, event->event);
717 * Create a listening RDMA service endpoint.
719 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
720 struct sockaddr *sa, int salen,
723 struct rdma_cm_id *listen_id;
724 struct svcxprt_rdma *cma_xprt;
725 struct svc_xprt *xprt;
728 dprintk("svcrdma: Creating RDMA socket\n");
730 cma_xprt = rdma_create_xprt(serv, 1);
732 return ERR_PTR(-ENOMEM);
733 xprt = &cma_xprt->sc_xprt;
735 listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
736 if (IS_ERR(listen_id)) {
737 ret = PTR_ERR(listen_id);
738 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
742 ret = rdma_bind_addr(listen_id, sa);
744 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
747 cma_xprt->sc_cm_id = listen_id;
749 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
751 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
756 * We need to use the address from the cm_id in case the
757 * caller specified 0 for the port number.
759 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
760 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
762 return &cma_xprt->sc_xprt;
765 rdma_destroy_id(listen_id);
772 * This is the xpo_recvfrom function for listening endpoints. Its
773 * purpose is to accept incoming connections. The CMA callback handler
774 * has already created a new transport and attached it to the new CMA
777 * There is a queue of pending connections hung on the listening
778 * transport. This queue contains the new svc_xprt structure. This
779 * function takes svc_xprt structures off the accept_q and completes
782 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
784 struct svcxprt_rdma *listen_rdma;
785 struct svcxprt_rdma *newxprt = NULL;
786 struct rdma_conn_param conn_param;
787 struct ib_qp_init_attr qp_attr;
788 struct ib_device_attr devattr;
792 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
793 clear_bit(XPT_CONN, &xprt->xpt_flags);
794 /* Get the next entry off the accept list */
795 spin_lock_bh(&listen_rdma->sc_lock);
796 if (!list_empty(&listen_rdma->sc_accept_q)) {
797 newxprt = list_entry(listen_rdma->sc_accept_q.next,
798 struct svcxprt_rdma, sc_accept_q);
799 list_del_init(&newxprt->sc_accept_q);
801 if (!list_empty(&listen_rdma->sc_accept_q))
802 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
803 spin_unlock_bh(&listen_rdma->sc_lock);
807 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
808 newxprt, newxprt->sc_cm_id);
810 ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
812 dprintk("svcrdma: could not query device attributes on "
813 "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
817 /* Qualify the transport resource defaults with the
818 * capabilities of this particular device */
819 newxprt->sc_max_sge = min((size_t)devattr.max_sge,
820 (size_t)RPCSVC_MAXPAGES);
821 newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
822 (size_t)svcrdma_max_requests);
823 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
825 newxprt->sc_ord = min((size_t)devattr.max_qp_rd_atom,
826 (size_t)svcrdma_ord);
828 newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
829 if (IS_ERR(newxprt->sc_pd)) {
830 dprintk("svcrdma: error creating PD for connect request\n");
833 newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
837 newxprt->sc_sq_depth,
839 if (IS_ERR(newxprt->sc_sq_cq)) {
840 dprintk("svcrdma: error creating SQ CQ for connect request\n");
843 newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
847 newxprt->sc_max_requests,
849 if (IS_ERR(newxprt->sc_rq_cq)) {
850 dprintk("svcrdma: error creating RQ CQ for connect request\n");
854 memset(&qp_attr, 0, sizeof qp_attr);
855 qp_attr.event_handler = qp_event_handler;
856 qp_attr.qp_context = &newxprt->sc_xprt;
857 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
858 qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
859 qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
860 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
861 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
862 qp_attr.qp_type = IB_QPT_RC;
863 qp_attr.send_cq = newxprt->sc_sq_cq;
864 qp_attr.recv_cq = newxprt->sc_rq_cq;
865 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
866 " cm_id->device=%p, sc_pd->device=%p\n"
867 " cap.max_send_wr = %d\n"
868 " cap.max_recv_wr = %d\n"
869 " cap.max_send_sge = %d\n"
870 " cap.max_recv_sge = %d\n",
871 newxprt->sc_cm_id, newxprt->sc_pd,
872 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
873 qp_attr.cap.max_send_wr,
874 qp_attr.cap.max_recv_wr,
875 qp_attr.cap.max_send_sge,
876 qp_attr.cap.max_recv_sge);
878 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
881 * XXX: This is a hack. We need a xx_request_qp interface
882 * that will adjust the qp_attr's with a best-effort
885 qp_attr.cap.max_send_sge -= 2;
886 qp_attr.cap.max_recv_sge -= 2;
887 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
890 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
893 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
894 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
895 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
896 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
898 newxprt->sc_qp = newxprt->sc_cm_id->qp;
900 /* Register all of physical memory */
901 newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
902 IB_ACCESS_LOCAL_WRITE |
903 IB_ACCESS_REMOTE_WRITE);
904 if (IS_ERR(newxprt->sc_phys_mr)) {
905 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
909 /* Post receive buffers */
910 for (i = 0; i < newxprt->sc_max_requests; i++) {
911 ret = svc_rdma_post_recv(newxprt);
913 dprintk("svcrdma: failure posting receive buffers\n");
918 /* Swap out the handler */
919 newxprt->sc_cm_id->event_handler = rdma_cma_handler;
922 * Arm the CQs for the SQ and RQ before accepting so we can't
923 * miss the first message
925 ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
926 ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
928 /* Accept Connection */
929 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
930 memset(&conn_param, 0, sizeof conn_param);
931 conn_param.responder_resources = 0;
932 conn_param.initiator_depth = newxprt->sc_ord;
933 ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
935 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
940 dprintk("svcrdma: new connection %p accepted with the following "
942 " local_ip : %d.%d.%d.%d\n"
944 " remote_ip : %d.%d.%d.%d\n"
945 " remote_port : %d\n"
948 " max_requests : %d\n"
951 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
952 route.addr.src_addr)->sin_addr.s_addr),
953 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
954 route.addr.src_addr)->sin_port),
955 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
956 route.addr.dst_addr)->sin_addr.s_addr),
957 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
958 route.addr.dst_addr)->sin_port),
960 newxprt->sc_sq_depth,
961 newxprt->sc_max_requests,
964 return &newxprt->sc_xprt;
967 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
968 /* Take a reference in case the DTO handler runs */
969 svc_xprt_get(&newxprt->sc_xprt);
970 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
971 ib_destroy_qp(newxprt->sc_qp);
972 rdma_destroy_id(newxprt->sc_cm_id);
973 /* This call to put will destroy the transport */
974 svc_xprt_put(&newxprt->sc_xprt);
978 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
983 * When connected, an svc_xprt has at least two references:
985 * - A reference held by the cm_id between the ESTABLISHED and
986 * DISCONNECTED events. If the remote peer disconnected first, this
987 * reference could be gone.
989 * - A reference held by the svc_recv code that called this function
990 * as part of close processing.
992 * At a minimum one references should still be held.
994 static void svc_rdma_detach(struct svc_xprt *xprt)
996 struct svcxprt_rdma *rdma =
997 container_of(xprt, struct svcxprt_rdma, sc_xprt);
998 dprintk("svc: svc_rdma_detach(%p)\n", xprt);
1000 /* Disconnect and flush posted WQE */
1001 rdma_disconnect(rdma->sc_cm_id);
1004 static void __svc_rdma_free(struct work_struct *work)
1006 struct svcxprt_rdma *rdma =
1007 container_of(work, struct svcxprt_rdma, sc_work);
1008 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
1010 /* We should only be called from kref_put */
1011 BUG_ON(atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0);
1014 * Destroy queued, but not processed read completions. Note
1015 * that this cleanup has to be done before destroying the
1016 * cm_id because the device ptr is needed to unmap the dma in
1017 * svc_rdma_put_context.
1019 spin_lock_bh(&rdma->sc_read_complete_lock);
1020 while (!list_empty(&rdma->sc_read_complete_q)) {
1021 struct svc_rdma_op_ctxt *ctxt;
1022 ctxt = list_entry(rdma->sc_read_complete_q.next,
1023 struct svc_rdma_op_ctxt,
1025 list_del_init(&ctxt->dto_q);
1026 svc_rdma_put_context(ctxt, 1);
1028 spin_unlock_bh(&rdma->sc_read_complete_lock);
1030 /* Destroy queued, but not processed recv completions */
1031 spin_lock_bh(&rdma->sc_rq_dto_lock);
1032 while (!list_empty(&rdma->sc_rq_dto_q)) {
1033 struct svc_rdma_op_ctxt *ctxt;
1034 ctxt = list_entry(rdma->sc_rq_dto_q.next,
1035 struct svc_rdma_op_ctxt,
1037 list_del_init(&ctxt->dto_q);
1038 svc_rdma_put_context(ctxt, 1);
1040 spin_unlock_bh(&rdma->sc_rq_dto_lock);
1042 /* Warn if we leaked a resource or under-referenced */
1043 WARN_ON(atomic_read(&rdma->sc_ctxt_used) != 0);
1045 /* Destroy the QP if present (not a listener) */
1046 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1047 ib_destroy_qp(rdma->sc_qp);
1049 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
1050 ib_destroy_cq(rdma->sc_sq_cq);
1052 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
1053 ib_destroy_cq(rdma->sc_rq_cq);
1055 if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
1056 ib_dereg_mr(rdma->sc_phys_mr);
1058 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
1059 ib_dealloc_pd(rdma->sc_pd);
1061 /* Destroy the CM ID */
1062 rdma_destroy_id(rdma->sc_cm_id);
1064 destroy_context_cache(rdma);
1068 static void svc_rdma_free(struct svc_xprt *xprt)
1070 struct svcxprt_rdma *rdma =
1071 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1072 INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1073 schedule_work(&rdma->sc_work);
1076 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1078 struct svcxprt_rdma *rdma =
1079 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1082 * If there are fewer SQ WR available than required to send a
1083 * simple response, return false.
1085 if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
1089 * ...or there are already waiters on the SQ,
1092 if (waitqueue_active(&rdma->sc_send_wait))
1095 /* Otherwise return true. */
1099 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1101 struct ib_send_wr *bad_wr;
1104 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1107 BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1108 BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1110 /* If the SQ is full, wait until an SQ entry is available */
1112 spin_lock_bh(&xprt->sc_lock);
1113 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1114 spin_unlock_bh(&xprt->sc_lock);
1115 atomic_inc(&rdma_stat_sq_starve);
1117 /* See if we can opportunistically reap SQ WR to make room */
1120 /* Wait until SQ WR available if SQ still full */
1121 wait_event(xprt->sc_send_wait,
1122 atomic_read(&xprt->sc_sq_count) <
1124 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1128 /* Bumped used SQ WR count and post */
1129 svc_xprt_get(&xprt->sc_xprt);
1130 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1132 atomic_inc(&xprt->sc_sq_count);
1134 svc_xprt_put(&xprt->sc_xprt);
1135 dprintk("svcrdma: failed to post SQ WR rc=%d, "
1136 "sc_sq_count=%d, sc_sq_depth=%d\n",
1137 ret, atomic_read(&xprt->sc_sq_count),
1140 spin_unlock_bh(&xprt->sc_lock);
1146 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1147 enum rpcrdma_errcode err)
1149 struct ib_send_wr err_wr;
1152 struct svc_rdma_op_ctxt *ctxt;
1157 p = svc_rdma_get_page();
1158 va = page_address(p);
1160 /* XDR encode error */
1161 length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1163 /* Prepare SGE for local address */
1164 sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1165 p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1166 sge.lkey = xprt->sc_phys_mr->lkey;
1167 sge.length = length;
1169 ctxt = svc_rdma_get_context(xprt);
1173 /* Prepare SEND WR */
1174 memset(&err_wr, 0, sizeof err_wr);
1175 ctxt->wr_op = IB_WR_SEND;
1176 err_wr.wr_id = (unsigned long)ctxt;
1177 err_wr.sg_list = &sge;
1179 err_wr.opcode = IB_WR_SEND;
1180 err_wr.send_flags = IB_SEND_SIGNALED;
1183 ret = svc_rdma_send(xprt, &err_wr);
1185 dprintk("svcrdma: Error %d posting send for protocol error\n",
1187 svc_rdma_put_context(ctxt, 1);