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/slab.h>
97 #include <linux/spinlock.h>
98 #include <asm/unaligned.h>
99 #include <rdma/ib_verbs.h>
100 #include <rdma/rdma_cm.h>
101 
102 #include <linux/sunrpc/xdr.h>
103 #include <linux/sunrpc/debug.h>
104 #include <linux/sunrpc/rpc_rdma.h>
105 #include <linux/sunrpc/svc_rdma.h>
106 
107 #include "xprt_rdma.h"
108 #include <trace/events/rpcrdma.h>
109 
110 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
111 
112 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc);
113 
114 static inline struct svc_rdma_recv_ctxt *
115 svc_rdma_next_recv_ctxt(struct list_head *list)
116 {
117 	return list_first_entry_or_null(list, struct svc_rdma_recv_ctxt,
118 					rc_list);
119 }
120 
121 static void svc_rdma_recv_cid_init(struct svcxprt_rdma *rdma,
122 				   struct rpc_rdma_cid *cid)
123 {
124 	cid->ci_queue_id = rdma->sc_rq_cq->res.id;
125 	cid->ci_completion_id = atomic_inc_return(&rdma->sc_completion_ids);
126 }
127 
128 static struct svc_rdma_recv_ctxt *
129 svc_rdma_recv_ctxt_alloc(struct svcxprt_rdma *rdma)
130 {
131 	struct svc_rdma_recv_ctxt *ctxt;
132 	dma_addr_t addr;
133 	void *buffer;
134 
135 	ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
136 	if (!ctxt)
137 		goto fail0;
138 	buffer = kmalloc(rdma->sc_max_req_size, GFP_KERNEL);
139 	if (!buffer)
140 		goto fail1;
141 	addr = ib_dma_map_single(rdma->sc_pd->device, buffer,
142 				 rdma->sc_max_req_size, DMA_FROM_DEVICE);
143 	if (ib_dma_mapping_error(rdma->sc_pd->device, addr))
144 		goto fail2;
145 
146 	svc_rdma_recv_cid_init(rdma, &ctxt->rc_cid);
147 	pcl_init(&ctxt->rc_call_pcl);
148 	pcl_init(&ctxt->rc_read_pcl);
149 	pcl_init(&ctxt->rc_write_pcl);
150 	pcl_init(&ctxt->rc_reply_pcl);
151 
152 	ctxt->rc_recv_wr.next = NULL;
153 	ctxt->rc_recv_wr.wr_cqe = &ctxt->rc_cqe;
154 	ctxt->rc_recv_wr.sg_list = &ctxt->rc_recv_sge;
155 	ctxt->rc_recv_wr.num_sge = 1;
156 	ctxt->rc_cqe.done = svc_rdma_wc_receive;
157 	ctxt->rc_recv_sge.addr = addr;
158 	ctxt->rc_recv_sge.length = rdma->sc_max_req_size;
159 	ctxt->rc_recv_sge.lkey = rdma->sc_pd->local_dma_lkey;
160 	ctxt->rc_recv_buf = buffer;
161 	ctxt->rc_temp = false;
162 	return ctxt;
163 
164 fail2:
165 	kfree(buffer);
166 fail1:
167 	kfree(ctxt);
168 fail0:
169 	return NULL;
170 }
171 
172 static void svc_rdma_recv_ctxt_destroy(struct svcxprt_rdma *rdma,
173 				       struct svc_rdma_recv_ctxt *ctxt)
174 {
175 	ib_dma_unmap_single(rdma->sc_pd->device, ctxt->rc_recv_sge.addr,
176 			    ctxt->rc_recv_sge.length, DMA_FROM_DEVICE);
177 	kfree(ctxt->rc_recv_buf);
178 	kfree(ctxt);
179 }
180 
181 /**
182  * svc_rdma_recv_ctxts_destroy - Release all recv_ctxt's for an xprt
183  * @rdma: svcxprt_rdma being torn down
184  *
185  */
186 void svc_rdma_recv_ctxts_destroy(struct svcxprt_rdma *rdma)
187 {
188 	struct svc_rdma_recv_ctxt *ctxt;
189 	struct llist_node *node;
190 
191 	while ((node = llist_del_first(&rdma->sc_recv_ctxts))) {
192 		ctxt = llist_entry(node, struct svc_rdma_recv_ctxt, rc_node);
193 		svc_rdma_recv_ctxt_destroy(rdma, ctxt);
194 	}
195 }
196 
197 /**
198  * svc_rdma_recv_ctxt_get - Allocate a recv_ctxt
199  * @rdma: controlling svcxprt_rdma
200  *
201  * Returns a recv_ctxt or (rarely) NULL if none are available.
202  */
203 struct svc_rdma_recv_ctxt *svc_rdma_recv_ctxt_get(struct svcxprt_rdma *rdma)
204 {
205 	struct svc_rdma_recv_ctxt *ctxt;
206 	struct llist_node *node;
207 
208 	node = llist_del_first(&rdma->sc_recv_ctxts);
209 	if (!node)
210 		goto out_empty;
211 	ctxt = llist_entry(node, struct svc_rdma_recv_ctxt, rc_node);
212 
213 out:
214 	ctxt->rc_page_count = 0;
215 	return ctxt;
216 
217 out_empty:
218 	ctxt = svc_rdma_recv_ctxt_alloc(rdma);
219 	if (!ctxt)
220 		return NULL;
221 	goto out;
222 }
223 
224 /**
225  * svc_rdma_recv_ctxt_put - Return recv_ctxt to free list
226  * @rdma: controlling svcxprt_rdma
227  * @ctxt: object to return to the free list
228  *
229  */
230 void svc_rdma_recv_ctxt_put(struct svcxprt_rdma *rdma,
231 			    struct svc_rdma_recv_ctxt *ctxt)
232 {
233 	unsigned int i;
234 
235 	for (i = 0; i < ctxt->rc_page_count; i++)
236 		put_page(ctxt->rc_pages[i]);
237 
238 	pcl_free(&ctxt->rc_call_pcl);
239 	pcl_free(&ctxt->rc_read_pcl);
240 	pcl_free(&ctxt->rc_write_pcl);
241 	pcl_free(&ctxt->rc_reply_pcl);
242 
243 	if (!ctxt->rc_temp)
244 		llist_add(&ctxt->rc_node, &rdma->sc_recv_ctxts);
245 	else
246 		svc_rdma_recv_ctxt_destroy(rdma, ctxt);
247 }
248 
249 /**
250  * svc_rdma_release_rqst - Release transport-specific per-rqst resources
251  * @rqstp: svc_rqst being released
252  *
253  * Ensure that the recv_ctxt is released whether or not a Reply
254  * was sent. For example, the client could close the connection,
255  * or svc_process could drop an RPC, before the Reply is sent.
256  */
257 void svc_rdma_release_rqst(struct svc_rqst *rqstp)
258 {
259 	struct svc_rdma_recv_ctxt *ctxt = rqstp->rq_xprt_ctxt;
260 	struct svc_xprt *xprt = rqstp->rq_xprt;
261 	struct svcxprt_rdma *rdma =
262 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
263 
264 	rqstp->rq_xprt_ctxt = NULL;
265 	if (ctxt)
266 		svc_rdma_recv_ctxt_put(rdma, ctxt);
267 }
268 
269 static int __svc_rdma_post_recv(struct svcxprt_rdma *rdma,
270 				struct svc_rdma_recv_ctxt *ctxt)
271 {
272 	int ret;
273 
274 	trace_svcrdma_post_recv(ctxt);
275 	ret = ib_post_recv(rdma->sc_qp, &ctxt->rc_recv_wr, NULL);
276 	if (ret)
277 		goto err_post;
278 	return 0;
279 
280 err_post:
281 	trace_svcrdma_rq_post_err(rdma, ret);
282 	svc_rdma_recv_ctxt_put(rdma, ctxt);
283 	return ret;
284 }
285 
286 static int svc_rdma_post_recv(struct svcxprt_rdma *rdma)
287 {
288 	struct svc_rdma_recv_ctxt *ctxt;
289 
290 	if (test_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags))
291 		return 0;
292 	ctxt = svc_rdma_recv_ctxt_get(rdma);
293 	if (!ctxt)
294 		return -ENOMEM;
295 	return __svc_rdma_post_recv(rdma, ctxt);
296 }
297 
298 /**
299  * svc_rdma_post_recvs - Post initial set of Recv WRs
300  * @rdma: fresh svcxprt_rdma
301  *
302  * Returns true if successful, otherwise false.
303  */
304 bool svc_rdma_post_recvs(struct svcxprt_rdma *rdma)
305 {
306 	struct svc_rdma_recv_ctxt *ctxt;
307 	unsigned int i;
308 	int ret;
309 
310 	for (i = 0; i < rdma->sc_max_requests; i++) {
311 		ctxt = svc_rdma_recv_ctxt_get(rdma);
312 		if (!ctxt)
313 			return false;
314 		ctxt->rc_temp = true;
315 		ret = __svc_rdma_post_recv(rdma, ctxt);
316 		if (ret)
317 			return false;
318 	}
319 	return true;
320 }
321 
322 /**
323  * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
324  * @cq: Completion Queue context
325  * @wc: Work Completion object
326  *
327  * NB: The svc_xprt/svcxprt_rdma is pinned whenever it's possible that
328  * the Receive completion handler could be running.
329  */
330 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
331 {
332 	struct svcxprt_rdma *rdma = cq->cq_context;
333 	struct ib_cqe *cqe = wc->wr_cqe;
334 	struct svc_rdma_recv_ctxt *ctxt;
335 
336 	/* WARNING: Only wc->wr_cqe and wc->status are reliable */
337 	ctxt = container_of(cqe, struct svc_rdma_recv_ctxt, rc_cqe);
338 
339 	trace_svcrdma_wc_receive(wc, &ctxt->rc_cid);
340 	if (wc->status != IB_WC_SUCCESS)
341 		goto flushed;
342 
343 	if (svc_rdma_post_recv(rdma))
344 		goto post_err;
345 
346 	/* All wc fields are now known to be valid */
347 	ctxt->rc_byte_len = wc->byte_len;
348 
349 	spin_lock(&rdma->sc_rq_dto_lock);
350 	list_add_tail(&ctxt->rc_list, &rdma->sc_rq_dto_q);
351 	/* Note the unlock pairs with the smp_rmb in svc_xprt_ready: */
352 	set_bit(XPT_DATA, &rdma->sc_xprt.xpt_flags);
353 	spin_unlock(&rdma->sc_rq_dto_lock);
354 	if (!test_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags))
355 		svc_xprt_enqueue(&rdma->sc_xprt);
356 	return;
357 
358 flushed:
359 post_err:
360 	svc_rdma_recv_ctxt_put(rdma, ctxt);
361 	set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
362 	svc_xprt_enqueue(&rdma->sc_xprt);
363 }
364 
365 /**
366  * svc_rdma_flush_recv_queues - Drain pending Receive work
367  * @rdma: svcxprt_rdma being shut down
368  *
369  */
370 void svc_rdma_flush_recv_queues(struct svcxprt_rdma *rdma)
371 {
372 	struct svc_rdma_recv_ctxt *ctxt;
373 
374 	while ((ctxt = svc_rdma_next_recv_ctxt(&rdma->sc_read_complete_q))) {
375 		list_del(&ctxt->rc_list);
376 		svc_rdma_recv_ctxt_put(rdma, ctxt);
377 	}
378 	while ((ctxt = svc_rdma_next_recv_ctxt(&rdma->sc_rq_dto_q))) {
379 		list_del(&ctxt->rc_list);
380 		svc_rdma_recv_ctxt_put(rdma, ctxt);
381 	}
382 }
383 
384 static void svc_rdma_build_arg_xdr(struct svc_rqst *rqstp,
385 				   struct svc_rdma_recv_ctxt *ctxt)
386 {
387 	struct xdr_buf *arg = &rqstp->rq_arg;
388 
389 	arg->head[0].iov_base = ctxt->rc_recv_buf;
390 	arg->head[0].iov_len = ctxt->rc_byte_len;
391 	arg->tail[0].iov_base = NULL;
392 	arg->tail[0].iov_len = 0;
393 	arg->page_len = 0;
394 	arg->page_base = 0;
395 	arg->buflen = ctxt->rc_byte_len;
396 	arg->len = ctxt->rc_byte_len;
397 }
398 
399 /**
400  * xdr_count_read_segments - Count number of Read segments in Read list
401  * @rctxt: Ingress receive context
402  * @p: Start of an un-decoded Read list
403  *
404  * Before allocating anything, ensure the ingress Read list is safe
405  * to use.
406  *
407  * The segment count is limited to how many segments can fit in the
408  * transport header without overflowing the buffer. That's about 40
409  * Read segments for a 1KB inline threshold.
410  *
411  * Return values:
412  *   %true: Read list is valid. @rctxt's xdr_stream is updated to point
413  *	    to the first byte past the Read list. rc_read_pcl and
414  *	    rc_call_pcl cl_count fields are set to the number of
415  *	    Read segments in the list.
416  *  %false: Read list is corrupt. @rctxt's xdr_stream is left in an
417  *	    unknown state.
418  */
419 static bool xdr_count_read_segments(struct svc_rdma_recv_ctxt *rctxt, __be32 *p)
420 {
421 	rctxt->rc_call_pcl.cl_count = 0;
422 	rctxt->rc_read_pcl.cl_count = 0;
423 	while (xdr_item_is_present(p)) {
424 		u32 position, handle, length;
425 		u64 offset;
426 
427 		p = xdr_inline_decode(&rctxt->rc_stream,
428 				      rpcrdma_readseg_maxsz * sizeof(*p));
429 		if (!p)
430 			return false;
431 
432 		xdr_decode_read_segment(p, &position, &handle,
433 					    &length, &offset);
434 		if (position) {
435 			if (position & 3)
436 				return false;
437 			++rctxt->rc_read_pcl.cl_count;
438 		} else {
439 			++rctxt->rc_call_pcl.cl_count;
440 		}
441 
442 		p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
443 		if (!p)
444 			return false;
445 	}
446 	return true;
447 }
448 
449 /* Sanity check the Read list.
450  *
451  * Sanity checks:
452  * - Read list does not overflow Receive buffer.
453  * - Chunk size limited by largest NFS data payload.
454  *
455  * Return values:
456  *   %true: Read list is valid. @rctxt's xdr_stream is updated
457  *	    to point to the first byte past the Read list.
458  *  %false: Read list is corrupt. @rctxt's xdr_stream is left
459  *	    in an unknown state.
460  */
461 static bool xdr_check_read_list(struct svc_rdma_recv_ctxt *rctxt)
462 {
463 	__be32 *p;
464 
465 	p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
466 	if (!p)
467 		return false;
468 	if (!xdr_count_read_segments(rctxt, p))
469 		return false;
470 	if (!pcl_alloc_call(rctxt, p))
471 		return false;
472 	return pcl_alloc_read(rctxt, p);
473 }
474 
475 static bool xdr_check_write_chunk(struct svc_rdma_recv_ctxt *rctxt)
476 {
477 	u32 segcount;
478 	__be32 *p;
479 
480 	if (xdr_stream_decode_u32(&rctxt->rc_stream, &segcount))
481 		return false;
482 
483 	/* A bogus segcount causes this buffer overflow check to fail. */
484 	p = xdr_inline_decode(&rctxt->rc_stream,
485 			      segcount * rpcrdma_segment_maxsz * sizeof(*p));
486 	return p != NULL;
487 }
488 
489 /**
490  * xdr_count_write_chunks - Count number of Write chunks in Write list
491  * @rctxt: Received header and decoding state
492  * @p: start of an un-decoded Write list
493  *
494  * Before allocating anything, ensure the ingress Write list is
495  * safe to use.
496  *
497  * Return values:
498  *       %true: Write list is valid. @rctxt's xdr_stream is updated
499  *		to point to the first byte past the Write list, and
500  *		the number of Write chunks is in rc_write_pcl.cl_count.
501  *      %false: Write list is corrupt. @rctxt's xdr_stream is left
502  *		in an indeterminate state.
503  */
504 static bool xdr_count_write_chunks(struct svc_rdma_recv_ctxt *rctxt, __be32 *p)
505 {
506 	rctxt->rc_write_pcl.cl_count = 0;
507 	while (xdr_item_is_present(p)) {
508 		if (!xdr_check_write_chunk(rctxt))
509 			return false;
510 		++rctxt->rc_write_pcl.cl_count;
511 		p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
512 		if (!p)
513 			return false;
514 	}
515 	return true;
516 }
517 
518 /* Sanity check the Write list.
519  *
520  * Implementation limits:
521  * - This implementation currently supports only one Write chunk.
522  *
523  * Sanity checks:
524  * - Write list does not overflow Receive buffer.
525  * - Chunk size limited by largest NFS data payload.
526  *
527  * Return values:
528  *       %true: Write list is valid. @rctxt's xdr_stream is updated
529  *		to point to the first byte past the Write list.
530  *      %false: Write list is corrupt. @rctxt's xdr_stream is left
531  *		in an unknown state.
532  */
533 static bool xdr_check_write_list(struct svc_rdma_recv_ctxt *rctxt)
534 {
535 	__be32 *p;
536 
537 	p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
538 	if (!p)
539 		return false;
540 	if (!xdr_count_write_chunks(rctxt, p))
541 		return false;
542 	if (!pcl_alloc_write(rctxt, &rctxt->rc_write_pcl, p))
543 		return false;
544 
545 	rctxt->rc_cur_result_payload = pcl_first_chunk(&rctxt->rc_write_pcl);
546 	return true;
547 }
548 
549 /* Sanity check the Reply chunk.
550  *
551  * Sanity checks:
552  * - Reply chunk does not overflow Receive buffer.
553  * - Chunk size limited by largest NFS data payload.
554  *
555  * Return values:
556  *       %true: Reply chunk is valid. @rctxt's xdr_stream is updated
557  *		to point to the first byte past the Reply chunk.
558  *      %false: Reply chunk is corrupt. @rctxt's xdr_stream is left
559  *		in an unknown state.
560  */
561 static bool xdr_check_reply_chunk(struct svc_rdma_recv_ctxt *rctxt)
562 {
563 	__be32 *p;
564 
565 	p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
566 	if (!p)
567 		return false;
568 
569 	if (!xdr_item_is_present(p))
570 		return true;
571 	if (!xdr_check_write_chunk(rctxt))
572 		return false;
573 
574 	rctxt->rc_reply_pcl.cl_count = 1;
575 	return pcl_alloc_write(rctxt, &rctxt->rc_reply_pcl, p);
576 }
577 
578 /* RPC-over-RDMA Version One private extension: Remote Invalidation.
579  * Responder's choice: requester signals it can handle Send With
580  * Invalidate, and responder chooses one R_key to invalidate.
581  *
582  * If there is exactly one distinct R_key in the received transport
583  * header, set rc_inv_rkey to that R_key. Otherwise, set it to zero.
584  */
585 static void svc_rdma_get_inv_rkey(struct svcxprt_rdma *rdma,
586 				  struct svc_rdma_recv_ctxt *ctxt)
587 {
588 	struct svc_rdma_segment *segment;
589 	struct svc_rdma_chunk *chunk;
590 	u32 inv_rkey;
591 
592 	ctxt->rc_inv_rkey = 0;
593 
594 	if (!rdma->sc_snd_w_inv)
595 		return;
596 
597 	inv_rkey = 0;
598 	pcl_for_each_chunk(chunk, &ctxt->rc_call_pcl) {
599 		pcl_for_each_segment(segment, chunk) {
600 			if (inv_rkey == 0)
601 				inv_rkey = segment->rs_handle;
602 			else if (inv_rkey != segment->rs_handle)
603 				return;
604 		}
605 	}
606 	pcl_for_each_chunk(chunk, &ctxt->rc_read_pcl) {
607 		pcl_for_each_segment(segment, chunk) {
608 			if (inv_rkey == 0)
609 				inv_rkey = segment->rs_handle;
610 			else if (inv_rkey != segment->rs_handle)
611 				return;
612 		}
613 	}
614 	pcl_for_each_chunk(chunk, &ctxt->rc_write_pcl) {
615 		pcl_for_each_segment(segment, chunk) {
616 			if (inv_rkey == 0)
617 				inv_rkey = segment->rs_handle;
618 			else if (inv_rkey != segment->rs_handle)
619 				return;
620 		}
621 	}
622 	pcl_for_each_chunk(chunk, &ctxt->rc_reply_pcl) {
623 		pcl_for_each_segment(segment, chunk) {
624 			if (inv_rkey == 0)
625 				inv_rkey = segment->rs_handle;
626 			else if (inv_rkey != segment->rs_handle)
627 				return;
628 		}
629 	}
630 	ctxt->rc_inv_rkey = inv_rkey;
631 }
632 
633 /**
634  * svc_rdma_xdr_decode_req - Decode the transport header
635  * @rq_arg: xdr_buf containing ingress RPC/RDMA message
636  * @rctxt: state of decoding
637  *
638  * On entry, xdr->head[0].iov_base points to first byte of the
639  * RPC-over-RDMA transport header.
640  *
641  * On successful exit, head[0] points to first byte past the
642  * RPC-over-RDMA header. For RDMA_MSG, this is the RPC message.
643  *
644  * The length of the RPC-over-RDMA header is returned.
645  *
646  * Assumptions:
647  * - The transport header is entirely contained in the head iovec.
648  */
649 static int svc_rdma_xdr_decode_req(struct xdr_buf *rq_arg,
650 				   struct svc_rdma_recv_ctxt *rctxt)
651 {
652 	__be32 *p, *rdma_argp;
653 	unsigned int hdr_len;
654 
655 	rdma_argp = rq_arg->head[0].iov_base;
656 	xdr_init_decode(&rctxt->rc_stream, rq_arg, rdma_argp, NULL);
657 
658 	p = xdr_inline_decode(&rctxt->rc_stream,
659 			      rpcrdma_fixed_maxsz * sizeof(*p));
660 	if (unlikely(!p))
661 		goto out_short;
662 	p++;
663 	if (*p != rpcrdma_version)
664 		goto out_version;
665 	p += 2;
666 	rctxt->rc_msgtype = *p;
667 	switch (rctxt->rc_msgtype) {
668 	case rdma_msg:
669 		break;
670 	case rdma_nomsg:
671 		break;
672 	case rdma_done:
673 		goto out_drop;
674 	case rdma_error:
675 		goto out_drop;
676 	default:
677 		goto out_proc;
678 	}
679 
680 	if (!xdr_check_read_list(rctxt))
681 		goto out_inval;
682 	if (!xdr_check_write_list(rctxt))
683 		goto out_inval;
684 	if (!xdr_check_reply_chunk(rctxt))
685 		goto out_inval;
686 
687 	rq_arg->head[0].iov_base = rctxt->rc_stream.p;
688 	hdr_len = xdr_stream_pos(&rctxt->rc_stream);
689 	rq_arg->head[0].iov_len -= hdr_len;
690 	rq_arg->len -= hdr_len;
691 	trace_svcrdma_decode_rqst(rctxt, rdma_argp, hdr_len);
692 	return hdr_len;
693 
694 out_short:
695 	trace_svcrdma_decode_short_err(rctxt, rq_arg->len);
696 	return -EINVAL;
697 
698 out_version:
699 	trace_svcrdma_decode_badvers_err(rctxt, rdma_argp);
700 	return -EPROTONOSUPPORT;
701 
702 out_drop:
703 	trace_svcrdma_decode_drop_err(rctxt, rdma_argp);
704 	return 0;
705 
706 out_proc:
707 	trace_svcrdma_decode_badproc_err(rctxt, rdma_argp);
708 	return -EINVAL;
709 
710 out_inval:
711 	trace_svcrdma_decode_parse_err(rctxt, rdma_argp);
712 	return -EINVAL;
713 }
714 
715 static void rdma_read_complete(struct svc_rqst *rqstp,
716 			       struct svc_rdma_recv_ctxt *head)
717 {
718 	int page_no;
719 
720 	/* Move Read chunk pages to rqstp so that they will be released
721 	 * when svc_process is done with them.
722 	 */
723 	for (page_no = 0; page_no < head->rc_page_count; page_no++) {
724 		put_page(rqstp->rq_pages[page_no]);
725 		rqstp->rq_pages[page_no] = head->rc_pages[page_no];
726 	}
727 	head->rc_page_count = 0;
728 
729 	/* Point rq_arg.pages past header */
730 	rqstp->rq_arg.pages = &rqstp->rq_pages[head->rc_hdr_count];
731 	rqstp->rq_arg.page_len = head->rc_arg.page_len;
732 
733 	/* rq_respages starts after the last arg page */
734 	rqstp->rq_respages = &rqstp->rq_pages[page_no];
735 	rqstp->rq_next_page = rqstp->rq_respages + 1;
736 
737 	/* Rebuild rq_arg head and tail. */
738 	rqstp->rq_arg.head[0] = head->rc_arg.head[0];
739 	rqstp->rq_arg.tail[0] = head->rc_arg.tail[0];
740 	rqstp->rq_arg.len = head->rc_arg.len;
741 	rqstp->rq_arg.buflen = head->rc_arg.buflen;
742 }
743 
744 static void svc_rdma_send_error(struct svcxprt_rdma *rdma,
745 				struct svc_rdma_recv_ctxt *rctxt,
746 				int status)
747 {
748 	struct svc_rdma_send_ctxt *sctxt;
749 
750 	sctxt = svc_rdma_send_ctxt_get(rdma);
751 	if (!sctxt)
752 		return;
753 	svc_rdma_send_error_msg(rdma, sctxt, rctxt, status);
754 }
755 
756 /* By convention, backchannel calls arrive via rdma_msg type
757  * messages, and never populate the chunk lists. This makes
758  * the RPC/RDMA header small and fixed in size, so it is
759  * straightforward to check the RPC header's direction field.
760  */
761 static bool svc_rdma_is_reverse_direction_reply(struct svc_xprt *xprt,
762 						struct svc_rdma_recv_ctxt *rctxt)
763 {
764 	__be32 *p = rctxt->rc_recv_buf;
765 
766 	if (!xprt->xpt_bc_xprt)
767 		return false;
768 
769 	if (rctxt->rc_msgtype != rdma_msg)
770 		return false;
771 
772 	if (!pcl_is_empty(&rctxt->rc_call_pcl))
773 		return false;
774 	if (!pcl_is_empty(&rctxt->rc_read_pcl))
775 		return false;
776 	if (!pcl_is_empty(&rctxt->rc_write_pcl))
777 		return false;
778 	if (!pcl_is_empty(&rctxt->rc_reply_pcl))
779 		return false;
780 
781 	/* RPC call direction */
782 	if (*(p + 8) == cpu_to_be32(RPC_CALL))
783 		return false;
784 
785 	return true;
786 }
787 
788 /**
789  * svc_rdma_recvfrom - Receive an RPC call
790  * @rqstp: request structure into which to receive an RPC Call
791  *
792  * Returns:
793  *	The positive number of bytes in the RPC Call message,
794  *	%0 if there were no Calls ready to return,
795  *	%-EINVAL if the Read chunk data is too large,
796  *	%-ENOMEM if rdma_rw context pool was exhausted,
797  *	%-ENOTCONN if posting failed (connection is lost),
798  *	%-EIO if rdma_rw initialization failed (DMA mapping, etc).
799  *
800  * Called in a loop when XPT_DATA is set. XPT_DATA is cleared only
801  * when there are no remaining ctxt's to process.
802  *
803  * The next ctxt is removed from the "receive" lists.
804  *
805  * - If the ctxt completes a Read, then finish assembling the Call
806  *   message and return the number of bytes in the message.
807  *
808  * - If the ctxt completes a Receive, then construct the Call
809  *   message from the contents of the Receive buffer.
810  *
811  *   - If there are no Read chunks in this message, then finish
812  *     assembling the Call message and return the number of bytes
813  *     in the message.
814  *
815  *   - If there are Read chunks in this message, post Read WRs to
816  *     pull that payload and return 0.
817  */
818 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
819 {
820 	struct svc_xprt *xprt = rqstp->rq_xprt;
821 	struct svcxprt_rdma *rdma_xprt =
822 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
823 	struct svc_rdma_recv_ctxt *ctxt;
824 	int ret;
825 
826 	rqstp->rq_xprt_ctxt = NULL;
827 
828 	spin_lock(&rdma_xprt->sc_rq_dto_lock);
829 	ctxt = svc_rdma_next_recv_ctxt(&rdma_xprt->sc_read_complete_q);
830 	if (ctxt) {
831 		list_del(&ctxt->rc_list);
832 		spin_unlock(&rdma_xprt->sc_rq_dto_lock);
833 		rdma_read_complete(rqstp, ctxt);
834 		goto complete;
835 	}
836 	ctxt = svc_rdma_next_recv_ctxt(&rdma_xprt->sc_rq_dto_q);
837 	if (!ctxt) {
838 		/* No new incoming requests, terminate the loop */
839 		clear_bit(XPT_DATA, &xprt->xpt_flags);
840 		spin_unlock(&rdma_xprt->sc_rq_dto_lock);
841 		return 0;
842 	}
843 	list_del(&ctxt->rc_list);
844 	spin_unlock(&rdma_xprt->sc_rq_dto_lock);
845 	percpu_counter_inc(&svcrdma_stat_recv);
846 
847 	ib_dma_sync_single_for_cpu(rdma_xprt->sc_pd->device,
848 				   ctxt->rc_recv_sge.addr, ctxt->rc_byte_len,
849 				   DMA_FROM_DEVICE);
850 	svc_rdma_build_arg_xdr(rqstp, ctxt);
851 
852 	/* Prevent svc_xprt_release from releasing pages in rq_pages
853 	 * if we return 0 or an error.
854 	 */
855 	rqstp->rq_respages = rqstp->rq_pages;
856 	rqstp->rq_next_page = rqstp->rq_respages;
857 
858 	ret = svc_rdma_xdr_decode_req(&rqstp->rq_arg, ctxt);
859 	if (ret < 0)
860 		goto out_err;
861 	if (ret == 0)
862 		goto out_drop;
863 	rqstp->rq_xprt_hlen = ret;
864 
865 	if (svc_rdma_is_reverse_direction_reply(xprt, ctxt))
866 		goto out_backchannel;
867 
868 	svc_rdma_get_inv_rkey(rdma_xprt, ctxt);
869 
870 	if (!pcl_is_empty(&ctxt->rc_read_pcl) ||
871 	    !pcl_is_empty(&ctxt->rc_call_pcl))
872 		goto out_readlist;
873 
874 complete:
875 	rqstp->rq_xprt_ctxt = ctxt;
876 	rqstp->rq_prot = IPPROTO_MAX;
877 	svc_xprt_copy_addrs(rqstp, xprt);
878 	return rqstp->rq_arg.len;
879 
880 out_readlist:
881 	ret = svc_rdma_process_read_list(rdma_xprt, rqstp, ctxt);
882 	if (ret < 0)
883 		goto out_readfail;
884 	return 0;
885 
886 out_err:
887 	svc_rdma_send_error(rdma_xprt, ctxt, ret);
888 	svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
889 	return 0;
890 
891 out_readfail:
892 	if (ret == -EINVAL)
893 		svc_rdma_send_error(rdma_xprt, ctxt, ret);
894 	svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
895 	return ret;
896 
897 out_backchannel:
898 	svc_rdma_handle_bc_reply(rqstp, ctxt);
899 out_drop:
900 	svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
901 	return 0;
902 }
903