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/pci.h> /* for Tavor hack below */
51 #include <linux/slab.h>
53 #include "xprt_rdma.h"
60 # define RPCDBG_FACILITY RPCDBG_TRANS
68 * handle replies in tasklet context, using a single, global list
69 * rdma tasklet function -- just turn around and call the func
70 * for all replies on the list
73 static DEFINE_SPINLOCK(rpcrdma_tk_lock_g);
74 static LIST_HEAD(rpcrdma_tasklets_g);
77 rpcrdma_run_tasklet(unsigned long data)
79 struct rpcrdma_rep *rep;
80 void (*func)(struct rpcrdma_rep *);
84 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
85 while (!list_empty(&rpcrdma_tasklets_g)) {
86 rep = list_entry(rpcrdma_tasklets_g.next,
87 struct rpcrdma_rep, rr_list);
88 list_del(&rep->rr_list);
91 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
96 rpcrdma_recv_buffer_put(rep);
98 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
100 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
103 static DECLARE_TASKLET(rpcrdma_tasklet_g, rpcrdma_run_tasklet, 0UL);
106 rpcrdma_schedule_tasklet(struct rpcrdma_rep *rep)
110 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
111 list_add_tail(&rep->rr_list, &rpcrdma_tasklets_g);
112 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
113 tasklet_schedule(&rpcrdma_tasklet_g);
117 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
119 struct rpcrdma_ep *ep = context;
121 dprintk("RPC: %s: QP error %X on device %s ep %p\n",
122 __func__, event->event, event->device->name, context);
123 if (ep->rep_connected == 1) {
124 ep->rep_connected = -EIO;
126 wake_up_all(&ep->rep_connect_wait);
131 rpcrdma_cq_async_error_upcall(struct ib_event *event, void *context)
133 struct rpcrdma_ep *ep = context;
135 dprintk("RPC: %s: CQ error %X on device %s ep %p\n",
136 __func__, event->event, event->device->name, context);
137 if (ep->rep_connected == 1) {
138 ep->rep_connected = -EIO;
140 wake_up_all(&ep->rep_connect_wait);
145 void rpcrdma_event_process(struct ib_wc *wc)
147 struct rpcrdma_rep *rep =
148 (struct rpcrdma_rep *)(unsigned long) wc->wr_id;
150 dprintk("RPC: %s: event rep %p status %X opcode %X length %u\n",
151 __func__, rep, wc->status, wc->opcode, wc->byte_len);
153 if (!rep) /* send or bind completion that we don't care about */
156 if (IB_WC_SUCCESS != wc->status) {
157 dprintk("RPC: %s: %s WC status %X, connection lost\n",
158 __func__, (wc->opcode & IB_WC_RECV) ? "recv" : "send",
161 rpcrdma_schedule_tasklet(rep);
165 switch (wc->opcode) {
167 rep->rr_len = wc->byte_len;
168 ib_dma_sync_single_for_cpu(
169 rdmab_to_ia(rep->rr_buffer)->ri_id->device,
170 rep->rr_iov.addr, rep->rr_len, DMA_FROM_DEVICE);
171 /* Keep (only) the most recent credits, after check validity */
172 if (rep->rr_len >= 16) {
173 struct rpcrdma_msg *p =
174 (struct rpcrdma_msg *) rep->rr_base;
175 unsigned int credits = ntohl(p->rm_credit);
177 dprintk("RPC: %s: server"
178 " dropped credits to 0!\n", __func__);
181 } else if (credits > rep->rr_buffer->rb_max_requests) {
182 dprintk("RPC: %s: server"
183 " over-crediting: %d (%d)\n",
185 rep->rr_buffer->rb_max_requests);
186 credits = rep->rr_buffer->rb_max_requests;
188 atomic_set(&rep->rr_buffer->rb_credits, credits);
192 rpcrdma_schedule_tasklet(rep);
195 dprintk("RPC: %s: unexpected WC event %X\n",
196 __func__, wc->opcode);
202 rpcrdma_cq_poll(struct ib_cq *cq)
208 rc = ib_poll_cq(cq, 1, &wc);
210 dprintk("RPC: %s: ib_poll_cq failed %i\n",
217 rpcrdma_event_process(&wc);
224 * rpcrdma_cq_event_upcall
226 * This upcall handles recv, send, bind and unbind events.
227 * It is reentrant but processes single events in order to maintain
228 * ordering of receives to keep server credits.
230 * It is the responsibility of the scheduled tasklet to return
231 * recv buffers to the pool. NOTE: this affects synchronization of
232 * connection shutdown. That is, the structures required for
233 * the completion of the reply handler must remain intact until
234 * all memory has been reclaimed.
236 * Note that send events are suppressed and do not result in an upcall.
239 rpcrdma_cq_event_upcall(struct ib_cq *cq, void *context)
243 rc = rpcrdma_cq_poll(cq);
247 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
249 dprintk("RPC: %s: ib_req_notify_cq failed %i\n",
258 static const char * const conn[] = {
275 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
277 struct rpcrdma_xprt *xprt = id->context;
278 struct rpcrdma_ia *ia = &xprt->rx_ia;
279 struct rpcrdma_ep *ep = &xprt->rx_ep;
281 struct sockaddr_in *addr = (struct sockaddr_in *) &ep->rep_remote_addr;
283 struct ib_qp_attr attr;
284 struct ib_qp_init_attr iattr;
287 switch (event->event) {
288 case RDMA_CM_EVENT_ADDR_RESOLVED:
289 case RDMA_CM_EVENT_ROUTE_RESOLVED:
291 complete(&ia->ri_done);
293 case RDMA_CM_EVENT_ADDR_ERROR:
294 ia->ri_async_rc = -EHOSTUNREACH;
295 dprintk("RPC: %s: CM address resolution error, ep 0x%p\n",
297 complete(&ia->ri_done);
299 case RDMA_CM_EVENT_ROUTE_ERROR:
300 ia->ri_async_rc = -ENETUNREACH;
301 dprintk("RPC: %s: CM route resolution error, ep 0x%p\n",
303 complete(&ia->ri_done);
305 case RDMA_CM_EVENT_ESTABLISHED:
307 ib_query_qp(ia->ri_id->qp, &attr,
308 IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
310 dprintk("RPC: %s: %d responder resources"
312 __func__, attr.max_dest_rd_atomic, attr.max_rd_atomic);
314 case RDMA_CM_EVENT_CONNECT_ERROR:
315 connstate = -ENOTCONN;
317 case RDMA_CM_EVENT_UNREACHABLE:
318 connstate = -ENETDOWN;
320 case RDMA_CM_EVENT_REJECTED:
321 connstate = -ECONNREFUSED;
323 case RDMA_CM_EVENT_DISCONNECTED:
324 connstate = -ECONNABORTED;
326 case RDMA_CM_EVENT_DEVICE_REMOVAL:
329 dprintk("RPC: %s: %s: %pI4:%u (ep 0x%p event 0x%x)\n",
331 (event->event <= 11) ? conn[event->event] :
332 "unknown connection error",
333 &addr->sin_addr.s_addr,
334 ntohs(addr->sin_port),
336 atomic_set(&rpcx_to_rdmax(ep->rep_xprt)->rx_buf.rb_credits, 1);
337 dprintk("RPC: %s: %sconnected\n",
338 __func__, connstate > 0 ? "" : "dis");
339 ep->rep_connected = connstate;
341 wake_up_all(&ep->rep_connect_wait);
344 dprintk("RPC: %s: unexpected CM event %d\n",
345 __func__, event->event);
350 if (connstate == 1) {
351 int ird = attr.max_dest_rd_atomic;
352 int tird = ep->rep_remote_cma.responder_resources;
353 printk(KERN_INFO "rpcrdma: connection to %pI4:%u "
354 "on %s, memreg %d slots %d ird %d%s\n",
355 &addr->sin_addr.s_addr,
356 ntohs(addr->sin_port),
357 ia->ri_id->device->name,
358 ia->ri_memreg_strategy,
359 xprt->rx_buf.rb_max_requests,
360 ird, ird < 4 && ird < tird / 2 ? " (low!)" : "");
361 } else if (connstate < 0) {
362 printk(KERN_INFO "rpcrdma: connection to %pI4:%u closed (%d)\n",
363 &addr->sin_addr.s_addr,
364 ntohs(addr->sin_port),
372 static struct rdma_cm_id *
373 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
374 struct rpcrdma_ia *ia, struct sockaddr *addr)
376 struct rdma_cm_id *id;
379 init_completion(&ia->ri_done);
381 id = rdma_create_id(rpcrdma_conn_upcall, xprt, RDMA_PS_TCP);
384 dprintk("RPC: %s: rdma_create_id() failed %i\n",
389 ia->ri_async_rc = -ETIMEDOUT;
390 rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
392 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
396 wait_for_completion_interruptible_timeout(&ia->ri_done,
397 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
398 rc = ia->ri_async_rc;
402 ia->ri_async_rc = -ETIMEDOUT;
403 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
405 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
409 wait_for_completion_interruptible_timeout(&ia->ri_done,
410 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
411 rc = ia->ri_async_rc;
423 * Drain any cq, prior to teardown.
426 rpcrdma_clean_cq(struct ib_cq *cq)
431 while (1 == ib_poll_cq(cq, 1, &wc))
435 dprintk("RPC: %s: flushed %d events (last 0x%x)\n",
436 __func__, count, wc.opcode);
440 * Exported functions.
444 * Open and initialize an Interface Adapter.
445 * o initializes fields of struct rpcrdma_ia, including
446 * interface and provider attributes and protection zone.
449 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
452 struct ib_device_attr devattr;
453 struct rpcrdma_ia *ia = &xprt->rx_ia;
455 ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
456 if (IS_ERR(ia->ri_id)) {
457 rc = PTR_ERR(ia->ri_id);
461 ia->ri_pd = ib_alloc_pd(ia->ri_id->device);
462 if (IS_ERR(ia->ri_pd)) {
463 rc = PTR_ERR(ia->ri_pd);
464 dprintk("RPC: %s: ib_alloc_pd() failed %i\n",
470 * Query the device to determine if the requested memory
471 * registration strategy is supported. If it isn't, set the
472 * strategy to a globally supported model.
474 rc = ib_query_device(ia->ri_id->device, &devattr);
476 dprintk("RPC: %s: ib_query_device failed %d\n",
481 if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY) {
482 ia->ri_have_dma_lkey = 1;
483 ia->ri_dma_lkey = ia->ri_id->device->local_dma_lkey;
487 case RPCRDMA_MEMWINDOWS:
488 case RPCRDMA_MEMWINDOWS_ASYNC:
489 if (!(devattr.device_cap_flags & IB_DEVICE_MEM_WINDOW)) {
490 dprintk("RPC: %s: MEMWINDOWS registration "
491 "specified but not supported by adapter, "
492 "using slower RPCRDMA_REGISTER\n",
494 memreg = RPCRDMA_REGISTER;
497 case RPCRDMA_MTHCAFMR:
498 if (!ia->ri_id->device->alloc_fmr) {
499 #if RPCRDMA_PERSISTENT_REGISTRATION
500 dprintk("RPC: %s: MTHCAFMR registration "
501 "specified but not supported by adapter, "
502 "using riskier RPCRDMA_ALLPHYSICAL\n",
504 memreg = RPCRDMA_ALLPHYSICAL;
506 dprintk("RPC: %s: MTHCAFMR registration "
507 "specified but not supported by adapter, "
508 "using slower RPCRDMA_REGISTER\n",
510 memreg = RPCRDMA_REGISTER;
515 /* Requires both frmr reg and local dma lkey */
516 if ((devattr.device_cap_flags &
517 (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) !=
518 (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) {
519 #if RPCRDMA_PERSISTENT_REGISTRATION
520 dprintk("RPC: %s: FRMR registration "
521 "specified but not supported by adapter, "
522 "using riskier RPCRDMA_ALLPHYSICAL\n",
524 memreg = RPCRDMA_ALLPHYSICAL;
526 dprintk("RPC: %s: FRMR registration "
527 "specified but not supported by adapter, "
528 "using slower RPCRDMA_REGISTER\n",
530 memreg = RPCRDMA_REGISTER;
537 * Optionally obtain an underlying physical identity mapping in
538 * order to do a memory window-based bind. This base registration
539 * is protected from remote access - that is enabled only by binding
540 * for the specific bytes targeted during each RPC operation, and
541 * revoked after the corresponding completion similar to a storage
545 case RPCRDMA_BOUNCEBUFFERS:
546 case RPCRDMA_REGISTER:
549 #if RPCRDMA_PERSISTENT_REGISTRATION
550 case RPCRDMA_ALLPHYSICAL:
551 mem_priv = IB_ACCESS_LOCAL_WRITE |
552 IB_ACCESS_REMOTE_WRITE |
553 IB_ACCESS_REMOTE_READ;
556 case RPCRDMA_MEMWINDOWS_ASYNC:
557 case RPCRDMA_MEMWINDOWS:
558 mem_priv = IB_ACCESS_LOCAL_WRITE |
561 case RPCRDMA_MTHCAFMR:
562 if (ia->ri_have_dma_lkey)
564 mem_priv = IB_ACCESS_LOCAL_WRITE;
566 ia->ri_bind_mem = ib_get_dma_mr(ia->ri_pd, mem_priv);
567 if (IS_ERR(ia->ri_bind_mem)) {
568 printk(KERN_ALERT "%s: ib_get_dma_mr for "
569 "phys register failed with %lX\n\t"
570 "Will continue with degraded performance\n",
571 __func__, PTR_ERR(ia->ri_bind_mem));
572 memreg = RPCRDMA_REGISTER;
573 ia->ri_bind_mem = NULL;
577 printk(KERN_ERR "%s: invalid memory registration mode %d\n",
582 dprintk("RPC: %s: memory registration strategy is %d\n",
585 /* Else will do memory reg/dereg for each chunk */
586 ia->ri_memreg_strategy = memreg;
590 rdma_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)
606 dprintk("RPC: %s: entering\n", __func__);
607 if (ia->ri_bind_mem != NULL) {
608 rc = ib_dereg_mr(ia->ri_bind_mem);
609 dprintk("RPC: %s: ib_dereg_mr returned %i\n",
612 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
614 rdma_destroy_qp(ia->ri_id);
615 rdma_destroy_id(ia->ri_id);
618 if (ia->ri_pd != NULL && !IS_ERR(ia->ri_pd)) {
619 rc = ib_dealloc_pd(ia->ri_pd);
620 dprintk("RPC: %s: ib_dealloc_pd returned %i\n",
626 * Create unconnected endpoint.
629 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
630 struct rpcrdma_create_data_internal *cdata)
632 struct ib_device_attr devattr;
635 rc = ib_query_device(ia->ri_id->device, &devattr);
637 dprintk("RPC: %s: ib_query_device failed %d\n",
642 /* check provider's send/recv wr limits */
643 if (cdata->max_requests > devattr.max_qp_wr)
644 cdata->max_requests = devattr.max_qp_wr;
646 ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
647 ep->rep_attr.qp_context = ep;
648 /* send_cq and recv_cq initialized below */
649 ep->rep_attr.srq = NULL;
650 ep->rep_attr.cap.max_send_wr = cdata->max_requests;
651 switch (ia->ri_memreg_strategy) {
653 /* Add room for frmr register and invalidate WRs.
654 * 1. FRMR reg WR for head
655 * 2. FRMR invalidate WR for head
656 * 3. FRMR reg WR for pagelist
657 * 4. FRMR invalidate WR for pagelist
658 * 5. FRMR reg WR for tail
659 * 6. FRMR invalidate WR for tail
660 * 7. The RDMA_SEND WR
662 ep->rep_attr.cap.max_send_wr *= 7;
663 if (ep->rep_attr.cap.max_send_wr > devattr.max_qp_wr) {
664 cdata->max_requests = devattr.max_qp_wr / 7;
665 if (!cdata->max_requests)
667 ep->rep_attr.cap.max_send_wr = cdata->max_requests * 7;
670 case RPCRDMA_MEMWINDOWS_ASYNC:
671 case RPCRDMA_MEMWINDOWS:
672 /* Add room for mw_binds+unbinds - overkill! */
673 ep->rep_attr.cap.max_send_wr++;
674 ep->rep_attr.cap.max_send_wr *= (2 * RPCRDMA_MAX_SEGS);
675 if (ep->rep_attr.cap.max_send_wr > devattr.max_qp_wr)
681 ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
682 ep->rep_attr.cap.max_send_sge = (cdata->padding ? 4 : 2);
683 ep->rep_attr.cap.max_recv_sge = 1;
684 ep->rep_attr.cap.max_inline_data = 0;
685 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
686 ep->rep_attr.qp_type = IB_QPT_RC;
687 ep->rep_attr.port_num = ~0;
689 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
690 "iovs: send %d recv %d\n",
692 ep->rep_attr.cap.max_send_wr,
693 ep->rep_attr.cap.max_recv_wr,
694 ep->rep_attr.cap.max_send_sge,
695 ep->rep_attr.cap.max_recv_sge);
697 /* set trigger for requesting send completion */
698 ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 /* - 1*/;
699 switch (ia->ri_memreg_strategy) {
700 case RPCRDMA_MEMWINDOWS_ASYNC:
701 case RPCRDMA_MEMWINDOWS:
702 ep->rep_cqinit -= RPCRDMA_MAX_SEGS;
707 if (ep->rep_cqinit <= 2)
711 init_waitqueue_head(&ep->rep_connect_wait);
714 * Create a single cq for receive dto and mw_bind (only ever
715 * care about unbind, really). Send completions are suppressed.
716 * Use single threaded tasklet upcalls to maintain ordering.
718 ep->rep_cq = ib_create_cq(ia->ri_id->device, rpcrdma_cq_event_upcall,
719 rpcrdma_cq_async_error_upcall, NULL,
720 ep->rep_attr.cap.max_recv_wr +
721 ep->rep_attr.cap.max_send_wr + 1, 0);
722 if (IS_ERR(ep->rep_cq)) {
723 rc = PTR_ERR(ep->rep_cq);
724 dprintk("RPC: %s: ib_create_cq failed: %i\n",
729 rc = ib_req_notify_cq(ep->rep_cq, IB_CQ_NEXT_COMP);
731 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
736 ep->rep_attr.send_cq = ep->rep_cq;
737 ep->rep_attr.recv_cq = ep->rep_cq;
739 /* Initialize cma parameters */
741 /* RPC/RDMA does not use private data */
742 ep->rep_remote_cma.private_data = NULL;
743 ep->rep_remote_cma.private_data_len = 0;
745 /* Client offers RDMA Read but does not initiate */
746 ep->rep_remote_cma.initiator_depth = 0;
747 if (ia->ri_memreg_strategy == RPCRDMA_BOUNCEBUFFERS)
748 ep->rep_remote_cma.responder_resources = 0;
749 else if (devattr.max_qp_rd_atom > 32) /* arbitrary but <= 255 */
750 ep->rep_remote_cma.responder_resources = 32;
752 ep->rep_remote_cma.responder_resources = devattr.max_qp_rd_atom;
754 ep->rep_remote_cma.retry_count = 7;
755 ep->rep_remote_cma.flow_control = 0;
756 ep->rep_remote_cma.rnr_retry_count = 0;
761 err = ib_destroy_cq(ep->rep_cq);
763 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
772 * Disconnect and destroy endpoint. After this, the only
773 * valid operations on the ep are to free it (if dynamically
774 * allocated) or re-create it.
776 * The caller's error handling must be sure to not leak the endpoint
777 * if this function fails.
780 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
784 dprintk("RPC: %s: entering, connected is %d\n",
785 __func__, ep->rep_connected);
788 rc = rpcrdma_ep_disconnect(ep, ia);
790 dprintk("RPC: %s: rpcrdma_ep_disconnect"
791 " returned %i\n", __func__, rc);
792 rdma_destroy_qp(ia->ri_id);
793 ia->ri_id->qp = NULL;
796 /* padding - could be done in rpcrdma_buffer_destroy... */
797 if (ep->rep_pad_mr) {
798 rpcrdma_deregister_internal(ia, ep->rep_pad_mr, &ep->rep_pad);
799 ep->rep_pad_mr = NULL;
802 rpcrdma_clean_cq(ep->rep_cq);
803 rc = ib_destroy_cq(ep->rep_cq);
805 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
812 * Connect unconnected endpoint.
815 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
817 struct rdma_cm_id *id;
821 if (ep->rep_connected != 0) {
822 struct rpcrdma_xprt *xprt;
824 rc = rpcrdma_ep_disconnect(ep, ia);
825 if (rc && rc != -ENOTCONN)
826 dprintk("RPC: %s: rpcrdma_ep_disconnect"
827 " status %i\n", __func__, rc);
828 rpcrdma_clean_cq(ep->rep_cq);
830 xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
831 id = rpcrdma_create_id(xprt, ia,
832 (struct sockaddr *)&xprt->rx_data.addr);
837 /* TEMP TEMP TEMP - fail if new device:
838 * Deregister/remarshal *all* requests!
839 * Close and recreate adapter, pd, etc!
840 * Re-determine all attributes still sane!
841 * More stuff I haven't thought of!
844 if (ia->ri_id->device != id->device) {
845 printk("RPC: %s: can't reconnect on "
846 "different device!\n", __func__);
852 rdma_destroy_qp(ia->ri_id);
853 rdma_destroy_id(ia->ri_id);
857 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
859 dprintk("RPC: %s: rdma_create_qp failed %i\n",
864 /* XXX Tavor device performs badly with 2K MTU! */
865 if (strnicmp(ia->ri_id->device->dma_device->bus->name, "pci", 3) == 0) {
866 struct pci_dev *pcid = to_pci_dev(ia->ri_id->device->dma_device);
867 if (pcid->device == PCI_DEVICE_ID_MELLANOX_TAVOR &&
868 (pcid->vendor == PCI_VENDOR_ID_MELLANOX ||
869 pcid->vendor == PCI_VENDOR_ID_TOPSPIN)) {
870 struct ib_qp_attr attr = {
871 .path_mtu = IB_MTU_1024
873 rc = ib_modify_qp(ia->ri_id->qp, &attr, IB_QP_PATH_MTU);
877 ep->rep_connected = 0;
879 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
881 dprintk("RPC: %s: rdma_connect() failed with %i\n",
886 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
889 * Check state. A non-peer reject indicates no listener
890 * (ECONNREFUSED), which may be a transient state. All
891 * others indicate a transport condition which has already
892 * undergone a best-effort.
894 if (ep->rep_connected == -ECONNREFUSED &&
895 ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
896 dprintk("RPC: %s: non-peer_reject, retry\n", __func__);
899 if (ep->rep_connected <= 0) {
900 /* Sometimes, the only way to reliably connect to remote
901 * CMs is to use same nonzero values for ORD and IRD. */
902 if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
903 (ep->rep_remote_cma.responder_resources == 0 ||
904 ep->rep_remote_cma.initiator_depth !=
905 ep->rep_remote_cma.responder_resources)) {
906 if (ep->rep_remote_cma.responder_resources == 0)
907 ep->rep_remote_cma.responder_resources = 1;
908 ep->rep_remote_cma.initiator_depth =
909 ep->rep_remote_cma.responder_resources;
912 rc = ep->rep_connected;
914 dprintk("RPC: %s: connected\n", __func__);
919 ep->rep_connected = rc;
924 * rpcrdma_ep_disconnect
926 * This is separate from destroy to facilitate the ability
927 * to reconnect without recreating the endpoint.
929 * This call is not reentrant, and must not be made in parallel
930 * on the same endpoint.
933 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
937 rpcrdma_clean_cq(ep->rep_cq);
938 rc = rdma_disconnect(ia->ri_id);
940 /* returns without wait if not connected */
941 wait_event_interruptible(ep->rep_connect_wait,
942 ep->rep_connected != 1);
943 dprintk("RPC: %s: after wait, %sconnected\n", __func__,
944 (ep->rep_connected == 1) ? "still " : "dis");
946 dprintk("RPC: %s: rdma_disconnect %i\n", __func__, rc);
947 ep->rep_connected = rc;
953 * Initialize buffer memory
956 rpcrdma_buffer_create(struct rpcrdma_buffer *buf, struct rpcrdma_ep *ep,
957 struct rpcrdma_ia *ia, struct rpcrdma_create_data_internal *cdata)
962 struct rpcrdma_mw *r;
964 buf->rb_max_requests = cdata->max_requests;
965 spin_lock_init(&buf->rb_lock);
966 atomic_set(&buf->rb_credits, 1);
969 * 1. arrays for send and recv pointers
970 * 2. arrays of struct rpcrdma_req to fill in pointers
971 * 3. array of struct rpcrdma_rep for replies
973 * 5. mw's, fmr's or frmr's, if any
974 * Send/recv buffers in req/rep need to be registered
977 len = buf->rb_max_requests *
978 (sizeof(struct rpcrdma_req *) + sizeof(struct rpcrdma_rep *));
979 len += cdata->padding;
980 switch (ia->ri_memreg_strategy) {
982 len += buf->rb_max_requests * RPCRDMA_MAX_SEGS *
983 sizeof(struct rpcrdma_mw);
985 case RPCRDMA_MTHCAFMR:
986 /* TBD we are perhaps overallocating here */
987 len += (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS *
988 sizeof(struct rpcrdma_mw);
990 case RPCRDMA_MEMWINDOWS_ASYNC:
991 case RPCRDMA_MEMWINDOWS:
992 len += (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS *
993 sizeof(struct rpcrdma_mw);
999 /* allocate 1, 4 and 5 in one shot */
1000 p = kzalloc(len, GFP_KERNEL);
1002 dprintk("RPC: %s: req_t/rep_t/pad kzalloc(%zd) failed\n",
1007 buf->rb_pool = p; /* for freeing it later */
1009 buf->rb_send_bufs = (struct rpcrdma_req **) p;
1010 p = (char *) &buf->rb_send_bufs[buf->rb_max_requests];
1011 buf->rb_recv_bufs = (struct rpcrdma_rep **) p;
1012 p = (char *) &buf->rb_recv_bufs[buf->rb_max_requests];
1015 * Register the zeroed pad buffer, if any.
1017 if (cdata->padding) {
1018 rc = rpcrdma_register_internal(ia, p, cdata->padding,
1019 &ep->rep_pad_mr, &ep->rep_pad);
1023 p += cdata->padding;
1026 * Allocate the fmr's, or mw's for mw_bind chunk registration.
1027 * We "cycle" the mw's in order to minimize rkey reuse,
1028 * and also reduce unbind-to-bind collision.
1030 INIT_LIST_HEAD(&buf->rb_mws);
1031 r = (struct rpcrdma_mw *)p;
1032 switch (ia->ri_memreg_strategy) {
1034 for (i = buf->rb_max_requests * RPCRDMA_MAX_SEGS; i; i--) {
1035 r->r.frmr.fr_mr = ib_alloc_fast_reg_mr(ia->ri_pd,
1037 if (IS_ERR(r->r.frmr.fr_mr)) {
1038 rc = PTR_ERR(r->r.frmr.fr_mr);
1039 dprintk("RPC: %s: ib_alloc_fast_reg_mr"
1040 " failed %i\n", __func__, rc);
1044 ib_alloc_fast_reg_page_list(ia->ri_id->device,
1046 if (IS_ERR(r->r.frmr.fr_pgl)) {
1047 rc = PTR_ERR(r->r.frmr.fr_pgl);
1049 "ib_alloc_fast_reg_page_list "
1050 "failed %i\n", __func__, rc);
1053 list_add(&r->mw_list, &buf->rb_mws);
1057 case RPCRDMA_MTHCAFMR:
1058 /* TBD we are perhaps overallocating here */
1059 for (i = (buf->rb_max_requests+1) * RPCRDMA_MAX_SEGS; i; i--) {
1060 static struct ib_fmr_attr fa =
1061 { RPCRDMA_MAX_DATA_SEGS, 1, PAGE_SHIFT };
1062 r->r.fmr = ib_alloc_fmr(ia->ri_pd,
1063 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ,
1065 if (IS_ERR(r->r.fmr)) {
1066 rc = PTR_ERR(r->r.fmr);
1067 dprintk("RPC: %s: ib_alloc_fmr"
1068 " failed %i\n", __func__, rc);
1071 list_add(&r->mw_list, &buf->rb_mws);
1075 case RPCRDMA_MEMWINDOWS_ASYNC:
1076 case RPCRDMA_MEMWINDOWS:
1077 /* Allocate one extra request's worth, for full cycling */
1078 for (i = (buf->rb_max_requests+1) * RPCRDMA_MAX_SEGS; i; i--) {
1079 r->r.mw = ib_alloc_mw(ia->ri_pd);
1080 if (IS_ERR(r->r.mw)) {
1081 rc = PTR_ERR(r->r.mw);
1082 dprintk("RPC: %s: ib_alloc_mw"
1083 " failed %i\n", __func__, rc);
1086 list_add(&r->mw_list, &buf->rb_mws);
1095 * Allocate/init the request/reply buffers. Doing this
1096 * using kmalloc for now -- one for each buf.
1098 for (i = 0; i < buf->rb_max_requests; i++) {
1099 struct rpcrdma_req *req;
1100 struct rpcrdma_rep *rep;
1102 len = cdata->inline_wsize + sizeof(struct rpcrdma_req);
1103 /* RPC layer requests *double* size + 1K RPC_SLACK_SPACE! */
1104 /* Typical ~2400b, so rounding up saves work later */
1107 req = kmalloc(len, GFP_KERNEL);
1109 dprintk("RPC: %s: request buffer %d alloc"
1110 " failed\n", __func__, i);
1114 memset(req, 0, sizeof(struct rpcrdma_req));
1115 buf->rb_send_bufs[i] = req;
1116 buf->rb_send_bufs[i]->rl_buffer = buf;
1118 rc = rpcrdma_register_internal(ia, req->rl_base,
1119 len - offsetof(struct rpcrdma_req, rl_base),
1120 &buf->rb_send_bufs[i]->rl_handle,
1121 &buf->rb_send_bufs[i]->rl_iov);
1125 buf->rb_send_bufs[i]->rl_size = len-sizeof(struct rpcrdma_req);
1127 len = cdata->inline_rsize + sizeof(struct rpcrdma_rep);
1128 rep = kmalloc(len, GFP_KERNEL);
1130 dprintk("RPC: %s: reply buffer %d alloc failed\n",
1135 memset(rep, 0, sizeof(struct rpcrdma_rep));
1136 buf->rb_recv_bufs[i] = rep;
1137 buf->rb_recv_bufs[i]->rr_buffer = buf;
1138 init_waitqueue_head(&rep->rr_unbind);
1140 rc = rpcrdma_register_internal(ia, rep->rr_base,
1141 len - offsetof(struct rpcrdma_rep, rr_base),
1142 &buf->rb_recv_bufs[i]->rr_handle,
1143 &buf->rb_recv_bufs[i]->rr_iov);
1148 dprintk("RPC: %s: max_requests %d\n",
1149 __func__, buf->rb_max_requests);
1153 rpcrdma_buffer_destroy(buf);
1158 * Unregister and destroy buffer memory. Need to deal with
1159 * partial initialization, so it's callable from failed create.
1160 * Must be called before destroying endpoint, as registrations
1164 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1167 struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1168 struct rpcrdma_mw *r;
1170 /* clean up in reverse order from create
1171 * 1. recv mr memory (mr free, then kfree)
1172 * 1a. bind mw memory
1173 * 2. send mr memory (mr free, then kfree)
1174 * 3. padding (if any) [moved to rpcrdma_ep_destroy]
1177 dprintk("RPC: %s: entering\n", __func__);
1179 for (i = 0; i < buf->rb_max_requests; i++) {
1180 if (buf->rb_recv_bufs && buf->rb_recv_bufs[i]) {
1181 rpcrdma_deregister_internal(ia,
1182 buf->rb_recv_bufs[i]->rr_handle,
1183 &buf->rb_recv_bufs[i]->rr_iov);
1184 kfree(buf->rb_recv_bufs[i]);
1186 if (buf->rb_send_bufs && buf->rb_send_bufs[i]) {
1187 while (!list_empty(&buf->rb_mws)) {
1188 r = list_entry(buf->rb_mws.next,
1189 struct rpcrdma_mw, mw_list);
1190 list_del(&r->mw_list);
1191 switch (ia->ri_memreg_strategy) {
1193 rc = ib_dereg_mr(r->r.frmr.fr_mr);
1199 ib_free_fast_reg_page_list(r->r.frmr.fr_pgl);
1201 case RPCRDMA_MTHCAFMR:
1202 rc = ib_dealloc_fmr(r->r.fmr);
1209 case RPCRDMA_MEMWINDOWS_ASYNC:
1210 case RPCRDMA_MEMWINDOWS:
1211 rc = ib_dealloc_mw(r->r.mw);
1222 rpcrdma_deregister_internal(ia,
1223 buf->rb_send_bufs[i]->rl_handle,
1224 &buf->rb_send_bufs[i]->rl_iov);
1225 kfree(buf->rb_send_bufs[i]);
1229 kfree(buf->rb_pool);
1233 * Get a set of request/reply buffers.
1235 * Reply buffer (if needed) is attached to send buffer upon return.
1237 * rb_send_index and rb_recv_index MUST always be pointing to the
1238 * *next* available buffer (non-NULL). They are incremented after
1239 * removing buffers, and decremented *before* returning them.
1241 struct rpcrdma_req *
1242 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1244 struct rpcrdma_req *req;
1245 unsigned long flags;
1247 struct rpcrdma_mw *r;
1249 spin_lock_irqsave(&buffers->rb_lock, flags);
1250 if (buffers->rb_send_index == buffers->rb_max_requests) {
1251 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1252 dprintk("RPC: %s: out of request buffers\n", __func__);
1253 return ((struct rpcrdma_req *)NULL);
1256 req = buffers->rb_send_bufs[buffers->rb_send_index];
1257 if (buffers->rb_send_index < buffers->rb_recv_index) {
1258 dprintk("RPC: %s: %d extra receives outstanding (ok)\n",
1260 buffers->rb_recv_index - buffers->rb_send_index);
1261 req->rl_reply = NULL;
1263 req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1264 buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1266 buffers->rb_send_bufs[buffers->rb_send_index++] = NULL;
1267 if (!list_empty(&buffers->rb_mws)) {
1268 i = RPCRDMA_MAX_SEGS - 1;
1270 r = list_entry(buffers->rb_mws.next,
1271 struct rpcrdma_mw, mw_list);
1272 list_del(&r->mw_list);
1273 req->rl_segments[i].mr_chunk.rl_mw = r;
1276 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1281 * Put request/reply buffers back into pool.
1282 * Pre-decrement counter/array index.
1285 rpcrdma_buffer_put(struct rpcrdma_req *req)
1287 struct rpcrdma_buffer *buffers = req->rl_buffer;
1288 struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
1290 unsigned long flags;
1292 BUG_ON(req->rl_nchunks != 0);
1293 spin_lock_irqsave(&buffers->rb_lock, flags);
1294 buffers->rb_send_bufs[--buffers->rb_send_index] = req;
1296 if (req->rl_reply) {
1297 buffers->rb_recv_bufs[--buffers->rb_recv_index] = req->rl_reply;
1298 init_waitqueue_head(&req->rl_reply->rr_unbind);
1299 req->rl_reply->rr_func = NULL;
1300 req->rl_reply = NULL;
1302 switch (ia->ri_memreg_strategy) {
1304 case RPCRDMA_MTHCAFMR:
1305 case RPCRDMA_MEMWINDOWS_ASYNC:
1306 case RPCRDMA_MEMWINDOWS:
1308 * Cycle mw's back in reverse order, and "spin" them.
1309 * This delays and scrambles reuse as much as possible.
1313 struct rpcrdma_mw **mw;
1314 mw = &req->rl_segments[i].mr_chunk.rl_mw;
1315 list_add_tail(&(*mw)->mw_list, &buffers->rb_mws);
1317 } while (++i < RPCRDMA_MAX_SEGS);
1318 list_add_tail(&req->rl_segments[0].mr_chunk.rl_mw->mw_list,
1320 req->rl_segments[0].mr_chunk.rl_mw = NULL;
1325 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1329 * Recover reply buffers from pool.
1330 * This happens when recovering from error conditions.
1331 * Post-increment counter/array index.
1334 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1336 struct rpcrdma_buffer *buffers = req->rl_buffer;
1337 unsigned long flags;
1339 if (req->rl_iov.length == 0) /* special case xprt_rdma_allocate() */
1340 buffers = ((struct rpcrdma_req *) buffers)->rl_buffer;
1341 spin_lock_irqsave(&buffers->rb_lock, flags);
1342 if (buffers->rb_recv_index < buffers->rb_max_requests) {
1343 req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1344 buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1346 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1350 * Put reply buffers back into pool when not attached to
1351 * request. This happens in error conditions, and when
1352 * aborting unbinds. Pre-decrement counter/array index.
1355 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1357 struct rpcrdma_buffer *buffers = rep->rr_buffer;
1358 unsigned long flags;
1360 rep->rr_func = NULL;
1361 spin_lock_irqsave(&buffers->rb_lock, flags);
1362 buffers->rb_recv_bufs[--buffers->rb_recv_index] = rep;
1363 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1367 * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1371 rpcrdma_register_internal(struct rpcrdma_ia *ia, void *va, int len,
1372 struct ib_mr **mrp, struct ib_sge *iov)
1374 struct ib_phys_buf ipb;
1379 * All memory passed here was kmalloc'ed, therefore phys-contiguous.
1381 iov->addr = ib_dma_map_single(ia->ri_id->device,
1382 va, len, DMA_BIDIRECTIONAL);
1385 if (ia->ri_have_dma_lkey) {
1387 iov->lkey = ia->ri_dma_lkey;
1389 } else if (ia->ri_bind_mem != NULL) {
1391 iov->lkey = ia->ri_bind_mem->lkey;
1395 ipb.addr = iov->addr;
1396 ipb.size = iov->length;
1397 mr = ib_reg_phys_mr(ia->ri_pd, &ipb, 1,
1398 IB_ACCESS_LOCAL_WRITE, &iov->addr);
1400 dprintk("RPC: %s: phys convert: 0x%llx "
1401 "registered 0x%llx length %d\n",
1402 __func__, (unsigned long long)ipb.addr,
1403 (unsigned long long)iov->addr, len);
1408 dprintk("RPC: %s: failed with %i\n", __func__, rc);
1411 iov->lkey = mr->lkey;
1419 rpcrdma_deregister_internal(struct rpcrdma_ia *ia,
1420 struct ib_mr *mr, struct ib_sge *iov)
1424 ib_dma_unmap_single(ia->ri_id->device,
1425 iov->addr, iov->length, DMA_BIDIRECTIONAL);
1430 rc = ib_dereg_mr(mr);
1432 dprintk("RPC: %s: ib_dereg_mr failed %i\n", __func__, rc);
1437 * Wrappers for chunk registration, shared by read/write chunk code.
1441 rpcrdma_map_one(struct rpcrdma_ia *ia, struct rpcrdma_mr_seg *seg, int writing)
1443 seg->mr_dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
1444 seg->mr_dmalen = seg->mr_len;
1446 seg->mr_dma = ib_dma_map_page(ia->ri_id->device,
1447 seg->mr_page, offset_in_page(seg->mr_offset),
1448 seg->mr_dmalen, seg->mr_dir);
1450 seg->mr_dma = ib_dma_map_single(ia->ri_id->device,
1452 seg->mr_dmalen, seg->mr_dir);
1456 rpcrdma_unmap_one(struct rpcrdma_ia *ia, struct rpcrdma_mr_seg *seg)
1459 ib_dma_unmap_page(ia->ri_id->device,
1460 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
1462 ib_dma_unmap_single(ia->ri_id->device,
1463 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
1467 rpcrdma_register_frmr_external(struct rpcrdma_mr_seg *seg,
1468 int *nsegs, int writing, struct rpcrdma_ia *ia,
1469 struct rpcrdma_xprt *r_xprt)
1471 struct rpcrdma_mr_seg *seg1 = seg;
1472 struct ib_send_wr frmr_wr, *bad_wr;
1477 pageoff = offset_in_page(seg1->mr_offset);
1478 seg1->mr_offset -= pageoff; /* start of page */
1479 seg1->mr_len += pageoff;
1481 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1482 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1483 for (i = 0; i < *nsegs;) {
1484 rpcrdma_map_one(ia, seg, writing);
1485 seg1->mr_chunk.rl_mw->r.frmr.fr_pgl->page_list[i] = seg->mr_dma;
1489 /* Check for holes */
1490 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1491 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
1494 dprintk("RPC: %s: Using frmr %p to map %d segments\n",
1495 __func__, seg1->mr_chunk.rl_mw, i);
1498 key = (u8)(seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey & 0x000000FF);
1499 ib_update_fast_reg_key(seg1->mr_chunk.rl_mw->r.frmr.fr_mr, ++key);
1501 /* Prepare FRMR WR */
1502 memset(&frmr_wr, 0, sizeof frmr_wr);
1503 frmr_wr.opcode = IB_WR_FAST_REG_MR;
1504 frmr_wr.send_flags = 0; /* unsignaled */
1505 frmr_wr.wr.fast_reg.iova_start = seg1->mr_dma;
1506 frmr_wr.wr.fast_reg.page_list = seg1->mr_chunk.rl_mw->r.frmr.fr_pgl;
1507 frmr_wr.wr.fast_reg.page_list_len = i;
1508 frmr_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1509 frmr_wr.wr.fast_reg.length = i << PAGE_SHIFT;
1510 frmr_wr.wr.fast_reg.access_flags = (writing ?
1511 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
1512 IB_ACCESS_REMOTE_READ);
1513 frmr_wr.wr.fast_reg.rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1514 DECR_CQCOUNT(&r_xprt->rx_ep);
1516 rc = ib_post_send(ia->ri_id->qp, &frmr_wr, &bad_wr);
1519 dprintk("RPC: %s: failed ib_post_send for register,"
1520 " status %i\n", __func__, rc);
1522 rpcrdma_unmap_one(ia, --seg);
1524 seg1->mr_rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1525 seg1->mr_base = seg1->mr_dma + pageoff;
1534 rpcrdma_deregister_frmr_external(struct rpcrdma_mr_seg *seg,
1535 struct rpcrdma_ia *ia, struct rpcrdma_xprt *r_xprt)
1537 struct rpcrdma_mr_seg *seg1 = seg;
1538 struct ib_send_wr invalidate_wr, *bad_wr;
1541 while (seg1->mr_nsegs--)
1542 rpcrdma_unmap_one(ia, seg++);
1544 memset(&invalidate_wr, 0, sizeof invalidate_wr);
1545 invalidate_wr.opcode = IB_WR_LOCAL_INV;
1546 invalidate_wr.send_flags = 0; /* unsignaled */
1547 invalidate_wr.ex.invalidate_rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1548 DECR_CQCOUNT(&r_xprt->rx_ep);
1550 rc = ib_post_send(ia->ri_id->qp, &invalidate_wr, &bad_wr);
1552 dprintk("RPC: %s: failed ib_post_send for invalidate,"
1553 " status %i\n", __func__, rc);
1558 rpcrdma_register_fmr_external(struct rpcrdma_mr_seg *seg,
1559 int *nsegs, int writing, struct rpcrdma_ia *ia)
1561 struct rpcrdma_mr_seg *seg1 = seg;
1562 u64 physaddrs[RPCRDMA_MAX_DATA_SEGS];
1563 int len, pageoff, i, rc;
1565 pageoff = offset_in_page(seg1->mr_offset);
1566 seg1->mr_offset -= pageoff; /* start of page */
1567 seg1->mr_len += pageoff;
1569 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1570 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1571 for (i = 0; i < *nsegs;) {
1572 rpcrdma_map_one(ia, seg, writing);
1573 physaddrs[i] = seg->mr_dma;
1577 /* Check for holes */
1578 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1579 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
1582 rc = ib_map_phys_fmr(seg1->mr_chunk.rl_mw->r.fmr,
1583 physaddrs, i, seg1->mr_dma);
1585 dprintk("RPC: %s: failed ib_map_phys_fmr "
1586 "%u@0x%llx+%i (%d)... status %i\n", __func__,
1587 len, (unsigned long long)seg1->mr_dma,
1590 rpcrdma_unmap_one(ia, --seg);
1592 seg1->mr_rkey = seg1->mr_chunk.rl_mw->r.fmr->rkey;
1593 seg1->mr_base = seg1->mr_dma + pageoff;
1602 rpcrdma_deregister_fmr_external(struct rpcrdma_mr_seg *seg,
1603 struct rpcrdma_ia *ia)
1605 struct rpcrdma_mr_seg *seg1 = seg;
1609 list_add(&seg1->mr_chunk.rl_mw->r.fmr->list, &l);
1610 rc = ib_unmap_fmr(&l);
1611 while (seg1->mr_nsegs--)
1612 rpcrdma_unmap_one(ia, seg++);
1614 dprintk("RPC: %s: failed ib_unmap_fmr,"
1615 " status %i\n", __func__, rc);
1620 rpcrdma_register_memwin_external(struct rpcrdma_mr_seg *seg,
1621 int *nsegs, int writing, struct rpcrdma_ia *ia,
1622 struct rpcrdma_xprt *r_xprt)
1624 int mem_priv = (writing ? IB_ACCESS_REMOTE_WRITE :
1625 IB_ACCESS_REMOTE_READ);
1626 struct ib_mw_bind param;
1630 rpcrdma_map_one(ia, seg, writing);
1631 param.mr = ia->ri_bind_mem;
1632 param.wr_id = 0ULL; /* no send cookie */
1633 param.addr = seg->mr_dma;
1634 param.length = seg->mr_len;
1635 param.send_flags = 0;
1636 param.mw_access_flags = mem_priv;
1638 DECR_CQCOUNT(&r_xprt->rx_ep);
1639 rc = ib_bind_mw(ia->ri_id->qp, seg->mr_chunk.rl_mw->r.mw, ¶m);
1641 dprintk("RPC: %s: failed ib_bind_mw "
1642 "%u@0x%llx status %i\n",
1643 __func__, seg->mr_len,
1644 (unsigned long long)seg->mr_dma, rc);
1645 rpcrdma_unmap_one(ia, seg);
1647 seg->mr_rkey = seg->mr_chunk.rl_mw->r.mw->rkey;
1648 seg->mr_base = param.addr;
1655 rpcrdma_deregister_memwin_external(struct rpcrdma_mr_seg *seg,
1656 struct rpcrdma_ia *ia,
1657 struct rpcrdma_xprt *r_xprt, void **r)
1659 struct ib_mw_bind param;
1663 BUG_ON(seg->mr_nsegs != 1);
1664 param.mr = ia->ri_bind_mem;
1665 param.addr = 0ULL; /* unbind */
1667 param.mw_access_flags = 0;
1669 param.wr_id = (u64) (unsigned long) *r;
1670 param.send_flags = IB_SEND_SIGNALED;
1671 INIT_CQCOUNT(&r_xprt->rx_ep);
1674 param.send_flags = 0;
1675 DECR_CQCOUNT(&r_xprt->rx_ep);
1677 rc = ib_bind_mw(ia->ri_id->qp, seg->mr_chunk.rl_mw->r.mw, ¶m);
1678 rpcrdma_unmap_one(ia, seg);
1680 dprintk("RPC: %s: failed ib_(un)bind_mw,"
1681 " status %i\n", __func__, rc);
1683 *r = NULL; /* will upcall on completion */
1688 rpcrdma_register_default_external(struct rpcrdma_mr_seg *seg,
1689 int *nsegs, int writing, struct rpcrdma_ia *ia)
1691 int mem_priv = (writing ? IB_ACCESS_REMOTE_WRITE :
1692 IB_ACCESS_REMOTE_READ);
1693 struct rpcrdma_mr_seg *seg1 = seg;
1694 struct ib_phys_buf ipb[RPCRDMA_MAX_DATA_SEGS];
1697 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1698 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1699 for (len = 0, i = 0; i < *nsegs;) {
1700 rpcrdma_map_one(ia, seg, writing);
1701 ipb[i].addr = seg->mr_dma;
1702 ipb[i].size = seg->mr_len;
1706 /* Check for holes */
1707 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1708 offset_in_page((seg-1)->mr_offset+(seg-1)->mr_len))
1711 seg1->mr_base = seg1->mr_dma;
1712 seg1->mr_chunk.rl_mr = ib_reg_phys_mr(ia->ri_pd,
1713 ipb, i, mem_priv, &seg1->mr_base);
1714 if (IS_ERR(seg1->mr_chunk.rl_mr)) {
1715 rc = PTR_ERR(seg1->mr_chunk.rl_mr);
1716 dprintk("RPC: %s: failed ib_reg_phys_mr "
1717 "%u@0x%llx (%d)... status %i\n",
1719 (unsigned long long)seg1->mr_dma, i, rc);
1721 rpcrdma_unmap_one(ia, --seg);
1723 seg1->mr_rkey = seg1->mr_chunk.rl_mr->rkey;
1732 rpcrdma_deregister_default_external(struct rpcrdma_mr_seg *seg,
1733 struct rpcrdma_ia *ia)
1735 struct rpcrdma_mr_seg *seg1 = seg;
1738 rc = ib_dereg_mr(seg1->mr_chunk.rl_mr);
1739 seg1->mr_chunk.rl_mr = NULL;
1740 while (seg1->mr_nsegs--)
1741 rpcrdma_unmap_one(ia, seg++);
1743 dprintk("RPC: %s: failed ib_dereg_mr,"
1744 " status %i\n", __func__, rc);
1749 rpcrdma_register_external(struct rpcrdma_mr_seg *seg,
1750 int nsegs, int writing, struct rpcrdma_xprt *r_xprt)
1752 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1755 switch (ia->ri_memreg_strategy) {
1757 #if RPCRDMA_PERSISTENT_REGISTRATION
1758 case RPCRDMA_ALLPHYSICAL:
1759 rpcrdma_map_one(ia, seg, writing);
1760 seg->mr_rkey = ia->ri_bind_mem->rkey;
1761 seg->mr_base = seg->mr_dma;
1767 /* Registration using frmr registration */
1769 rc = rpcrdma_register_frmr_external(seg, &nsegs, writing, ia, r_xprt);
1772 /* Registration using fmr memory registration */
1773 case RPCRDMA_MTHCAFMR:
1774 rc = rpcrdma_register_fmr_external(seg, &nsegs, writing, ia);
1777 /* Registration using memory windows */
1778 case RPCRDMA_MEMWINDOWS_ASYNC:
1779 case RPCRDMA_MEMWINDOWS:
1780 rc = rpcrdma_register_memwin_external(seg, &nsegs, writing, ia, r_xprt);
1783 /* Default registration each time */
1785 rc = rpcrdma_register_default_external(seg, &nsegs, writing, ia);
1795 rpcrdma_deregister_external(struct rpcrdma_mr_seg *seg,
1796 struct rpcrdma_xprt *r_xprt, void *r)
1798 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1799 int nsegs = seg->mr_nsegs, rc;
1801 switch (ia->ri_memreg_strategy) {
1803 #if RPCRDMA_PERSISTENT_REGISTRATION
1804 case RPCRDMA_ALLPHYSICAL:
1806 rpcrdma_unmap_one(ia, seg);
1812 rc = rpcrdma_deregister_frmr_external(seg, ia, r_xprt);
1815 case RPCRDMA_MTHCAFMR:
1816 rc = rpcrdma_deregister_fmr_external(seg, ia);
1819 case RPCRDMA_MEMWINDOWS_ASYNC:
1820 case RPCRDMA_MEMWINDOWS:
1821 rc = rpcrdma_deregister_memwin_external(seg, ia, r_xprt, &r);
1825 rc = rpcrdma_deregister_default_external(seg, ia);
1829 struct rpcrdma_rep *rep = r;
1830 void (*func)(struct rpcrdma_rep *) = rep->rr_func;
1831 rep->rr_func = NULL;
1832 func(rep); /* dereg done, callback now */
1838 * Prepost any receive buffer, then post send.
1840 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1843 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1844 struct rpcrdma_ep *ep,
1845 struct rpcrdma_req *req)
1847 struct ib_send_wr send_wr, *send_wr_fail;
1848 struct rpcrdma_rep *rep = req->rl_reply;
1852 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1855 req->rl_reply = NULL;
1858 send_wr.next = NULL;
1859 send_wr.wr_id = 0ULL; /* no send cookie */
1860 send_wr.sg_list = req->rl_send_iov;
1861 send_wr.num_sge = req->rl_niovs;
1862 send_wr.opcode = IB_WR_SEND;
1863 if (send_wr.num_sge == 4) /* no need to sync any pad (constant) */
1864 ib_dma_sync_single_for_device(ia->ri_id->device,
1865 req->rl_send_iov[3].addr, req->rl_send_iov[3].length,
1867 ib_dma_sync_single_for_device(ia->ri_id->device,
1868 req->rl_send_iov[1].addr, req->rl_send_iov[1].length,
1870 ib_dma_sync_single_for_device(ia->ri_id->device,
1871 req->rl_send_iov[0].addr, req->rl_send_iov[0].length,
1874 if (DECR_CQCOUNT(ep) > 0)
1875 send_wr.send_flags = 0;
1876 else { /* Provider must take a send completion every now and then */
1878 send_wr.send_flags = IB_SEND_SIGNALED;
1881 rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
1883 dprintk("RPC: %s: ib_post_send returned %i\n", __func__,
1890 * (Re)post a receive buffer.
1893 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1894 struct rpcrdma_ep *ep,
1895 struct rpcrdma_rep *rep)
1897 struct ib_recv_wr recv_wr, *recv_wr_fail;
1900 recv_wr.next = NULL;
1901 recv_wr.wr_id = (u64) (unsigned long) rep;
1902 recv_wr.sg_list = &rep->rr_iov;
1903 recv_wr.num_sge = 1;
1905 ib_dma_sync_single_for_cpu(ia->ri_id->device,
1906 rep->rr_iov.addr, rep->rr_iov.length, DMA_BIDIRECTIONAL);
1909 rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
1912 dprintk("RPC: %s: ib_post_recv returned %i\n", __func__,