xref: /openbmc/linux/net/sunrpc/xprtrdma/rpc_rdma.c (revision 4f3db074)
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 #if IS_ENABLED(CONFIG_SUNRPC_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 #if IS_ENABLED(CONFIG_SUNRPC_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  * Returns positive number of segments converted, or a negative errno.
82  */
83 
84 static int
85 rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos,
86 	enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs)
87 {
88 	int len, n = 0, p;
89 	int page_base;
90 	struct page **ppages;
91 
92 	if (pos == 0 && xdrbuf->head[0].iov_len) {
93 		seg[n].mr_page = NULL;
94 		seg[n].mr_offset = xdrbuf->head[0].iov_base;
95 		seg[n].mr_len = xdrbuf->head[0].iov_len;
96 		++n;
97 	}
98 
99 	len = xdrbuf->page_len;
100 	ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
101 	page_base = xdrbuf->page_base & ~PAGE_MASK;
102 	p = 0;
103 	while (len && n < nsegs) {
104 		if (!ppages[p]) {
105 			/* alloc the pagelist for receiving buffer */
106 			ppages[p] = alloc_page(GFP_ATOMIC);
107 			if (!ppages[p])
108 				return -ENOMEM;
109 		}
110 		seg[n].mr_page = ppages[p];
111 		seg[n].mr_offset = (void *)(unsigned long) page_base;
112 		seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len);
113 		if (seg[n].mr_len > PAGE_SIZE)
114 			return -EIO;
115 		len -= seg[n].mr_len;
116 		++n;
117 		++p;
118 		page_base = 0;	/* page offset only applies to first page */
119 	}
120 
121 	/* Message overflows the seg array */
122 	if (len && n == nsegs)
123 		return -EIO;
124 
125 	if (xdrbuf->tail[0].iov_len) {
126 		/* the rpcrdma protocol allows us to omit any trailing
127 		 * xdr pad bytes, saving the server an RDMA operation. */
128 		if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize)
129 			return n;
130 		if (n == nsegs)
131 			/* Tail remains, but we're out of segments */
132 			return -EIO;
133 		seg[n].mr_page = NULL;
134 		seg[n].mr_offset = xdrbuf->tail[0].iov_base;
135 		seg[n].mr_len = xdrbuf->tail[0].iov_len;
136 		++n;
137 	}
138 
139 	return n;
140 }
141 
142 /*
143  * Create read/write chunk lists, and reply chunks, for RDMA
144  *
145  *   Assume check against THRESHOLD has been done, and chunks are required.
146  *   Assume only encoding one list entry for read|write chunks. The NFSv3
147  *     protocol is simple enough to allow this as it only has a single "bulk
148  *     result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The
149  *     RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.)
150  *
151  * When used for a single reply chunk (which is a special write
152  * chunk used for the entire reply, rather than just the data), it
153  * is used primarily for READDIR and READLINK which would otherwise
154  * be severely size-limited by a small rdma inline read max. The server
155  * response will come back as an RDMA Write, followed by a message
156  * of type RDMA_NOMSG carrying the xid and length. As a result, reply
157  * chunks do not provide data alignment, however they do not require
158  * "fixup" (moving the response to the upper layer buffer) either.
159  *
160  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
161  *
162  *  Read chunklist (a linked list):
163  *   N elements, position P (same P for all chunks of same arg!):
164  *    1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
165  *
166  *  Write chunklist (a list of (one) counted array):
167  *   N elements:
168  *    1 - N - HLOO - HLOO - ... - HLOO - 0
169  *
170  *  Reply chunk (a counted array):
171  *   N elements:
172  *    1 - N - HLOO - HLOO - ... - HLOO
173  *
174  * Returns positive RPC/RDMA header size, or negative errno.
175  */
176 
177 static ssize_t
178 rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target,
179 		struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type)
180 {
181 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
182 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
183 	int n, nsegs, nchunks = 0;
184 	unsigned int pos;
185 	struct rpcrdma_mr_seg *seg = req->rl_segments;
186 	struct rpcrdma_read_chunk *cur_rchunk = NULL;
187 	struct rpcrdma_write_array *warray = NULL;
188 	struct rpcrdma_write_chunk *cur_wchunk = NULL;
189 	__be32 *iptr = headerp->rm_body.rm_chunks;
190 	int (*map)(struct rpcrdma_xprt *, struct rpcrdma_mr_seg *, int, bool);
191 
192 	if (type == rpcrdma_readch || type == rpcrdma_areadch) {
193 		/* a read chunk - server will RDMA Read our memory */
194 		cur_rchunk = (struct rpcrdma_read_chunk *) iptr;
195 	} else {
196 		/* a write or reply chunk - server will RDMA Write our memory */
197 		*iptr++ = xdr_zero;	/* encode a NULL read chunk list */
198 		if (type == rpcrdma_replych)
199 			*iptr++ = xdr_zero;	/* a NULL write chunk list */
200 		warray = (struct rpcrdma_write_array *) iptr;
201 		cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1);
202 	}
203 
204 	if (type == rpcrdma_replych || type == rpcrdma_areadch)
205 		pos = 0;
206 	else
207 		pos = target->head[0].iov_len;
208 
209 	nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS);
210 	if (nsegs < 0)
211 		return nsegs;
212 
213 	map = r_xprt->rx_ia.ri_ops->ro_map;
214 	do {
215 		n = map(r_xprt, seg, nsegs, cur_wchunk != NULL);
216 		if (n <= 0)
217 			goto out;
218 		if (cur_rchunk) {	/* read */
219 			cur_rchunk->rc_discrim = xdr_one;
220 			/* all read chunks have the same "position" */
221 			cur_rchunk->rc_position = cpu_to_be32(pos);
222 			cur_rchunk->rc_target.rs_handle =
223 						cpu_to_be32(seg->mr_rkey);
224 			cur_rchunk->rc_target.rs_length =
225 						cpu_to_be32(seg->mr_len);
226 			xdr_encode_hyper(
227 					(__be32 *)&cur_rchunk->rc_target.rs_offset,
228 					seg->mr_base);
229 			dprintk("RPC:       %s: read chunk "
230 				"elem %d@0x%llx:0x%x pos %u (%s)\n", __func__,
231 				seg->mr_len, (unsigned long long)seg->mr_base,
232 				seg->mr_rkey, pos, n < nsegs ? "more" : "last");
233 			cur_rchunk++;
234 			r_xprt->rx_stats.read_chunk_count++;
235 		} else {		/* write/reply */
236 			cur_wchunk->wc_target.rs_handle =
237 						cpu_to_be32(seg->mr_rkey);
238 			cur_wchunk->wc_target.rs_length =
239 						cpu_to_be32(seg->mr_len);
240 			xdr_encode_hyper(
241 					(__be32 *)&cur_wchunk->wc_target.rs_offset,
242 					seg->mr_base);
243 			dprintk("RPC:       %s: %s chunk "
244 				"elem %d@0x%llx:0x%x (%s)\n", __func__,
245 				(type == rpcrdma_replych) ? "reply" : "write",
246 				seg->mr_len, (unsigned long long)seg->mr_base,
247 				seg->mr_rkey, n < nsegs ? "more" : "last");
248 			cur_wchunk++;
249 			if (type == rpcrdma_replych)
250 				r_xprt->rx_stats.reply_chunk_count++;
251 			else
252 				r_xprt->rx_stats.write_chunk_count++;
253 			r_xprt->rx_stats.total_rdma_request += seg->mr_len;
254 		}
255 		nchunks++;
256 		seg   += n;
257 		nsegs -= n;
258 	} while (nsegs);
259 
260 	/* success. all failures return above */
261 	req->rl_nchunks = nchunks;
262 
263 	/*
264 	 * finish off header. If write, marshal discrim and nchunks.
265 	 */
266 	if (cur_rchunk) {
267 		iptr = (__be32 *) cur_rchunk;
268 		*iptr++ = xdr_zero;	/* finish the read chunk list */
269 		*iptr++ = xdr_zero;	/* encode a NULL write chunk list */
270 		*iptr++ = xdr_zero;	/* encode a NULL reply chunk */
271 	} else {
272 		warray->wc_discrim = xdr_one;
273 		warray->wc_nchunks = cpu_to_be32(nchunks);
274 		iptr = (__be32 *) cur_wchunk;
275 		if (type == rpcrdma_writech) {
276 			*iptr++ = xdr_zero; /* finish the write chunk list */
277 			*iptr++ = xdr_zero; /* encode a NULL reply chunk */
278 		}
279 	}
280 
281 	/*
282 	 * Return header size.
283 	 */
284 	return (unsigned char *)iptr - (unsigned char *)headerp;
285 
286 out:
287 	if (r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_FRMR)
288 		return n;
289 
290 	for (pos = 0; nchunks--;)
291 		pos += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt,
292 						      &req->rl_segments[pos]);
293 	return n;
294 }
295 
296 /*
297  * Copy write data inline.
298  * This function is used for "small" requests. Data which is passed
299  * to RPC via iovecs (or page list) is copied directly into the
300  * pre-registered memory buffer for this request. For small amounts
301  * of data, this is efficient. The cutoff value is tunable.
302  */
303 static int
304 rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad)
305 {
306 	int i, npages, curlen;
307 	int copy_len;
308 	unsigned char *srcp, *destp;
309 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
310 	int page_base;
311 	struct page **ppages;
312 
313 	destp = rqst->rq_svec[0].iov_base;
314 	curlen = rqst->rq_svec[0].iov_len;
315 	destp += curlen;
316 	/*
317 	 * Do optional padding where it makes sense. Alignment of write
318 	 * payload can help the server, if our setting is accurate.
319 	 */
320 	pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/);
321 	if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH)
322 		pad = 0;	/* don't pad this request */
323 
324 	dprintk("RPC:       %s: pad %d destp 0x%p len %d hdrlen %d\n",
325 		__func__, pad, destp, rqst->rq_slen, curlen);
326 
327 	copy_len = rqst->rq_snd_buf.page_len;
328 
329 	if (rqst->rq_snd_buf.tail[0].iov_len) {
330 		curlen = rqst->rq_snd_buf.tail[0].iov_len;
331 		if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) {
332 			memmove(destp + copy_len,
333 				rqst->rq_snd_buf.tail[0].iov_base, curlen);
334 			r_xprt->rx_stats.pullup_copy_count += curlen;
335 		}
336 		dprintk("RPC:       %s: tail destp 0x%p len %d\n",
337 			__func__, destp + copy_len, curlen);
338 		rqst->rq_svec[0].iov_len += curlen;
339 	}
340 	r_xprt->rx_stats.pullup_copy_count += copy_len;
341 
342 	page_base = rqst->rq_snd_buf.page_base;
343 	ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT);
344 	page_base &= ~PAGE_MASK;
345 	npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT;
346 	for (i = 0; copy_len && i < npages; i++) {
347 		curlen = PAGE_SIZE - page_base;
348 		if (curlen > copy_len)
349 			curlen = copy_len;
350 		dprintk("RPC:       %s: page %d destp 0x%p len %d curlen %d\n",
351 			__func__, i, destp, copy_len, curlen);
352 		srcp = kmap_atomic(ppages[i]);
353 		memcpy(destp, srcp+page_base, curlen);
354 		kunmap_atomic(srcp);
355 		rqst->rq_svec[0].iov_len += curlen;
356 		destp += curlen;
357 		copy_len -= curlen;
358 		page_base = 0;
359 	}
360 	/* header now contains entire send message */
361 	return pad;
362 }
363 
364 /*
365  * Marshal a request: the primary job of this routine is to choose
366  * the transfer modes. See comments below.
367  *
368  * Uses multiple RDMA IOVs for a request:
369  *  [0] -- RPC RDMA header, which uses memory from the *start* of the
370  *         preregistered buffer that already holds the RPC data in
371  *         its middle.
372  *  [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
373  *  [2] -- optional padding.
374  *  [3] -- if padded, header only in [1] and data here.
375  *
376  * Returns zero on success, otherwise a negative errno.
377  */
378 
379 int
380 rpcrdma_marshal_req(struct rpc_rqst *rqst)
381 {
382 	struct rpc_xprt *xprt = rqst->rq_xprt;
383 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
384 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
385 	char *base;
386 	size_t rpclen, padlen;
387 	ssize_t hdrlen;
388 	enum rpcrdma_chunktype rtype, wtype;
389 	struct rpcrdma_msg *headerp;
390 
391 	/*
392 	 * rpclen gets amount of data in first buffer, which is the
393 	 * pre-registered buffer.
394 	 */
395 	base = rqst->rq_svec[0].iov_base;
396 	rpclen = rqst->rq_svec[0].iov_len;
397 
398 	headerp = rdmab_to_msg(req->rl_rdmabuf);
399 	/* don't byte-swap XID, it's already done in request */
400 	headerp->rm_xid = rqst->rq_xid;
401 	headerp->rm_vers = rpcrdma_version;
402 	headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_max_requests);
403 	headerp->rm_type = rdma_msg;
404 
405 	/*
406 	 * Chunks needed for results?
407 	 *
408 	 * o If the expected result is under the inline threshold, all ops
409 	 *   return as inline (but see later).
410 	 * o Large non-read ops return as a single reply chunk.
411 	 * o Large read ops return data as write chunk(s), header as inline.
412 	 *
413 	 * Note: the NFS code sending down multiple result segments implies
414 	 * the op is one of read, readdir[plus], readlink or NFSv4 getacl.
415 	 */
416 
417 	/*
418 	 * This code can handle read chunks, write chunks OR reply
419 	 * chunks -- only one type. If the request is too big to fit
420 	 * inline, then we will choose read chunks. If the request is
421 	 * a READ, then use write chunks to separate the file data
422 	 * into pages; otherwise use reply chunks.
423 	 */
424 	if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst))
425 		wtype = rpcrdma_noch;
426 	else if (rqst->rq_rcv_buf.page_len == 0)
427 		wtype = rpcrdma_replych;
428 	else if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
429 		wtype = rpcrdma_writech;
430 	else
431 		wtype = rpcrdma_replych;
432 
433 	/*
434 	 * Chunks needed for arguments?
435 	 *
436 	 * o If the total request is under the inline threshold, all ops
437 	 *   are sent as inline.
438 	 * o Large non-write ops are sent with the entire message as a
439 	 *   single read chunk (protocol 0-position special case).
440 	 * o Large write ops transmit data as read chunk(s), header as
441 	 *   inline.
442 	 *
443 	 * Note: the NFS code sending down multiple argument segments
444 	 * implies the op is a write.
445 	 * TBD check NFSv4 setacl
446 	 */
447 	if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst))
448 		rtype = rpcrdma_noch;
449 	else if (rqst->rq_snd_buf.page_len == 0)
450 		rtype = rpcrdma_areadch;
451 	else
452 		rtype = rpcrdma_readch;
453 
454 	/* The following simplification is not true forever */
455 	if (rtype != rpcrdma_noch && wtype == rpcrdma_replych)
456 		wtype = rpcrdma_noch;
457 	if (rtype != rpcrdma_noch && wtype != rpcrdma_noch) {
458 		dprintk("RPC:       %s: cannot marshal multiple chunk lists\n",
459 			__func__);
460 		return -EIO;
461 	}
462 
463 	hdrlen = RPCRDMA_HDRLEN_MIN;
464 	padlen = 0;
465 
466 	/*
467 	 * Pull up any extra send data into the preregistered buffer.
468 	 * When padding is in use and applies to the transfer, insert
469 	 * it and change the message type.
470 	 */
471 	if (rtype == rpcrdma_noch) {
472 
473 		padlen = rpcrdma_inline_pullup(rqst,
474 						RPCRDMA_INLINE_PAD_VALUE(rqst));
475 
476 		if (padlen) {
477 			headerp->rm_type = rdma_msgp;
478 			headerp->rm_body.rm_padded.rm_align =
479 				cpu_to_be32(RPCRDMA_INLINE_PAD_VALUE(rqst));
480 			headerp->rm_body.rm_padded.rm_thresh =
481 				cpu_to_be32(RPCRDMA_INLINE_PAD_THRESH);
482 			headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero;
483 			headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero;
484 			headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero;
485 			hdrlen += 2 * sizeof(u32); /* extra words in padhdr */
486 			if (wtype != rpcrdma_noch) {
487 				dprintk("RPC:       %s: invalid chunk list\n",
488 					__func__);
489 				return -EIO;
490 			}
491 		} else {
492 			headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
493 			headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
494 			headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
495 			/* new length after pullup */
496 			rpclen = rqst->rq_svec[0].iov_len;
497 			/*
498 			 * Currently we try to not actually use read inline.
499 			 * Reply chunks have the desirable property that
500 			 * they land, packed, directly in the target buffers
501 			 * without headers, so they require no fixup. The
502 			 * additional RDMA Write op sends the same amount
503 			 * of data, streams on-the-wire and adds no overhead
504 			 * on receive. Therefore, we request a reply chunk
505 			 * for non-writes wherever feasible and efficient.
506 			 */
507 			if (wtype == rpcrdma_noch)
508 				wtype = rpcrdma_replych;
509 		}
510 	}
511 
512 	if (rtype != rpcrdma_noch) {
513 		hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_snd_buf,
514 					       headerp, rtype);
515 		wtype = rtype;	/* simplify dprintk */
516 
517 	} else if (wtype != rpcrdma_noch) {
518 		hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_rcv_buf,
519 					       headerp, wtype);
520 	}
521 	if (hdrlen < 0)
522 		return hdrlen;
523 
524 	dprintk("RPC:       %s: %s: hdrlen %zd rpclen %zd padlen %zd"
525 		" headerp 0x%p base 0x%p lkey 0x%x\n",
526 		__func__, transfertypes[wtype], hdrlen, rpclen, padlen,
527 		headerp, base, rdmab_lkey(req->rl_rdmabuf));
528 
529 	/*
530 	 * initialize send_iov's - normally only two: rdma chunk header and
531 	 * single preregistered RPC header buffer, but if padding is present,
532 	 * then use a preregistered (and zeroed) pad buffer between the RPC
533 	 * header and any write data. In all non-rdma cases, any following
534 	 * data has been copied into the RPC header buffer.
535 	 */
536 	req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf);
537 	req->rl_send_iov[0].length = hdrlen;
538 	req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf);
539 
540 	req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf);
541 	req->rl_send_iov[1].length = rpclen;
542 	req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf);
543 
544 	req->rl_niovs = 2;
545 
546 	if (padlen) {
547 		struct rpcrdma_ep *ep = &r_xprt->rx_ep;
548 
549 		req->rl_send_iov[2].addr = rdmab_addr(ep->rep_padbuf);
550 		req->rl_send_iov[2].length = padlen;
551 		req->rl_send_iov[2].lkey = rdmab_lkey(ep->rep_padbuf);
552 
553 		req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen;
554 		req->rl_send_iov[3].length = rqst->rq_slen - rpclen;
555 		req->rl_send_iov[3].lkey = rdmab_lkey(req->rl_sendbuf);
556 
557 		req->rl_niovs = 4;
558 	}
559 
560 	return 0;
561 }
562 
563 /*
564  * Chase down a received write or reply chunklist to get length
565  * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
566  */
567 static int
568 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
569 {
570 	unsigned int i, total_len;
571 	struct rpcrdma_write_chunk *cur_wchunk;
572 	char *base = (char *)rdmab_to_msg(rep->rr_rdmabuf);
573 
574 	i = be32_to_cpu(**iptrp);
575 	if (i > max)
576 		return -1;
577 	cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
578 	total_len = 0;
579 	while (i--) {
580 		struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
581 		ifdebug(FACILITY) {
582 			u64 off;
583 			xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
584 			dprintk("RPC:       %s: chunk %d@0x%llx:0x%x\n",
585 				__func__,
586 				be32_to_cpu(seg->rs_length),
587 				(unsigned long long)off,
588 				be32_to_cpu(seg->rs_handle));
589 		}
590 		total_len += be32_to_cpu(seg->rs_length);
591 		++cur_wchunk;
592 	}
593 	/* check and adjust for properly terminated write chunk */
594 	if (wrchunk) {
595 		__be32 *w = (__be32 *) cur_wchunk;
596 		if (*w++ != xdr_zero)
597 			return -1;
598 		cur_wchunk = (struct rpcrdma_write_chunk *) w;
599 	}
600 	if ((char *)cur_wchunk > base + rep->rr_len)
601 		return -1;
602 
603 	*iptrp = (__be32 *) cur_wchunk;
604 	return total_len;
605 }
606 
607 /*
608  * Scatter inline received data back into provided iov's.
609  */
610 static void
611 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
612 {
613 	int i, npages, curlen, olen;
614 	char *destp;
615 	struct page **ppages;
616 	int page_base;
617 
618 	curlen = rqst->rq_rcv_buf.head[0].iov_len;
619 	if (curlen > copy_len) {	/* write chunk header fixup */
620 		curlen = copy_len;
621 		rqst->rq_rcv_buf.head[0].iov_len = curlen;
622 	}
623 
624 	dprintk("RPC:       %s: srcp 0x%p len %d hdrlen %d\n",
625 		__func__, srcp, copy_len, curlen);
626 
627 	/* Shift pointer for first receive segment only */
628 	rqst->rq_rcv_buf.head[0].iov_base = srcp;
629 	srcp += curlen;
630 	copy_len -= curlen;
631 
632 	olen = copy_len;
633 	i = 0;
634 	rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen;
635 	page_base = rqst->rq_rcv_buf.page_base;
636 	ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT);
637 	page_base &= ~PAGE_MASK;
638 
639 	if (copy_len && rqst->rq_rcv_buf.page_len) {
640 		npages = PAGE_ALIGN(page_base +
641 			rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT;
642 		for (; i < npages; i++) {
643 			curlen = PAGE_SIZE - page_base;
644 			if (curlen > copy_len)
645 				curlen = copy_len;
646 			dprintk("RPC:       %s: page %d"
647 				" srcp 0x%p len %d curlen %d\n",
648 				__func__, i, srcp, copy_len, curlen);
649 			destp = kmap_atomic(ppages[i]);
650 			memcpy(destp + page_base, srcp, curlen);
651 			flush_dcache_page(ppages[i]);
652 			kunmap_atomic(destp);
653 			srcp += curlen;
654 			copy_len -= curlen;
655 			if (copy_len == 0)
656 				break;
657 			page_base = 0;
658 		}
659 	}
660 
661 	if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) {
662 		curlen = copy_len;
663 		if (curlen > rqst->rq_rcv_buf.tail[0].iov_len)
664 			curlen = rqst->rq_rcv_buf.tail[0].iov_len;
665 		if (rqst->rq_rcv_buf.tail[0].iov_base != srcp)
666 			memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen);
667 		dprintk("RPC:       %s: tail srcp 0x%p len %d curlen %d\n",
668 			__func__, srcp, copy_len, curlen);
669 		rqst->rq_rcv_buf.tail[0].iov_len = curlen;
670 		copy_len -= curlen; ++i;
671 	} else
672 		rqst->rq_rcv_buf.tail[0].iov_len = 0;
673 
674 	if (pad) {
675 		/* implicit padding on terminal chunk */
676 		unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base;
677 		while (pad--)
678 			p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0;
679 	}
680 
681 	if (copy_len)
682 		dprintk("RPC:       %s: %d bytes in"
683 			" %d extra segments (%d lost)\n",
684 			__func__, olen, i, copy_len);
685 
686 	/* TBD avoid a warning from call_decode() */
687 	rqst->rq_private_buf = rqst->rq_rcv_buf;
688 }
689 
690 void
691 rpcrdma_connect_worker(struct work_struct *work)
692 {
693 	struct rpcrdma_ep *ep =
694 		container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
695 	struct rpcrdma_xprt *r_xprt =
696 		container_of(ep, struct rpcrdma_xprt, rx_ep);
697 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
698 
699 	spin_lock_bh(&xprt->transport_lock);
700 	if (++xprt->connect_cookie == 0)	/* maintain a reserved value */
701 		++xprt->connect_cookie;
702 	if (ep->rep_connected > 0) {
703 		if (!xprt_test_and_set_connected(xprt))
704 			xprt_wake_pending_tasks(xprt, 0);
705 	} else {
706 		if (xprt_test_and_clear_connected(xprt))
707 			xprt_wake_pending_tasks(xprt, -ENOTCONN);
708 	}
709 	spin_unlock_bh(&xprt->transport_lock);
710 }
711 
712 /*
713  * This function is called when an async event is posted to
714  * the connection which changes the connection state. All it
715  * does at this point is mark the connection up/down, the rpc
716  * timers do the rest.
717  */
718 void
719 rpcrdma_conn_func(struct rpcrdma_ep *ep)
720 {
721 	schedule_delayed_work(&ep->rep_connect_worker, 0);
722 }
723 
724 /*
725  * Called as a tasklet to do req/reply match and complete a request
726  * Errors must result in the RPC task either being awakened, or
727  * allowed to timeout, to discover the errors at that time.
728  */
729 void
730 rpcrdma_reply_handler(struct rpcrdma_rep *rep)
731 {
732 	struct rpcrdma_msg *headerp;
733 	struct rpcrdma_req *req;
734 	struct rpc_rqst *rqst;
735 	struct rpc_xprt *xprt = rep->rr_xprt;
736 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
737 	__be32 *iptr;
738 	int rdmalen, status;
739 	unsigned long cwnd;
740 	u32 credits;
741 
742 	/* Check status. If bad, signal disconnect and return rep to pool */
743 	if (rep->rr_len == ~0U) {
744 		rpcrdma_recv_buffer_put(rep);
745 		if (r_xprt->rx_ep.rep_connected == 1) {
746 			r_xprt->rx_ep.rep_connected = -EIO;
747 			rpcrdma_conn_func(&r_xprt->rx_ep);
748 		}
749 		return;
750 	}
751 	if (rep->rr_len < RPCRDMA_HDRLEN_MIN) {
752 		dprintk("RPC:       %s: short/invalid reply\n", __func__);
753 		goto repost;
754 	}
755 	headerp = rdmab_to_msg(rep->rr_rdmabuf);
756 	if (headerp->rm_vers != rpcrdma_version) {
757 		dprintk("RPC:       %s: invalid version %d\n",
758 			__func__, be32_to_cpu(headerp->rm_vers));
759 		goto repost;
760 	}
761 
762 	/* Get XID and try for a match. */
763 	spin_lock(&xprt->transport_lock);
764 	rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
765 	if (rqst == NULL) {
766 		spin_unlock(&xprt->transport_lock);
767 		dprintk("RPC:       %s: reply 0x%p failed "
768 			"to match any request xid 0x%08x len %d\n",
769 			__func__, rep, be32_to_cpu(headerp->rm_xid),
770 			rep->rr_len);
771 repost:
772 		r_xprt->rx_stats.bad_reply_count++;
773 		rep->rr_func = rpcrdma_reply_handler;
774 		if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
775 			rpcrdma_recv_buffer_put(rep);
776 
777 		return;
778 	}
779 
780 	/* get request object */
781 	req = rpcr_to_rdmar(rqst);
782 	if (req->rl_reply) {
783 		spin_unlock(&xprt->transport_lock);
784 		dprintk("RPC:       %s: duplicate reply 0x%p to RPC "
785 			"request 0x%p: xid 0x%08x\n", __func__, rep, req,
786 			be32_to_cpu(headerp->rm_xid));
787 		goto repost;
788 	}
789 
790 	dprintk("RPC:       %s: reply 0x%p completes request 0x%p\n"
791 		"                   RPC request 0x%p xid 0x%08x\n",
792 			__func__, rep, req, rqst,
793 			be32_to_cpu(headerp->rm_xid));
794 
795 	/* from here on, the reply is no longer an orphan */
796 	req->rl_reply = rep;
797 	xprt->reestablish_timeout = 0;
798 
799 	/* check for expected message types */
800 	/* The order of some of these tests is important. */
801 	switch (headerp->rm_type) {
802 	case rdma_msg:
803 		/* never expect read chunks */
804 		/* never expect reply chunks (two ways to check) */
805 		/* never expect write chunks without having offered RDMA */
806 		if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
807 		    (headerp->rm_body.rm_chunks[1] == xdr_zero &&
808 		     headerp->rm_body.rm_chunks[2] != xdr_zero) ||
809 		    (headerp->rm_body.rm_chunks[1] != xdr_zero &&
810 		     req->rl_nchunks == 0))
811 			goto badheader;
812 		if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
813 			/* count any expected write chunks in read reply */
814 			/* start at write chunk array count */
815 			iptr = &headerp->rm_body.rm_chunks[2];
816 			rdmalen = rpcrdma_count_chunks(rep,
817 						req->rl_nchunks, 1, &iptr);
818 			/* check for validity, and no reply chunk after */
819 			if (rdmalen < 0 || *iptr++ != xdr_zero)
820 				goto badheader;
821 			rep->rr_len -=
822 			    ((unsigned char *)iptr - (unsigned char *)headerp);
823 			status = rep->rr_len + rdmalen;
824 			r_xprt->rx_stats.total_rdma_reply += rdmalen;
825 			/* special case - last chunk may omit padding */
826 			if (rdmalen &= 3) {
827 				rdmalen = 4 - rdmalen;
828 				status += rdmalen;
829 			}
830 		} else {
831 			/* else ordinary inline */
832 			rdmalen = 0;
833 			iptr = (__be32 *)((unsigned char *)headerp +
834 							RPCRDMA_HDRLEN_MIN);
835 			rep->rr_len -= RPCRDMA_HDRLEN_MIN;
836 			status = rep->rr_len;
837 		}
838 		/* Fix up the rpc results for upper layer */
839 		rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen);
840 		break;
841 
842 	case rdma_nomsg:
843 		/* never expect read or write chunks, always reply chunks */
844 		if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
845 		    headerp->rm_body.rm_chunks[1] != xdr_zero ||
846 		    headerp->rm_body.rm_chunks[2] != xdr_one ||
847 		    req->rl_nchunks == 0)
848 			goto badheader;
849 		iptr = (__be32 *)((unsigned char *)headerp +
850 							RPCRDMA_HDRLEN_MIN);
851 		rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
852 		if (rdmalen < 0)
853 			goto badheader;
854 		r_xprt->rx_stats.total_rdma_reply += rdmalen;
855 		/* Reply chunk buffer already is the reply vector - no fixup. */
856 		status = rdmalen;
857 		break;
858 
859 badheader:
860 	default:
861 		dprintk("%s: invalid rpcrdma reply header (type %d):"
862 				" chunks[012] == %d %d %d"
863 				" expected chunks <= %d\n",
864 				__func__, be32_to_cpu(headerp->rm_type),
865 				headerp->rm_body.rm_chunks[0],
866 				headerp->rm_body.rm_chunks[1],
867 				headerp->rm_body.rm_chunks[2],
868 				req->rl_nchunks);
869 		status = -EIO;
870 		r_xprt->rx_stats.bad_reply_count++;
871 		break;
872 	}
873 
874 	credits = be32_to_cpu(headerp->rm_credit);
875 	if (credits == 0)
876 		credits = 1;	/* don't deadlock */
877 	else if (credits > r_xprt->rx_buf.rb_max_requests)
878 		credits = r_xprt->rx_buf.rb_max_requests;
879 
880 	cwnd = xprt->cwnd;
881 	xprt->cwnd = credits << RPC_CWNDSHIFT;
882 	if (xprt->cwnd > cwnd)
883 		xprt_release_rqst_cong(rqst->rq_task);
884 
885 	dprintk("RPC:       %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
886 			__func__, xprt, rqst, status);
887 	xprt_complete_rqst(rqst->rq_task, status);
888 	spin_unlock(&xprt->transport_lock);
889 }
890