xref: /openbmc/linux/net/sunrpc/xprtrdma/rpc_rdma.c (revision f15cbe6f1a4b4d9df59142fc8e4abb973302cf44)
1 /*
2  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the BSD-type
8  * license below:
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  *
14  *      Redistributions of source code must retain the above copyright
15  *      notice, this list of conditions and the following disclaimer.
16  *
17  *      Redistributions in binary form must reproduce the above
18  *      copyright notice, this list of conditions and the following
19  *      disclaimer in the documentation and/or other materials provided
20  *      with the distribution.
21  *
22  *      Neither the name of the Network Appliance, Inc. nor the names of
23  *      its contributors may be used to endorse or promote products
24  *      derived from this software without specific prior written
25  *      permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
30  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
31  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
32  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
33  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
34  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
35  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
36  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
37  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
38  */
39 
40 /*
41  * rpc_rdma.c
42  *
43  * This file contains the guts of the RPC RDMA protocol, and
44  * does marshaling/unmarshaling, etc. It is also where interfacing
45  * to the Linux RPC framework lives.
46  */
47 
48 #include "xprt_rdma.h"
49 
50 #include <linux/highmem.h>
51 
52 #ifdef RPC_DEBUG
53 # define RPCDBG_FACILITY	RPCDBG_TRANS
54 #endif
55 
56 enum rpcrdma_chunktype {
57 	rpcrdma_noch = 0,
58 	rpcrdma_readch,
59 	rpcrdma_areadch,
60 	rpcrdma_writech,
61 	rpcrdma_replych
62 };
63 
64 #ifdef RPC_DEBUG
65 static const char transfertypes[][12] = {
66 	"pure inline",	/* no chunks */
67 	" read chunk",	/* some argument via rdma read */
68 	"*read chunk",	/* entire request via rdma read */
69 	"write chunk",	/* some result via rdma write */
70 	"reply chunk"	/* entire reply via rdma write */
71 };
72 #endif
73 
74 /*
75  * Chunk assembly from upper layer xdr_buf.
76  *
77  * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk
78  * elements. Segments are then coalesced when registered, if possible
79  * within the selected memreg mode.
80  *
81  * Note, this routine is never called if the connection's memory
82  * registration strategy is 0 (bounce buffers).
83  */
84 
85 static int
86 rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos,
87 	enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs)
88 {
89 	int len, n = 0, p;
90 
91 	if (pos == 0 && xdrbuf->head[0].iov_len) {
92 		seg[n].mr_page = NULL;
93 		seg[n].mr_offset = xdrbuf->head[0].iov_base;
94 		seg[n].mr_len = xdrbuf->head[0].iov_len;
95 		++n;
96 	}
97 
98 	if (xdrbuf->page_len && (xdrbuf->pages[0] != NULL)) {
99 		if (n == nsegs)
100 			return 0;
101 		seg[n].mr_page = xdrbuf->pages[0];
102 		seg[n].mr_offset = (void *)(unsigned long) xdrbuf->page_base;
103 		seg[n].mr_len = min_t(u32,
104 			PAGE_SIZE - xdrbuf->page_base, xdrbuf->page_len);
105 		len = xdrbuf->page_len - seg[n].mr_len;
106 		++n;
107 		p = 1;
108 		while (len > 0) {
109 			if (n == nsegs)
110 				return 0;
111 			seg[n].mr_page = xdrbuf->pages[p];
112 			seg[n].mr_offset = NULL;
113 			seg[n].mr_len = min_t(u32, PAGE_SIZE, len);
114 			len -= seg[n].mr_len;
115 			++n;
116 			++p;
117 		}
118 	}
119 
120 	if (xdrbuf->tail[0].iov_len) {
121 		if (n == nsegs)
122 			return 0;
123 		seg[n].mr_page = NULL;
124 		seg[n].mr_offset = xdrbuf->tail[0].iov_base;
125 		seg[n].mr_len = xdrbuf->tail[0].iov_len;
126 		++n;
127 	}
128 
129 	return n;
130 }
131 
132 /*
133  * Create read/write chunk lists, and reply chunks, for RDMA
134  *
135  *   Assume check against THRESHOLD has been done, and chunks are required.
136  *   Assume only encoding one list entry for read|write chunks. The NFSv3
137  *     protocol is simple enough to allow this as it only has a single "bulk
138  *     result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The
139  *     RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.)
140  *
141  * When used for a single reply chunk (which is a special write
142  * chunk used for the entire reply, rather than just the data), it
143  * is used primarily for READDIR and READLINK which would otherwise
144  * be severely size-limited by a small rdma inline read max. The server
145  * response will come back as an RDMA Write, followed by a message
146  * of type RDMA_NOMSG carrying the xid and length. As a result, reply
147  * chunks do not provide data alignment, however they do not require
148  * "fixup" (moving the response to the upper layer buffer) either.
149  *
150  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
151  *
152  *  Read chunklist (a linked list):
153  *   N elements, position P (same P for all chunks of same arg!):
154  *    1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
155  *
156  *  Write chunklist (a list of (one) counted array):
157  *   N elements:
158  *    1 - N - HLOO - HLOO - ... - HLOO - 0
159  *
160  *  Reply chunk (a counted array):
161  *   N elements:
162  *    1 - N - HLOO - HLOO - ... - HLOO
163  */
164 
165 static unsigned int
166 rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target,
167 		struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type)
168 {
169 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
170 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_task->tk_xprt);
171 	int nsegs, nchunks = 0;
172 	unsigned int pos;
173 	struct rpcrdma_mr_seg *seg = req->rl_segments;
174 	struct rpcrdma_read_chunk *cur_rchunk = NULL;
175 	struct rpcrdma_write_array *warray = NULL;
176 	struct rpcrdma_write_chunk *cur_wchunk = NULL;
177 	__be32 *iptr = headerp->rm_body.rm_chunks;
178 
179 	if (type == rpcrdma_readch || type == rpcrdma_areadch) {
180 		/* a read chunk - server will RDMA Read our memory */
181 		cur_rchunk = (struct rpcrdma_read_chunk *) iptr;
182 	} else {
183 		/* a write or reply chunk - server will RDMA Write our memory */
184 		*iptr++ = xdr_zero;	/* encode a NULL read chunk list */
185 		if (type == rpcrdma_replych)
186 			*iptr++ = xdr_zero;	/* a NULL write chunk list */
187 		warray = (struct rpcrdma_write_array *) iptr;
188 		cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1);
189 	}
190 
191 	if (type == rpcrdma_replych || type == rpcrdma_areadch)
192 		pos = 0;
193 	else
194 		pos = target->head[0].iov_len;
195 
196 	nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS);
197 	if (nsegs == 0)
198 		return 0;
199 
200 	do {
201 		/* bind/register the memory, then build chunk from result. */
202 		int n = rpcrdma_register_external(seg, nsegs,
203 						cur_wchunk != NULL, r_xprt);
204 		if (n <= 0)
205 			goto out;
206 		if (cur_rchunk) {	/* read */
207 			cur_rchunk->rc_discrim = xdr_one;
208 			/* all read chunks have the same "position" */
209 			cur_rchunk->rc_position = htonl(pos);
210 			cur_rchunk->rc_target.rs_handle = htonl(seg->mr_rkey);
211 			cur_rchunk->rc_target.rs_length = htonl(seg->mr_len);
212 			xdr_encode_hyper(
213 					(__be32 *)&cur_rchunk->rc_target.rs_offset,
214 					seg->mr_base);
215 			dprintk("RPC:       %s: read chunk "
216 				"elem %d@0x%llx:0x%x pos %u (%s)\n", __func__,
217 				seg->mr_len, (unsigned long long)seg->mr_base,
218 				seg->mr_rkey, pos, n < nsegs ? "more" : "last");
219 			cur_rchunk++;
220 			r_xprt->rx_stats.read_chunk_count++;
221 		} else {		/* write/reply */
222 			cur_wchunk->wc_target.rs_handle = htonl(seg->mr_rkey);
223 			cur_wchunk->wc_target.rs_length = htonl(seg->mr_len);
224 			xdr_encode_hyper(
225 					(__be32 *)&cur_wchunk->wc_target.rs_offset,
226 					seg->mr_base);
227 			dprintk("RPC:       %s: %s chunk "
228 				"elem %d@0x%llx:0x%x (%s)\n", __func__,
229 				(type == rpcrdma_replych) ? "reply" : "write",
230 				seg->mr_len, (unsigned long long)seg->mr_base,
231 				seg->mr_rkey, n < nsegs ? "more" : "last");
232 			cur_wchunk++;
233 			if (type == rpcrdma_replych)
234 				r_xprt->rx_stats.reply_chunk_count++;
235 			else
236 				r_xprt->rx_stats.write_chunk_count++;
237 			r_xprt->rx_stats.total_rdma_request += seg->mr_len;
238 		}
239 		nchunks++;
240 		seg   += n;
241 		nsegs -= n;
242 	} while (nsegs);
243 
244 	/* success. all failures return above */
245 	req->rl_nchunks = nchunks;
246 
247 	BUG_ON(nchunks == 0);
248 
249 	/*
250 	 * finish off header. If write, marshal discrim and nchunks.
251 	 */
252 	if (cur_rchunk) {
253 		iptr = (__be32 *) cur_rchunk;
254 		*iptr++ = xdr_zero;	/* finish the read chunk list */
255 		*iptr++ = xdr_zero;	/* encode a NULL write chunk list */
256 		*iptr++ = xdr_zero;	/* encode a NULL reply chunk */
257 	} else {
258 		warray->wc_discrim = xdr_one;
259 		warray->wc_nchunks = htonl(nchunks);
260 		iptr = (__be32 *) cur_wchunk;
261 		if (type == rpcrdma_writech) {
262 			*iptr++ = xdr_zero; /* finish the write chunk list */
263 			*iptr++ = xdr_zero; /* encode a NULL reply chunk */
264 		}
265 	}
266 
267 	/*
268 	 * Return header size.
269 	 */
270 	return (unsigned char *)iptr - (unsigned char *)headerp;
271 
272 out:
273 	for (pos = 0; nchunks--;)
274 		pos += rpcrdma_deregister_external(
275 				&req->rl_segments[pos], r_xprt, NULL);
276 	return 0;
277 }
278 
279 /*
280  * Copy write data inline.
281  * This function is used for "small" requests. Data which is passed
282  * to RPC via iovecs (or page list) is copied directly into the
283  * pre-registered memory buffer for this request. For small amounts
284  * of data, this is efficient. The cutoff value is tunable.
285  */
286 static int
287 rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad)
288 {
289 	int i, npages, curlen;
290 	int copy_len;
291 	unsigned char *srcp, *destp;
292 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
293 
294 	destp = rqst->rq_svec[0].iov_base;
295 	curlen = rqst->rq_svec[0].iov_len;
296 	destp += curlen;
297 	/*
298 	 * Do optional padding where it makes sense. Alignment of write
299 	 * payload can help the server, if our setting is accurate.
300 	 */
301 	pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/);
302 	if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH)
303 		pad = 0;	/* don't pad this request */
304 
305 	dprintk("RPC:       %s: pad %d destp 0x%p len %d hdrlen %d\n",
306 		__func__, pad, destp, rqst->rq_slen, curlen);
307 
308 	copy_len = rqst->rq_snd_buf.page_len;
309 	r_xprt->rx_stats.pullup_copy_count += copy_len;
310 	npages = PAGE_ALIGN(rqst->rq_snd_buf.page_base+copy_len) >> PAGE_SHIFT;
311 	for (i = 0; copy_len && i < npages; i++) {
312 		if (i == 0)
313 			curlen = PAGE_SIZE - rqst->rq_snd_buf.page_base;
314 		else
315 			curlen = PAGE_SIZE;
316 		if (curlen > copy_len)
317 			curlen = copy_len;
318 		dprintk("RPC:       %s: page %d destp 0x%p len %d curlen %d\n",
319 			__func__, i, destp, copy_len, curlen);
320 		srcp = kmap_atomic(rqst->rq_snd_buf.pages[i],
321 					KM_SKB_SUNRPC_DATA);
322 		if (i == 0)
323 			memcpy(destp, srcp+rqst->rq_snd_buf.page_base, curlen);
324 		else
325 			memcpy(destp, srcp, curlen);
326 		kunmap_atomic(srcp, KM_SKB_SUNRPC_DATA);
327 		rqst->rq_svec[0].iov_len += curlen;
328 		destp += curlen;
329 		copy_len -= curlen;
330 	}
331 	if (rqst->rq_snd_buf.tail[0].iov_len) {
332 		curlen = rqst->rq_snd_buf.tail[0].iov_len;
333 		if (destp != rqst->rq_snd_buf.tail[0].iov_base) {
334 			memcpy(destp,
335 				rqst->rq_snd_buf.tail[0].iov_base, curlen);
336 			r_xprt->rx_stats.pullup_copy_count += curlen;
337 		}
338 		dprintk("RPC:       %s: tail destp 0x%p len %d curlen %d\n",
339 			__func__, destp, copy_len, curlen);
340 		rqst->rq_svec[0].iov_len += curlen;
341 	}
342 	/* header now contains entire send message */
343 	return pad;
344 }
345 
346 /*
347  * Marshal a request: the primary job of this routine is to choose
348  * the transfer modes. See comments below.
349  *
350  * Uses multiple RDMA IOVs for a request:
351  *  [0] -- RPC RDMA header, which uses memory from the *start* of the
352  *         preregistered buffer that already holds the RPC data in
353  *         its middle.
354  *  [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
355  *  [2] -- optional padding.
356  *  [3] -- if padded, header only in [1] and data here.
357  */
358 
359 int
360 rpcrdma_marshal_req(struct rpc_rqst *rqst)
361 {
362 	struct rpc_xprt *xprt = rqst->rq_task->tk_xprt;
363 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
364 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
365 	char *base;
366 	size_t hdrlen, rpclen, padlen;
367 	enum rpcrdma_chunktype rtype, wtype;
368 	struct rpcrdma_msg *headerp;
369 
370 	/*
371 	 * rpclen gets amount of data in first buffer, which is the
372 	 * pre-registered buffer.
373 	 */
374 	base = rqst->rq_svec[0].iov_base;
375 	rpclen = rqst->rq_svec[0].iov_len;
376 
377 	/* build RDMA header in private area at front */
378 	headerp = (struct rpcrdma_msg *) req->rl_base;
379 	/* don't htonl XID, it's already done in request */
380 	headerp->rm_xid = rqst->rq_xid;
381 	headerp->rm_vers = xdr_one;
382 	headerp->rm_credit = htonl(r_xprt->rx_buf.rb_max_requests);
383 	headerp->rm_type = htonl(RDMA_MSG);
384 
385 	/*
386 	 * Chunks needed for results?
387 	 *
388 	 * o If the expected result is under the inline threshold, all ops
389 	 *   return as inline (but see later).
390 	 * o Large non-read ops return as a single reply chunk.
391 	 * o Large read ops return data as write chunk(s), header as inline.
392 	 *
393 	 * Note: the NFS code sending down multiple result segments implies
394 	 * the op is one of read, readdir[plus], readlink or NFSv4 getacl.
395 	 */
396 
397 	/*
398 	 * This code can handle read chunks, write chunks OR reply
399 	 * chunks -- only one type. If the request is too big to fit
400 	 * inline, then we will choose read chunks. If the request is
401 	 * a READ, then use write chunks to separate the file data
402 	 * into pages; otherwise use reply chunks.
403 	 */
404 	if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst))
405 		wtype = rpcrdma_noch;
406 	else if (rqst->rq_rcv_buf.page_len == 0)
407 		wtype = rpcrdma_replych;
408 	else if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
409 		wtype = rpcrdma_writech;
410 	else
411 		wtype = rpcrdma_replych;
412 
413 	/*
414 	 * Chunks needed for arguments?
415 	 *
416 	 * o If the total request is under the inline threshold, all ops
417 	 *   are sent as inline.
418 	 * o Large non-write ops are sent with the entire message as a
419 	 *   single read chunk (protocol 0-position special case).
420 	 * o Large write ops transmit data as read chunk(s), header as
421 	 *   inline.
422 	 *
423 	 * Note: the NFS code sending down multiple argument segments
424 	 * implies the op is a write.
425 	 * TBD check NFSv4 setacl
426 	 */
427 	if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst))
428 		rtype = rpcrdma_noch;
429 	else if (rqst->rq_snd_buf.page_len == 0)
430 		rtype = rpcrdma_areadch;
431 	else
432 		rtype = rpcrdma_readch;
433 
434 	/* The following simplification is not true forever */
435 	if (rtype != rpcrdma_noch && wtype == rpcrdma_replych)
436 		wtype = rpcrdma_noch;
437 	BUG_ON(rtype != rpcrdma_noch && wtype != rpcrdma_noch);
438 
439 	if (r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_BOUNCEBUFFERS &&
440 	    (rtype != rpcrdma_noch || wtype != rpcrdma_noch)) {
441 		/* forced to "pure inline"? */
442 		dprintk("RPC:       %s: too much data (%d/%d) for inline\n",
443 			__func__, rqst->rq_rcv_buf.len, rqst->rq_snd_buf.len);
444 		return -1;
445 	}
446 
447 	hdrlen = 28; /*sizeof *headerp;*/
448 	padlen = 0;
449 
450 	/*
451 	 * Pull up any extra send data into the preregistered buffer.
452 	 * When padding is in use and applies to the transfer, insert
453 	 * it and change the message type.
454 	 */
455 	if (rtype == rpcrdma_noch) {
456 
457 		padlen = rpcrdma_inline_pullup(rqst,
458 						RPCRDMA_INLINE_PAD_VALUE(rqst));
459 
460 		if (padlen) {
461 			headerp->rm_type = htonl(RDMA_MSGP);
462 			headerp->rm_body.rm_padded.rm_align =
463 				htonl(RPCRDMA_INLINE_PAD_VALUE(rqst));
464 			headerp->rm_body.rm_padded.rm_thresh =
465 				htonl(RPCRDMA_INLINE_PAD_THRESH);
466 			headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero;
467 			headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero;
468 			headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero;
469 			hdrlen += 2 * sizeof(u32); /* extra words in padhdr */
470 			BUG_ON(wtype != rpcrdma_noch);
471 
472 		} else {
473 			headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
474 			headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
475 			headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
476 			/* new length after pullup */
477 			rpclen = rqst->rq_svec[0].iov_len;
478 			/*
479 			 * Currently we try to not actually use read inline.
480 			 * Reply chunks have the desirable property that
481 			 * they land, packed, directly in the target buffers
482 			 * without headers, so they require no fixup. The
483 			 * additional RDMA Write op sends the same amount
484 			 * of data, streams on-the-wire and adds no overhead
485 			 * on receive. Therefore, we request a reply chunk
486 			 * for non-writes wherever feasible and efficient.
487 			 */
488 			if (wtype == rpcrdma_noch &&
489 			    r_xprt->rx_ia.ri_memreg_strategy > RPCRDMA_REGISTER)
490 				wtype = rpcrdma_replych;
491 		}
492 	}
493 
494 	/*
495 	 * Marshal chunks. This routine will return the header length
496 	 * consumed by marshaling.
497 	 */
498 	if (rtype != rpcrdma_noch) {
499 		hdrlen = rpcrdma_create_chunks(rqst,
500 					&rqst->rq_snd_buf, headerp, rtype);
501 		wtype = rtype;	/* simplify dprintk */
502 
503 	} else if (wtype != rpcrdma_noch) {
504 		hdrlen = rpcrdma_create_chunks(rqst,
505 					&rqst->rq_rcv_buf, headerp, wtype);
506 	}
507 
508 	if (hdrlen == 0)
509 		return -1;
510 
511 	dprintk("RPC:       %s: %s: hdrlen %zd rpclen %zd padlen %zd\n"
512 		"                   headerp 0x%p base 0x%p lkey 0x%x\n",
513 		__func__, transfertypes[wtype], hdrlen, rpclen, padlen,
514 		headerp, base, req->rl_iov.lkey);
515 
516 	/*
517 	 * initialize send_iov's - normally only two: rdma chunk header and
518 	 * single preregistered RPC header buffer, but if padding is present,
519 	 * then use a preregistered (and zeroed) pad buffer between the RPC
520 	 * header and any write data. In all non-rdma cases, any following
521 	 * data has been copied into the RPC header buffer.
522 	 */
523 	req->rl_send_iov[0].addr = req->rl_iov.addr;
524 	req->rl_send_iov[0].length = hdrlen;
525 	req->rl_send_iov[0].lkey = req->rl_iov.lkey;
526 
527 	req->rl_send_iov[1].addr = req->rl_iov.addr + (base - req->rl_base);
528 	req->rl_send_iov[1].length = rpclen;
529 	req->rl_send_iov[1].lkey = req->rl_iov.lkey;
530 
531 	req->rl_niovs = 2;
532 
533 	if (padlen) {
534 		struct rpcrdma_ep *ep = &r_xprt->rx_ep;
535 
536 		req->rl_send_iov[2].addr = ep->rep_pad.addr;
537 		req->rl_send_iov[2].length = padlen;
538 		req->rl_send_iov[2].lkey = ep->rep_pad.lkey;
539 
540 		req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen;
541 		req->rl_send_iov[3].length = rqst->rq_slen - rpclen;
542 		req->rl_send_iov[3].lkey = req->rl_iov.lkey;
543 
544 		req->rl_niovs = 4;
545 	}
546 
547 	return 0;
548 }
549 
550 /*
551  * Chase down a received write or reply chunklist to get length
552  * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
553  */
554 static int
555 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
556 {
557 	unsigned int i, total_len;
558 	struct rpcrdma_write_chunk *cur_wchunk;
559 
560 	i = ntohl(**iptrp);	/* get array count */
561 	if (i > max)
562 		return -1;
563 	cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
564 	total_len = 0;
565 	while (i--) {
566 		struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
567 		ifdebug(FACILITY) {
568 			u64 off;
569 			xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
570 			dprintk("RPC:       %s: chunk %d@0x%llx:0x%x\n",
571 				__func__,
572 				ntohl(seg->rs_length),
573 				(unsigned long long)off,
574 				ntohl(seg->rs_handle));
575 		}
576 		total_len += ntohl(seg->rs_length);
577 		++cur_wchunk;
578 	}
579 	/* check and adjust for properly terminated write chunk */
580 	if (wrchunk) {
581 		__be32 *w = (__be32 *) cur_wchunk;
582 		if (*w++ != xdr_zero)
583 			return -1;
584 		cur_wchunk = (struct rpcrdma_write_chunk *) w;
585 	}
586 	if ((char *) cur_wchunk > rep->rr_base + rep->rr_len)
587 		return -1;
588 
589 	*iptrp = (__be32 *) cur_wchunk;
590 	return total_len;
591 }
592 
593 /*
594  * Scatter inline received data back into provided iov's.
595  */
596 static void
597 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len)
598 {
599 	int i, npages, curlen, olen;
600 	char *destp;
601 
602 	curlen = rqst->rq_rcv_buf.head[0].iov_len;
603 	if (curlen > copy_len) {	/* write chunk header fixup */
604 		curlen = copy_len;
605 		rqst->rq_rcv_buf.head[0].iov_len = curlen;
606 	}
607 
608 	dprintk("RPC:       %s: srcp 0x%p len %d hdrlen %d\n",
609 		__func__, srcp, copy_len, curlen);
610 
611 	/* Shift pointer for first receive segment only */
612 	rqst->rq_rcv_buf.head[0].iov_base = srcp;
613 	srcp += curlen;
614 	copy_len -= curlen;
615 
616 	olen = copy_len;
617 	i = 0;
618 	rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen;
619 	if (copy_len && rqst->rq_rcv_buf.page_len) {
620 		npages = PAGE_ALIGN(rqst->rq_rcv_buf.page_base +
621 			rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT;
622 		for (; i < npages; i++) {
623 			if (i == 0)
624 				curlen = PAGE_SIZE - rqst->rq_rcv_buf.page_base;
625 			else
626 				curlen = PAGE_SIZE;
627 			if (curlen > copy_len)
628 				curlen = copy_len;
629 			dprintk("RPC:       %s: page %d"
630 				" srcp 0x%p len %d curlen %d\n",
631 				__func__, i, srcp, copy_len, curlen);
632 			destp = kmap_atomic(rqst->rq_rcv_buf.pages[i],
633 						KM_SKB_SUNRPC_DATA);
634 			if (i == 0)
635 				memcpy(destp + rqst->rq_rcv_buf.page_base,
636 						srcp, curlen);
637 			else
638 				memcpy(destp, srcp, curlen);
639 			flush_dcache_page(rqst->rq_rcv_buf.pages[i]);
640 			kunmap_atomic(destp, KM_SKB_SUNRPC_DATA);
641 			srcp += curlen;
642 			copy_len -= curlen;
643 			if (copy_len == 0)
644 				break;
645 		}
646 		rqst->rq_rcv_buf.page_len = olen - copy_len;
647 	} else
648 		rqst->rq_rcv_buf.page_len = 0;
649 
650 	if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) {
651 		curlen = copy_len;
652 		if (curlen > rqst->rq_rcv_buf.tail[0].iov_len)
653 			curlen = rqst->rq_rcv_buf.tail[0].iov_len;
654 		if (rqst->rq_rcv_buf.tail[0].iov_base != srcp)
655 			memcpy(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen);
656 		dprintk("RPC:       %s: tail srcp 0x%p len %d curlen %d\n",
657 			__func__, srcp, copy_len, curlen);
658 		rqst->rq_rcv_buf.tail[0].iov_len = curlen;
659 		copy_len -= curlen; ++i;
660 	} else
661 		rqst->rq_rcv_buf.tail[0].iov_len = 0;
662 
663 	if (copy_len)
664 		dprintk("RPC:       %s: %d bytes in"
665 			" %d extra segments (%d lost)\n",
666 			__func__, olen, i, copy_len);
667 
668 	/* TBD avoid a warning from call_decode() */
669 	rqst->rq_private_buf = rqst->rq_rcv_buf;
670 }
671 
672 /*
673  * This function is called when an async event is posted to
674  * the connection which changes the connection state. All it
675  * does at this point is mark the connection up/down, the rpc
676  * timers do the rest.
677  */
678 void
679 rpcrdma_conn_func(struct rpcrdma_ep *ep)
680 {
681 	struct rpc_xprt *xprt = ep->rep_xprt;
682 
683 	spin_lock_bh(&xprt->transport_lock);
684 	if (ep->rep_connected > 0) {
685 		if (!xprt_test_and_set_connected(xprt))
686 			xprt_wake_pending_tasks(xprt, 0);
687 	} else {
688 		if (xprt_test_and_clear_connected(xprt))
689 			xprt_wake_pending_tasks(xprt, ep->rep_connected);
690 	}
691 	spin_unlock_bh(&xprt->transport_lock);
692 }
693 
694 /*
695  * This function is called when memory window unbind which we are waiting
696  * for completes. Just use rr_func (zeroed by upcall) to signal completion.
697  */
698 static void
699 rpcrdma_unbind_func(struct rpcrdma_rep *rep)
700 {
701 	wake_up(&rep->rr_unbind);
702 }
703 
704 /*
705  * Called as a tasklet to do req/reply match and complete a request
706  * Errors must result in the RPC task either being awakened, or
707  * allowed to timeout, to discover the errors at that time.
708  */
709 void
710 rpcrdma_reply_handler(struct rpcrdma_rep *rep)
711 {
712 	struct rpcrdma_msg *headerp;
713 	struct rpcrdma_req *req;
714 	struct rpc_rqst *rqst;
715 	struct rpc_xprt *xprt = rep->rr_xprt;
716 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
717 	__be32 *iptr;
718 	int i, rdmalen, status;
719 
720 	/* Check status. If bad, signal disconnect and return rep to pool */
721 	if (rep->rr_len == ~0U) {
722 		rpcrdma_recv_buffer_put(rep);
723 		if (r_xprt->rx_ep.rep_connected == 1) {
724 			r_xprt->rx_ep.rep_connected = -EIO;
725 			rpcrdma_conn_func(&r_xprt->rx_ep);
726 		}
727 		return;
728 	}
729 	if (rep->rr_len < 28) {
730 		dprintk("RPC:       %s: short/invalid reply\n", __func__);
731 		goto repost;
732 	}
733 	headerp = (struct rpcrdma_msg *) rep->rr_base;
734 	if (headerp->rm_vers != xdr_one) {
735 		dprintk("RPC:       %s: invalid version %d\n",
736 			__func__, ntohl(headerp->rm_vers));
737 		goto repost;
738 	}
739 
740 	/* Get XID and try for a match. */
741 	spin_lock(&xprt->transport_lock);
742 	rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
743 	if (rqst == NULL) {
744 		spin_unlock(&xprt->transport_lock);
745 		dprintk("RPC:       %s: reply 0x%p failed "
746 			"to match any request xid 0x%08x len %d\n",
747 			__func__, rep, headerp->rm_xid, rep->rr_len);
748 repost:
749 		r_xprt->rx_stats.bad_reply_count++;
750 		rep->rr_func = rpcrdma_reply_handler;
751 		if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
752 			rpcrdma_recv_buffer_put(rep);
753 
754 		return;
755 	}
756 
757 	/* get request object */
758 	req = rpcr_to_rdmar(rqst);
759 
760 	dprintk("RPC:       %s: reply 0x%p completes request 0x%p\n"
761 		"                   RPC request 0x%p xid 0x%08x\n",
762 			__func__, rep, req, rqst, headerp->rm_xid);
763 
764 	BUG_ON(!req || req->rl_reply);
765 
766 	/* from here on, the reply is no longer an orphan */
767 	req->rl_reply = rep;
768 
769 	/* check for expected message types */
770 	/* The order of some of these tests is important. */
771 	switch (headerp->rm_type) {
772 	case __constant_htonl(RDMA_MSG):
773 		/* never expect read chunks */
774 		/* never expect reply chunks (two ways to check) */
775 		/* never expect write chunks without having offered RDMA */
776 		if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
777 		    (headerp->rm_body.rm_chunks[1] == xdr_zero &&
778 		     headerp->rm_body.rm_chunks[2] != xdr_zero) ||
779 		    (headerp->rm_body.rm_chunks[1] != xdr_zero &&
780 		     req->rl_nchunks == 0))
781 			goto badheader;
782 		if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
783 			/* count any expected write chunks in read reply */
784 			/* start at write chunk array count */
785 			iptr = &headerp->rm_body.rm_chunks[2];
786 			rdmalen = rpcrdma_count_chunks(rep,
787 						req->rl_nchunks, 1, &iptr);
788 			/* check for validity, and no reply chunk after */
789 			if (rdmalen < 0 || *iptr++ != xdr_zero)
790 				goto badheader;
791 			rep->rr_len -=
792 			    ((unsigned char *)iptr - (unsigned char *)headerp);
793 			status = rep->rr_len + rdmalen;
794 			r_xprt->rx_stats.total_rdma_reply += rdmalen;
795 		} else {
796 			/* else ordinary inline */
797 			iptr = (__be32 *)((unsigned char *)headerp + 28);
798 			rep->rr_len -= 28; /*sizeof *headerp;*/
799 			status = rep->rr_len;
800 		}
801 		/* Fix up the rpc results for upper layer */
802 		rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len);
803 		break;
804 
805 	case __constant_htonl(RDMA_NOMSG):
806 		/* never expect read or write chunks, always reply chunks */
807 		if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
808 		    headerp->rm_body.rm_chunks[1] != xdr_zero ||
809 		    headerp->rm_body.rm_chunks[2] != xdr_one ||
810 		    req->rl_nchunks == 0)
811 			goto badheader;
812 		iptr = (__be32 *)((unsigned char *)headerp + 28);
813 		rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
814 		if (rdmalen < 0)
815 			goto badheader;
816 		r_xprt->rx_stats.total_rdma_reply += rdmalen;
817 		/* Reply chunk buffer already is the reply vector - no fixup. */
818 		status = rdmalen;
819 		break;
820 
821 badheader:
822 	default:
823 		dprintk("%s: invalid rpcrdma reply header (type %d):"
824 				" chunks[012] == %d %d %d"
825 				" expected chunks <= %d\n",
826 				__func__, ntohl(headerp->rm_type),
827 				headerp->rm_body.rm_chunks[0],
828 				headerp->rm_body.rm_chunks[1],
829 				headerp->rm_body.rm_chunks[2],
830 				req->rl_nchunks);
831 		status = -EIO;
832 		r_xprt->rx_stats.bad_reply_count++;
833 		break;
834 	}
835 
836 	/* If using mw bind, start the deregister process now. */
837 	/* (Note: if mr_free(), cannot perform it here, in tasklet context) */
838 	if (req->rl_nchunks) switch (r_xprt->rx_ia.ri_memreg_strategy) {
839 	case RPCRDMA_MEMWINDOWS:
840 		for (i = 0; req->rl_nchunks-- > 1;)
841 			i += rpcrdma_deregister_external(
842 				&req->rl_segments[i], r_xprt, NULL);
843 		/* Optionally wait (not here) for unbinds to complete */
844 		rep->rr_func = rpcrdma_unbind_func;
845 		(void) rpcrdma_deregister_external(&req->rl_segments[i],
846 						   r_xprt, rep);
847 		break;
848 	case RPCRDMA_MEMWINDOWS_ASYNC:
849 		for (i = 0; req->rl_nchunks--;)
850 			i += rpcrdma_deregister_external(&req->rl_segments[i],
851 							 r_xprt, NULL);
852 		break;
853 	default:
854 		break;
855 	}
856 
857 	dprintk("RPC:       %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
858 			__func__, xprt, rqst, status);
859 	xprt_complete_rqst(rqst->rq_task, status);
860 	spin_unlock(&xprt->transport_lock);
861 }
862