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  */
94 
95 #include <linux/slab.h>
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 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc);
110 
111 static inline struct svc_rdma_recv_ctxt *
112 svc_rdma_next_recv_ctxt(struct list_head *list)
113 {
114 	return list_first_entry_or_null(list, struct svc_rdma_recv_ctxt,
115 					rc_list);
116 }
117 
118 static void svc_rdma_recv_cid_init(struct svcxprt_rdma *rdma,
119 				   struct rpc_rdma_cid *cid)
120 {
121 	cid->ci_queue_id = rdma->sc_rq_cq->res.id;
122 	cid->ci_completion_id = atomic_inc_return(&rdma->sc_completion_ids);
123 }
124 
125 static struct svc_rdma_recv_ctxt *
126 svc_rdma_recv_ctxt_alloc(struct svcxprt_rdma *rdma)
127 {
128 	struct svc_rdma_recv_ctxt *ctxt;
129 	dma_addr_t addr;
130 	void *buffer;
131 
132 	ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
133 	if (!ctxt)
134 		goto fail0;
135 	buffer = kmalloc(rdma->sc_max_req_size, GFP_KERNEL);
136 	if (!buffer)
137 		goto fail1;
138 	addr = ib_dma_map_single(rdma->sc_pd->device, buffer,
139 				 rdma->sc_max_req_size, DMA_FROM_DEVICE);
140 	if (ib_dma_mapping_error(rdma->sc_pd->device, addr))
141 		goto fail2;
142 
143 	svc_rdma_recv_cid_init(rdma, &ctxt->rc_cid);
144 	pcl_init(&ctxt->rc_call_pcl);
145 	pcl_init(&ctxt->rc_read_pcl);
146 	pcl_init(&ctxt->rc_write_pcl);
147 	pcl_init(&ctxt->rc_reply_pcl);
148 
149 	ctxt->rc_recv_wr.next = NULL;
150 	ctxt->rc_recv_wr.wr_cqe = &ctxt->rc_cqe;
151 	ctxt->rc_recv_wr.sg_list = &ctxt->rc_recv_sge;
152 	ctxt->rc_recv_wr.num_sge = 1;
153 	ctxt->rc_cqe.done = svc_rdma_wc_receive;
154 	ctxt->rc_recv_sge.addr = addr;
155 	ctxt->rc_recv_sge.length = rdma->sc_max_req_size;
156 	ctxt->rc_recv_sge.lkey = rdma->sc_pd->local_dma_lkey;
157 	ctxt->rc_recv_buf = buffer;
158 	ctxt->rc_temp = false;
159 	return ctxt;
160 
161 fail2:
162 	kfree(buffer);
163 fail1:
164 	kfree(ctxt);
165 fail0:
166 	return NULL;
167 }
168 
169 static void svc_rdma_recv_ctxt_destroy(struct svcxprt_rdma *rdma,
170 				       struct svc_rdma_recv_ctxt *ctxt)
171 {
172 	ib_dma_unmap_single(rdma->sc_pd->device, ctxt->rc_recv_sge.addr,
173 			    ctxt->rc_recv_sge.length, DMA_FROM_DEVICE);
174 	kfree(ctxt->rc_recv_buf);
175 	kfree(ctxt);
176 }
177 
178 /**
179  * svc_rdma_recv_ctxts_destroy - Release all recv_ctxt's for an xprt
180  * @rdma: svcxprt_rdma being torn down
181  *
182  */
183 void svc_rdma_recv_ctxts_destroy(struct svcxprt_rdma *rdma)
184 {
185 	struct svc_rdma_recv_ctxt *ctxt;
186 	struct llist_node *node;
187 
188 	while ((node = llist_del_first(&rdma->sc_recv_ctxts))) {
189 		ctxt = llist_entry(node, struct svc_rdma_recv_ctxt, rc_node);
190 		svc_rdma_recv_ctxt_destroy(rdma, ctxt);
191 	}
192 }
193 
194 /**
195  * svc_rdma_recv_ctxt_get - Allocate a recv_ctxt
196  * @rdma: controlling svcxprt_rdma
197  *
198  * Returns a recv_ctxt or (rarely) NULL if none are available.
199  */
200 struct svc_rdma_recv_ctxt *svc_rdma_recv_ctxt_get(struct svcxprt_rdma *rdma)
201 {
202 	struct svc_rdma_recv_ctxt *ctxt;
203 	struct llist_node *node;
204 
205 	node = llist_del_first(&rdma->sc_recv_ctxts);
206 	if (!node)
207 		goto out_empty;
208 	ctxt = llist_entry(node, struct svc_rdma_recv_ctxt, rc_node);
209 
210 out:
211 	ctxt->rc_page_count = 0;
212 	return ctxt;
213 
214 out_empty:
215 	ctxt = svc_rdma_recv_ctxt_alloc(rdma);
216 	if (!ctxt)
217 		return NULL;
218 	goto out;
219 }
220 
221 /**
222  * svc_rdma_recv_ctxt_put - Return recv_ctxt to free list
223  * @rdma: controlling svcxprt_rdma
224  * @ctxt: object to return to the free list
225  *
226  */
227 void svc_rdma_recv_ctxt_put(struct svcxprt_rdma *rdma,
228 			    struct svc_rdma_recv_ctxt *ctxt)
229 {
230 	pcl_free(&ctxt->rc_call_pcl);
231 	pcl_free(&ctxt->rc_read_pcl);
232 	pcl_free(&ctxt->rc_write_pcl);
233 	pcl_free(&ctxt->rc_reply_pcl);
234 
235 	if (!ctxt->rc_temp)
236 		llist_add(&ctxt->rc_node, &rdma->sc_recv_ctxts);
237 	else
238 		svc_rdma_recv_ctxt_destroy(rdma, ctxt);
239 }
240 
241 /**
242  * svc_rdma_release_ctxt - Release transport-specific per-rqst resources
243  * @xprt: the transport which owned the context
244  * @vctxt: the context from rqstp->rq_xprt_ctxt or dr->xprt_ctxt
245  *
246  * Ensure that the recv_ctxt is released whether or not a Reply
247  * was sent. For example, the client could close the connection,
248  * or svc_process could drop an RPC, before the Reply is sent.
249  */
250 void svc_rdma_release_ctxt(struct svc_xprt *xprt, void *vctxt)
251 {
252 	struct svc_rdma_recv_ctxt *ctxt = vctxt;
253 	struct svcxprt_rdma *rdma =
254 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
255 
256 	if (ctxt)
257 		svc_rdma_recv_ctxt_put(rdma, ctxt);
258 }
259 
260 static bool svc_rdma_refresh_recvs(struct svcxprt_rdma *rdma,
261 				   unsigned int wanted, bool temp)
262 {
263 	const struct ib_recv_wr *bad_wr = NULL;
264 	struct svc_rdma_recv_ctxt *ctxt;
265 	struct ib_recv_wr *recv_chain;
266 	int ret;
267 
268 	if (test_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags))
269 		return false;
270 
271 	recv_chain = NULL;
272 	while (wanted--) {
273 		ctxt = svc_rdma_recv_ctxt_get(rdma);
274 		if (!ctxt)
275 			break;
276 
277 		trace_svcrdma_post_recv(ctxt);
278 		ctxt->rc_temp = temp;
279 		ctxt->rc_recv_wr.next = recv_chain;
280 		recv_chain = &ctxt->rc_recv_wr;
281 		rdma->sc_pending_recvs++;
282 	}
283 	if (!recv_chain)
284 		return false;
285 
286 	ret = ib_post_recv(rdma->sc_qp, recv_chain, &bad_wr);
287 	if (ret)
288 		goto err_free;
289 	return true;
290 
291 err_free:
292 	trace_svcrdma_rq_post_err(rdma, ret);
293 	while (bad_wr) {
294 		ctxt = container_of(bad_wr, struct svc_rdma_recv_ctxt,
295 				    rc_recv_wr);
296 		bad_wr = bad_wr->next;
297 		svc_rdma_recv_ctxt_put(rdma, ctxt);
298 	}
299 	/* Since we're destroying the xprt, no need to reset
300 	 * sc_pending_recvs. */
301 	return false;
302 }
303 
304 /**
305  * svc_rdma_post_recvs - Post initial set of Recv WRs
306  * @rdma: fresh svcxprt_rdma
307  *
308  * Returns true if successful, otherwise false.
309  */
310 bool svc_rdma_post_recvs(struct svcxprt_rdma *rdma)
311 {
312 	return svc_rdma_refresh_recvs(rdma, rdma->sc_max_requests, true);
313 }
314 
315 /**
316  * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
317  * @cq: Completion Queue context
318  * @wc: Work Completion object
319  *
320  */
321 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
322 {
323 	struct svcxprt_rdma *rdma = cq->cq_context;
324 	struct ib_cqe *cqe = wc->wr_cqe;
325 	struct svc_rdma_recv_ctxt *ctxt;
326 
327 	rdma->sc_pending_recvs--;
328 
329 	/* WARNING: Only wc->wr_cqe and wc->status are reliable */
330 	ctxt = container_of(cqe, struct svc_rdma_recv_ctxt, rc_cqe);
331 
332 	if (wc->status != IB_WC_SUCCESS)
333 		goto flushed;
334 	trace_svcrdma_wc_recv(wc, &ctxt->rc_cid);
335 
336 	/* If receive posting fails, the connection is about to be
337 	 * lost anyway. The server will not be able to send a reply
338 	 * for this RPC, and the client will retransmit this RPC
339 	 * anyway when it reconnects.
340 	 *
341 	 * Therefore we drop the Receive, even if status was SUCCESS
342 	 * to reduce the likelihood of replayed requests once the
343 	 * client reconnects.
344 	 */
345 	if (rdma->sc_pending_recvs < rdma->sc_max_requests)
346 		if (!svc_rdma_refresh_recvs(rdma, rdma->sc_recv_batch, false))
347 			goto dropped;
348 
349 	/* All wc fields are now known to be valid */
350 	ctxt->rc_byte_len = wc->byte_len;
351 
352 	spin_lock(&rdma->sc_rq_dto_lock);
353 	list_add_tail(&ctxt->rc_list, &rdma->sc_rq_dto_q);
354 	/* Note the unlock pairs with the smp_rmb in svc_xprt_ready: */
355 	set_bit(XPT_DATA, &rdma->sc_xprt.xpt_flags);
356 	spin_unlock(&rdma->sc_rq_dto_lock);
357 	if (!test_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags))
358 		svc_xprt_enqueue(&rdma->sc_xprt);
359 	return;
360 
361 flushed:
362 	if (wc->status == IB_WC_WR_FLUSH_ERR)
363 		trace_svcrdma_wc_recv_flush(wc, &ctxt->rc_cid);
364 	else
365 		trace_svcrdma_wc_recv_err(wc, &ctxt->rc_cid);
366 dropped:
367 	svc_rdma_recv_ctxt_put(rdma, ctxt);
368 	svc_xprt_deferred_close(&rdma->sc_xprt);
369 }
370 
371 /**
372  * svc_rdma_flush_recv_queues - Drain pending Receive work
373  * @rdma: svcxprt_rdma being shut down
374  *
375  */
376 void svc_rdma_flush_recv_queues(struct svcxprt_rdma *rdma)
377 {
378 	struct svc_rdma_recv_ctxt *ctxt;
379 
380 	while ((ctxt = svc_rdma_next_recv_ctxt(&rdma->sc_rq_dto_q))) {
381 		list_del(&ctxt->rc_list);
382 		svc_rdma_recv_ctxt_put(rdma, ctxt);
383 	}
384 }
385 
386 static void svc_rdma_build_arg_xdr(struct svc_rqst *rqstp,
387 				   struct svc_rdma_recv_ctxt *ctxt)
388 {
389 	struct xdr_buf *arg = &rqstp->rq_arg;
390 
391 	arg->head[0].iov_base = ctxt->rc_recv_buf;
392 	arg->head[0].iov_len = ctxt->rc_byte_len;
393 	arg->tail[0].iov_base = NULL;
394 	arg->tail[0].iov_len = 0;
395 	arg->page_len = 0;
396 	arg->page_base = 0;
397 	arg->buflen = ctxt->rc_byte_len;
398 	arg->len = ctxt->rc_byte_len;
399 }
400 
401 /**
402  * xdr_count_read_segments - Count number of Read segments in Read list
403  * @rctxt: Ingress receive context
404  * @p: Start of an un-decoded Read list
405  *
406  * Before allocating anything, ensure the ingress Read list is safe
407  * to use.
408  *
409  * The segment count is limited to how many segments can fit in the
410  * transport header without overflowing the buffer. That's about 40
411  * Read segments for a 1KB inline threshold.
412  *
413  * Return values:
414  *   %true: Read list is valid. @rctxt's xdr_stream is updated to point
415  *	    to the first byte past the Read list. rc_read_pcl and
416  *	    rc_call_pcl cl_count fields are set to the number of
417  *	    Read segments in the list.
418  *  %false: Read list is corrupt. @rctxt's xdr_stream is left in an
419  *	    unknown state.
420  */
421 static bool xdr_count_read_segments(struct svc_rdma_recv_ctxt *rctxt, __be32 *p)
422 {
423 	rctxt->rc_call_pcl.cl_count = 0;
424 	rctxt->rc_read_pcl.cl_count = 0;
425 	while (xdr_item_is_present(p)) {
426 		u32 position, handle, length;
427 		u64 offset;
428 
429 		p = xdr_inline_decode(&rctxt->rc_stream,
430 				      rpcrdma_readseg_maxsz * sizeof(*p));
431 		if (!p)
432 			return false;
433 
434 		xdr_decode_read_segment(p, &position, &handle,
435 					    &length, &offset);
436 		if (position) {
437 			if (position & 3)
438 				return false;
439 			++rctxt->rc_read_pcl.cl_count;
440 		} else {
441 			++rctxt->rc_call_pcl.cl_count;
442 		}
443 
444 		p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
445 		if (!p)
446 			return false;
447 	}
448 	return true;
449 }
450 
451 /* Sanity check the Read list.
452  *
453  * Sanity checks:
454  * - Read list does not overflow Receive buffer.
455  * - Chunk size limited by largest NFS data payload.
456  *
457  * Return values:
458  *   %true: Read list is valid. @rctxt's xdr_stream is updated
459  *	    to point to the first byte past the Read list.
460  *  %false: Read list is corrupt. @rctxt's xdr_stream is left
461  *	    in an unknown state.
462  */
463 static bool xdr_check_read_list(struct svc_rdma_recv_ctxt *rctxt)
464 {
465 	__be32 *p;
466 
467 	p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
468 	if (!p)
469 		return false;
470 	if (!xdr_count_read_segments(rctxt, p))
471 		return false;
472 	if (!pcl_alloc_call(rctxt, p))
473 		return false;
474 	return pcl_alloc_read(rctxt, p);
475 }
476 
477 static bool xdr_check_write_chunk(struct svc_rdma_recv_ctxt *rctxt)
478 {
479 	u32 segcount;
480 	__be32 *p;
481 
482 	if (xdr_stream_decode_u32(&rctxt->rc_stream, &segcount))
483 		return false;
484 
485 	/* A bogus segcount causes this buffer overflow check to fail. */
486 	p = xdr_inline_decode(&rctxt->rc_stream,
487 			      segcount * rpcrdma_segment_maxsz * sizeof(*p));
488 	return p != NULL;
489 }
490 
491 /**
492  * xdr_count_write_chunks - Count number of Write chunks in Write list
493  * @rctxt: Received header and decoding state
494  * @p: start of an un-decoded Write list
495  *
496  * Before allocating anything, ensure the ingress Write list is
497  * safe to use.
498  *
499  * Return values:
500  *       %true: Write list is valid. @rctxt's xdr_stream is updated
501  *		to point to the first byte past the Write list, and
502  *		the number of Write chunks is in rc_write_pcl.cl_count.
503  *      %false: Write list is corrupt. @rctxt's xdr_stream is left
504  *		in an indeterminate state.
505  */
506 static bool xdr_count_write_chunks(struct svc_rdma_recv_ctxt *rctxt, __be32 *p)
507 {
508 	rctxt->rc_write_pcl.cl_count = 0;
509 	while (xdr_item_is_present(p)) {
510 		if (!xdr_check_write_chunk(rctxt))
511 			return false;
512 		++rctxt->rc_write_pcl.cl_count;
513 		p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
514 		if (!p)
515 			return false;
516 	}
517 	return true;
518 }
519 
520 /* Sanity check the Write list.
521  *
522  * Implementation limits:
523  * - This implementation currently supports only one Write chunk.
524  *
525  * Sanity checks:
526  * - Write list does not overflow Receive buffer.
527  * - Chunk size limited by largest NFS data payload.
528  *
529  * Return values:
530  *       %true: Write list is valid. @rctxt's xdr_stream is updated
531  *		to point to the first byte past the Write list.
532  *      %false: Write list is corrupt. @rctxt's xdr_stream is left
533  *		in an unknown state.
534  */
535 static bool xdr_check_write_list(struct svc_rdma_recv_ctxt *rctxt)
536 {
537 	__be32 *p;
538 
539 	p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
540 	if (!p)
541 		return false;
542 	if (!xdr_count_write_chunks(rctxt, p))
543 		return false;
544 	if (!pcl_alloc_write(rctxt, &rctxt->rc_write_pcl, p))
545 		return false;
546 
547 	rctxt->rc_cur_result_payload = pcl_first_chunk(&rctxt->rc_write_pcl);
548 	return true;
549 }
550 
551 /* Sanity check the Reply chunk.
552  *
553  * Sanity checks:
554  * - Reply chunk does not overflow Receive buffer.
555  * - Chunk size limited by largest NFS data payload.
556  *
557  * Return values:
558  *       %true: Reply chunk is valid. @rctxt's xdr_stream is updated
559  *		to point to the first byte past the Reply chunk.
560  *      %false: Reply chunk is corrupt. @rctxt's xdr_stream is left
561  *		in an unknown state.
562  */
563 static bool xdr_check_reply_chunk(struct svc_rdma_recv_ctxt *rctxt)
564 {
565 	__be32 *p;
566 
567 	p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p));
568 	if (!p)
569 		return false;
570 
571 	if (!xdr_item_is_present(p))
572 		return true;
573 	if (!xdr_check_write_chunk(rctxt))
574 		return false;
575 
576 	rctxt->rc_reply_pcl.cl_count = 1;
577 	return pcl_alloc_write(rctxt, &rctxt->rc_reply_pcl, p);
578 }
579 
580 /* RPC-over-RDMA Version One private extension: Remote Invalidation.
581  * Responder's choice: requester signals it can handle Send With
582  * Invalidate, and responder chooses one R_key to invalidate.
583  *
584  * If there is exactly one distinct R_key in the received transport
585  * header, set rc_inv_rkey to that R_key. Otherwise, set it to zero.
586  */
587 static void svc_rdma_get_inv_rkey(struct svcxprt_rdma *rdma,
588 				  struct svc_rdma_recv_ctxt *ctxt)
589 {
590 	struct svc_rdma_segment *segment;
591 	struct svc_rdma_chunk *chunk;
592 	u32 inv_rkey;
593 
594 	ctxt->rc_inv_rkey = 0;
595 
596 	if (!rdma->sc_snd_w_inv)
597 		return;
598 
599 	inv_rkey = 0;
600 	pcl_for_each_chunk(chunk, &ctxt->rc_call_pcl) {
601 		pcl_for_each_segment(segment, chunk) {
602 			if (inv_rkey == 0)
603 				inv_rkey = segment->rs_handle;
604 			else if (inv_rkey != segment->rs_handle)
605 				return;
606 		}
607 	}
608 	pcl_for_each_chunk(chunk, &ctxt->rc_read_pcl) {
609 		pcl_for_each_segment(segment, chunk) {
610 			if (inv_rkey == 0)
611 				inv_rkey = segment->rs_handle;
612 			else if (inv_rkey != segment->rs_handle)
613 				return;
614 		}
615 	}
616 	pcl_for_each_chunk(chunk, &ctxt->rc_write_pcl) {
617 		pcl_for_each_segment(segment, chunk) {
618 			if (inv_rkey == 0)
619 				inv_rkey = segment->rs_handle;
620 			else if (inv_rkey != segment->rs_handle)
621 				return;
622 		}
623 	}
624 	pcl_for_each_chunk(chunk, &ctxt->rc_reply_pcl) {
625 		pcl_for_each_segment(segment, chunk) {
626 			if (inv_rkey == 0)
627 				inv_rkey = segment->rs_handle;
628 			else if (inv_rkey != segment->rs_handle)
629 				return;
630 		}
631 	}
632 	ctxt->rc_inv_rkey = inv_rkey;
633 }
634 
635 /**
636  * svc_rdma_xdr_decode_req - Decode the transport header
637  * @rq_arg: xdr_buf containing ingress RPC/RDMA message
638  * @rctxt: state of decoding
639  *
640  * On entry, xdr->head[0].iov_base points to first byte of the
641  * RPC-over-RDMA transport header.
642  *
643  * On successful exit, head[0] points to first byte past the
644  * RPC-over-RDMA header. For RDMA_MSG, this is the RPC message.
645  *
646  * The length of the RPC-over-RDMA header is returned.
647  *
648  * Assumptions:
649  * - The transport header is entirely contained in the head iovec.
650  */
651 static int svc_rdma_xdr_decode_req(struct xdr_buf *rq_arg,
652 				   struct svc_rdma_recv_ctxt *rctxt)
653 {
654 	__be32 *p, *rdma_argp;
655 	unsigned int hdr_len;
656 
657 	rdma_argp = rq_arg->head[0].iov_base;
658 	xdr_init_decode(&rctxt->rc_stream, rq_arg, rdma_argp, NULL);
659 
660 	p = xdr_inline_decode(&rctxt->rc_stream,
661 			      rpcrdma_fixed_maxsz * sizeof(*p));
662 	if (unlikely(!p))
663 		goto out_short;
664 	p++;
665 	if (*p != rpcrdma_version)
666 		goto out_version;
667 	p += 2;
668 	rctxt->rc_msgtype = *p;
669 	switch (rctxt->rc_msgtype) {
670 	case rdma_msg:
671 		break;
672 	case rdma_nomsg:
673 		break;
674 	case rdma_done:
675 		goto out_drop;
676 	case rdma_error:
677 		goto out_drop;
678 	default:
679 		goto out_proc;
680 	}
681 
682 	if (!xdr_check_read_list(rctxt))
683 		goto out_inval;
684 	if (!xdr_check_write_list(rctxt))
685 		goto out_inval;
686 	if (!xdr_check_reply_chunk(rctxt))
687 		goto out_inval;
688 
689 	rq_arg->head[0].iov_base = rctxt->rc_stream.p;
690 	hdr_len = xdr_stream_pos(&rctxt->rc_stream);
691 	rq_arg->head[0].iov_len -= hdr_len;
692 	rq_arg->len -= hdr_len;
693 	trace_svcrdma_decode_rqst(rctxt, rdma_argp, hdr_len);
694 	return hdr_len;
695 
696 out_short:
697 	trace_svcrdma_decode_short_err(rctxt, rq_arg->len);
698 	return -EINVAL;
699 
700 out_version:
701 	trace_svcrdma_decode_badvers_err(rctxt, rdma_argp);
702 	return -EPROTONOSUPPORT;
703 
704 out_drop:
705 	trace_svcrdma_decode_drop_err(rctxt, rdma_argp);
706 	return 0;
707 
708 out_proc:
709 	trace_svcrdma_decode_badproc_err(rctxt, rdma_argp);
710 	return -EINVAL;
711 
712 out_inval:
713 	trace_svcrdma_decode_parse_err(rctxt, rdma_argp);
714 	return -EINVAL;
715 }
716 
717 static void svc_rdma_send_error(struct svcxprt_rdma *rdma,
718 				struct svc_rdma_recv_ctxt *rctxt,
719 				int status)
720 {
721 	struct svc_rdma_send_ctxt *sctxt;
722 
723 	sctxt = svc_rdma_send_ctxt_get(rdma);
724 	if (!sctxt)
725 		return;
726 	svc_rdma_send_error_msg(rdma, sctxt, rctxt, status);
727 }
728 
729 /* By convention, backchannel calls arrive via rdma_msg type
730  * messages, and never populate the chunk lists. This makes
731  * the RPC/RDMA header small and fixed in size, so it is
732  * straightforward to check the RPC header's direction field.
733  */
734 static bool svc_rdma_is_reverse_direction_reply(struct svc_xprt *xprt,
735 						struct svc_rdma_recv_ctxt *rctxt)
736 {
737 	__be32 *p = rctxt->rc_recv_buf;
738 
739 	if (!xprt->xpt_bc_xprt)
740 		return false;
741 
742 	if (rctxt->rc_msgtype != rdma_msg)
743 		return false;
744 
745 	if (!pcl_is_empty(&rctxt->rc_call_pcl))
746 		return false;
747 	if (!pcl_is_empty(&rctxt->rc_read_pcl))
748 		return false;
749 	if (!pcl_is_empty(&rctxt->rc_write_pcl))
750 		return false;
751 	if (!pcl_is_empty(&rctxt->rc_reply_pcl))
752 		return false;
753 
754 	/* RPC call direction */
755 	if (*(p + 8) == cpu_to_be32(RPC_CALL))
756 		return false;
757 
758 	return true;
759 }
760 
761 /**
762  * svc_rdma_recvfrom - Receive an RPC call
763  * @rqstp: request structure into which to receive an RPC Call
764  *
765  * Returns:
766  *	The positive number of bytes in the RPC Call message,
767  *	%0 if there were no Calls ready to return,
768  *	%-EINVAL if the Read chunk data is too large,
769  *	%-ENOMEM if rdma_rw context pool was exhausted,
770  *	%-ENOTCONN if posting failed (connection is lost),
771  *	%-EIO if rdma_rw initialization failed (DMA mapping, etc).
772  *
773  * Called in a loop when XPT_DATA is set. XPT_DATA is cleared only
774  * when there are no remaining ctxt's to process.
775  *
776  * The next ctxt is removed from the "receive" lists.
777  *
778  * - If the ctxt completes a Read, then finish assembling the Call
779  *   message and return the number of bytes in the message.
780  *
781  * - If the ctxt completes a Receive, then construct the Call
782  *   message from the contents of the Receive buffer.
783  *
784  *   - If there are no Read chunks in this message, then finish
785  *     assembling the Call message and return the number of bytes
786  *     in the message.
787  *
788  *   - If there are Read chunks in this message, post Read WRs to
789  *     pull that payload and return 0.
790  */
791 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
792 {
793 	struct svc_xprt *xprt = rqstp->rq_xprt;
794 	struct svcxprt_rdma *rdma_xprt =
795 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
796 	struct svc_rdma_recv_ctxt *ctxt;
797 	int ret;
798 
799 	rqstp->rq_xprt_ctxt = NULL;
800 
801 	ctxt = NULL;
802 	spin_lock(&rdma_xprt->sc_rq_dto_lock);
803 	ctxt = svc_rdma_next_recv_ctxt(&rdma_xprt->sc_rq_dto_q);
804 	if (ctxt)
805 		list_del(&ctxt->rc_list);
806 	else
807 		/* No new incoming requests, terminate the loop */
808 		clear_bit(XPT_DATA, &xprt->xpt_flags);
809 	spin_unlock(&rdma_xprt->sc_rq_dto_lock);
810 
811 	/* Unblock the transport for the next receive */
812 	svc_xprt_received(xprt);
813 	if (!ctxt)
814 		return 0;
815 
816 	percpu_counter_inc(&svcrdma_stat_recv);
817 	ib_dma_sync_single_for_cpu(rdma_xprt->sc_pd->device,
818 				   ctxt->rc_recv_sge.addr, ctxt->rc_byte_len,
819 				   DMA_FROM_DEVICE);
820 	svc_rdma_build_arg_xdr(rqstp, ctxt);
821 
822 	/* Prevent svc_xprt_release from releasing pages in rq_pages
823 	 * if we return 0 or an error.
824 	 */
825 	rqstp->rq_respages = rqstp->rq_pages;
826 	rqstp->rq_next_page = rqstp->rq_respages;
827 
828 	ret = svc_rdma_xdr_decode_req(&rqstp->rq_arg, ctxt);
829 	if (ret < 0)
830 		goto out_err;
831 	if (ret == 0)
832 		goto out_drop;
833 
834 	if (svc_rdma_is_reverse_direction_reply(xprt, ctxt))
835 		goto out_backchannel;
836 
837 	svc_rdma_get_inv_rkey(rdma_xprt, ctxt);
838 
839 	if (!pcl_is_empty(&ctxt->rc_read_pcl) ||
840 	    !pcl_is_empty(&ctxt->rc_call_pcl)) {
841 		ret = svc_rdma_process_read_list(rdma_xprt, rqstp, ctxt);
842 		if (ret < 0)
843 			goto out_readfail;
844 	}
845 
846 	rqstp->rq_xprt_ctxt = ctxt;
847 	rqstp->rq_prot = IPPROTO_MAX;
848 	svc_xprt_copy_addrs(rqstp, xprt);
849 	set_bit(RQ_SECURE, &rqstp->rq_flags);
850 	return rqstp->rq_arg.len;
851 
852 out_err:
853 	svc_rdma_send_error(rdma_xprt, ctxt, ret);
854 	svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
855 	return 0;
856 
857 out_readfail:
858 	if (ret == -EINVAL)
859 		svc_rdma_send_error(rdma_xprt, ctxt, ret);
860 	svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
861 	return ret;
862 
863 out_backchannel:
864 	svc_rdma_handle_bc_reply(rqstp, ctxt);
865 out_drop:
866 	svc_rdma_recv_ctxt_put(rdma_xprt, ctxt);
867 	return 0;
868 }
869