2 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
3 * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the BSD-type
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
15 * Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
18 * Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials provided
21 * with the distribution.
23 * Neither the name of the Network Appliance, Inc. nor the names of
24 * its contributors may be used to endorse or promote products
25 * derived from this software without specific prior written
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 * Author: Tom Tucker <tom@opengridcomputing.com>
43 #include <linux/sunrpc/svc_xprt.h>
44 #include <linux/sunrpc/debug.h>
45 #include <linux/sunrpc/rpc_rdma.h>
46 #include <linux/interrupt.h>
47 #include <linux/sched.h>
48 #include <linux/slab.h>
49 #include <linux/spinlock.h>
50 #include <linux/workqueue.h>
51 #include <rdma/ib_verbs.h>
52 #include <rdma/rdma_cm.h>
53 #include <linux/sunrpc/svc_rdma.h>
54 #include <linux/export.h>
55 #include "xprt_rdma.h"
57 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
59 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *, int);
60 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
62 struct sockaddr *sa, int salen,
64 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
65 static void svc_rdma_release_rqst(struct svc_rqst *);
66 static void dto_tasklet_func(unsigned long data);
67 static void svc_rdma_detach(struct svc_xprt *xprt);
68 static void svc_rdma_free(struct svc_xprt *xprt);
69 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
70 static int svc_rdma_secure_port(struct svc_rqst *);
71 static void rq_cq_reap(struct svcxprt_rdma *xprt);
72 static void sq_cq_reap(struct svcxprt_rdma *xprt);
74 static DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL);
75 static DEFINE_SPINLOCK(dto_lock);
76 static LIST_HEAD(dto_xprt_q);
78 static struct svc_xprt_ops svc_rdma_ops = {
79 .xpo_create = svc_rdma_create,
80 .xpo_recvfrom = svc_rdma_recvfrom,
81 .xpo_sendto = svc_rdma_sendto,
82 .xpo_release_rqst = svc_rdma_release_rqst,
83 .xpo_detach = svc_rdma_detach,
84 .xpo_free = svc_rdma_free,
85 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
86 .xpo_has_wspace = svc_rdma_has_wspace,
87 .xpo_accept = svc_rdma_accept,
88 .xpo_secure_port = svc_rdma_secure_port,
91 struct svc_xprt_class svc_rdma_class = {
93 .xcl_owner = THIS_MODULE,
94 .xcl_ops = &svc_rdma_ops,
95 .xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
96 .xcl_ident = XPRT_TRANSPORT_RDMA,
99 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
100 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *, struct net *,
101 struct sockaddr *, int, int);
102 static void svc_rdma_bc_detach(struct svc_xprt *);
103 static void svc_rdma_bc_free(struct svc_xprt *);
105 static struct svc_xprt_ops svc_rdma_bc_ops = {
106 .xpo_create = svc_rdma_bc_create,
107 .xpo_detach = svc_rdma_bc_detach,
108 .xpo_free = svc_rdma_bc_free,
109 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
110 .xpo_secure_port = svc_rdma_secure_port,
113 struct svc_xprt_class svc_rdma_bc_class = {
114 .xcl_name = "rdma-bc",
115 .xcl_owner = THIS_MODULE,
116 .xcl_ops = &svc_rdma_bc_ops,
117 .xcl_max_payload = (1024 - RPCRDMA_HDRLEN_MIN)
120 static struct svc_xprt *svc_rdma_bc_create(struct svc_serv *serv,
122 struct sockaddr *sa, int salen,
125 struct svcxprt_rdma *cma_xprt;
126 struct svc_xprt *xprt;
128 cma_xprt = rdma_create_xprt(serv, 0);
130 return ERR_PTR(-ENOMEM);
131 xprt = &cma_xprt->sc_xprt;
133 svc_xprt_init(net, &svc_rdma_bc_class, xprt, serv);
134 serv->sv_bc_xprt = xprt;
136 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
140 static void svc_rdma_bc_detach(struct svc_xprt *xprt)
142 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
145 static void svc_rdma_bc_free(struct svc_xprt *xprt)
147 struct svcxprt_rdma *rdma =
148 container_of(xprt, struct svcxprt_rdma, sc_xprt);
150 dprintk("svcrdma: %s(%p)\n", __func__, xprt);
154 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
156 static struct svc_rdma_op_ctxt *alloc_ctxt(struct svcxprt_rdma *xprt,
159 struct svc_rdma_op_ctxt *ctxt;
161 ctxt = kmalloc(sizeof(*ctxt), flags);
164 INIT_LIST_HEAD(&ctxt->free);
165 INIT_LIST_HEAD(&ctxt->dto_q);
170 static bool svc_rdma_prealloc_ctxts(struct svcxprt_rdma *xprt)
174 /* Each RPC/RDMA credit can consume a number of send
175 * and receive WQEs. One ctxt is allocated for each.
177 i = xprt->sc_sq_depth + xprt->sc_rq_depth;
180 struct svc_rdma_op_ctxt *ctxt;
182 ctxt = alloc_ctxt(xprt, GFP_KERNEL);
184 dprintk("svcrdma: No memory for RDMA ctxt\n");
187 list_add(&ctxt->free, &xprt->sc_ctxts);
192 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
194 struct svc_rdma_op_ctxt *ctxt = NULL;
196 spin_lock_bh(&xprt->sc_ctxt_lock);
197 xprt->sc_ctxt_used++;
198 if (list_empty(&xprt->sc_ctxts))
201 ctxt = list_first_entry(&xprt->sc_ctxts,
202 struct svc_rdma_op_ctxt, free);
203 list_del_init(&ctxt->free);
204 spin_unlock_bh(&xprt->sc_ctxt_lock);
212 /* Either pre-allocation missed the mark, or send
213 * queue accounting is broken.
215 spin_unlock_bh(&xprt->sc_ctxt_lock);
217 ctxt = alloc_ctxt(xprt, GFP_NOIO);
221 spin_lock_bh(&xprt->sc_ctxt_lock);
222 xprt->sc_ctxt_used--;
223 spin_unlock_bh(&xprt->sc_ctxt_lock);
224 WARN_ONCE(1, "svcrdma: empty RDMA ctxt list?\n");
228 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
230 struct svcxprt_rdma *xprt = ctxt->xprt;
232 for (i = 0; i < ctxt->count && ctxt->sge[i].length; i++) {
234 * Unmap the DMA addr in the SGE if the lkey matches
235 * the sc_dma_lkey, otherwise, ignore it since it is
236 * an FRMR lkey and will be unmapped later when the
237 * last WR that uses it completes.
239 if (ctxt->sge[i].lkey == xprt->sc_dma_lkey) {
240 atomic_dec(&xprt->sc_dma_used);
241 ib_dma_unmap_page(xprt->sc_cm_id->device,
249 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
251 struct svcxprt_rdma *xprt = ctxt->xprt;
255 for (i = 0; i < ctxt->count; i++)
256 put_page(ctxt->pages[i]);
258 spin_lock_bh(&xprt->sc_ctxt_lock);
259 xprt->sc_ctxt_used--;
260 list_add(&ctxt->free, &xprt->sc_ctxts);
261 spin_unlock_bh(&xprt->sc_ctxt_lock);
264 static void svc_rdma_destroy_ctxts(struct svcxprt_rdma *xprt)
266 while (!list_empty(&xprt->sc_ctxts)) {
267 struct svc_rdma_op_ctxt *ctxt;
269 ctxt = list_first_entry(&xprt->sc_ctxts,
270 struct svc_rdma_op_ctxt, free);
271 list_del(&ctxt->free);
276 static struct svc_rdma_req_map *alloc_req_map(gfp_t flags)
278 struct svc_rdma_req_map *map;
280 map = kmalloc(sizeof(*map), flags);
282 INIT_LIST_HEAD(&map->free);
286 static bool svc_rdma_prealloc_maps(struct svcxprt_rdma *xprt)
290 /* One for each receive buffer on this connection. */
291 i = xprt->sc_max_requests;
294 struct svc_rdma_req_map *map;
296 map = alloc_req_map(GFP_KERNEL);
298 dprintk("svcrdma: No memory for request map\n");
301 list_add(&map->free, &xprt->sc_maps);
306 struct svc_rdma_req_map *svc_rdma_get_req_map(struct svcxprt_rdma *xprt)
308 struct svc_rdma_req_map *map = NULL;
310 spin_lock(&xprt->sc_map_lock);
311 if (list_empty(&xprt->sc_maps))
314 map = list_first_entry(&xprt->sc_maps,
315 struct svc_rdma_req_map, free);
316 list_del_init(&map->free);
317 spin_unlock(&xprt->sc_map_lock);
324 spin_unlock(&xprt->sc_map_lock);
326 /* Pre-allocation amount was incorrect */
327 map = alloc_req_map(GFP_NOIO);
331 WARN_ONCE(1, "svcrdma: empty request map list?\n");
335 void svc_rdma_put_req_map(struct svcxprt_rdma *xprt,
336 struct svc_rdma_req_map *map)
338 spin_lock(&xprt->sc_map_lock);
339 list_add(&map->free, &xprt->sc_maps);
340 spin_unlock(&xprt->sc_map_lock);
343 static void svc_rdma_destroy_maps(struct svcxprt_rdma *xprt)
345 while (!list_empty(&xprt->sc_maps)) {
346 struct svc_rdma_req_map *map;
348 map = list_first_entry(&xprt->sc_maps,
349 struct svc_rdma_req_map, free);
350 list_del(&map->free);
355 /* ib_cq event handler */
356 static void cq_event_handler(struct ib_event *event, void *context)
358 struct svc_xprt *xprt = context;
359 dprintk("svcrdma: received CQ event %s (%d), context=%p\n",
360 ib_event_msg(event->event), event->event, context);
361 set_bit(XPT_CLOSE, &xprt->xpt_flags);
364 /* QP event handler */
365 static void qp_event_handler(struct ib_event *event, void *context)
367 struct svc_xprt *xprt = context;
369 switch (event->event) {
370 /* These are considered benign events */
371 case IB_EVENT_PATH_MIG:
372 case IB_EVENT_COMM_EST:
373 case IB_EVENT_SQ_DRAINED:
374 case IB_EVENT_QP_LAST_WQE_REACHED:
375 dprintk("svcrdma: QP event %s (%d) received for QP=%p\n",
376 ib_event_msg(event->event), event->event,
379 /* These are considered fatal events */
380 case IB_EVENT_PATH_MIG_ERR:
381 case IB_EVENT_QP_FATAL:
382 case IB_EVENT_QP_REQ_ERR:
383 case IB_EVENT_QP_ACCESS_ERR:
384 case IB_EVENT_DEVICE_FATAL:
386 dprintk("svcrdma: QP ERROR event %s (%d) received for QP=%p, "
387 "closing transport\n",
388 ib_event_msg(event->event), event->event,
390 set_bit(XPT_CLOSE, &xprt->xpt_flags);
396 * Data Transfer Operation Tasklet
398 * Walks a list of transports with I/O pending, removing entries as
399 * they are added to the server's I/O pending list. Two bits indicate
400 * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
401 * spinlock that serializes access to the transport list with the RQ
402 * and SQ interrupt handlers.
404 static void dto_tasklet_func(unsigned long data)
406 struct svcxprt_rdma *xprt;
409 spin_lock_irqsave(&dto_lock, flags);
410 while (!list_empty(&dto_xprt_q)) {
411 xprt = list_entry(dto_xprt_q.next,
412 struct svcxprt_rdma, sc_dto_q);
413 list_del_init(&xprt->sc_dto_q);
414 spin_unlock_irqrestore(&dto_lock, flags);
419 svc_xprt_put(&xprt->sc_xprt);
420 spin_lock_irqsave(&dto_lock, flags);
422 spin_unlock_irqrestore(&dto_lock, flags);
426 * Receive Queue Completion Handler
428 * Since an RQ completion handler is called on interrupt context, we
429 * need to defer the handling of the I/O to a tasklet
431 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
433 struct svcxprt_rdma *xprt = cq_context;
436 /* Guard against unconditional flush call for destroyed QP */
437 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
441 * Set the bit regardless of whether or not it's on the list
442 * because it may be on the list already due to an SQ
445 set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
448 * If this transport is not already on the DTO transport queue,
451 spin_lock_irqsave(&dto_lock, flags);
452 if (list_empty(&xprt->sc_dto_q)) {
453 svc_xprt_get(&xprt->sc_xprt);
454 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
456 spin_unlock_irqrestore(&dto_lock, flags);
458 /* Tasklet does all the work to avoid irqsave locks. */
459 tasklet_schedule(&dto_tasklet);
463 * rq_cq_reap - Process the RQ CQ.
465 * Take all completing WC off the CQE and enqueue the associated DTO
466 * context on the dto_q for the transport.
468 * Note that caller must hold a transport reference.
470 static void rq_cq_reap(struct svcxprt_rdma *xprt)
474 struct svc_rdma_op_ctxt *ctxt = NULL;
476 if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
479 ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
480 atomic_inc(&rdma_stat_rq_poll);
482 while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
483 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
484 ctxt->wc_status = wc.status;
485 ctxt->byte_len = wc.byte_len;
486 svc_rdma_unmap_dma(ctxt);
487 if (wc.status != IB_WC_SUCCESS) {
488 /* Close the transport */
489 dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt);
490 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
491 svc_rdma_put_context(ctxt, 1);
492 svc_xprt_put(&xprt->sc_xprt);
495 spin_lock_bh(&xprt->sc_rq_dto_lock);
496 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
497 spin_unlock_bh(&xprt->sc_rq_dto_lock);
498 svc_xprt_put(&xprt->sc_xprt);
502 atomic_inc(&rdma_stat_rq_prod);
504 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
506 * If data arrived before established event,
507 * don't enqueue. This defers RPC I/O until the
508 * RDMA connection is complete.
510 if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
511 svc_xprt_enqueue(&xprt->sc_xprt);
515 * Process a completion context
517 static void process_context(struct svcxprt_rdma *xprt,
518 struct svc_rdma_op_ctxt *ctxt)
520 struct svc_rdma_op_ctxt *read_hdr;
523 svc_rdma_unmap_dma(ctxt);
525 switch (ctxt->wr_op) {
530 case IB_WR_RDMA_WRITE:
533 case IB_WR_RDMA_READ:
534 case IB_WR_RDMA_READ_WITH_INV:
535 svc_rdma_put_frmr(xprt, ctxt->frmr);
537 if (!test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags))
540 read_hdr = ctxt->read_hdr;
541 svc_rdma_put_context(ctxt, 0);
543 spin_lock_bh(&xprt->sc_rq_dto_lock);
544 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
545 list_add_tail(&read_hdr->dto_q,
546 &xprt->sc_read_complete_q);
547 spin_unlock_bh(&xprt->sc_rq_dto_lock);
548 svc_xprt_enqueue(&xprt->sc_xprt);
552 dprintk("svcrdma: unexpected completion opcode=%d\n",
557 svc_rdma_put_context(ctxt, free_pages);
561 * Send Queue Completion Handler - potentially called on interrupt context.
563 * Note that caller must hold a transport reference.
565 static void sq_cq_reap(struct svcxprt_rdma *xprt)
567 struct svc_rdma_op_ctxt *ctxt = NULL;
568 struct ib_wc wc_a[6];
570 struct ib_cq *cq = xprt->sc_sq_cq;
573 memset(wc_a, 0, sizeof(wc_a));
575 if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
578 ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
579 atomic_inc(&rdma_stat_sq_poll);
580 while ((ret = ib_poll_cq(cq, ARRAY_SIZE(wc_a), wc_a)) > 0) {
583 for (i = 0; i < ret; i++) {
585 if (wc->status != IB_WC_SUCCESS) {
586 dprintk("svcrdma: sq wc err status %s (%d)\n",
587 ib_wc_status_msg(wc->status),
590 /* Close the transport */
591 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
594 /* Decrement used SQ WR count */
595 atomic_dec(&xprt->sc_sq_count);
596 wake_up(&xprt->sc_send_wait);
598 ctxt = (struct svc_rdma_op_ctxt *)
599 (unsigned long)wc->wr_id;
601 process_context(xprt, ctxt);
603 svc_xprt_put(&xprt->sc_xprt);
608 atomic_inc(&rdma_stat_sq_prod);
611 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
613 struct svcxprt_rdma *xprt = cq_context;
616 /* Guard against unconditional flush call for destroyed QP */
617 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
621 * Set the bit regardless of whether or not it's on the list
622 * because it may be on the list already due to an RQ
625 set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
628 * If this transport is not already on the DTO transport queue,
631 spin_lock_irqsave(&dto_lock, flags);
632 if (list_empty(&xprt->sc_dto_q)) {
633 svc_xprt_get(&xprt->sc_xprt);
634 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
636 spin_unlock_irqrestore(&dto_lock, flags);
638 /* Tasklet does all the work to avoid irqsave locks. */
639 tasklet_schedule(&dto_tasklet);
642 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
645 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
649 svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
650 INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
651 INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
652 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
653 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
654 INIT_LIST_HEAD(&cma_xprt->sc_frmr_q);
655 INIT_LIST_HEAD(&cma_xprt->sc_ctxts);
656 INIT_LIST_HEAD(&cma_xprt->sc_maps);
657 init_waitqueue_head(&cma_xprt->sc_send_wait);
659 spin_lock_init(&cma_xprt->sc_lock);
660 spin_lock_init(&cma_xprt->sc_rq_dto_lock);
661 spin_lock_init(&cma_xprt->sc_frmr_q_lock);
662 spin_lock_init(&cma_xprt->sc_ctxt_lock);
663 spin_lock_init(&cma_xprt->sc_map_lock);
666 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
671 int svc_rdma_post_recv(struct svcxprt_rdma *xprt, gfp_t flags)
673 struct ib_recv_wr recv_wr, *bad_recv_wr;
674 struct svc_rdma_op_ctxt *ctxt;
681 ctxt = svc_rdma_get_context(xprt);
683 ctxt->direction = DMA_FROM_DEVICE;
684 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
685 if (sge_no >= xprt->sc_max_sge) {
686 pr_err("svcrdma: Too many sges (%d)\n", sge_no);
689 page = alloc_page(flags);
692 ctxt->pages[sge_no] = page;
693 pa = ib_dma_map_page(xprt->sc_cm_id->device,
696 if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa))
698 atomic_inc(&xprt->sc_dma_used);
699 ctxt->sge[sge_no].addr = pa;
700 ctxt->sge[sge_no].length = PAGE_SIZE;
701 ctxt->sge[sge_no].lkey = xprt->sc_dma_lkey;
702 ctxt->count = sge_no + 1;
706 recv_wr.sg_list = &ctxt->sge[0];
707 recv_wr.num_sge = ctxt->count;
708 recv_wr.wr_id = (u64)(unsigned long)ctxt;
710 svc_xprt_get(&xprt->sc_xprt);
711 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
713 svc_rdma_unmap_dma(ctxt);
714 svc_rdma_put_context(ctxt, 1);
715 svc_xprt_put(&xprt->sc_xprt);
720 svc_rdma_unmap_dma(ctxt);
721 svc_rdma_put_context(ctxt, 1);
726 * This function handles the CONNECT_REQUEST event on a listening
727 * endpoint. It is passed the cma_id for the _new_ connection. The context in
728 * this cma_id is inherited from the listening cma_id and is the svc_xprt
729 * structure for the listening endpoint.
731 * This function creates a new xprt for the new connection and enqueues it on
732 * the accept queue for the listent xprt. When the listen thread is kicked, it
733 * will call the recvfrom method on the listen xprt which will accept the new
736 static void handle_connect_req(struct rdma_cm_id *new_cma_id, size_t client_ird)
738 struct svcxprt_rdma *listen_xprt = new_cma_id->context;
739 struct svcxprt_rdma *newxprt;
742 /* Create a new transport */
743 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
745 dprintk("svcrdma: failed to create new transport\n");
748 newxprt->sc_cm_id = new_cma_id;
749 new_cma_id->context = newxprt;
750 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
751 newxprt, newxprt->sc_cm_id, listen_xprt);
753 /* Save client advertised inbound read limit for use later in accept. */
754 newxprt->sc_ord = client_ird;
756 /* Set the local and remote addresses in the transport */
757 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
758 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
759 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
760 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
763 * Enqueue the new transport on the accept queue of the listening
766 spin_lock_bh(&listen_xprt->sc_lock);
767 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
768 spin_unlock_bh(&listen_xprt->sc_lock);
770 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
771 svc_xprt_enqueue(&listen_xprt->sc_xprt);
775 * Handles events generated on the listening endpoint. These events will be
776 * either be incoming connect requests or adapter removal events.
778 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
779 struct rdma_cm_event *event)
781 struct svcxprt_rdma *xprt = cma_id->context;
784 switch (event->event) {
785 case RDMA_CM_EVENT_CONNECT_REQUEST:
786 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
787 "event = %s (%d)\n", cma_id, cma_id->context,
788 rdma_event_msg(event->event), event->event);
789 handle_connect_req(cma_id,
790 event->param.conn.initiator_depth);
793 case RDMA_CM_EVENT_ESTABLISHED:
794 /* Accept complete */
795 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
796 "cm_id=%p\n", xprt, cma_id);
799 case RDMA_CM_EVENT_DEVICE_REMOVAL:
800 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
803 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
807 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
808 "event = %s (%d)\n", cma_id,
809 rdma_event_msg(event->event), event->event);
816 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
817 struct rdma_cm_event *event)
819 struct svc_xprt *xprt = cma_id->context;
820 struct svcxprt_rdma *rdma =
821 container_of(xprt, struct svcxprt_rdma, sc_xprt);
822 switch (event->event) {
823 case RDMA_CM_EVENT_ESTABLISHED:
824 /* Accept complete */
826 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
827 "cm_id=%p\n", xprt, cma_id);
828 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
829 svc_xprt_enqueue(xprt);
831 case RDMA_CM_EVENT_DISCONNECTED:
832 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
835 set_bit(XPT_CLOSE, &xprt->xpt_flags);
836 svc_xprt_enqueue(xprt);
840 case RDMA_CM_EVENT_DEVICE_REMOVAL:
841 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
842 "event = %s (%d)\n", cma_id, xprt,
843 rdma_event_msg(event->event), event->event);
845 set_bit(XPT_CLOSE, &xprt->xpt_flags);
846 svc_xprt_enqueue(xprt);
851 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
852 "event = %s (%d)\n", cma_id,
853 rdma_event_msg(event->event), event->event);
860 * Create a listening RDMA service endpoint.
862 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
864 struct sockaddr *sa, int salen,
867 struct rdma_cm_id *listen_id;
868 struct svcxprt_rdma *cma_xprt;
871 dprintk("svcrdma: Creating RDMA socket\n");
872 if (sa->sa_family != AF_INET) {
873 dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family);
874 return ERR_PTR(-EAFNOSUPPORT);
876 cma_xprt = rdma_create_xprt(serv, 1);
878 return ERR_PTR(-ENOMEM);
880 listen_id = rdma_create_id(&init_net, rdma_listen_handler, cma_xprt,
881 RDMA_PS_TCP, IB_QPT_RC);
882 if (IS_ERR(listen_id)) {
883 ret = PTR_ERR(listen_id);
884 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
888 ret = rdma_bind_addr(listen_id, sa);
890 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
893 cma_xprt->sc_cm_id = listen_id;
895 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
897 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
902 * We need to use the address from the cm_id in case the
903 * caller specified 0 for the port number.
905 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
906 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
908 return &cma_xprt->sc_xprt;
911 rdma_destroy_id(listen_id);
917 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt)
920 struct scatterlist *sg;
921 struct svc_rdma_fastreg_mr *frmr;
924 frmr = kmalloc(sizeof(*frmr), GFP_KERNEL);
928 num_sg = min_t(u32, RPCSVC_MAXPAGES, xprt->sc_frmr_pg_list_len);
929 mr = ib_alloc_mr(xprt->sc_pd, IB_MR_TYPE_MEM_REG, num_sg);
933 sg = kcalloc(RPCSVC_MAXPAGES, sizeof(*sg), GFP_KERNEL);
937 sg_init_table(sg, RPCSVC_MAXPAGES);
941 INIT_LIST_HEAD(&frmr->frmr_list);
949 return ERR_PTR(-ENOMEM);
952 static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt)
954 struct svc_rdma_fastreg_mr *frmr;
956 while (!list_empty(&xprt->sc_frmr_q)) {
957 frmr = list_entry(xprt->sc_frmr_q.next,
958 struct svc_rdma_fastreg_mr, frmr_list);
959 list_del_init(&frmr->frmr_list);
961 ib_dereg_mr(frmr->mr);
966 struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma)
968 struct svc_rdma_fastreg_mr *frmr = NULL;
970 spin_lock_bh(&rdma->sc_frmr_q_lock);
971 if (!list_empty(&rdma->sc_frmr_q)) {
972 frmr = list_entry(rdma->sc_frmr_q.next,
973 struct svc_rdma_fastreg_mr, frmr_list);
974 list_del_init(&frmr->frmr_list);
977 spin_unlock_bh(&rdma->sc_frmr_q_lock);
981 return rdma_alloc_frmr(rdma);
984 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma,
985 struct svc_rdma_fastreg_mr *frmr)
988 ib_dma_unmap_sg(rdma->sc_cm_id->device,
989 frmr->sg, frmr->sg_nents, frmr->direction);
990 atomic_dec(&rdma->sc_dma_used);
991 spin_lock_bh(&rdma->sc_frmr_q_lock);
992 WARN_ON_ONCE(!list_empty(&frmr->frmr_list));
993 list_add(&frmr->frmr_list, &rdma->sc_frmr_q);
994 spin_unlock_bh(&rdma->sc_frmr_q_lock);
999 * This is the xpo_recvfrom function for listening endpoints. Its
1000 * purpose is to accept incoming connections. The CMA callback handler
1001 * has already created a new transport and attached it to the new CMA
1004 * There is a queue of pending connections hung on the listening
1005 * transport. This queue contains the new svc_xprt structure. This
1006 * function takes svc_xprt structures off the accept_q and completes
1009 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
1011 struct svcxprt_rdma *listen_rdma;
1012 struct svcxprt_rdma *newxprt = NULL;
1013 struct rdma_conn_param conn_param;
1014 struct ib_cq_init_attr cq_attr = {};
1015 struct ib_qp_init_attr qp_attr;
1016 struct ib_device *dev;
1017 int uninitialized_var(dma_mr_acc);
1018 int need_dma_mr = 0;
1022 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
1023 clear_bit(XPT_CONN, &xprt->xpt_flags);
1024 /* Get the next entry off the accept list */
1025 spin_lock_bh(&listen_rdma->sc_lock);
1026 if (!list_empty(&listen_rdma->sc_accept_q)) {
1027 newxprt = list_entry(listen_rdma->sc_accept_q.next,
1028 struct svcxprt_rdma, sc_accept_q);
1029 list_del_init(&newxprt->sc_accept_q);
1031 if (!list_empty(&listen_rdma->sc_accept_q))
1032 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
1033 spin_unlock_bh(&listen_rdma->sc_lock);
1037 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
1038 newxprt, newxprt->sc_cm_id);
1040 dev = newxprt->sc_cm_id->device;
1042 /* Qualify the transport resource defaults with the
1043 * capabilities of this particular device */
1044 newxprt->sc_max_sge = min((size_t)dev->attrs.max_sge,
1045 (size_t)RPCSVC_MAXPAGES);
1046 newxprt->sc_max_sge_rd = min_t(size_t, dev->attrs.max_sge_rd,
1048 newxprt->sc_max_req_size = svcrdma_max_req_size;
1049 newxprt->sc_max_requests = min_t(u32, dev->attrs.max_qp_wr,
1050 svcrdma_max_requests);
1051 newxprt->sc_max_bc_requests = min_t(u32, dev->attrs.max_qp_wr,
1052 svcrdma_max_bc_requests);
1053 newxprt->sc_rq_depth = newxprt->sc_max_requests +
1054 newxprt->sc_max_bc_requests;
1055 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_rq_depth;
1057 if (!svc_rdma_prealloc_ctxts(newxprt))
1059 if (!svc_rdma_prealloc_maps(newxprt))
1063 * Limit ORD based on client limit, local device limit, and
1064 * configured svcrdma limit.
1066 newxprt->sc_ord = min_t(size_t, dev->attrs.max_qp_rd_atom, newxprt->sc_ord);
1067 newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord);
1069 newxprt->sc_pd = ib_alloc_pd(dev);
1070 if (IS_ERR(newxprt->sc_pd)) {
1071 dprintk("svcrdma: error creating PD for connect request\n");
1074 cq_attr.cqe = newxprt->sc_sq_depth;
1075 newxprt->sc_sq_cq = ib_create_cq(dev,
1080 if (IS_ERR(newxprt->sc_sq_cq)) {
1081 dprintk("svcrdma: error creating SQ CQ for connect request\n");
1084 cq_attr.cqe = newxprt->sc_rq_depth;
1085 newxprt->sc_rq_cq = ib_create_cq(dev,
1090 if (IS_ERR(newxprt->sc_rq_cq)) {
1091 dprintk("svcrdma: error creating RQ CQ for connect request\n");
1095 memset(&qp_attr, 0, sizeof qp_attr);
1096 qp_attr.event_handler = qp_event_handler;
1097 qp_attr.qp_context = &newxprt->sc_xprt;
1098 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
1099 qp_attr.cap.max_recv_wr = newxprt->sc_rq_depth;
1100 qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
1101 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
1102 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1103 qp_attr.qp_type = IB_QPT_RC;
1104 qp_attr.send_cq = newxprt->sc_sq_cq;
1105 qp_attr.recv_cq = newxprt->sc_rq_cq;
1106 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
1107 " cm_id->device=%p, sc_pd->device=%p\n"
1108 " cap.max_send_wr = %d\n"
1109 " cap.max_recv_wr = %d\n"
1110 " cap.max_send_sge = %d\n"
1111 " cap.max_recv_sge = %d\n",
1112 newxprt->sc_cm_id, newxprt->sc_pd,
1113 dev, newxprt->sc_pd->device,
1114 qp_attr.cap.max_send_wr,
1115 qp_attr.cap.max_recv_wr,
1116 qp_attr.cap.max_send_sge,
1117 qp_attr.cap.max_recv_sge);
1119 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
1121 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
1124 newxprt->sc_qp = newxprt->sc_cm_id->qp;
1127 * Use the most secure set of MR resources based on the
1128 * transport type and available memory management features in
1129 * the device. Here's the table implemented below:
1131 * Fast Global DMA Remote WR
1132 * Reg LKEY MR Access
1133 * Sup'd Sup'd Needed Needed
1145 * NB: iWARP requires remote write access for the data sink
1146 * of an RDMA_READ. IB does not.
1148 newxprt->sc_reader = rdma_read_chunk_lcl;
1149 if (dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1150 newxprt->sc_frmr_pg_list_len =
1151 dev->attrs.max_fast_reg_page_list_len;
1152 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG;
1153 newxprt->sc_reader = rdma_read_chunk_frmr;
1157 * Determine if a DMA MR is required and if so, what privs are required
1159 if (!rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num) &&
1160 !rdma_ib_or_roce(dev, newxprt->sc_cm_id->port_num))
1163 if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG) ||
1164 !(dev->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) {
1166 dma_mr_acc = IB_ACCESS_LOCAL_WRITE;
1167 if (rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num) &&
1168 !(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG))
1169 dma_mr_acc |= IB_ACCESS_REMOTE_WRITE;
1172 if (rdma_protocol_iwarp(dev, newxprt->sc_cm_id->port_num))
1173 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV;
1175 /* Create the DMA MR if needed, otherwise, use the DMA LKEY */
1177 /* Register all of physical memory */
1178 newxprt->sc_phys_mr =
1179 ib_get_dma_mr(newxprt->sc_pd, dma_mr_acc);
1180 if (IS_ERR(newxprt->sc_phys_mr)) {
1181 dprintk("svcrdma: Failed to create DMA MR ret=%d\n",
1185 newxprt->sc_dma_lkey = newxprt->sc_phys_mr->lkey;
1187 newxprt->sc_dma_lkey = dev->local_dma_lkey;
1189 /* Post receive buffers */
1190 for (i = 0; i < newxprt->sc_rq_depth; i++) {
1191 ret = svc_rdma_post_recv(newxprt, GFP_KERNEL);
1193 dprintk("svcrdma: failure posting receive buffers\n");
1198 /* Swap out the handler */
1199 newxprt->sc_cm_id->event_handler = rdma_cma_handler;
1202 * Arm the CQs for the SQ and RQ before accepting so we can't
1203 * miss the first message
1205 ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
1206 ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
1208 /* Accept Connection */
1209 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
1210 memset(&conn_param, 0, sizeof conn_param);
1211 conn_param.responder_resources = 0;
1212 conn_param.initiator_depth = newxprt->sc_ord;
1213 ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
1215 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
1220 dprintk("svcrdma: new connection %p accepted with the following "
1222 " local_ip : %pI4\n"
1223 " local_port : %d\n"
1224 " remote_ip : %pI4\n"
1225 " remote_port : %d\n"
1227 " max_sge_rd : %d\n"
1229 " max_requests : %d\n"
1232 &((struct sockaddr_in *)&newxprt->sc_cm_id->
1233 route.addr.src_addr)->sin_addr.s_addr,
1234 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
1235 route.addr.src_addr)->sin_port),
1236 &((struct sockaddr_in *)&newxprt->sc_cm_id->
1237 route.addr.dst_addr)->sin_addr.s_addr,
1238 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
1239 route.addr.dst_addr)->sin_port),
1240 newxprt->sc_max_sge,
1241 newxprt->sc_max_sge_rd,
1242 newxprt->sc_sq_depth,
1243 newxprt->sc_max_requests,
1246 return &newxprt->sc_xprt;
1249 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
1250 /* Take a reference in case the DTO handler runs */
1251 svc_xprt_get(&newxprt->sc_xprt);
1252 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
1253 ib_destroy_qp(newxprt->sc_qp);
1254 rdma_destroy_id(newxprt->sc_cm_id);
1255 /* This call to put will destroy the transport */
1256 svc_xprt_put(&newxprt->sc_xprt);
1260 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
1265 * When connected, an svc_xprt has at least two references:
1267 * - A reference held by the cm_id between the ESTABLISHED and
1268 * DISCONNECTED events. If the remote peer disconnected first, this
1269 * reference could be gone.
1271 * - A reference held by the svc_recv code that called this function
1272 * as part of close processing.
1274 * At a minimum one references should still be held.
1276 static void svc_rdma_detach(struct svc_xprt *xprt)
1278 struct svcxprt_rdma *rdma =
1279 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1280 dprintk("svc: svc_rdma_detach(%p)\n", xprt);
1282 /* Disconnect and flush posted WQE */
1283 rdma_disconnect(rdma->sc_cm_id);
1286 static void __svc_rdma_free(struct work_struct *work)
1288 struct svcxprt_rdma *rdma =
1289 container_of(work, struct svcxprt_rdma, sc_work);
1290 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
1292 /* We should only be called from kref_put */
1293 if (atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0)
1294 pr_err("svcrdma: sc_xprt still in use? (%d)\n",
1295 atomic_read(&rdma->sc_xprt.xpt_ref.refcount));
1298 * Destroy queued, but not processed read completions. Note
1299 * that this cleanup has to be done before destroying the
1300 * cm_id because the device ptr is needed to unmap the dma in
1301 * svc_rdma_put_context.
1303 while (!list_empty(&rdma->sc_read_complete_q)) {
1304 struct svc_rdma_op_ctxt *ctxt;
1305 ctxt = list_entry(rdma->sc_read_complete_q.next,
1306 struct svc_rdma_op_ctxt,
1308 list_del_init(&ctxt->dto_q);
1309 svc_rdma_put_context(ctxt, 1);
1312 /* Destroy queued, but not processed recv completions */
1313 while (!list_empty(&rdma->sc_rq_dto_q)) {
1314 struct svc_rdma_op_ctxt *ctxt;
1315 ctxt = list_entry(rdma->sc_rq_dto_q.next,
1316 struct svc_rdma_op_ctxt,
1318 list_del_init(&ctxt->dto_q);
1319 svc_rdma_put_context(ctxt, 1);
1322 /* Warn if we leaked a resource or under-referenced */
1323 if (rdma->sc_ctxt_used != 0)
1324 pr_err("svcrdma: ctxt still in use? (%d)\n",
1325 rdma->sc_ctxt_used);
1326 if (atomic_read(&rdma->sc_dma_used) != 0)
1327 pr_err("svcrdma: dma still in use? (%d)\n",
1328 atomic_read(&rdma->sc_dma_used));
1330 rdma_dealloc_frmr_q(rdma);
1331 svc_rdma_destroy_ctxts(rdma);
1332 svc_rdma_destroy_maps(rdma);
1334 /* Destroy the QP if present (not a listener) */
1335 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1336 ib_destroy_qp(rdma->sc_qp);
1338 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
1339 ib_destroy_cq(rdma->sc_sq_cq);
1341 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
1342 ib_destroy_cq(rdma->sc_rq_cq);
1344 if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
1345 ib_dereg_mr(rdma->sc_phys_mr);
1347 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
1348 ib_dealloc_pd(rdma->sc_pd);
1350 /* Destroy the CM ID */
1351 rdma_destroy_id(rdma->sc_cm_id);
1356 static void svc_rdma_free(struct svc_xprt *xprt)
1358 struct svcxprt_rdma *rdma =
1359 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1360 INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1361 queue_work(svc_rdma_wq, &rdma->sc_work);
1364 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1366 struct svcxprt_rdma *rdma =
1367 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1370 * If there are already waiters on the SQ,
1373 if (waitqueue_active(&rdma->sc_send_wait))
1376 /* Otherwise return true. */
1380 static int svc_rdma_secure_port(struct svc_rqst *rqstp)
1385 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1387 struct ib_send_wr *bad_wr, *n_wr;
1392 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1396 for (n_wr = wr->next; n_wr; n_wr = n_wr->next)
1399 /* If the SQ is full, wait until an SQ entry is available */
1401 spin_lock_bh(&xprt->sc_lock);
1402 if (xprt->sc_sq_depth < atomic_read(&xprt->sc_sq_count) + wr_count) {
1403 spin_unlock_bh(&xprt->sc_lock);
1404 atomic_inc(&rdma_stat_sq_starve);
1406 /* See if we can opportunistically reap SQ WR to make room */
1409 /* Wait until SQ WR available if SQ still full */
1410 wait_event(xprt->sc_send_wait,
1411 atomic_read(&xprt->sc_sq_count) <
1413 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1417 /* Take a transport ref for each WR posted */
1418 for (i = 0; i < wr_count; i++)
1419 svc_xprt_get(&xprt->sc_xprt);
1421 /* Bump used SQ WR count and post */
1422 atomic_add(wr_count, &xprt->sc_sq_count);
1423 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1425 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
1426 atomic_sub(wr_count, &xprt->sc_sq_count);
1427 for (i = 0; i < wr_count; i ++)
1428 svc_xprt_put(&xprt->sc_xprt);
1429 dprintk("svcrdma: failed to post SQ WR rc=%d, "
1430 "sc_sq_count=%d, sc_sq_depth=%d\n",
1431 ret, atomic_read(&xprt->sc_sq_count),
1434 spin_unlock_bh(&xprt->sc_lock);
1436 wake_up(&xprt->sc_send_wait);
1442 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1443 enum rpcrdma_errcode err)
1445 struct ib_send_wr err_wr;
1447 struct svc_rdma_op_ctxt *ctxt;
1452 p = alloc_page(GFP_KERNEL);
1455 va = page_address(p);
1457 /* XDR encode error */
1458 length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1460 ctxt = svc_rdma_get_context(xprt);
1461 ctxt->direction = DMA_FROM_DEVICE;
1465 /* Prepare SGE for local address */
1466 ctxt->sge[0].addr = ib_dma_map_page(xprt->sc_cm_id->device,
1467 p, 0, length, DMA_FROM_DEVICE);
1468 if (ib_dma_mapping_error(xprt->sc_cm_id->device, ctxt->sge[0].addr)) {
1470 svc_rdma_put_context(ctxt, 1);
1473 atomic_inc(&xprt->sc_dma_used);
1474 ctxt->sge[0].lkey = xprt->sc_dma_lkey;
1475 ctxt->sge[0].length = length;
1477 /* Prepare SEND WR */
1478 memset(&err_wr, 0, sizeof err_wr);
1479 ctxt->wr_op = IB_WR_SEND;
1480 err_wr.wr_id = (unsigned long)ctxt;
1481 err_wr.sg_list = ctxt->sge;
1483 err_wr.opcode = IB_WR_SEND;
1484 err_wr.send_flags = IB_SEND_SIGNALED;
1487 ret = svc_rdma_send(xprt, &err_wr);
1489 dprintk("svcrdma: Error %d posting send for protocol error\n",
1491 svc_rdma_unmap_dma(ctxt);
1492 svc_rdma_put_context(ctxt, 1);