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 <linux/sunrpc/svc_rdma.h>
55 #include <asm/bitops.h>
56 #include <linux/module.h> /* try_module_get()/module_put() */
58 #include "xprt_rdma.h"
64 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
65 # define RPCDBG_FACILITY RPCDBG_TRANS
72 static struct workqueue_struct *rpcrdma_receive_wq;
75 rpcrdma_alloc_wq(void)
77 struct workqueue_struct *recv_wq;
79 recv_wq = alloc_workqueue("xprtrdma_receive",
80 WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_HIGHPRI,
85 rpcrdma_receive_wq = recv_wq;
90 rpcrdma_destroy_wq(void)
92 struct workqueue_struct *wq;
94 if (rpcrdma_receive_wq) {
95 wq = rpcrdma_receive_wq;
96 rpcrdma_receive_wq = NULL;
97 destroy_workqueue(wq);
102 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
104 struct rpcrdma_ep *ep = context;
106 pr_err("RPC: %s: %s on device %s ep %p\n",
107 __func__, ib_event_msg(event->event),
108 event->device->name, context);
109 if (ep->rep_connected == 1) {
110 ep->rep_connected = -EIO;
111 rpcrdma_conn_func(ep);
112 wake_up_all(&ep->rep_connect_wait);
117 * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
118 * @cq: completion queue (ignored)
123 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
125 /* WARNING: Only wr_cqe and status are reliable at this point */
126 if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
127 pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
128 ib_wc_status_msg(wc->status),
129 wc->status, wc->vendor_err);
133 rpcrdma_receive_worker(struct work_struct *work)
135 struct rpcrdma_rep *rep =
136 container_of(work, struct rpcrdma_rep, rr_work);
138 rpcrdma_reply_handler(rep);
141 /* Perform basic sanity checking to avoid using garbage
142 * to update the credit grant value.
145 rpcrdma_update_granted_credits(struct rpcrdma_rep *rep)
147 struct rpcrdma_msg *rmsgp = rdmab_to_msg(rep->rr_rdmabuf);
148 struct rpcrdma_buffer *buffer = &rep->rr_rxprt->rx_buf;
151 if (rep->rr_len < RPCRDMA_HDRLEN_ERR)
154 credits = be32_to_cpu(rmsgp->rm_credit);
156 credits = 1; /* don't deadlock */
157 else if (credits > buffer->rb_max_requests)
158 credits = buffer->rb_max_requests;
160 atomic_set(&buffer->rb_credits, credits);
164 * rpcrdma_receive_wc - Invoked by RDMA provider for each polled Receive WC
165 * @cq: completion queue (ignored)
170 rpcrdma_receive_wc(struct ib_cq *cq, struct ib_wc *wc)
172 struct ib_cqe *cqe = wc->wr_cqe;
173 struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
176 /* WARNING: Only wr_id and status are reliable at this point */
177 if (wc->status != IB_WC_SUCCESS)
180 /* status == SUCCESS means all fields in wc are trustworthy */
181 if (wc->opcode != IB_WC_RECV)
184 dprintk("RPC: %s: rep %p opcode 'recv', length %u: success\n",
185 __func__, rep, wc->byte_len);
187 rep->rr_len = wc->byte_len;
188 ib_dma_sync_single_for_cpu(rep->rr_device,
189 rdmab_addr(rep->rr_rdmabuf),
190 rep->rr_len, DMA_FROM_DEVICE);
192 rpcrdma_update_granted_credits(rep);
195 queue_work(rpcrdma_receive_wq, &rep->rr_work);
199 if (wc->status != IB_WC_WR_FLUSH_ERR)
200 pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
201 ib_wc_status_msg(wc->status),
202 wc->status, wc->vendor_err);
203 rep->rr_len = RPCRDMA_BAD_LEN;
208 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
210 struct rpcrdma_xprt *xprt = id->context;
211 struct rpcrdma_ia *ia = &xprt->rx_ia;
212 struct rpcrdma_ep *ep = &xprt->rx_ep;
213 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
214 struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr;
216 struct ib_qp_attr *attr = &ia->ri_qp_attr;
217 struct ib_qp_init_attr *iattr = &ia->ri_qp_init_attr;
220 switch (event->event) {
221 case RDMA_CM_EVENT_ADDR_RESOLVED:
222 case RDMA_CM_EVENT_ROUTE_RESOLVED:
224 complete(&ia->ri_done);
226 case RDMA_CM_EVENT_ADDR_ERROR:
227 ia->ri_async_rc = -EHOSTUNREACH;
228 dprintk("RPC: %s: CM address resolution error, ep 0x%p\n",
230 complete(&ia->ri_done);
232 case RDMA_CM_EVENT_ROUTE_ERROR:
233 ia->ri_async_rc = -ENETUNREACH;
234 dprintk("RPC: %s: CM route resolution error, ep 0x%p\n",
236 complete(&ia->ri_done);
238 case RDMA_CM_EVENT_ESTABLISHED:
240 ib_query_qp(ia->ri_id->qp, attr,
241 IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
243 dprintk("RPC: %s: %d responder resources"
245 __func__, attr->max_dest_rd_atomic,
246 attr->max_rd_atomic);
248 case RDMA_CM_EVENT_CONNECT_ERROR:
249 connstate = -ENOTCONN;
251 case RDMA_CM_EVENT_UNREACHABLE:
252 connstate = -ENETDOWN;
254 case RDMA_CM_EVENT_REJECTED:
255 connstate = -ECONNREFUSED;
257 case RDMA_CM_EVENT_DISCONNECTED:
258 connstate = -ECONNABORTED;
260 case RDMA_CM_EVENT_DEVICE_REMOVAL:
263 dprintk("RPC: %s: %sconnected\n",
264 __func__, connstate > 0 ? "" : "dis");
265 atomic_set(&xprt->rx_buf.rb_credits, 1);
266 ep->rep_connected = connstate;
267 rpcrdma_conn_func(ep);
268 wake_up_all(&ep->rep_connect_wait);
271 dprintk("RPC: %s: %pIS:%u (ep 0x%p): %s\n",
272 __func__, sap, rpc_get_port(sap), ep,
273 rdma_event_msg(event->event));
277 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
278 if (connstate == 1) {
279 int ird = attr->max_dest_rd_atomic;
280 int tird = ep->rep_remote_cma.responder_resources;
282 pr_info("rpcrdma: connection to %pIS:%u on %s, memreg '%s', %d credits, %d responders%s\n",
283 sap, rpc_get_port(sap),
285 ia->ri_ops->ro_displayname,
286 xprt->rx_buf.rb_max_requests,
287 ird, ird < 4 && ird < tird / 2 ? " (low!)" : "");
288 } else if (connstate < 0) {
289 pr_info("rpcrdma: connection to %pIS:%u closed (%d)\n",
290 sap, rpc_get_port(sap), connstate);
297 static void rpcrdma_destroy_id(struct rdma_cm_id *id)
300 module_put(id->device->owner);
305 static struct rdma_cm_id *
306 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
307 struct rpcrdma_ia *ia, struct sockaddr *addr)
309 struct rdma_cm_id *id;
312 init_completion(&ia->ri_done);
314 id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP,
318 dprintk("RPC: %s: rdma_create_id() failed %i\n",
323 ia->ri_async_rc = -ETIMEDOUT;
324 rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
326 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
330 wait_for_completion_interruptible_timeout(&ia->ri_done,
331 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
334 * Until xprtrdma supports DEVICE_REMOVAL, the provider must
335 * be pinned while there are active NFS/RDMA mounts to prevent
336 * hangs and crashes at umount time.
338 if (!ia->ri_async_rc && !try_module_get(id->device->owner)) {
339 dprintk("RPC: %s: Failed to get device module\n",
341 ia->ri_async_rc = -ENODEV;
343 rc = ia->ri_async_rc;
347 ia->ri_async_rc = -ETIMEDOUT;
348 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
350 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
354 wait_for_completion_interruptible_timeout(&ia->ri_done,
355 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
356 rc = ia->ri_async_rc;
362 module_put(id->device->owner);
369 * Exported functions.
373 * Open and initialize an Interface Adapter.
374 * o initializes fields of struct rpcrdma_ia, including
375 * interface and provider attributes and protection zone.
378 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
380 struct rpcrdma_ia *ia = &xprt->rx_ia;
383 ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
384 if (IS_ERR(ia->ri_id)) {
385 rc = PTR_ERR(ia->ri_id);
388 ia->ri_device = ia->ri_id->device;
390 ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
391 if (IS_ERR(ia->ri_pd)) {
392 rc = PTR_ERR(ia->ri_pd);
393 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
399 if (frwr_is_supported(ia)) {
400 ia->ri_ops = &rpcrdma_frwr_memreg_ops;
404 case RPCRDMA_MTHCAFMR:
405 if (fmr_is_supported(ia)) {
406 ia->ri_ops = &rpcrdma_fmr_memreg_ops;
411 pr_err("rpcrdma: Unsupported memory registration mode: %d\n",
420 ib_dealloc_pd(ia->ri_pd);
423 rpcrdma_destroy_id(ia->ri_id);
430 * Clean up/close an IA.
431 * o if event handles and PD have been initialized, free them.
435 rpcrdma_ia_close(struct rpcrdma_ia *ia)
437 dprintk("RPC: %s: entering\n", __func__);
438 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
440 rdma_destroy_qp(ia->ri_id);
441 rpcrdma_destroy_id(ia->ri_id);
445 /* If the pd is still busy, xprtrdma missed freeing a resource */
446 if (ia->ri_pd && !IS_ERR(ia->ri_pd))
447 ib_dealloc_pd(ia->ri_pd);
451 * Create unconnected endpoint.
454 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
455 struct rpcrdma_create_data_internal *cdata)
457 struct ib_cq *sendcq, *recvcq;
458 unsigned int max_qp_wr;
461 if (ia->ri_device->attrs.max_sge < RPCRDMA_MAX_IOVS) {
462 dprintk("RPC: %s: insufficient sge's available\n",
467 if (ia->ri_device->attrs.max_qp_wr <= RPCRDMA_BACKWARD_WRS) {
468 dprintk("RPC: %s: insufficient wqe's available\n",
472 max_qp_wr = ia->ri_device->attrs.max_qp_wr - RPCRDMA_BACKWARD_WRS - 1;
474 /* check provider's send/recv wr limits */
475 if (cdata->max_requests > max_qp_wr)
476 cdata->max_requests = max_qp_wr;
478 ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
479 ep->rep_attr.qp_context = ep;
480 ep->rep_attr.srq = NULL;
481 ep->rep_attr.cap.max_send_wr = cdata->max_requests;
482 ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
483 ep->rep_attr.cap.max_send_wr += 1; /* drain cqe */
484 rc = ia->ri_ops->ro_open(ia, ep, cdata);
487 ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
488 ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
489 ep->rep_attr.cap.max_recv_wr += 1; /* drain cqe */
490 ep->rep_attr.cap.max_send_sge = RPCRDMA_MAX_IOVS;
491 ep->rep_attr.cap.max_recv_sge = 1;
492 ep->rep_attr.cap.max_inline_data = 0;
493 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
494 ep->rep_attr.qp_type = IB_QPT_RC;
495 ep->rep_attr.port_num = ~0;
497 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
498 "iovs: send %d recv %d\n",
500 ep->rep_attr.cap.max_send_wr,
501 ep->rep_attr.cap.max_recv_wr,
502 ep->rep_attr.cap.max_send_sge,
503 ep->rep_attr.cap.max_recv_sge);
505 /* set trigger for requesting send completion */
506 ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1;
507 if (ep->rep_cqinit <= 2)
508 ep->rep_cqinit = 0; /* always signal? */
510 init_waitqueue_head(&ep->rep_connect_wait);
511 INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
513 sendcq = ib_alloc_cq(ia->ri_device, NULL,
514 ep->rep_attr.cap.max_send_wr + 1,
516 if (IS_ERR(sendcq)) {
517 rc = PTR_ERR(sendcq);
518 dprintk("RPC: %s: failed to create send CQ: %i\n",
523 recvcq = ib_alloc_cq(ia->ri_device, NULL,
524 ep->rep_attr.cap.max_recv_wr + 1,
526 if (IS_ERR(recvcq)) {
527 rc = PTR_ERR(recvcq);
528 dprintk("RPC: %s: failed to create recv CQ: %i\n",
533 ep->rep_attr.send_cq = sendcq;
534 ep->rep_attr.recv_cq = recvcq;
536 /* Initialize cma parameters */
537 memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
539 /* RPC/RDMA does not use private data */
540 ep->rep_remote_cma.private_data = NULL;
541 ep->rep_remote_cma.private_data_len = 0;
543 /* Client offers RDMA Read but does not initiate */
544 ep->rep_remote_cma.initiator_depth = 0;
545 if (ia->ri_device->attrs.max_qp_rd_atom > 32) /* arbitrary but <= 255 */
546 ep->rep_remote_cma.responder_resources = 32;
548 ep->rep_remote_cma.responder_resources =
549 ia->ri_device->attrs.max_qp_rd_atom;
551 /* Limit transport retries so client can detect server
552 * GID changes quickly. RPC layer handles re-establishing
553 * transport connection and retransmission.
555 ep->rep_remote_cma.retry_count = 6;
557 /* RPC-over-RDMA handles its own flow control. In addition,
558 * make all RNR NAKs visible so we know that RPC-over-RDMA
559 * flow control is working correctly (no NAKs should be seen).
561 ep->rep_remote_cma.flow_control = 0;
562 ep->rep_remote_cma.rnr_retry_count = 0;
575 * Disconnect and destroy endpoint. After this, the only
576 * valid operations on the ep are to free it (if dynamically
577 * allocated) or re-create it.
580 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
582 dprintk("RPC: %s: entering, connected is %d\n",
583 __func__, ep->rep_connected);
585 cancel_delayed_work_sync(&ep->rep_connect_worker);
588 rpcrdma_ep_disconnect(ep, ia);
589 rdma_destroy_qp(ia->ri_id);
590 ia->ri_id->qp = NULL;
593 ib_free_cq(ep->rep_attr.recv_cq);
594 ib_free_cq(ep->rep_attr.send_cq);
598 * Connect unconnected endpoint.
601 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
603 struct rdma_cm_id *id, *old;
607 if (ep->rep_connected != 0) {
608 struct rpcrdma_xprt *xprt;
610 dprintk("RPC: %s: reconnecting...\n", __func__);
612 rpcrdma_ep_disconnect(ep, ia);
614 xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
615 id = rpcrdma_create_id(xprt, ia,
616 (struct sockaddr *)&xprt->rx_data.addr);
621 /* TEMP TEMP TEMP - fail if new device:
622 * Deregister/remarshal *all* requests!
623 * Close and recreate adapter, pd, etc!
624 * Re-determine all attributes still sane!
625 * More stuff I haven't thought of!
628 if (ia->ri_device != id->device) {
629 printk("RPC: %s: can't reconnect on "
630 "different device!\n", __func__);
631 rpcrdma_destroy_id(id);
636 rc = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
638 dprintk("RPC: %s: rdma_create_qp failed %i\n",
640 rpcrdma_destroy_id(id);
648 rdma_destroy_qp(old);
649 rpcrdma_destroy_id(old);
651 dprintk("RPC: %s: connecting...\n", __func__);
652 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
654 dprintk("RPC: %s: rdma_create_qp failed %i\n",
656 /* do not update ep->rep_connected */
661 ep->rep_connected = 0;
663 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
665 dprintk("RPC: %s: rdma_connect() failed with %i\n",
670 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
673 * Check state. A non-peer reject indicates no listener
674 * (ECONNREFUSED), which may be a transient state. All
675 * others indicate a transport condition which has already
676 * undergone a best-effort.
678 if (ep->rep_connected == -ECONNREFUSED &&
679 ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
680 dprintk("RPC: %s: non-peer_reject, retry\n", __func__);
683 if (ep->rep_connected <= 0) {
684 /* Sometimes, the only way to reliably connect to remote
685 * CMs is to use same nonzero values for ORD and IRD. */
686 if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
687 (ep->rep_remote_cma.responder_resources == 0 ||
688 ep->rep_remote_cma.initiator_depth !=
689 ep->rep_remote_cma.responder_resources)) {
690 if (ep->rep_remote_cma.responder_resources == 0)
691 ep->rep_remote_cma.responder_resources = 1;
692 ep->rep_remote_cma.initiator_depth =
693 ep->rep_remote_cma.responder_resources;
696 rc = ep->rep_connected;
698 struct rpcrdma_xprt *r_xprt;
701 dprintk("RPC: %s: connected\n", __func__);
703 r_xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
704 extras = r_xprt->rx_buf.rb_bc_srv_max_requests;
707 rc = rpcrdma_ep_post_extra_recv(r_xprt, extras);
709 pr_warn("%s: rpcrdma_ep_post_extra_recv: %i\n",
718 ep->rep_connected = rc;
723 * rpcrdma_ep_disconnect
725 * This is separate from destroy to facilitate the ability
726 * to reconnect without recreating the endpoint.
728 * This call is not reentrant, and must not be made in parallel
729 * on the same endpoint.
732 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
736 rc = rdma_disconnect(ia->ri_id);
738 /* returns without wait if not connected */
739 wait_event_interruptible(ep->rep_connect_wait,
740 ep->rep_connected != 1);
741 dprintk("RPC: %s: after wait, %sconnected\n", __func__,
742 (ep->rep_connected == 1) ? "still " : "dis");
744 dprintk("RPC: %s: rdma_disconnect %i\n", __func__, rc);
745 ep->rep_connected = rc;
748 ib_drain_qp(ia->ri_id->qp);
752 rpcrdma_mr_recovery_worker(struct work_struct *work)
754 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
755 rb_recovery_worker.work);
756 struct rpcrdma_mw *mw;
758 spin_lock(&buf->rb_recovery_lock);
759 while (!list_empty(&buf->rb_stale_mrs)) {
760 mw = list_first_entry(&buf->rb_stale_mrs,
761 struct rpcrdma_mw, mw_list);
762 list_del_init(&mw->mw_list);
763 spin_unlock(&buf->rb_recovery_lock);
765 dprintk("RPC: %s: recovering MR %p\n", __func__, mw);
766 mw->mw_xprt->rx_ia.ri_ops->ro_recover_mr(mw);
768 spin_lock(&buf->rb_recovery_lock);
770 spin_unlock(&buf->rb_recovery_lock);
774 rpcrdma_defer_mr_recovery(struct rpcrdma_mw *mw)
776 struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
777 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
779 spin_lock(&buf->rb_recovery_lock);
780 list_add(&mw->mw_list, &buf->rb_stale_mrs);
781 spin_unlock(&buf->rb_recovery_lock);
783 schedule_delayed_work(&buf->rb_recovery_worker, 0);
787 rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt)
789 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
790 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
795 for (count = 0; count < 32; count++) {
796 struct rpcrdma_mw *mw;
799 mw = kzalloc(sizeof(*mw), GFP_KERNEL);
803 rc = ia->ri_ops->ro_init_mr(ia, mw);
809 mw->mw_xprt = r_xprt;
811 list_add(&mw->mw_list, &free);
812 list_add(&mw->mw_all, &all);
815 spin_lock(&buf->rb_mwlock);
816 list_splice(&free, &buf->rb_mws);
817 list_splice(&all, &buf->rb_all);
818 r_xprt->rx_stats.mrs_allocated += count;
819 spin_unlock(&buf->rb_mwlock);
821 dprintk("RPC: %s: created %u MRs\n", __func__, count);
825 rpcrdma_mr_refresh_worker(struct work_struct *work)
827 struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
828 rb_refresh_worker.work);
829 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
832 rpcrdma_create_mrs(r_xprt);
836 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
838 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
839 struct rpcrdma_req *req;
841 req = kzalloc(sizeof(*req), GFP_KERNEL);
843 return ERR_PTR(-ENOMEM);
845 INIT_LIST_HEAD(&req->rl_free);
846 spin_lock(&buffer->rb_reqslock);
847 list_add(&req->rl_all, &buffer->rb_allreqs);
848 spin_unlock(&buffer->rb_reqslock);
849 req->rl_cqe.done = rpcrdma_wc_send;
850 req->rl_buffer = &r_xprt->rx_buf;
851 INIT_LIST_HEAD(&req->rl_registered);
856 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
858 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
859 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
860 struct rpcrdma_rep *rep;
864 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
868 rep->rr_rdmabuf = rpcrdma_alloc_regbuf(ia, cdata->inline_rsize,
870 if (IS_ERR(rep->rr_rdmabuf)) {
871 rc = PTR_ERR(rep->rr_rdmabuf);
875 rep->rr_device = ia->ri_device;
876 rep->rr_cqe.done = rpcrdma_receive_wc;
877 rep->rr_rxprt = r_xprt;
878 INIT_WORK(&rep->rr_work, rpcrdma_receive_worker);
888 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
890 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
893 buf->rb_max_requests = r_xprt->rx_data.max_requests;
894 buf->rb_bc_srv_max_requests = 0;
895 atomic_set(&buf->rb_credits, 1);
896 spin_lock_init(&buf->rb_mwlock);
897 spin_lock_init(&buf->rb_lock);
898 spin_lock_init(&buf->rb_recovery_lock);
899 INIT_LIST_HEAD(&buf->rb_mws);
900 INIT_LIST_HEAD(&buf->rb_all);
901 INIT_LIST_HEAD(&buf->rb_stale_mrs);
902 INIT_DELAYED_WORK(&buf->rb_refresh_worker,
903 rpcrdma_mr_refresh_worker);
904 INIT_DELAYED_WORK(&buf->rb_recovery_worker,
905 rpcrdma_mr_recovery_worker);
907 rpcrdma_create_mrs(r_xprt);
909 INIT_LIST_HEAD(&buf->rb_send_bufs);
910 INIT_LIST_HEAD(&buf->rb_allreqs);
911 spin_lock_init(&buf->rb_reqslock);
912 for (i = 0; i < buf->rb_max_requests; i++) {
913 struct rpcrdma_req *req;
915 req = rpcrdma_create_req(r_xprt);
917 dprintk("RPC: %s: request buffer %d alloc"
918 " failed\n", __func__, i);
922 req->rl_backchannel = false;
923 list_add(&req->rl_free, &buf->rb_send_bufs);
926 INIT_LIST_HEAD(&buf->rb_recv_bufs);
927 for (i = 0; i < buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS; i++) {
928 struct rpcrdma_rep *rep;
930 rep = rpcrdma_create_rep(r_xprt);
932 dprintk("RPC: %s: reply buffer %d alloc failed\n",
937 list_add(&rep->rr_list, &buf->rb_recv_bufs);
942 rpcrdma_buffer_destroy(buf);
946 static struct rpcrdma_req *
947 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf)
949 struct rpcrdma_req *req;
951 req = list_first_entry(&buf->rb_send_bufs,
952 struct rpcrdma_req, rl_free);
953 list_del(&req->rl_free);
957 static struct rpcrdma_rep *
958 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf)
960 struct rpcrdma_rep *rep;
962 rep = list_first_entry(&buf->rb_recv_bufs,
963 struct rpcrdma_rep, rr_list);
964 list_del(&rep->rr_list);
969 rpcrdma_destroy_rep(struct rpcrdma_ia *ia, struct rpcrdma_rep *rep)
971 rpcrdma_free_regbuf(ia, rep->rr_rdmabuf);
976 rpcrdma_destroy_req(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
978 rpcrdma_free_regbuf(ia, req->rl_sendbuf);
979 rpcrdma_free_regbuf(ia, req->rl_rdmabuf);
984 rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf)
986 struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
988 struct rpcrdma_ia *ia = rdmab_to_ia(buf);
989 struct rpcrdma_mw *mw;
993 spin_lock(&buf->rb_mwlock);
994 while (!list_empty(&buf->rb_all)) {
995 mw = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
996 list_del(&mw->mw_all);
998 spin_unlock(&buf->rb_mwlock);
999 ia->ri_ops->ro_release_mr(mw);
1001 spin_lock(&buf->rb_mwlock);
1003 spin_unlock(&buf->rb_mwlock);
1004 r_xprt->rx_stats.mrs_allocated = 0;
1006 dprintk("RPC: %s: released %u MRs\n", __func__, count);
1010 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1012 struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1014 cancel_delayed_work_sync(&buf->rb_recovery_worker);
1016 while (!list_empty(&buf->rb_recv_bufs)) {
1017 struct rpcrdma_rep *rep;
1019 rep = rpcrdma_buffer_get_rep_locked(buf);
1020 rpcrdma_destroy_rep(ia, rep);
1022 buf->rb_send_count = 0;
1024 spin_lock(&buf->rb_reqslock);
1025 while (!list_empty(&buf->rb_allreqs)) {
1026 struct rpcrdma_req *req;
1028 req = list_first_entry(&buf->rb_allreqs,
1029 struct rpcrdma_req, rl_all);
1030 list_del(&req->rl_all);
1032 spin_unlock(&buf->rb_reqslock);
1033 rpcrdma_destroy_req(ia, req);
1034 spin_lock(&buf->rb_reqslock);
1036 spin_unlock(&buf->rb_reqslock);
1037 buf->rb_recv_count = 0;
1039 rpcrdma_destroy_mrs(buf);
1043 rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
1045 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1046 struct rpcrdma_mw *mw = NULL;
1048 spin_lock(&buf->rb_mwlock);
1049 if (!list_empty(&buf->rb_mws)) {
1050 mw = list_first_entry(&buf->rb_mws,
1051 struct rpcrdma_mw, mw_list);
1052 list_del_init(&mw->mw_list);
1054 spin_unlock(&buf->rb_mwlock);
1061 dprintk("RPC: %s: no MWs available\n", __func__);
1062 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1064 /* Allow the reply handler and refresh worker to run */
1071 rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
1073 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1075 spin_lock(&buf->rb_mwlock);
1076 list_add_tail(&mw->mw_list, &buf->rb_mws);
1077 spin_unlock(&buf->rb_mwlock);
1080 static struct rpcrdma_rep *
1081 rpcrdma_buffer_get_rep(struct rpcrdma_buffer *buffers)
1083 /* If an RPC previously completed without a reply (say, a
1084 * credential problem or a soft timeout occurs) then hold off
1085 * on supplying more Receive buffers until the number of new
1086 * pending RPCs catches up to the number of posted Receives.
1088 if (unlikely(buffers->rb_send_count < buffers->rb_recv_count))
1091 if (unlikely(list_empty(&buffers->rb_recv_bufs)))
1093 buffers->rb_recv_count++;
1094 return rpcrdma_buffer_get_rep_locked(buffers);
1098 * Get a set of request/reply buffers.
1100 * Reply buffer (if available) is attached to send buffer upon return.
1102 struct rpcrdma_req *
1103 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1105 struct rpcrdma_req *req;
1107 spin_lock(&buffers->rb_lock);
1108 if (list_empty(&buffers->rb_send_bufs))
1110 buffers->rb_send_count++;
1111 req = rpcrdma_buffer_get_req_locked(buffers);
1112 req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1113 spin_unlock(&buffers->rb_lock);
1117 spin_unlock(&buffers->rb_lock);
1118 pr_warn("RPC: %s: out of request buffers\n", __func__);
1123 * Put request/reply buffers back into pool.
1124 * Pre-decrement counter/array index.
1127 rpcrdma_buffer_put(struct rpcrdma_req *req)
1129 struct rpcrdma_buffer *buffers = req->rl_buffer;
1130 struct rpcrdma_rep *rep = req->rl_reply;
1133 req->rl_reply = NULL;
1135 spin_lock(&buffers->rb_lock);
1136 buffers->rb_send_count--;
1137 list_add_tail(&req->rl_free, &buffers->rb_send_bufs);
1139 buffers->rb_recv_count--;
1140 list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1142 spin_unlock(&buffers->rb_lock);
1146 * Recover reply buffers from pool.
1147 * This happens when recovering from disconnect.
1150 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1152 struct rpcrdma_buffer *buffers = req->rl_buffer;
1154 spin_lock(&buffers->rb_lock);
1155 req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1156 spin_unlock(&buffers->rb_lock);
1160 * Put reply buffers back into pool when not attached to
1161 * request. This happens in error conditions.
1164 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1166 struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1168 spin_lock(&buffers->rb_lock);
1169 buffers->rb_recv_count--;
1170 list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1171 spin_unlock(&buffers->rb_lock);
1175 * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1179 * rpcrdma_alloc_regbuf - kmalloc and register memory for SEND/RECV buffers
1180 * @ia: controlling rpcrdma_ia
1181 * @size: size of buffer to be allocated, in bytes
1184 * Returns pointer to private header of an area of internally
1185 * registered memory, or an ERR_PTR. The registered buffer follows
1186 * the end of the private header.
1188 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1189 * receiving the payload of RDMA RECV operations. regbufs are not
1190 * used for RDMA READ/WRITE operations, thus are registered only for
1193 struct rpcrdma_regbuf *
1194 rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size, gfp_t flags)
1196 struct rpcrdma_regbuf *rb;
1199 rb = kmalloc(sizeof(*rb) + size, flags);
1204 iov->addr = ib_dma_map_single(ia->ri_device,
1205 (void *)rb->rg_base, size,
1207 if (ib_dma_mapping_error(ia->ri_device, iov->addr))
1211 iov->lkey = ia->ri_pd->local_dma_lkey;
1213 rb->rg_owner = NULL;
1219 return ERR_PTR(-ENOMEM);
1223 * rpcrdma_free_regbuf - deregister and free registered buffer
1224 * @ia: controlling rpcrdma_ia
1225 * @rb: regbuf to be deregistered and freed
1228 rpcrdma_free_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1236 ib_dma_unmap_single(ia->ri_device,
1237 iov->addr, iov->length, DMA_BIDIRECTIONAL);
1242 * Prepost any receive buffer, then post send.
1244 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1247 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1248 struct rpcrdma_ep *ep,
1249 struct rpcrdma_req *req)
1251 struct ib_device *device = ia->ri_device;
1252 struct ib_send_wr send_wr, *send_wr_fail;
1253 struct rpcrdma_rep *rep = req->rl_reply;
1254 struct ib_sge *iov = req->rl_send_iov;
1258 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1261 req->rl_reply = NULL;
1264 send_wr.next = NULL;
1265 send_wr.wr_cqe = &req->rl_cqe;
1266 send_wr.sg_list = iov;
1267 send_wr.num_sge = req->rl_niovs;
1268 send_wr.opcode = IB_WR_SEND;
1270 for (i = 0; i < send_wr.num_sge; i++)
1271 ib_dma_sync_single_for_device(device, iov[i].addr,
1272 iov[i].length, DMA_TO_DEVICE);
1273 dprintk("RPC: %s: posting %d s/g entries\n",
1274 __func__, send_wr.num_sge);
1276 if (DECR_CQCOUNT(ep) > 0)
1277 send_wr.send_flags = 0;
1278 else { /* Provider must take a send completion every now and then */
1280 send_wr.send_flags = IB_SEND_SIGNALED;
1283 rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
1285 goto out_postsend_err;
1289 pr_err("rpcrdma: RDMA Send ib_post_send returned %i\n", rc);
1294 * (Re)post a receive buffer.
1297 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1298 struct rpcrdma_ep *ep,
1299 struct rpcrdma_rep *rep)
1301 struct ib_recv_wr recv_wr, *recv_wr_fail;
1304 recv_wr.next = NULL;
1305 recv_wr.wr_cqe = &rep->rr_cqe;
1306 recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1307 recv_wr.num_sge = 1;
1309 ib_dma_sync_single_for_cpu(ia->ri_device,
1310 rdmab_addr(rep->rr_rdmabuf),
1311 rdmab_length(rep->rr_rdmabuf),
1314 rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
1320 pr_err("rpcrdma: ib_post_recv returned %i\n", rc);
1325 * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1326 * @r_xprt: transport associated with these backchannel resources
1327 * @min_reqs: minimum number of incoming requests expected
1329 * Returns zero if all requested buffers were posted, or a negative errno.
1332 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count)
1334 struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
1335 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1336 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1337 struct rpcrdma_rep *rep;
1341 spin_lock(&buffers->rb_lock);
1342 if (list_empty(&buffers->rb_recv_bufs))
1344 rep = rpcrdma_buffer_get_rep_locked(buffers);
1345 spin_unlock(&buffers->rb_lock);
1347 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1355 spin_unlock(&buffers->rb_lock);
1356 pr_warn("%s: no extra receive buffers\n", __func__);
1360 rpcrdma_recv_buffer_put(rep);