1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
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.
6  *
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
11  * license below:
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  *
17  *      Redistributions of source code must retain the above copyright
18  *      notice, this list of conditions and the following disclaimer.
19  *
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.
24  *
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
28  *      permission.
29  *
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.
41  *
42  * Author: Tom Tucker <tom@opengridcomputing.com>
43  */
44 
45 /* Operation
46  *
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.
51  *
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.
54  *
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.
59  *
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.
67  *
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
73  * svc_rdma_recv_ctxt.
74  *
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.
80  *
81  * Page Management
82  *
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.
85  *
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.
90  *
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).
94  */
95 
96 #include <linux/spinlock.h>
97 #include <asm/unaligned.h>
98 #include <rdma/ib_verbs.h>
99 #include <rdma/rdma_cm.h>
100 
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>
105 
106 #include "xprt_rdma.h"
107 #include <trace/events/rpcrdma.h>
108 
109 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
110 
111 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc);
112 
113 static inline struct svc_rdma_recv_ctxt *
114 svc_rdma_next_recv_ctxt(struct list_head *list)
115 {
116 	return list_first_entry_or_null(list, struct svc_rdma_recv_ctxt,
117 					rc_list);
118 }
119 
120 static struct svc_rdma_recv_ctxt *
121 svc_rdma_recv_ctxt_alloc(struct svcxprt_rdma *rdma)
122 {
123 	struct svc_rdma_recv_ctxt *ctxt;
124 	dma_addr_t addr;
125 	void *buffer;
126 
127 	ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
128 	if (!ctxt)
129 		goto fail0;
130 	buffer = kmalloc(rdma->sc_max_req_size, GFP_KERNEL);
131 	if (!buffer)
132 		goto fail1;
133 	addr = ib_dma_map_single(rdma->sc_pd->device, buffer,
134 				 rdma->sc_max_req_size, DMA_FROM_DEVICE);
135 	if (ib_dma_mapping_error(rdma->sc_pd->device, addr))
136 		goto fail2;
137 
138 	ctxt->rc_recv_wr.next = NULL;
139 	ctxt->rc_recv_wr.wr_cqe = &ctxt->rc_cqe;
140 	ctxt->rc_recv_wr.sg_list = &ctxt->rc_recv_sge;
141 	ctxt->rc_recv_wr.num_sge = 1;
142 	ctxt->rc_cqe.done = svc_rdma_wc_receive;
143 	ctxt->rc_recv_sge.addr = addr;
144 	ctxt->rc_recv_sge.length = rdma->sc_max_req_size;
145 	ctxt->rc_recv_sge.lkey = rdma->sc_pd->local_dma_lkey;
146 	ctxt->rc_recv_buf = buffer;
147 	ctxt->rc_temp = false;
148 	return ctxt;
149 
150 fail2:
151 	kfree(buffer);
152 fail1:
153 	kfree(ctxt);
154 fail0:
155 	return NULL;
156 }
157 
158 static void svc_rdma_recv_ctxt_destroy(struct svcxprt_rdma *rdma,
159 				       struct svc_rdma_recv_ctxt *ctxt)
160 {
161 	ib_dma_unmap_single(rdma->sc_pd->device, ctxt->rc_recv_sge.addr,
162 			    ctxt->rc_recv_sge.length, DMA_FROM_DEVICE);
163 	kfree(ctxt->rc_recv_buf);
164 	kfree(ctxt);
165 }
166 
167 /**
168  * svc_rdma_recv_ctxts_destroy - Release all recv_ctxt's for an xprt
169  * @rdma: svcxprt_rdma being torn down
170  *
171  */
172 void svc_rdma_recv_ctxts_destroy(struct svcxprt_rdma *rdma)
173 {
174 	struct svc_rdma_recv_ctxt *ctxt;
175 	struct llist_node *node;
176 
177 	while ((node = llist_del_first(&rdma->sc_recv_ctxts))) {
178 		ctxt = llist_entry(node, struct svc_rdma_recv_ctxt, rc_node);
179 		svc_rdma_recv_ctxt_destroy(rdma, ctxt);
180 	}
181 }
182 
183 static struct svc_rdma_recv_ctxt *
184 svc_rdma_recv_ctxt_get(struct svcxprt_rdma *rdma)
185 {
186 	struct svc_rdma_recv_ctxt *ctxt;
187 	struct llist_node *node;
188 
189 	node = llist_del_first(&rdma->sc_recv_ctxts);
190 	if (!node)
191 		goto out_empty;
192 	ctxt = llist_entry(node, struct svc_rdma_recv_ctxt, rc_node);
193 
194 out:
195 	ctxt->rc_page_count = 0;
196 	ctxt->rc_read_payload_length = 0;
197 	return ctxt;
198 
199 out_empty:
200 	ctxt = svc_rdma_recv_ctxt_alloc(rdma);
201 	if (!ctxt)
202 		return NULL;
203 	goto out;
204 }
205 
206 /**
207  * svc_rdma_recv_ctxt_put - Return recv_ctxt to free list
208  * @rdma: controlling svcxprt_rdma
209  * @ctxt: object to return to the free list
210  *
211  */
212 void svc_rdma_recv_ctxt_put(struct svcxprt_rdma *rdma,
213 			    struct svc_rdma_recv_ctxt *ctxt)
214 {
215 	unsigned int i;
216 
217 	for (i = 0; i < ctxt->rc_page_count; i++)
218 		put_page(ctxt->rc_pages[i]);
219 
220 	if (!ctxt->rc_temp)
221 		llist_add(&ctxt->rc_node, &rdma->sc_recv_ctxts);
222 	else
223 		svc_rdma_recv_ctxt_destroy(rdma, ctxt);
224 }
225 
226 static int __svc_rdma_post_recv(struct svcxprt_rdma *rdma,
227 				struct svc_rdma_recv_ctxt *ctxt)
228 {
229 	int ret;
230 
231 	svc_xprt_get(&rdma->sc_xprt);
232 	ret = ib_post_recv(rdma->sc_qp, &ctxt->rc_recv_wr, NULL);
233 	trace_svcrdma_post_recv(&ctxt->rc_recv_wr, ret);
234 	if (ret)
235 		goto err_post;
236 	return 0;
237 
238 err_post:
239 	svc_rdma_recv_ctxt_put(rdma, ctxt);
240 	svc_xprt_put(&rdma->sc_xprt);
241 	return ret;
242 }
243 
244 static int svc_rdma_post_recv(struct svcxprt_rdma *rdma)
245 {
246 	struct svc_rdma_recv_ctxt *ctxt;
247 
248 	ctxt = svc_rdma_recv_ctxt_get(rdma);
249 	if (!ctxt)
250 		return -ENOMEM;
251 	return __svc_rdma_post_recv(rdma, ctxt);
252 }
253 
254 /**
255  * svc_rdma_post_recvs - Post initial set of Recv WRs
256  * @rdma: fresh svcxprt_rdma
257  *
258  * Returns true if successful, otherwise false.
259  */
260 bool svc_rdma_post_recvs(struct svcxprt_rdma *rdma)
261 {
262 	struct svc_rdma_recv_ctxt *ctxt;
263 	unsigned int i;
264 	int ret;
265 
266 	for (i = 0; i < rdma->sc_max_requests; i++) {
267 		ctxt = svc_rdma_recv_ctxt_get(rdma);
268 		if (!ctxt)
269 			return false;
270 		ctxt->rc_temp = true;
271 		ret = __svc_rdma_post_recv(rdma, ctxt);
272 		if (ret)
273 			return false;
274 	}
275 	return true;
276 }
277 
278 /**
279  * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
280  * @cq: Completion Queue context
281  * @wc: Work Completion object
282  *
283  * NB: The svc_xprt/svcxprt_rdma is pinned whenever it's possible that
284  * the Receive completion handler could be running.
285  */
286 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
287 {
288 	struct svcxprt_rdma *rdma = cq->cq_context;
289 	struct ib_cqe *cqe = wc->wr_cqe;
290 	struct svc_rdma_recv_ctxt *ctxt;
291 
292 	trace_svcrdma_wc_receive(wc);
293 
294 	/* WARNING: Only wc->wr_cqe and wc->status are reliable */
295 	ctxt = container_of(cqe, struct svc_rdma_recv_ctxt, rc_cqe);
296 
297 	if (wc->status != IB_WC_SUCCESS)
298 		goto flushed;
299 
300 	if (svc_rdma_post_recv(rdma))
301 		goto post_err;
302 
303 	/* All wc fields are now known to be valid */
304 	ctxt->rc_byte_len = wc->byte_len;
305 	ib_dma_sync_single_for_cpu(rdma->sc_pd->device,
306 				   ctxt->rc_recv_sge.addr,
307 				   wc->byte_len, DMA_FROM_DEVICE);
308 
309 	spin_lock(&rdma->sc_rq_dto_lock);
310 	list_add_tail(&ctxt->rc_list, &rdma->sc_rq_dto_q);
311 	/* Note the unlock pairs with the smp_rmb in svc_xprt_ready: */
312 	set_bit(XPT_DATA, &rdma->sc_xprt.xpt_flags);
313 	spin_unlock(&rdma->sc_rq_dto_lock);
314 	if (!test_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags))
315 		svc_xprt_enqueue(&rdma->sc_xprt);
316 	goto out;
317 
318 flushed:
319 post_err:
320 	svc_rdma_recv_ctxt_put(rdma, ctxt);
321 	set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
322 	svc_xprt_enqueue(&rdma->sc_xprt);
323 out:
324 	svc_xprt_put(&rdma->sc_xprt);
325 }
326 
327 /**
328  * svc_rdma_flush_recv_queues - Drain pending Receive work
329  * @rdma: svcxprt_rdma being shut down
330  *
331  */
332 void svc_rdma_flush_recv_queues(struct svcxprt_rdma *rdma)
333 {
334 	struct svc_rdma_recv_ctxt *ctxt;
335 
336 	while ((ctxt = svc_rdma_next_recv_ctxt(&rdma->sc_read_complete_q))) {
337 		list_del(&ctxt->rc_list);
338 		svc_rdma_recv_ctxt_put(rdma, ctxt);
339 	}
340 	while ((ctxt = svc_rdma_next_recv_ctxt(&rdma->sc_rq_dto_q))) {
341 		list_del(&ctxt->rc_list);
342 		svc_rdma_recv_ctxt_put(rdma, ctxt);
343 	}
344 }
345 
346 static void svc_rdma_build_arg_xdr(struct svc_rqst *rqstp,
347 				   struct svc_rdma_recv_ctxt *ctxt)
348 {
349 	struct xdr_buf *arg = &rqstp->rq_arg;
350 
351 	arg->head[0].iov_base = ctxt->rc_recv_buf;
352 	arg->head[0].iov_len = ctxt->rc_byte_len;
353 	arg->tail[0].iov_base = NULL;
354 	arg->tail[0].iov_len = 0;
355 	arg->page_len = 0;
356 	arg->page_base = 0;
357 	arg->buflen = ctxt->rc_byte_len;
358 	arg->len = ctxt->rc_byte_len;
359 }
360 
361 /* This accommodates the largest possible Write chunk.
362  */
363 #define MAX_BYTES_WRITE_CHUNK ((u32)(RPCSVC_MAXPAGES << PAGE_SHIFT))
364 
365 /* This accommodates the largest possible Position-Zero
366  * Read chunk or Reply chunk.
367  */
368 #define MAX_BYTES_SPECIAL_CHUNK ((u32)((RPCSVC_MAXPAGES + 2) << PAGE_SHIFT))
369 
370 /* Sanity check the Read list.
371  *
372  * Implementation limits:
373  * - This implementation supports only one Read chunk.
374  *
375  * Sanity checks:
376  * - Read list does not overflow Receive buffer.
377  * - Segment size limited by largest NFS data payload.
378  *
379  * The segment count is limited to how many segments can
380  * fit in the transport header without overflowing the
381  * buffer. That's about 40 Read segments for a 1KB inline
382  * threshold.
383  *
384  * Return values:
385  *       %true: Read list is valid. @rctxt's xdr_stream is updated
386  *		to point to the first byte past the Read list.
387  *      %false: Read list is corrupt. @rctxt's xdr_stream is left
388  *		in an unknown state.
389  */
390 static bool xdr_check_read_list(struct svc_rdma_recv_ctxt *rctxt)
391 {
392 	u32 position, len;
393 	bool first;
394 	__be32 *p;
395 
396 	p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
397 	if (!p)
398 		return false;
399 
400 	len = 0;
401 	first = true;
402 	while (*p != xdr_zero) {
403 		p = xdr_inline_decode(&rctxt->rc_stream,
404 				      rpcrdma_readseg_maxsz * sizeof(*p));
405 		if (!p)
406 			return false;
407 
408 		if (first) {
409 			position = be32_to_cpup(p);
410 			first = false;
411 		} else if (be32_to_cpup(p) != position) {
412 			return false;
413 		}
414 		p += 2;
415 		len += be32_to_cpup(p);
416 
417 		p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
418 		if (!p)
419 			return false;
420 	}
421 	return len <= MAX_BYTES_SPECIAL_CHUNK;
422 }
423 
424 /* The segment count is limited to how many segments can
425  * fit in the transport header without overflowing the
426  * buffer. That's about 60 Write segments for a 1KB inline
427  * threshold.
428  */
429 static bool xdr_check_write_chunk(struct svc_rdma_recv_ctxt *rctxt, u32 maxlen)
430 {
431 	u32 i, segcount, total;
432 	__be32 *p;
433 
434 	p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
435 	if (!p)
436 		return false;
437 	segcount = be32_to_cpup(p);
438 
439 	total = 0;
440 	for (i = 0; i < segcount; i++) {
441 		u32 handle, length;
442 		u64 offset;
443 
444 		p = xdr_inline_decode(&rctxt->rc_stream,
445 				      rpcrdma_segment_maxsz * sizeof(*p));
446 		if (!p)
447 			return false;
448 
449 		handle = be32_to_cpup(p++);
450 		length = be32_to_cpup(p++);
451 		xdr_decode_hyper(p, &offset);
452 		trace_svcrdma_decode_wseg(handle, length, offset);
453 
454 		total += length;
455 	}
456 	return total <= maxlen;
457 }
458 
459 /* Sanity check the Write list.
460  *
461  * Implementation limits:
462  * - This implementation currently supports only one Write chunk.
463  *
464  * Sanity checks:
465  * - Write list does not overflow Receive buffer.
466  * - Chunk size limited by largest NFS data payload.
467  *
468  * Return values:
469  *       %true: Write list is valid. @rctxt's xdr_stream is updated
470  *		to point to the first byte past the Write list.
471  *      %false: Write list is corrupt. @rctxt's xdr_stream is left
472  *		in an unknown state.
473  */
474 static bool xdr_check_write_list(struct svc_rdma_recv_ctxt *rctxt)
475 {
476 	u32 chcount = 0;
477 	__be32 *p;
478 
479 	p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
480 	if (!p)
481 		return false;
482 	rctxt->rc_write_list = p;
483 	while (*p != xdr_zero) {
484 		if (!xdr_check_write_chunk(rctxt, MAX_BYTES_WRITE_CHUNK))
485 			return false;
486 		++chcount;
487 		p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
488 		if (!p)
489 			return false;
490 	}
491 	if (!chcount)
492 		rctxt->rc_write_list = NULL;
493 	return chcount < 2;
494 }
495 
496 /* Sanity check the Reply chunk.
497  *
498  * Sanity checks:
499  * - Reply chunk does not overflow Receive buffer.
500  * - Chunk size limited by largest NFS data payload.
501  *
502  * Return values:
503  *       %true: Reply chunk is valid. @rctxt's xdr_stream is updated
504  *		to point to the first byte past the Reply chunk.
505  *      %false: Reply chunk is corrupt. @rctxt's xdr_stream is left
506  *		in an unknown state.
507  */
508 static bool xdr_check_reply_chunk(struct svc_rdma_recv_ctxt *rctxt)
509 {
510 	__be32 *p;
511 
512 	p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
513 	if (!p)
514 		return false;
515 	rctxt->rc_reply_chunk = p;
516 	if (*p != xdr_zero) {
517 		if (!xdr_check_write_chunk(rctxt, MAX_BYTES_SPECIAL_CHUNK))
518 			return false;
519 	} else {
520 		rctxt->rc_reply_chunk = NULL;
521 	}
522 	return true;
523 }
524 
525 /* RPC-over-RDMA Version One private extension: Remote Invalidation.
526  * Responder's choice: requester signals it can handle Send With
527  * Invalidate, and responder chooses one R_key to invalidate.
528  *
529  * If there is exactly one distinct R_key in the received transport
530  * header, set rc_inv_rkey to that R_key. Otherwise, set it to zero.
531  *
532  * Perform this operation while the received transport header is
533  * still in the CPU cache.
534  */
535 static void svc_rdma_get_inv_rkey(struct svcxprt_rdma *rdma,
536 				  struct svc_rdma_recv_ctxt *ctxt)
537 {
538 	__be32 inv_rkey, *p;
539 	u32 i, segcount;
540 
541 	ctxt->rc_inv_rkey = 0;
542 
543 	if (!rdma->sc_snd_w_inv)
544 		return;
545 
546 	inv_rkey = xdr_zero;
547 	p = ctxt->rc_recv_buf;
548 	p += rpcrdma_fixed_maxsz;
549 
550 	/* Read list */
551 	while (*p++ != xdr_zero) {
552 		p++;	/* position */
553 		if (inv_rkey == xdr_zero)
554 			inv_rkey = *p;
555 		else if (inv_rkey != *p)
556 			return;
557 		p += 4;
558 	}
559 
560 	/* Write list */
561 	while (*p++ != xdr_zero) {
562 		segcount = be32_to_cpup(p++);
563 		for (i = 0; i < segcount; i++) {
564 			if (inv_rkey == xdr_zero)
565 				inv_rkey = *p;
566 			else if (inv_rkey != *p)
567 				return;
568 			p += 4;
569 		}
570 	}
571 
572 	/* Reply chunk */
573 	if (*p++ != xdr_zero) {
574 		segcount = be32_to_cpup(p++);
575 		for (i = 0; i < segcount; i++) {
576 			if (inv_rkey == xdr_zero)
577 				inv_rkey = *p;
578 			else if (inv_rkey != *p)
579 				return;
580 			p += 4;
581 		}
582 	}
583 
584 	ctxt->rc_inv_rkey = be32_to_cpu(inv_rkey);
585 }
586 
587 /**
588  * svc_rdma_xdr_decode_req - Decode the transport header
589  * @rq_arg: xdr_buf containing ingress RPC/RDMA message
590  * @rctxt: state of decoding
591  *
592  * On entry, xdr->head[0].iov_base points to first byte of the
593  * RPC-over-RDMA transport header.
594  *
595  * On successful exit, head[0] points to first byte past the
596  * RPC-over-RDMA header. For RDMA_MSG, this is the RPC message.
597  *
598  * The length of the RPC-over-RDMA header is returned.
599  *
600  * Assumptions:
601  * - The transport header is entirely contained in the head iovec.
602  */
603 static int svc_rdma_xdr_decode_req(struct xdr_buf *rq_arg,
604 				   struct svc_rdma_recv_ctxt *rctxt)
605 {
606 	__be32 *p, *rdma_argp;
607 	unsigned int hdr_len;
608 
609 	rdma_argp = rq_arg->head[0].iov_base;
610 	xdr_init_decode(&rctxt->rc_stream, rq_arg, rdma_argp, NULL);
611 
612 	p = xdr_inline_decode(&rctxt->rc_stream,
613 			      rpcrdma_fixed_maxsz * sizeof(*p));
614 	if (unlikely(!p))
615 		goto out_short;
616 	p++;
617 	if (*p != rpcrdma_version)
618 		goto out_version;
619 	p += 2;
620 	switch (*p) {
621 	case rdma_msg:
622 		break;
623 	case rdma_nomsg:
624 		break;
625 	case rdma_done:
626 		goto out_drop;
627 	case rdma_error:
628 		goto out_drop;
629 	default:
630 		goto out_proc;
631 	}
632 
633 	if (!xdr_check_read_list(rctxt))
634 		goto out_inval;
635 	if (!xdr_check_write_list(rctxt))
636 		goto out_inval;
637 	if (!xdr_check_reply_chunk(rctxt))
638 		goto out_inval;
639 
640 	rq_arg->head[0].iov_base = rctxt->rc_stream.p;
641 	hdr_len = xdr_stream_pos(&rctxt->rc_stream);
642 	rq_arg->head[0].iov_len -= hdr_len;
643 	rq_arg->len -= hdr_len;
644 	trace_svcrdma_decode_rqst(rdma_argp, hdr_len);
645 	return hdr_len;
646 
647 out_short:
648 	trace_svcrdma_decode_short(rq_arg->len);
649 	return -EINVAL;
650 
651 out_version:
652 	trace_svcrdma_decode_badvers(rdma_argp);
653 	return -EPROTONOSUPPORT;
654 
655 out_drop:
656 	trace_svcrdma_decode_drop(rdma_argp);
657 	return 0;
658 
659 out_proc:
660 	trace_svcrdma_decode_badproc(rdma_argp);
661 	return -EINVAL;
662 
663 out_inval:
664 	trace_svcrdma_decode_parse(rdma_argp);
665 	return -EINVAL;
666 }
667 
668 static void rdma_read_complete(struct svc_rqst *rqstp,
669 			       struct svc_rdma_recv_ctxt *head)
670 {
671 	int page_no;
672 
673 	/* Move Read chunk pages to rqstp so that they will be released
674 	 * when svc_process is done with them.
675 	 */
676 	for (page_no = 0; page_no < head->rc_page_count; page_no++) {
677 		put_page(rqstp->rq_pages[page_no]);
678 		rqstp->rq_pages[page_no] = head->rc_pages[page_no];
679 	}
680 	head->rc_page_count = 0;
681 
682 	/* Point rq_arg.pages past header */
683 	rqstp->rq_arg.pages = &rqstp->rq_pages[head->rc_hdr_count];
684 	rqstp->rq_arg.page_len = head->rc_arg.page_len;
685 
686 	/* rq_respages starts after the last arg page */
687 	rqstp->rq_respages = &rqstp->rq_pages[page_no];
688 	rqstp->rq_next_page = rqstp->rq_respages + 1;
689 
690 	/* Rebuild rq_arg head and tail. */
691 	rqstp->rq_arg.head[0] = head->rc_arg.head[0];
692 	rqstp->rq_arg.tail[0] = head->rc_arg.tail[0];
693 	rqstp->rq_arg.len = head->rc_arg.len;
694 	rqstp->rq_arg.buflen = head->rc_arg.buflen;
695 }
696 
697 static void svc_rdma_send_error(struct svcxprt_rdma *xprt,
698 				__be32 *rdma_argp, int status)
699 {
700 	struct svc_rdma_send_ctxt *ctxt;
701 	__be32 *p;
702 	int ret;
703 
704 	ctxt = svc_rdma_send_ctxt_get(xprt);
705 	if (!ctxt)
706 		return;
707 
708 	p = xdr_reserve_space(&ctxt->sc_stream,
709 			      rpcrdma_fixed_maxsz * sizeof(*p));
710 	if (!p)
711 		goto put_ctxt;
712 
713 	*p++ = *rdma_argp;
714 	*p++ = *(rdma_argp + 1);
715 	*p++ = xprt->sc_fc_credits;
716 	*p = rdma_error;
717 
718 	switch (status) {
719 	case -EPROTONOSUPPORT:
720 		p = xdr_reserve_space(&ctxt->sc_stream, 3 * sizeof(*p));
721 		if (!p)
722 			goto put_ctxt;
723 
724 		*p++ = err_vers;
725 		*p++ = rpcrdma_version;
726 		*p = rpcrdma_version;
727 		trace_svcrdma_err_vers(*rdma_argp);
728 		break;
729 	default:
730 		p = xdr_reserve_space(&ctxt->sc_stream, sizeof(*p));
731 		if (!p)
732 			goto put_ctxt;
733 
734 		*p = err_chunk;
735 		trace_svcrdma_err_chunk(*rdma_argp);
736 	}
737 
738 	ctxt->sc_send_wr.num_sge = 1;
739 	ctxt->sc_send_wr.opcode = IB_WR_SEND;
740 	ctxt->sc_sges[0].length = ctxt->sc_hdrbuf.len;
741 	ret = svc_rdma_send(xprt, &ctxt->sc_send_wr);
742 	if (ret)
743 		goto put_ctxt;
744 	return;
745 
746 put_ctxt:
747 	svc_rdma_send_ctxt_put(xprt, ctxt);
748 }
749 
750 /* By convention, backchannel calls arrive via rdma_msg type
751  * messages, and never populate the chunk lists. This makes
752  * the RPC/RDMA header small and fixed in size, so it is
753  * straightforward to check the RPC header's direction field.
754  */
755 static bool svc_rdma_is_backchannel_reply(struct svc_xprt *xprt,
756 					  __be32 *rdma_resp)
757 {
758 	__be32 *p;
759 
760 	if (!xprt->xpt_bc_xprt)
761 		return false;
762 
763 	p = rdma_resp + 3;
764 	if (*p++ != rdma_msg)
765 		return false;
766 
767 	if (*p++ != xdr_zero)
768 		return false;
769 	if (*p++ != xdr_zero)
770 		return false;
771 	if (*p++ != xdr_zero)
772 		return false;
773 
774 	/* XID sanity */
775 	if (*p++ != *rdma_resp)
776 		return false;
777 	/* call direction */
778 	if (*p == cpu_to_be32(RPC_CALL))
779 		return false;
780 
781 	return true;
782 }
783 
784 /**
785  * svc_rdma_recvfrom - Receive an RPC call
786  * @rqstp: request structure into which to receive an RPC Call
787  *
788  * Returns:
789  *	The positive number of bytes in the RPC Call message,
790  *	%0 if there were no Calls ready to return,
791  *	%-EINVAL if the Read chunk data is too large,
792  *	%-ENOMEM if rdma_rw context pool was exhausted,
793  *	%-ENOTCONN if posting failed (connection is lost),
794  *	%-EIO if rdma_rw initialization failed (DMA mapping, etc).
795  *
796  * Called in a loop when XPT_DATA is set. XPT_DATA is cleared only
797  * when there are no remaining ctxt's to process.
798  *
799  * The next ctxt is removed from the "receive" lists.
800  *
801  * - If the ctxt completes a Read, then finish assembling the Call
802  *   message and return the number of bytes in the message.
803  *
804  * - If the ctxt completes a Receive, then construct the Call
805  *   message from the contents of the Receive buffer.
806  *
807  *   - If there are no Read chunks in this message, then finish
808  *     assembling the Call message and return the number of bytes
809  *     in the message.
810  *
811  *   - If there are Read chunks in this message, post Read WRs to
812  *     pull that payload and return 0.
813  */
814 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
815 {
816 	struct svc_xprt *xprt = rqstp->rq_xprt;
817 	struct svcxprt_rdma *rdma_xprt =
818 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
819 	struct svc_rdma_recv_ctxt *ctxt;
820 	__be32 *p;
821 	int ret;
822 
823 	spin_lock(&rdma_xprt->sc_rq_dto_lock);
824 	ctxt = svc_rdma_next_recv_ctxt(&rdma_xprt->sc_read_complete_q);
825 	if (ctxt) {
826 		list_del(&ctxt->rc_list);
827 		spin_unlock(&rdma_xprt->sc_rq_dto_lock);
828 		rdma_read_complete(rqstp, ctxt);
829 		goto complete;
830 	}
831 	ctxt = svc_rdma_next_recv_ctxt(&rdma_xprt->sc_rq_dto_q);
832 	if (!ctxt) {
833 		/* No new incoming requests, terminate the loop */
834 		clear_bit(XPT_DATA, &xprt->xpt_flags);
835 		spin_unlock(&rdma_xprt->sc_rq_dto_lock);
836 		return 0;
837 	}
838 	list_del(&ctxt->rc_list);
839 	spin_unlock(&rdma_xprt->sc_rq_dto_lock);
840 
841 	atomic_inc(&rdma_stat_recv);
842 
843 	svc_rdma_build_arg_xdr(rqstp, ctxt);
844 
845 	/* Prevent svc_xprt_release from releasing pages in rq_pages
846 	 * if we return 0 or an error.
847 	 */
848 	rqstp->rq_respages = rqstp->rq_pages;
849 	rqstp->rq_next_page = rqstp->rq_respages;
850 
851 	p = (__be32 *)rqstp->rq_arg.head[0].iov_base;
852 	ret = svc_rdma_xdr_decode_req(&rqstp->rq_arg, ctxt);
853 	if (ret < 0)
854 		goto out_err;
855 	if (ret == 0)
856 		goto out_drop;
857 	rqstp->rq_xprt_hlen = ret;
858 
859 	if (svc_rdma_is_backchannel_reply(xprt, p)) {
860 		ret = svc_rdma_handle_bc_reply(xprt->xpt_bc_xprt, p,
861 					       &rqstp->rq_arg);
862 		svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
863 		return ret;
864 	}
865 	svc_rdma_get_inv_rkey(rdma_xprt, ctxt);
866 
867 	p += rpcrdma_fixed_maxsz;
868 	if (*p != xdr_zero)
869 		goto out_readchunk;
870 
871 complete:
872 	rqstp->rq_xprt_ctxt = ctxt;
873 	rqstp->rq_prot = IPPROTO_MAX;
874 	svc_xprt_copy_addrs(rqstp, xprt);
875 	return rqstp->rq_arg.len;
876 
877 out_readchunk:
878 	ret = svc_rdma_recv_read_chunk(rdma_xprt, rqstp, ctxt, p);
879 	if (ret < 0)
880 		goto out_postfail;
881 	return 0;
882 
883 out_err:
884 	svc_rdma_send_error(rdma_xprt, p, ret);
885 	svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
886 	return 0;
887 
888 out_postfail:
889 	if (ret == -EINVAL)
890 		svc_rdma_send_error(rdma_xprt, p, ret);
891 	svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
892 	return ret;
893 
894 out_drop:
895 	svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
896 	return 0;
897 }
898