2 * Copyright (c) 2003-2007 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.
43 * Encapsulates the major functions managing:
50 #include <linux/interrupt.h>
51 #include <linux/slab.h>
52 #include <linux/prefetch.h>
53 #include <linux/sunrpc/addr.h>
54 #include <asm/bitops.h>
55 #include <linux/module.h> /* try_module_get()/module_put() */
57 #include "xprt_rdma.h"
63 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
64 # define RPCDBG_FACILITY RPCDBG_TRANS
72 * handle replies in tasklet context, using a single, global list
73 * rdma tasklet function -- just turn around and call the func
74 * for all replies on the list
77 static DEFINE_SPINLOCK(rpcrdma_tk_lock_g);
78 static LIST_HEAD(rpcrdma_tasklets_g);
81 rpcrdma_run_tasklet(unsigned long data)
83 struct rpcrdma_rep *rep;
87 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
88 while (!list_empty(&rpcrdma_tasklets_g)) {
89 rep = list_entry(rpcrdma_tasklets_g.next,
90 struct rpcrdma_rep, rr_list);
91 list_del(&rep->rr_list);
92 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
94 rpcrdma_reply_handler(rep);
96 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
98 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
101 static DECLARE_TASKLET(rpcrdma_tasklet_g, rpcrdma_run_tasklet, 0UL);
104 rpcrdma_schedule_tasklet(struct list_head *sched_list)
108 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
109 list_splice_tail(sched_list, &rpcrdma_tasklets_g);
110 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
111 tasklet_schedule(&rpcrdma_tasklet_g);
115 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
117 struct rpcrdma_ep *ep = context;
119 pr_err("RPC: %s: %s on device %s ep %p\n",
120 __func__, ib_event_msg(event->event),
121 event->device->name, context);
122 if (ep->rep_connected == 1) {
123 ep->rep_connected = -EIO;
124 rpcrdma_conn_func(ep);
125 wake_up_all(&ep->rep_connect_wait);
130 rpcrdma_cq_async_error_upcall(struct ib_event *event, void *context)
132 struct rpcrdma_ep *ep = context;
134 pr_err("RPC: %s: %s on device %s ep %p\n",
135 __func__, ib_event_msg(event->event),
136 event->device->name, context);
137 if (ep->rep_connected == 1) {
138 ep->rep_connected = -EIO;
139 rpcrdma_conn_func(ep);
140 wake_up_all(&ep->rep_connect_wait);
145 rpcrdma_sendcq_process_wc(struct ib_wc *wc)
147 /* WARNING: Only wr_id and status are reliable at this point */
148 if (wc->wr_id == RPCRDMA_IGNORE_COMPLETION) {
149 if (wc->status != IB_WC_SUCCESS &&
150 wc->status != IB_WC_WR_FLUSH_ERR)
151 pr_err("RPC: %s: SEND: %s\n",
152 __func__, ib_wc_status_msg(wc->status));
154 struct rpcrdma_mw *r;
156 r = (struct rpcrdma_mw *)(unsigned long)wc->wr_id;
157 r->mw_sendcompletion(wc);
162 rpcrdma_sendcq_poll(struct ib_cq *cq, struct rpcrdma_ep *ep)
165 int budget, count, rc;
167 budget = RPCRDMA_WC_BUDGET / RPCRDMA_POLLSIZE;
169 wcs = ep->rep_send_wcs;
171 rc = ib_poll_cq(cq, RPCRDMA_POLLSIZE, wcs);
177 rpcrdma_sendcq_process_wc(wcs++);
178 } while (rc == RPCRDMA_POLLSIZE && --budget);
183 * Handle send, fast_reg_mr, and local_inv completions.
185 * Send events are typically suppressed and thus do not result
186 * in an upcall. Occasionally one is signaled, however. This
187 * prevents the provider's completion queue from wrapping and
188 * losing a completion.
191 rpcrdma_sendcq_upcall(struct ib_cq *cq, void *cq_context)
193 struct rpcrdma_ep *ep = (struct rpcrdma_ep *)cq_context;
196 rc = rpcrdma_sendcq_poll(cq, ep);
198 dprintk("RPC: %s: ib_poll_cq failed: %i\n",
203 rc = ib_req_notify_cq(cq,
204 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
208 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
213 rpcrdma_sendcq_poll(cq, ep);
217 rpcrdma_recvcq_process_wc(struct ib_wc *wc, struct list_head *sched_list)
219 struct rpcrdma_rep *rep =
220 (struct rpcrdma_rep *)(unsigned long)wc->wr_id;
222 /* WARNING: Only wr_id and status are reliable at this point */
223 if (wc->status != IB_WC_SUCCESS)
226 /* status == SUCCESS means all fields in wc are trustworthy */
227 if (wc->opcode != IB_WC_RECV)
230 dprintk("RPC: %s: rep %p opcode 'recv', length %u: success\n",
231 __func__, rep, wc->byte_len);
233 rep->rr_len = wc->byte_len;
234 ib_dma_sync_single_for_cpu(rep->rr_device,
235 rdmab_addr(rep->rr_rdmabuf),
236 rep->rr_len, DMA_FROM_DEVICE);
237 prefetch(rdmab_to_msg(rep->rr_rdmabuf));
240 list_add_tail(&rep->rr_list, sched_list);
243 if (wc->status != IB_WC_WR_FLUSH_ERR)
244 pr_err("RPC: %s: rep %p: %s\n",
245 __func__, rep, ib_wc_status_msg(wc->status));
251 rpcrdma_recvcq_poll(struct ib_cq *cq, struct rpcrdma_ep *ep)
253 struct list_head sched_list;
255 int budget, count, rc;
257 INIT_LIST_HEAD(&sched_list);
258 budget = RPCRDMA_WC_BUDGET / RPCRDMA_POLLSIZE;
260 wcs = ep->rep_recv_wcs;
262 rc = ib_poll_cq(cq, RPCRDMA_POLLSIZE, wcs);
268 rpcrdma_recvcq_process_wc(wcs++, &sched_list);
269 } while (rc == RPCRDMA_POLLSIZE && --budget);
273 rpcrdma_schedule_tasklet(&sched_list);
278 * Handle receive completions.
280 * It is reentrant but processes single events in order to maintain
281 * ordering of receives to keep server credits.
283 * It is the responsibility of the scheduled tasklet to return
284 * recv buffers to the pool. NOTE: this affects synchronization of
285 * connection shutdown. That is, the structures required for
286 * the completion of the reply handler must remain intact until
287 * all memory has been reclaimed.
290 rpcrdma_recvcq_upcall(struct ib_cq *cq, void *cq_context)
292 struct rpcrdma_ep *ep = (struct rpcrdma_ep *)cq_context;
295 rc = rpcrdma_recvcq_poll(cq, ep);
297 dprintk("RPC: %s: ib_poll_cq failed: %i\n",
302 rc = ib_req_notify_cq(cq,
303 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
307 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
312 rpcrdma_recvcq_poll(cq, ep);
316 rpcrdma_flush_cqs(struct rpcrdma_ep *ep)
319 LIST_HEAD(sched_list);
321 while (ib_poll_cq(ep->rep_attr.recv_cq, 1, &wc) > 0)
322 rpcrdma_recvcq_process_wc(&wc, &sched_list);
323 if (!list_empty(&sched_list))
324 rpcrdma_schedule_tasklet(&sched_list);
325 while (ib_poll_cq(ep->rep_attr.send_cq, 1, &wc) > 0)
326 rpcrdma_sendcq_process_wc(&wc);
330 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
332 struct rpcrdma_xprt *xprt = id->context;
333 struct rpcrdma_ia *ia = &xprt->rx_ia;
334 struct rpcrdma_ep *ep = &xprt->rx_ep;
335 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
336 struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr;
338 struct ib_qp_attr *attr = &ia->ri_qp_attr;
339 struct ib_qp_init_attr *iattr = &ia->ri_qp_init_attr;
342 switch (event->event) {
343 case RDMA_CM_EVENT_ADDR_RESOLVED:
344 case RDMA_CM_EVENT_ROUTE_RESOLVED:
346 complete(&ia->ri_done);
348 case RDMA_CM_EVENT_ADDR_ERROR:
349 ia->ri_async_rc = -EHOSTUNREACH;
350 dprintk("RPC: %s: CM address resolution error, ep 0x%p\n",
352 complete(&ia->ri_done);
354 case RDMA_CM_EVENT_ROUTE_ERROR:
355 ia->ri_async_rc = -ENETUNREACH;
356 dprintk("RPC: %s: CM route resolution error, ep 0x%p\n",
358 complete(&ia->ri_done);
360 case RDMA_CM_EVENT_ESTABLISHED:
362 ib_query_qp(ia->ri_id->qp, attr,
363 IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
365 dprintk("RPC: %s: %d responder resources"
367 __func__, attr->max_dest_rd_atomic,
368 attr->max_rd_atomic);
370 case RDMA_CM_EVENT_CONNECT_ERROR:
371 connstate = -ENOTCONN;
373 case RDMA_CM_EVENT_UNREACHABLE:
374 connstate = -ENETDOWN;
376 case RDMA_CM_EVENT_REJECTED:
377 connstate = -ECONNREFUSED;
379 case RDMA_CM_EVENT_DISCONNECTED:
380 connstate = -ECONNABORTED;
382 case RDMA_CM_EVENT_DEVICE_REMOVAL:
385 dprintk("RPC: %s: %sconnected\n",
386 __func__, connstate > 0 ? "" : "dis");
387 ep->rep_connected = connstate;
388 rpcrdma_conn_func(ep);
389 wake_up_all(&ep->rep_connect_wait);
392 dprintk("RPC: %s: %pIS:%u (ep 0x%p): %s\n",
393 __func__, sap, rpc_get_port(sap), ep,
394 rdma_event_msg(event->event));
398 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
399 if (connstate == 1) {
400 int ird = attr->max_dest_rd_atomic;
401 int tird = ep->rep_remote_cma.responder_resources;
403 pr_info("rpcrdma: connection to %pIS:%u on %s, memreg '%s', %d credits, %d responders%s\n",
404 sap, rpc_get_port(sap),
406 ia->ri_ops->ro_displayname,
407 xprt->rx_buf.rb_max_requests,
408 ird, ird < 4 && ird < tird / 2 ? " (low!)" : "");
409 } else if (connstate < 0) {
410 pr_info("rpcrdma: connection to %pIS:%u closed (%d)\n",
411 sap, rpc_get_port(sap), connstate);
418 static void rpcrdma_destroy_id(struct rdma_cm_id *id)
421 module_put(id->device->owner);
426 static struct rdma_cm_id *
427 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
428 struct rpcrdma_ia *ia, struct sockaddr *addr)
430 struct rdma_cm_id *id;
433 init_completion(&ia->ri_done);
435 id = rdma_create_id(rpcrdma_conn_upcall, xprt, RDMA_PS_TCP, IB_QPT_RC);
438 dprintk("RPC: %s: rdma_create_id() failed %i\n",
443 ia->ri_async_rc = -ETIMEDOUT;
444 rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
446 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
450 wait_for_completion_interruptible_timeout(&ia->ri_done,
451 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
454 * Until xprtrdma supports DEVICE_REMOVAL, the provider must
455 * be pinned while there are active NFS/RDMA mounts to prevent
456 * hangs and crashes at umount time.
458 if (!ia->ri_async_rc && !try_module_get(id->device->owner)) {
459 dprintk("RPC: %s: Failed to get device module\n",
461 ia->ri_async_rc = -ENODEV;
463 rc = ia->ri_async_rc;
467 ia->ri_async_rc = -ETIMEDOUT;
468 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
470 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
474 wait_for_completion_interruptible_timeout(&ia->ri_done,
475 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
476 rc = ia->ri_async_rc;
482 module_put(id->device->owner);
489 * Drain any cq, prior to teardown.
492 rpcrdma_clean_cq(struct ib_cq *cq)
497 while (1 == ib_poll_cq(cq, 1, &wc))
501 dprintk("RPC: %s: flushed %d events (last 0x%x)\n",
502 __func__, count, wc.opcode);
506 * Exported functions.
510 * Open and initialize an Interface Adapter.
511 * o initializes fields of struct rpcrdma_ia, including
512 * interface and provider attributes and protection zone.
515 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
517 struct rpcrdma_ia *ia = &xprt->rx_ia;
518 struct ib_device_attr *devattr = &ia->ri_devattr;
521 ia->ri_dma_mr = NULL;
523 ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
524 if (IS_ERR(ia->ri_id)) {
525 rc = PTR_ERR(ia->ri_id);
528 ia->ri_device = ia->ri_id->device;
530 ia->ri_pd = ib_alloc_pd(ia->ri_device);
531 if (IS_ERR(ia->ri_pd)) {
532 rc = PTR_ERR(ia->ri_pd);
533 dprintk("RPC: %s: ib_alloc_pd() failed %i\n",
538 rc = ib_query_device(ia->ri_device, devattr);
540 dprintk("RPC: %s: ib_query_device failed %d\n",
545 if (memreg == RPCRDMA_FRMR) {
546 /* Requires both frmr reg and local dma lkey */
547 if (((devattr->device_cap_flags &
548 (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) !=
549 (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) ||
550 (devattr->max_fast_reg_page_list_len == 0)) {
551 dprintk("RPC: %s: FRMR registration "
552 "not supported by HCA\n", __func__);
553 memreg = RPCRDMA_MTHCAFMR;
556 if (memreg == RPCRDMA_MTHCAFMR) {
557 if (!ia->ri_device->alloc_fmr) {
558 dprintk("RPC: %s: MTHCAFMR registration "
559 "not supported by HCA\n", __func__);
566 ia->ri_ops = &rpcrdma_frwr_memreg_ops;
568 case RPCRDMA_ALLPHYSICAL:
569 ia->ri_ops = &rpcrdma_physical_memreg_ops;
571 case RPCRDMA_MTHCAFMR:
572 ia->ri_ops = &rpcrdma_fmr_memreg_ops;
575 printk(KERN_ERR "RPC: Unsupported memory "
576 "registration mode: %d\n", memreg);
580 dprintk("RPC: %s: memory registration strategy is '%s'\n",
581 __func__, ia->ri_ops->ro_displayname);
583 rwlock_init(&ia->ri_qplock);
587 ib_dealloc_pd(ia->ri_pd);
590 rpcrdma_destroy_id(ia->ri_id);
597 * Clean up/close an IA.
598 * o if event handles and PD have been initialized, free them.
602 rpcrdma_ia_close(struct rpcrdma_ia *ia)
604 dprintk("RPC: %s: entering\n", __func__);
605 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
607 rdma_destroy_qp(ia->ri_id);
608 rpcrdma_destroy_id(ia->ri_id);
612 /* If the pd is still busy, xprtrdma missed freeing a resource */
613 if (ia->ri_pd && !IS_ERR(ia->ri_pd))
614 ib_dealloc_pd(ia->ri_pd);
618 * Create unconnected endpoint.
621 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
622 struct rpcrdma_create_data_internal *cdata)
624 struct ib_device_attr *devattr = &ia->ri_devattr;
625 struct ib_cq *sendcq, *recvcq;
626 struct ib_cq_init_attr cq_attr = {};
629 if (devattr->max_sge < RPCRDMA_MAX_IOVS) {
630 dprintk("RPC: %s: insufficient sge's available\n",
635 /* check provider's send/recv wr limits */
636 if (cdata->max_requests > devattr->max_qp_wr)
637 cdata->max_requests = devattr->max_qp_wr;
639 ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
640 ep->rep_attr.qp_context = ep;
641 ep->rep_attr.srq = NULL;
642 ep->rep_attr.cap.max_send_wr = cdata->max_requests;
643 rc = ia->ri_ops->ro_open(ia, ep, cdata);
646 ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
647 ep->rep_attr.cap.max_send_sge = RPCRDMA_MAX_IOVS;
648 ep->rep_attr.cap.max_recv_sge = 1;
649 ep->rep_attr.cap.max_inline_data = 0;
650 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
651 ep->rep_attr.qp_type = IB_QPT_RC;
652 ep->rep_attr.port_num = ~0;
654 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
655 "iovs: send %d recv %d\n",
657 ep->rep_attr.cap.max_send_wr,
658 ep->rep_attr.cap.max_recv_wr,
659 ep->rep_attr.cap.max_send_sge,
660 ep->rep_attr.cap.max_recv_sge);
662 /* set trigger for requesting send completion */
663 ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1;
664 if (ep->rep_cqinit > RPCRDMA_MAX_UNSIGNALED_SENDS)
665 ep->rep_cqinit = RPCRDMA_MAX_UNSIGNALED_SENDS;
666 else if (ep->rep_cqinit <= 2)
669 init_waitqueue_head(&ep->rep_connect_wait);
670 INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
672 cq_attr.cqe = ep->rep_attr.cap.max_send_wr + 1;
673 sendcq = ib_create_cq(ia->ri_device, rpcrdma_sendcq_upcall,
674 rpcrdma_cq_async_error_upcall, ep, &cq_attr);
675 if (IS_ERR(sendcq)) {
676 rc = PTR_ERR(sendcq);
677 dprintk("RPC: %s: failed to create send CQ: %i\n",
682 rc = ib_req_notify_cq(sendcq, IB_CQ_NEXT_COMP);
684 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
689 cq_attr.cqe = ep->rep_attr.cap.max_recv_wr + 1;
690 recvcq = ib_create_cq(ia->ri_device, rpcrdma_recvcq_upcall,
691 rpcrdma_cq_async_error_upcall, ep, &cq_attr);
692 if (IS_ERR(recvcq)) {
693 rc = PTR_ERR(recvcq);
694 dprintk("RPC: %s: failed to create recv CQ: %i\n",
699 rc = ib_req_notify_cq(recvcq, IB_CQ_NEXT_COMP);
701 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
703 ib_destroy_cq(recvcq);
707 ep->rep_attr.send_cq = sendcq;
708 ep->rep_attr.recv_cq = recvcq;
710 /* Initialize cma parameters */
712 /* RPC/RDMA does not use private data */
713 ep->rep_remote_cma.private_data = NULL;
714 ep->rep_remote_cma.private_data_len = 0;
716 /* Client offers RDMA Read but does not initiate */
717 ep->rep_remote_cma.initiator_depth = 0;
718 if (devattr->max_qp_rd_atom > 32) /* arbitrary but <= 255 */
719 ep->rep_remote_cma.responder_resources = 32;
721 ep->rep_remote_cma.responder_resources =
722 devattr->max_qp_rd_atom;
724 ep->rep_remote_cma.retry_count = 7;
725 ep->rep_remote_cma.flow_control = 0;
726 ep->rep_remote_cma.rnr_retry_count = 0;
731 err = ib_destroy_cq(sendcq);
733 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
737 ib_dereg_mr(ia->ri_dma_mr);
744 * Disconnect and destroy endpoint. After this, the only
745 * valid operations on the ep are to free it (if dynamically
746 * allocated) or re-create it.
749 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
753 dprintk("RPC: %s: entering, connected is %d\n",
754 __func__, ep->rep_connected);
756 cancel_delayed_work_sync(&ep->rep_connect_worker);
759 rpcrdma_ep_disconnect(ep, ia);
761 rpcrdma_clean_cq(ep->rep_attr.recv_cq);
762 rpcrdma_clean_cq(ep->rep_attr.send_cq);
765 rdma_destroy_qp(ia->ri_id);
766 ia->ri_id->qp = NULL;
769 rc = ib_destroy_cq(ep->rep_attr.recv_cq);
771 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
774 rc = ib_destroy_cq(ep->rep_attr.send_cq);
776 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
780 rc = ib_dereg_mr(ia->ri_dma_mr);
781 dprintk("RPC: %s: ib_dereg_mr returned %i\n",
787 * Connect unconnected endpoint.
790 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
792 struct rdma_cm_id *id, *old;
796 if (ep->rep_connected != 0) {
797 struct rpcrdma_xprt *xprt;
799 dprintk("RPC: %s: reconnecting...\n", __func__);
801 rpcrdma_ep_disconnect(ep, ia);
802 rpcrdma_flush_cqs(ep);
804 xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
805 id = rpcrdma_create_id(xprt, ia,
806 (struct sockaddr *)&xprt->rx_data.addr);
811 /* TEMP TEMP TEMP - fail if new device:
812 * Deregister/remarshal *all* requests!
813 * Close and recreate adapter, pd, etc!
814 * Re-determine all attributes still sane!
815 * More stuff I haven't thought of!
818 if (ia->ri_device != id->device) {
819 printk("RPC: %s: can't reconnect on "
820 "different device!\n", __func__);
821 rpcrdma_destroy_id(id);
826 rc = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
828 dprintk("RPC: %s: rdma_create_qp failed %i\n",
830 rpcrdma_destroy_id(id);
835 write_lock(&ia->ri_qplock);
838 write_unlock(&ia->ri_qplock);
840 rdma_destroy_qp(old);
841 rpcrdma_destroy_id(old);
843 dprintk("RPC: %s: connecting...\n", __func__);
844 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
846 dprintk("RPC: %s: rdma_create_qp failed %i\n",
848 /* do not update ep->rep_connected */
853 ep->rep_connected = 0;
855 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
857 dprintk("RPC: %s: rdma_connect() failed with %i\n",
862 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
865 * Check state. A non-peer reject indicates no listener
866 * (ECONNREFUSED), which may be a transient state. All
867 * others indicate a transport condition which has already
868 * undergone a best-effort.
870 if (ep->rep_connected == -ECONNREFUSED &&
871 ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
872 dprintk("RPC: %s: non-peer_reject, retry\n", __func__);
875 if (ep->rep_connected <= 0) {
876 /* Sometimes, the only way to reliably connect to remote
877 * CMs is to use same nonzero values for ORD and IRD. */
878 if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
879 (ep->rep_remote_cma.responder_resources == 0 ||
880 ep->rep_remote_cma.initiator_depth !=
881 ep->rep_remote_cma.responder_resources)) {
882 if (ep->rep_remote_cma.responder_resources == 0)
883 ep->rep_remote_cma.responder_resources = 1;
884 ep->rep_remote_cma.initiator_depth =
885 ep->rep_remote_cma.responder_resources;
888 rc = ep->rep_connected;
890 dprintk("RPC: %s: connected\n", __func__);
895 ep->rep_connected = rc;
900 * rpcrdma_ep_disconnect
902 * This is separate from destroy to facilitate the ability
903 * to reconnect without recreating the endpoint.
905 * This call is not reentrant, and must not be made in parallel
906 * on the same endpoint.
909 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
913 rpcrdma_flush_cqs(ep);
914 rc = rdma_disconnect(ia->ri_id);
916 /* returns without wait if not connected */
917 wait_event_interruptible(ep->rep_connect_wait,
918 ep->rep_connected != 1);
919 dprintk("RPC: %s: after wait, %sconnected\n", __func__,
920 (ep->rep_connected == 1) ? "still " : "dis");
922 dprintk("RPC: %s: rdma_disconnect %i\n", __func__, rc);
923 ep->rep_connected = rc;
927 static struct rpcrdma_req *
928 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
930 struct rpcrdma_req *req;
932 req = kzalloc(sizeof(*req), GFP_KERNEL);
934 return ERR_PTR(-ENOMEM);
936 req->rl_buffer = &r_xprt->rx_buf;
940 static struct rpcrdma_rep *
941 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
943 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
944 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
945 struct rpcrdma_rep *rep;
949 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
953 rep->rr_rdmabuf = rpcrdma_alloc_regbuf(ia, cdata->inline_rsize,
955 if (IS_ERR(rep->rr_rdmabuf)) {
956 rc = PTR_ERR(rep->rr_rdmabuf);
960 rep->rr_device = ia->ri_device;
961 rep->rr_rxprt = r_xprt;
971 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
973 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
974 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
975 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
980 buf->rb_max_requests = cdata->max_requests;
981 spin_lock_init(&buf->rb_lock);
984 * 1. arrays for send and recv pointers
985 * 2. arrays of struct rpcrdma_req to fill in pointers
986 * 3. array of struct rpcrdma_rep for replies
987 * Send/recv buffers in req/rep need to be registered
989 len = buf->rb_max_requests *
990 (sizeof(struct rpcrdma_req *) + sizeof(struct rpcrdma_rep *));
992 p = kzalloc(len, GFP_KERNEL);
994 dprintk("RPC: %s: req_t/rep_t/pad kzalloc(%zd) failed\n",
999 buf->rb_pool = p; /* for freeing it later */
1001 buf->rb_send_bufs = (struct rpcrdma_req **) p;
1002 p = (char *) &buf->rb_send_bufs[buf->rb_max_requests];
1003 buf->rb_recv_bufs = (struct rpcrdma_rep **) p;
1004 p = (char *) &buf->rb_recv_bufs[buf->rb_max_requests];
1006 rc = ia->ri_ops->ro_init(r_xprt);
1010 for (i = 0; i < buf->rb_max_requests; i++) {
1011 struct rpcrdma_req *req;
1012 struct rpcrdma_rep *rep;
1014 req = rpcrdma_create_req(r_xprt);
1016 dprintk("RPC: %s: request buffer %d alloc"
1017 " failed\n", __func__, i);
1021 buf->rb_send_bufs[i] = req;
1023 rep = rpcrdma_create_rep(r_xprt);
1025 dprintk("RPC: %s: reply buffer %d alloc failed\n",
1030 buf->rb_recv_bufs[i] = rep;
1035 rpcrdma_buffer_destroy(buf);
1040 rpcrdma_destroy_rep(struct rpcrdma_ia *ia, struct rpcrdma_rep *rep)
1045 rpcrdma_free_regbuf(ia, rep->rr_rdmabuf);
1050 rpcrdma_destroy_req(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
1055 rpcrdma_free_regbuf(ia, req->rl_sendbuf);
1056 rpcrdma_free_regbuf(ia, req->rl_rdmabuf);
1061 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1063 struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1066 /* clean up in reverse order from create
1067 * 1. recv mr memory (mr free, then kfree)
1068 * 2. send mr memory (mr free, then kfree)
1071 dprintk("RPC: %s: entering\n", __func__);
1073 for (i = 0; i < buf->rb_max_requests; i++) {
1074 if (buf->rb_recv_bufs)
1075 rpcrdma_destroy_rep(ia, buf->rb_recv_bufs[i]);
1076 if (buf->rb_send_bufs)
1077 rpcrdma_destroy_req(ia, buf->rb_send_bufs[i]);
1080 ia->ri_ops->ro_destroy(buf);
1082 kfree(buf->rb_pool);
1086 rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
1088 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1089 struct rpcrdma_mw *mw = NULL;
1091 spin_lock(&buf->rb_mwlock);
1092 if (!list_empty(&buf->rb_mws)) {
1093 mw = list_first_entry(&buf->rb_mws,
1094 struct rpcrdma_mw, mw_list);
1095 list_del_init(&mw->mw_list);
1097 spin_unlock(&buf->rb_mwlock);
1100 pr_err("RPC: %s: no MWs available\n", __func__);
1105 rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
1107 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1109 spin_lock(&buf->rb_mwlock);
1110 list_add_tail(&mw->mw_list, &buf->rb_mws);
1111 spin_unlock(&buf->rb_mwlock);
1115 rpcrdma_buffer_put_sendbuf(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
1117 buf->rb_send_bufs[--buf->rb_send_index] = req;
1119 if (req->rl_reply) {
1120 buf->rb_recv_bufs[--buf->rb_recv_index] = req->rl_reply;
1121 req->rl_reply = NULL;
1126 * Get a set of request/reply buffers.
1128 * Reply buffer (if needed) is attached to send buffer upon return.
1130 * rb_send_index and rb_recv_index MUST always be pointing to the
1131 * *next* available buffer (non-NULL). They are incremented after
1132 * removing buffers, and decremented *before* returning them.
1134 struct rpcrdma_req *
1135 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1137 struct rpcrdma_req *req;
1138 unsigned long flags;
1140 spin_lock_irqsave(&buffers->rb_lock, flags);
1142 if (buffers->rb_send_index == buffers->rb_max_requests) {
1143 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1144 dprintk("RPC: %s: out of request buffers\n", __func__);
1145 return ((struct rpcrdma_req *)NULL);
1148 req = buffers->rb_send_bufs[buffers->rb_send_index];
1149 if (buffers->rb_send_index < buffers->rb_recv_index) {
1150 dprintk("RPC: %s: %d extra receives outstanding (ok)\n",
1152 buffers->rb_recv_index - buffers->rb_send_index);
1153 req->rl_reply = NULL;
1155 req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1156 buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1158 buffers->rb_send_bufs[buffers->rb_send_index++] = NULL;
1160 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1165 * Put request/reply buffers back into pool.
1166 * Pre-decrement counter/array index.
1169 rpcrdma_buffer_put(struct rpcrdma_req *req)
1171 struct rpcrdma_buffer *buffers = req->rl_buffer;
1172 unsigned long flags;
1174 spin_lock_irqsave(&buffers->rb_lock, flags);
1175 rpcrdma_buffer_put_sendbuf(req, buffers);
1176 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1180 * Recover reply buffers from pool.
1181 * This happens when recovering from error conditions.
1182 * Post-increment counter/array index.
1185 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1187 struct rpcrdma_buffer *buffers = req->rl_buffer;
1188 unsigned long flags;
1190 spin_lock_irqsave(&buffers->rb_lock, flags);
1191 if (buffers->rb_recv_index < buffers->rb_max_requests) {
1192 req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1193 buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1195 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1199 * Put reply buffers back into pool when not attached to
1200 * request. This happens in error conditions.
1203 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1205 struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1206 unsigned long flags;
1208 spin_lock_irqsave(&buffers->rb_lock, flags);
1209 buffers->rb_recv_bufs[--buffers->rb_recv_index] = rep;
1210 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1214 * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1218 rpcrdma_mapping_error(struct rpcrdma_mr_seg *seg)
1220 dprintk("RPC: map_one: offset %p iova %llx len %zu\n",
1222 (unsigned long long)seg->mr_dma, seg->mr_dmalen);
1226 * rpcrdma_alloc_regbuf - kmalloc and register memory for SEND/RECV buffers
1227 * @ia: controlling rpcrdma_ia
1228 * @size: size of buffer to be allocated, in bytes
1231 * Returns pointer to private header of an area of internally
1232 * registered memory, or an ERR_PTR. The registered buffer follows
1233 * the end of the private header.
1235 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1236 * receiving the payload of RDMA RECV operations. regbufs are not
1237 * used for RDMA READ/WRITE operations, thus are registered only for
1240 struct rpcrdma_regbuf *
1241 rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size, gfp_t flags)
1243 struct rpcrdma_regbuf *rb;
1246 rb = kmalloc(sizeof(*rb) + size, flags);
1251 iov->addr = ib_dma_map_single(ia->ri_device,
1252 (void *)rb->rg_base, size,
1254 if (ib_dma_mapping_error(ia->ri_device, iov->addr))
1258 iov->lkey = ia->ri_pd->local_dma_lkey;
1260 rb->rg_owner = NULL;
1266 return ERR_PTR(-ENOMEM);
1270 * rpcrdma_free_regbuf - deregister and free registered buffer
1271 * @ia: controlling rpcrdma_ia
1272 * @rb: regbuf to be deregistered and freed
1275 rpcrdma_free_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1283 ib_dma_unmap_single(ia->ri_device,
1284 iov->addr, iov->length, DMA_BIDIRECTIONAL);
1289 * Prepost any receive buffer, then post send.
1291 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1294 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1295 struct rpcrdma_ep *ep,
1296 struct rpcrdma_req *req)
1298 struct ib_device *device = ia->ri_device;
1299 struct ib_send_wr send_wr, *send_wr_fail;
1300 struct rpcrdma_rep *rep = req->rl_reply;
1301 struct ib_sge *iov = req->rl_send_iov;
1305 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1308 req->rl_reply = NULL;
1311 send_wr.next = NULL;
1312 send_wr.wr_id = RPCRDMA_IGNORE_COMPLETION;
1313 send_wr.sg_list = iov;
1314 send_wr.num_sge = req->rl_niovs;
1315 send_wr.opcode = IB_WR_SEND;
1317 for (i = 0; i < send_wr.num_sge; i++)
1318 ib_dma_sync_single_for_device(device, iov[i].addr,
1319 iov[i].length, DMA_TO_DEVICE);
1320 dprintk("RPC: %s: posting %d s/g entries\n",
1321 __func__, send_wr.num_sge);
1323 if (DECR_CQCOUNT(ep) > 0)
1324 send_wr.send_flags = 0;
1325 else { /* Provider must take a send completion every now and then */
1327 send_wr.send_flags = IB_SEND_SIGNALED;
1330 rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
1332 dprintk("RPC: %s: ib_post_send returned %i\n", __func__,
1339 * (Re)post a receive buffer.
1342 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1343 struct rpcrdma_ep *ep,
1344 struct rpcrdma_rep *rep)
1346 struct ib_recv_wr recv_wr, *recv_wr_fail;
1349 recv_wr.next = NULL;
1350 recv_wr.wr_id = (u64) (unsigned long) rep;
1351 recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1352 recv_wr.num_sge = 1;
1354 ib_dma_sync_single_for_cpu(ia->ri_device,
1355 rdmab_addr(rep->rr_rdmabuf),
1356 rdmab_length(rep->rr_rdmabuf),
1359 rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
1362 dprintk("RPC: %s: ib_post_recv returned %i\n", __func__,
1367 /* How many chunk list items fit within our inline buffers?
1370 rpcrdma_max_segments(struct rpcrdma_xprt *r_xprt)
1372 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1373 int bytes, segments;
1375 bytes = min_t(unsigned int, cdata->inline_wsize, cdata->inline_rsize);
1376 bytes -= RPCRDMA_HDRLEN_MIN;
1377 if (bytes < sizeof(struct rpcrdma_segment) * 2) {
1378 pr_warn("RPC: %s: inline threshold too small\n",
1383 segments = 1 << (fls(bytes / sizeof(struct rpcrdma_segment)) - 1);
1384 dprintk("RPC: %s: max chunk list size = %d segments\n",
1385 __func__, segments);