1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
3 * Copyright (c) 2016-2018 Oracle. All rights reserved.
4 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
5 * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the BSD-type
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
17 * Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
20 * Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials provided
23 * with the distribution.
25 * Neither the name of the Network Appliance, Inc. nor the names of
26 * its contributors may be used to endorse or promote products
27 * derived from this software without specific prior written
30 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
31 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
32 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
33 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
34 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
35 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
36 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
37 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
38 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
39 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
40 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
42 * Author: Tom Tucker <tom@opengridcomputing.com>
47 * The main entry point is svc_rdma_recvfrom. This is called from
48 * svc_recv when the transport indicates there is incoming data to
49 * be read. "Data Ready" is signaled when an RDMA Receive completes,
50 * or when a set of RDMA Reads complete.
52 * An svc_rqst is passed in. This structure contains an array of
53 * free pages (rq_pages) that will contain the incoming RPC message.
55 * Short messages are moved directly into svc_rqst::rq_arg, and
56 * the RPC Call is ready to be processed by the Upper Layer.
57 * svc_rdma_recvfrom returns the length of the RPC Call message,
58 * completing the reception of the RPC Call.
60 * However, when an incoming message has Read chunks,
61 * svc_rdma_recvfrom must post RDMA Reads to pull the RPC Call's
62 * data payload from the client. svc_rdma_recvfrom sets up the
63 * RDMA Reads using pages in svc_rqst::rq_pages, which are
64 * transferred to an svc_rdma_recv_ctxt for the duration of the
65 * I/O. svc_rdma_recvfrom then returns zero, since the RPC message
66 * is still not yet ready.
68 * When the Read chunk payloads have become available on the
69 * server, "Data Ready" is raised again, and svc_recv calls
70 * svc_rdma_recvfrom again. This second call may use a different
71 * svc_rqst than the first one, thus any information that needs
72 * to be preserved across these two calls is kept in an
75 * The second call to svc_rdma_recvfrom performs final assembly
76 * of the RPC Call message, using the RDMA Read sink pages kept in
77 * the svc_rdma_recv_ctxt. The xdr_buf is copied from the
78 * svc_rdma_recv_ctxt to the second svc_rqst. The second call returns
79 * the length of the completed RPC Call message.
83 * Pages under I/O must be transferred from the first svc_rqst to an
84 * svc_rdma_recv_ctxt before the first svc_rdma_recvfrom call returns.
86 * The first svc_rqst supplies pages for RDMA Reads. These are moved
87 * from rqstp::rq_pages into ctxt::pages. The consumed elements of
88 * the rq_pages array are set to NULL and refilled with the first
89 * svc_rdma_recvfrom call returns.
91 * During the second svc_rdma_recvfrom call, RDMA Read sink pages
92 * are transferred from the svc_rdma_recv_ctxt to the second svc_rqst
93 * (see rdma_read_complete() below).
96 #include <linux/spinlock.h>
97 #include <asm/unaligned.h>
98 #include <rdma/ib_verbs.h>
99 #include <rdma/rdma_cm.h>
101 #include <linux/sunrpc/xdr.h>
102 #include <linux/sunrpc/debug.h>
103 #include <linux/sunrpc/rpc_rdma.h>
104 #include <linux/sunrpc/svc_rdma.h>
106 #include "xprt_rdma.h"
107 #include <trace/events/rpcrdma.h>
109 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
111 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc);
113 static inline struct svc_rdma_recv_ctxt *
114 svc_rdma_next_recv_ctxt(struct list_head *list)
116 return list_first_entry_or_null(list, struct svc_rdma_recv_ctxt,
120 static void svc_rdma_recv_cid_init(struct svcxprt_rdma *rdma,
121 struct rpc_rdma_cid *cid)
123 cid->ci_queue_id = rdma->sc_rq_cq->res.id;
124 cid->ci_completion_id = atomic_inc_return(&rdma->sc_completion_ids);
127 static struct svc_rdma_recv_ctxt *
128 svc_rdma_recv_ctxt_alloc(struct svcxprt_rdma *rdma)
130 struct svc_rdma_recv_ctxt *ctxt;
134 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
137 buffer = kmalloc(rdma->sc_max_req_size, GFP_KERNEL);
140 addr = ib_dma_map_single(rdma->sc_pd->device, buffer,
141 rdma->sc_max_req_size, DMA_FROM_DEVICE);
142 if (ib_dma_mapping_error(rdma->sc_pd->device, addr))
145 svc_rdma_recv_cid_init(rdma, &ctxt->rc_cid);
147 ctxt->rc_recv_wr.next = NULL;
148 ctxt->rc_recv_wr.wr_cqe = &ctxt->rc_cqe;
149 ctxt->rc_recv_wr.sg_list = &ctxt->rc_recv_sge;
150 ctxt->rc_recv_wr.num_sge = 1;
151 ctxt->rc_cqe.done = svc_rdma_wc_receive;
152 ctxt->rc_recv_sge.addr = addr;
153 ctxt->rc_recv_sge.length = rdma->sc_max_req_size;
154 ctxt->rc_recv_sge.lkey = rdma->sc_pd->local_dma_lkey;
155 ctxt->rc_recv_buf = buffer;
156 ctxt->rc_temp = false;
167 static void svc_rdma_recv_ctxt_destroy(struct svcxprt_rdma *rdma,
168 struct svc_rdma_recv_ctxt *ctxt)
170 ib_dma_unmap_single(rdma->sc_pd->device, ctxt->rc_recv_sge.addr,
171 ctxt->rc_recv_sge.length, DMA_FROM_DEVICE);
172 kfree(ctxt->rc_recv_buf);
177 * svc_rdma_recv_ctxts_destroy - Release all recv_ctxt's for an xprt
178 * @rdma: svcxprt_rdma being torn down
181 void svc_rdma_recv_ctxts_destroy(struct svcxprt_rdma *rdma)
183 struct svc_rdma_recv_ctxt *ctxt;
184 struct llist_node *node;
186 while ((node = llist_del_first(&rdma->sc_recv_ctxts))) {
187 ctxt = llist_entry(node, struct svc_rdma_recv_ctxt, rc_node);
188 svc_rdma_recv_ctxt_destroy(rdma, ctxt);
192 static struct svc_rdma_recv_ctxt *
193 svc_rdma_recv_ctxt_get(struct svcxprt_rdma *rdma)
195 struct svc_rdma_recv_ctxt *ctxt;
196 struct llist_node *node;
198 node = llist_del_first(&rdma->sc_recv_ctxts);
201 ctxt = llist_entry(node, struct svc_rdma_recv_ctxt, rc_node);
204 ctxt->rc_page_count = 0;
205 ctxt->rc_read_payload_length = 0;
209 ctxt = svc_rdma_recv_ctxt_alloc(rdma);
216 * svc_rdma_recv_ctxt_put - Return recv_ctxt to free list
217 * @rdma: controlling svcxprt_rdma
218 * @ctxt: object to return to the free list
221 void svc_rdma_recv_ctxt_put(struct svcxprt_rdma *rdma,
222 struct svc_rdma_recv_ctxt *ctxt)
226 for (i = 0; i < ctxt->rc_page_count; i++)
227 put_page(ctxt->rc_pages[i]);
230 llist_add(&ctxt->rc_node, &rdma->sc_recv_ctxts);
232 svc_rdma_recv_ctxt_destroy(rdma, ctxt);
236 * svc_rdma_release_rqst - Release transport-specific per-rqst resources
237 * @rqstp: svc_rqst being released
239 * Ensure that the recv_ctxt is released whether or not a Reply
240 * was sent. For example, the client could close the connection,
241 * or svc_process could drop an RPC, before the Reply is sent.
243 void svc_rdma_release_rqst(struct svc_rqst *rqstp)
245 struct svc_rdma_recv_ctxt *ctxt = rqstp->rq_xprt_ctxt;
246 struct svc_xprt *xprt = rqstp->rq_xprt;
247 struct svcxprt_rdma *rdma =
248 container_of(xprt, struct svcxprt_rdma, sc_xprt);
250 rqstp->rq_xprt_ctxt = NULL;
252 svc_rdma_recv_ctxt_put(rdma, ctxt);
255 static int __svc_rdma_post_recv(struct svcxprt_rdma *rdma,
256 struct svc_rdma_recv_ctxt *ctxt)
260 trace_svcrdma_post_recv(ctxt);
261 ret = ib_post_recv(rdma->sc_qp, &ctxt->rc_recv_wr, NULL);
267 trace_svcrdma_rq_post_err(rdma, ret);
268 svc_rdma_recv_ctxt_put(rdma, ctxt);
272 static int svc_rdma_post_recv(struct svcxprt_rdma *rdma)
274 struct svc_rdma_recv_ctxt *ctxt;
276 if (test_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags))
278 ctxt = svc_rdma_recv_ctxt_get(rdma);
281 return __svc_rdma_post_recv(rdma, ctxt);
285 * svc_rdma_post_recvs - Post initial set of Recv WRs
286 * @rdma: fresh svcxprt_rdma
288 * Returns true if successful, otherwise false.
290 bool svc_rdma_post_recvs(struct svcxprt_rdma *rdma)
292 struct svc_rdma_recv_ctxt *ctxt;
296 for (i = 0; i < rdma->sc_max_requests; i++) {
297 ctxt = svc_rdma_recv_ctxt_get(rdma);
300 ctxt->rc_temp = true;
301 ret = __svc_rdma_post_recv(rdma, ctxt);
309 * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
310 * @cq: Completion Queue context
311 * @wc: Work Completion object
313 * NB: The svc_xprt/svcxprt_rdma is pinned whenever it's possible that
314 * the Receive completion handler could be running.
316 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
318 struct svcxprt_rdma *rdma = cq->cq_context;
319 struct ib_cqe *cqe = wc->wr_cqe;
320 struct svc_rdma_recv_ctxt *ctxt;
322 /* WARNING: Only wc->wr_cqe and wc->status are reliable */
323 ctxt = container_of(cqe, struct svc_rdma_recv_ctxt, rc_cqe);
325 trace_svcrdma_wc_receive(wc, &ctxt->rc_cid);
326 if (wc->status != IB_WC_SUCCESS)
329 if (svc_rdma_post_recv(rdma))
332 /* All wc fields are now known to be valid */
333 ctxt->rc_byte_len = wc->byte_len;
334 ib_dma_sync_single_for_cpu(rdma->sc_pd->device,
335 ctxt->rc_recv_sge.addr,
336 wc->byte_len, DMA_FROM_DEVICE);
338 spin_lock(&rdma->sc_rq_dto_lock);
339 list_add_tail(&ctxt->rc_list, &rdma->sc_rq_dto_q);
340 /* Note the unlock pairs with the smp_rmb in svc_xprt_ready: */
341 set_bit(XPT_DATA, &rdma->sc_xprt.xpt_flags);
342 spin_unlock(&rdma->sc_rq_dto_lock);
343 if (!test_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags))
344 svc_xprt_enqueue(&rdma->sc_xprt);
349 svc_rdma_recv_ctxt_put(rdma, ctxt);
350 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
351 svc_xprt_enqueue(&rdma->sc_xprt);
355 * svc_rdma_flush_recv_queues - Drain pending Receive work
356 * @rdma: svcxprt_rdma being shut down
359 void svc_rdma_flush_recv_queues(struct svcxprt_rdma *rdma)
361 struct svc_rdma_recv_ctxt *ctxt;
363 while ((ctxt = svc_rdma_next_recv_ctxt(&rdma->sc_read_complete_q))) {
364 list_del(&ctxt->rc_list);
365 svc_rdma_recv_ctxt_put(rdma, ctxt);
367 while ((ctxt = svc_rdma_next_recv_ctxt(&rdma->sc_rq_dto_q))) {
368 list_del(&ctxt->rc_list);
369 svc_rdma_recv_ctxt_put(rdma, ctxt);
373 static void svc_rdma_build_arg_xdr(struct svc_rqst *rqstp,
374 struct svc_rdma_recv_ctxt *ctxt)
376 struct xdr_buf *arg = &rqstp->rq_arg;
378 arg->head[0].iov_base = ctxt->rc_recv_buf;
379 arg->head[0].iov_len = ctxt->rc_byte_len;
380 arg->tail[0].iov_base = NULL;
381 arg->tail[0].iov_len = 0;
384 arg->buflen = ctxt->rc_byte_len;
385 arg->len = ctxt->rc_byte_len;
388 /* This accommodates the largest possible Write chunk.
390 #define MAX_BYTES_WRITE_CHUNK ((u32)(RPCSVC_MAXPAGES << PAGE_SHIFT))
392 /* This accommodates the largest possible Position-Zero
393 * Read chunk or Reply chunk.
395 #define MAX_BYTES_SPECIAL_CHUNK ((u32)((RPCSVC_MAXPAGES + 2) << PAGE_SHIFT))
397 /* Sanity check the Read list.
399 * Implementation limits:
400 * - This implementation supports only one Read chunk.
403 * - Read list does not overflow Receive buffer.
404 * - Segment size limited by largest NFS data payload.
406 * The segment count is limited to how many segments can
407 * fit in the transport header without overflowing the
408 * buffer. That's about 40 Read segments for a 1KB inline
412 * %true: Read list is valid. @rctxt's xdr_stream is updated
413 * to point to the first byte past the Read list.
414 * %false: Read list is corrupt. @rctxt's xdr_stream is left
415 * in an unknown state.
417 static bool xdr_check_read_list(struct svc_rdma_recv_ctxt *rctxt)
423 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
429 while (xdr_item_is_present(p)) {
430 p = xdr_inline_decode(&rctxt->rc_stream,
431 rpcrdma_readseg_maxsz * sizeof(*p));
436 position = be32_to_cpup(p);
438 } else if (be32_to_cpup(p) != position) {
442 len += be32_to_cpup(p);
444 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
448 return len <= MAX_BYTES_SPECIAL_CHUNK;
451 /* The segment count is limited to how many segments can
452 * fit in the transport header without overflowing the
453 * buffer. That's about 60 Write segments for a 1KB inline
456 static bool xdr_check_write_chunk(struct svc_rdma_recv_ctxt *rctxt, u32 maxlen)
458 u32 i, segcount, total;
461 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
464 segcount = be32_to_cpup(p);
467 for (i = 0; i < segcount; i++) {
471 p = xdr_inline_decode(&rctxt->rc_stream,
472 rpcrdma_segment_maxsz * sizeof(*p));
476 xdr_decode_rdma_segment(p, &handle, &length, &offset);
477 trace_svcrdma_decode_wseg(handle, length, offset);
481 return total <= maxlen;
484 /* Sanity check the Write list.
486 * Implementation limits:
487 * - This implementation currently supports only one Write chunk.
490 * - Write list does not overflow Receive buffer.
491 * - Chunk size limited by largest NFS data payload.
494 * %true: Write list is valid. @rctxt's xdr_stream is updated
495 * to point to the first byte past the Write list.
496 * %false: Write list is corrupt. @rctxt's xdr_stream is left
497 * in an unknown state.
499 static bool xdr_check_write_list(struct svc_rdma_recv_ctxt *rctxt)
504 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
507 rctxt->rc_write_list = p;
508 while (xdr_item_is_present(p)) {
509 if (!xdr_check_write_chunk(rctxt, MAX_BYTES_WRITE_CHUNK))
512 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
517 rctxt->rc_write_list = NULL;
521 /* Sanity check the Reply chunk.
524 * - Reply chunk does not overflow Receive buffer.
525 * - Chunk size limited by largest NFS data payload.
528 * %true: Reply chunk is valid. @rctxt's xdr_stream is updated
529 * to point to the first byte past the Reply chunk.
530 * %false: Reply chunk is corrupt. @rctxt's xdr_stream is left
531 * in an unknown state.
533 static bool xdr_check_reply_chunk(struct svc_rdma_recv_ctxt *rctxt)
537 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
540 rctxt->rc_reply_chunk = NULL;
541 if (xdr_item_is_present(p)) {
542 if (!xdr_check_write_chunk(rctxt, MAX_BYTES_SPECIAL_CHUNK))
544 rctxt->rc_reply_chunk = p;
549 /* RPC-over-RDMA Version One private extension: Remote Invalidation.
550 * Responder's choice: requester signals it can handle Send With
551 * Invalidate, and responder chooses one R_key to invalidate.
553 * If there is exactly one distinct R_key in the received transport
554 * header, set rc_inv_rkey to that R_key. Otherwise, set it to zero.
556 * Perform this operation while the received transport header is
557 * still in the CPU cache.
559 static void svc_rdma_get_inv_rkey(struct svcxprt_rdma *rdma,
560 struct svc_rdma_recv_ctxt *ctxt)
565 ctxt->rc_inv_rkey = 0;
567 if (!rdma->sc_snd_w_inv)
571 p = ctxt->rc_recv_buf;
572 p += rpcrdma_fixed_maxsz;
575 while (xdr_item_is_present(p++)) {
577 if (inv_rkey == xdr_zero)
579 else if (inv_rkey != *p)
585 while (xdr_item_is_present(p++)) {
586 segcount = be32_to_cpup(p++);
587 for (i = 0; i < segcount; i++) {
588 if (inv_rkey == xdr_zero)
590 else if (inv_rkey != *p)
597 if (xdr_item_is_present(p++)) {
598 segcount = be32_to_cpup(p++);
599 for (i = 0; i < segcount; i++) {
600 if (inv_rkey == xdr_zero)
602 else if (inv_rkey != *p)
608 ctxt->rc_inv_rkey = be32_to_cpu(inv_rkey);
612 * svc_rdma_xdr_decode_req - Decode the transport header
613 * @rq_arg: xdr_buf containing ingress RPC/RDMA message
614 * @rctxt: state of decoding
616 * On entry, xdr->head[0].iov_base points to first byte of the
617 * RPC-over-RDMA transport header.
619 * On successful exit, head[0] points to first byte past the
620 * RPC-over-RDMA header. For RDMA_MSG, this is the RPC message.
622 * The length of the RPC-over-RDMA header is returned.
625 * - The transport header is entirely contained in the head iovec.
627 static int svc_rdma_xdr_decode_req(struct xdr_buf *rq_arg,
628 struct svc_rdma_recv_ctxt *rctxt)
630 __be32 *p, *rdma_argp;
631 unsigned int hdr_len;
633 rdma_argp = rq_arg->head[0].iov_base;
634 xdr_init_decode(&rctxt->rc_stream, rq_arg, rdma_argp, NULL);
636 p = xdr_inline_decode(&rctxt->rc_stream,
637 rpcrdma_fixed_maxsz * sizeof(*p));
641 if (*p != rpcrdma_version)
657 if (!xdr_check_read_list(rctxt))
659 if (!xdr_check_write_list(rctxt))
661 if (!xdr_check_reply_chunk(rctxt))
664 rq_arg->head[0].iov_base = rctxt->rc_stream.p;
665 hdr_len = xdr_stream_pos(&rctxt->rc_stream);
666 rq_arg->head[0].iov_len -= hdr_len;
667 rq_arg->len -= hdr_len;
668 trace_svcrdma_decode_rqst(rctxt, rdma_argp, hdr_len);
672 trace_svcrdma_decode_short_err(rctxt, rq_arg->len);
676 trace_svcrdma_decode_badvers_err(rctxt, rdma_argp);
677 return -EPROTONOSUPPORT;
680 trace_svcrdma_decode_drop_err(rctxt, rdma_argp);
684 trace_svcrdma_decode_badproc_err(rctxt, rdma_argp);
688 trace_svcrdma_decode_parse_err(rctxt, rdma_argp);
692 static void rdma_read_complete(struct svc_rqst *rqstp,
693 struct svc_rdma_recv_ctxt *head)
697 /* Move Read chunk pages to rqstp so that they will be released
698 * when svc_process is done with them.
700 for (page_no = 0; page_no < head->rc_page_count; page_no++) {
701 put_page(rqstp->rq_pages[page_no]);
702 rqstp->rq_pages[page_no] = head->rc_pages[page_no];
704 head->rc_page_count = 0;
706 /* Point rq_arg.pages past header */
707 rqstp->rq_arg.pages = &rqstp->rq_pages[head->rc_hdr_count];
708 rqstp->rq_arg.page_len = head->rc_arg.page_len;
710 /* rq_respages starts after the last arg page */
711 rqstp->rq_respages = &rqstp->rq_pages[page_no];
712 rqstp->rq_next_page = rqstp->rq_respages + 1;
714 /* Rebuild rq_arg head and tail. */
715 rqstp->rq_arg.head[0] = head->rc_arg.head[0];
716 rqstp->rq_arg.tail[0] = head->rc_arg.tail[0];
717 rqstp->rq_arg.len = head->rc_arg.len;
718 rqstp->rq_arg.buflen = head->rc_arg.buflen;
721 static void svc_rdma_send_error(struct svcxprt_rdma *rdma,
722 struct svc_rdma_recv_ctxt *rctxt,
725 struct svc_rdma_send_ctxt *sctxt;
727 sctxt = svc_rdma_send_ctxt_get(rdma);
730 svc_rdma_send_error_msg(rdma, sctxt, rctxt, status);
733 /* By convention, backchannel calls arrive via rdma_msg type
734 * messages, and never populate the chunk lists. This makes
735 * the RPC/RDMA header small and fixed in size, so it is
736 * straightforward to check the RPC header's direction field.
738 static bool svc_rdma_is_backchannel_reply(struct svc_xprt *xprt,
743 if (!xprt->xpt_bc_xprt)
747 if (*p++ != rdma_msg)
750 if (*p++ != xdr_zero)
752 if (*p++ != xdr_zero)
754 if (*p++ != xdr_zero)
758 if (*p++ != *rdma_resp)
761 if (*p == cpu_to_be32(RPC_CALL))
768 * svc_rdma_recvfrom - Receive an RPC call
769 * @rqstp: request structure into which to receive an RPC Call
772 * The positive number of bytes in the RPC Call message,
773 * %0 if there were no Calls ready to return,
774 * %-EINVAL if the Read chunk data is too large,
775 * %-ENOMEM if rdma_rw context pool was exhausted,
776 * %-ENOTCONN if posting failed (connection is lost),
777 * %-EIO if rdma_rw initialization failed (DMA mapping, etc).
779 * Called in a loop when XPT_DATA is set. XPT_DATA is cleared only
780 * when there are no remaining ctxt's to process.
782 * The next ctxt is removed from the "receive" lists.
784 * - If the ctxt completes a Read, then finish assembling the Call
785 * message and return the number of bytes in the message.
787 * - If the ctxt completes a Receive, then construct the Call
788 * message from the contents of the Receive buffer.
790 * - If there are no Read chunks in this message, then finish
791 * assembling the Call message and return the number of bytes
794 * - If there are Read chunks in this message, post Read WRs to
795 * pull that payload and return 0.
797 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
799 struct svc_xprt *xprt = rqstp->rq_xprt;
800 struct svcxprt_rdma *rdma_xprt =
801 container_of(xprt, struct svcxprt_rdma, sc_xprt);
802 struct svc_rdma_recv_ctxt *ctxt;
806 rqstp->rq_xprt_ctxt = NULL;
808 spin_lock(&rdma_xprt->sc_rq_dto_lock);
809 ctxt = svc_rdma_next_recv_ctxt(&rdma_xprt->sc_read_complete_q);
811 list_del(&ctxt->rc_list);
812 spin_unlock(&rdma_xprt->sc_rq_dto_lock);
813 rdma_read_complete(rqstp, ctxt);
816 ctxt = svc_rdma_next_recv_ctxt(&rdma_xprt->sc_rq_dto_q);
818 /* No new incoming requests, terminate the loop */
819 clear_bit(XPT_DATA, &xprt->xpt_flags);
820 spin_unlock(&rdma_xprt->sc_rq_dto_lock);
823 list_del(&ctxt->rc_list);
824 spin_unlock(&rdma_xprt->sc_rq_dto_lock);
826 atomic_inc(&rdma_stat_recv);
828 svc_rdma_build_arg_xdr(rqstp, ctxt);
830 /* Prevent svc_xprt_release from releasing pages in rq_pages
831 * if we return 0 or an error.
833 rqstp->rq_respages = rqstp->rq_pages;
834 rqstp->rq_next_page = rqstp->rq_respages;
836 p = (__be32 *)rqstp->rq_arg.head[0].iov_base;
837 ret = svc_rdma_xdr_decode_req(&rqstp->rq_arg, ctxt);
842 rqstp->rq_xprt_hlen = ret;
844 if (svc_rdma_is_backchannel_reply(xprt, p))
845 goto out_backchannel;
847 svc_rdma_get_inv_rkey(rdma_xprt, ctxt);
849 p += rpcrdma_fixed_maxsz;
854 rqstp->rq_xprt_ctxt = ctxt;
855 rqstp->rq_prot = IPPROTO_MAX;
856 svc_xprt_copy_addrs(rqstp, xprt);
857 return rqstp->rq_arg.len;
860 ret = svc_rdma_recv_read_chunk(rdma_xprt, rqstp, ctxt, p);
866 svc_rdma_send_error(rdma_xprt, ctxt, ret);
867 svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
872 svc_rdma_send_error(rdma_xprt, ctxt, ret);
873 svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
877 svc_rdma_handle_bc_reply(rqstp, ctxt);
879 svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);