xref: /openbmc/linux/net/sunrpc/xprtrdma/rpc_rdma.c (revision 3e26a691)
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 /* The client can send a request inline as long as the RPCRDMA header
75  * plus the RPC call fit under the transport's inline limit. If the
76  * combined call message size exceeds that limit, the client must use
77  * the read chunk list for this operation.
78  */
79 static bool rpcrdma_args_inline(struct rpc_rqst *rqst)
80 {
81 	unsigned int callsize = RPCRDMA_HDRLEN_MIN + rqst->rq_snd_buf.len;
82 
83 	return callsize <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst);
84 }
85 
86 /* The client can't know how large the actual reply will be. Thus it
87  * plans for the largest possible reply for that particular ULP
88  * operation. If the maximum combined reply message size exceeds that
89  * limit, the client must provide a write list or a reply chunk for
90  * this request.
91  */
92 static bool rpcrdma_results_inline(struct rpc_rqst *rqst)
93 {
94 	unsigned int repsize = RPCRDMA_HDRLEN_MIN + rqst->rq_rcv_buf.buflen;
95 
96 	return repsize <= RPCRDMA_INLINE_READ_THRESHOLD(rqst);
97 }
98 
99 static int
100 rpcrdma_tail_pullup(struct xdr_buf *buf)
101 {
102 	size_t tlen = buf->tail[0].iov_len;
103 	size_t skip = tlen & 3;
104 
105 	/* Do not include the tail if it is only an XDR pad */
106 	if (tlen < 4)
107 		return 0;
108 
109 	/* xdr_write_pages() adds a pad at the beginning of the tail
110 	 * if the content in "buf->pages" is unaligned. Force the
111 	 * tail's actual content to land at the next XDR position
112 	 * after the head instead.
113 	 */
114 	if (skip) {
115 		unsigned char *src, *dst;
116 		unsigned int count;
117 
118 		src = buf->tail[0].iov_base;
119 		dst = buf->head[0].iov_base;
120 		dst += buf->head[0].iov_len;
121 
122 		src += skip;
123 		tlen -= skip;
124 
125 		dprintk("RPC:       %s: skip=%zu, memmove(%p, %p, %zu)\n",
126 			__func__, skip, dst, src, tlen);
127 
128 		for (count = tlen; count; count--)
129 			*dst++ = *src++;
130 	}
131 
132 	return tlen;
133 }
134 
135 /* Split "vec" on page boundaries into segments. FMR registers pages,
136  * not a byte range. Other modes coalesce these segments into a single
137  * MR when they can.
138  */
139 static int
140 rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg,
141 		     int n, int nsegs)
142 {
143 	size_t page_offset;
144 	u32 remaining;
145 	char *base;
146 
147 	base = vec->iov_base;
148 	page_offset = offset_in_page(base);
149 	remaining = vec->iov_len;
150 	while (remaining && n < nsegs) {
151 		seg[n].mr_page = NULL;
152 		seg[n].mr_offset = base;
153 		seg[n].mr_len = min_t(u32, PAGE_SIZE - page_offset, remaining);
154 		remaining -= seg[n].mr_len;
155 		base += seg[n].mr_len;
156 		++n;
157 		page_offset = 0;
158 	}
159 	return n;
160 }
161 
162 /*
163  * Chunk assembly from upper layer xdr_buf.
164  *
165  * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk
166  * elements. Segments are then coalesced when registered, if possible
167  * within the selected memreg mode.
168  *
169  * Returns positive number of segments converted, or a negative errno.
170  */
171 
172 static int
173 rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos,
174 	enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs)
175 {
176 	int len, n = 0, p;
177 	int page_base;
178 	struct page **ppages;
179 
180 	if (pos == 0) {
181 		n = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, n, nsegs);
182 		if (n == nsegs)
183 			return -EIO;
184 	}
185 
186 	len = xdrbuf->page_len;
187 	ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT);
188 	page_base = xdrbuf->page_base & ~PAGE_MASK;
189 	p = 0;
190 	while (len && n < nsegs) {
191 		if (!ppages[p]) {
192 			/* alloc the pagelist for receiving buffer */
193 			ppages[p] = alloc_page(GFP_ATOMIC);
194 			if (!ppages[p])
195 				return -ENOMEM;
196 		}
197 		seg[n].mr_page = ppages[p];
198 		seg[n].mr_offset = (void *)(unsigned long) page_base;
199 		seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len);
200 		if (seg[n].mr_len > PAGE_SIZE)
201 			return -EIO;
202 		len -= seg[n].mr_len;
203 		++n;
204 		++p;
205 		page_base = 0;	/* page offset only applies to first page */
206 	}
207 
208 	/* Message overflows the seg array */
209 	if (len && n == nsegs)
210 		return -EIO;
211 
212 	/* When encoding the read list, the tail is always sent inline */
213 	if (type == rpcrdma_readch)
214 		return n;
215 
216 	if (xdrbuf->tail[0].iov_len) {
217 		/* the rpcrdma protocol allows us to omit any trailing
218 		 * xdr pad bytes, saving the server an RDMA operation. */
219 		if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize)
220 			return n;
221 		n = rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, n, nsegs);
222 		if (n == nsegs)
223 			return -EIO;
224 	}
225 
226 	return n;
227 }
228 
229 /*
230  * Create read/write chunk lists, and reply chunks, for RDMA
231  *
232  *   Assume check against THRESHOLD has been done, and chunks are required.
233  *   Assume only encoding one list entry for read|write chunks. The NFSv3
234  *     protocol is simple enough to allow this as it only has a single "bulk
235  *     result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The
236  *     RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.)
237  *
238  * When used for a single reply chunk (which is a special write
239  * chunk used for the entire reply, rather than just the data), it
240  * is used primarily for READDIR and READLINK which would otherwise
241  * be severely size-limited by a small rdma inline read max. The server
242  * response will come back as an RDMA Write, followed by a message
243  * of type RDMA_NOMSG carrying the xid and length. As a result, reply
244  * chunks do not provide data alignment, however they do not require
245  * "fixup" (moving the response to the upper layer buffer) either.
246  *
247  * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
248  *
249  *  Read chunklist (a linked list):
250  *   N elements, position P (same P for all chunks of same arg!):
251  *    1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
252  *
253  *  Write chunklist (a list of (one) counted array):
254  *   N elements:
255  *    1 - N - HLOO - HLOO - ... - HLOO - 0
256  *
257  *  Reply chunk (a counted array):
258  *   N elements:
259  *    1 - N - HLOO - HLOO - ... - HLOO
260  *
261  * Returns positive RPC/RDMA header size, or negative errno.
262  */
263 
264 static ssize_t
265 rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target,
266 		struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type)
267 {
268 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
269 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
270 	int n, nsegs, nchunks = 0;
271 	unsigned int pos;
272 	struct rpcrdma_mr_seg *seg = req->rl_segments;
273 	struct rpcrdma_read_chunk *cur_rchunk = NULL;
274 	struct rpcrdma_write_array *warray = NULL;
275 	struct rpcrdma_write_chunk *cur_wchunk = NULL;
276 	__be32 *iptr = headerp->rm_body.rm_chunks;
277 	int (*map)(struct rpcrdma_xprt *, struct rpcrdma_mr_seg *, int, bool);
278 
279 	if (type == rpcrdma_readch || type == rpcrdma_areadch) {
280 		/* a read chunk - server will RDMA Read our memory */
281 		cur_rchunk = (struct rpcrdma_read_chunk *) iptr;
282 	} else {
283 		/* a write or reply chunk - server will RDMA Write our memory */
284 		*iptr++ = xdr_zero;	/* encode a NULL read chunk list */
285 		if (type == rpcrdma_replych)
286 			*iptr++ = xdr_zero;	/* a NULL write chunk list */
287 		warray = (struct rpcrdma_write_array *) iptr;
288 		cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1);
289 	}
290 
291 	if (type == rpcrdma_replych || type == rpcrdma_areadch)
292 		pos = 0;
293 	else
294 		pos = target->head[0].iov_len;
295 
296 	nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS);
297 	if (nsegs < 0)
298 		return nsegs;
299 
300 	map = r_xprt->rx_ia.ri_ops->ro_map;
301 	do {
302 		n = map(r_xprt, seg, nsegs, cur_wchunk != NULL);
303 		if (n <= 0)
304 			goto out;
305 		if (cur_rchunk) {	/* read */
306 			cur_rchunk->rc_discrim = xdr_one;
307 			/* all read chunks have the same "position" */
308 			cur_rchunk->rc_position = cpu_to_be32(pos);
309 			cur_rchunk->rc_target.rs_handle =
310 						cpu_to_be32(seg->mr_rkey);
311 			cur_rchunk->rc_target.rs_length =
312 						cpu_to_be32(seg->mr_len);
313 			xdr_encode_hyper(
314 					(__be32 *)&cur_rchunk->rc_target.rs_offset,
315 					seg->mr_base);
316 			dprintk("RPC:       %s: read chunk "
317 				"elem %d@0x%llx:0x%x pos %u (%s)\n", __func__,
318 				seg->mr_len, (unsigned long long)seg->mr_base,
319 				seg->mr_rkey, pos, n < nsegs ? "more" : "last");
320 			cur_rchunk++;
321 			r_xprt->rx_stats.read_chunk_count++;
322 		} else {		/* write/reply */
323 			cur_wchunk->wc_target.rs_handle =
324 						cpu_to_be32(seg->mr_rkey);
325 			cur_wchunk->wc_target.rs_length =
326 						cpu_to_be32(seg->mr_len);
327 			xdr_encode_hyper(
328 					(__be32 *)&cur_wchunk->wc_target.rs_offset,
329 					seg->mr_base);
330 			dprintk("RPC:       %s: %s chunk "
331 				"elem %d@0x%llx:0x%x (%s)\n", __func__,
332 				(type == rpcrdma_replych) ? "reply" : "write",
333 				seg->mr_len, (unsigned long long)seg->mr_base,
334 				seg->mr_rkey, n < nsegs ? "more" : "last");
335 			cur_wchunk++;
336 			if (type == rpcrdma_replych)
337 				r_xprt->rx_stats.reply_chunk_count++;
338 			else
339 				r_xprt->rx_stats.write_chunk_count++;
340 			r_xprt->rx_stats.total_rdma_request += seg->mr_len;
341 		}
342 		nchunks++;
343 		seg   += n;
344 		nsegs -= n;
345 	} while (nsegs);
346 
347 	/* success. all failures return above */
348 	req->rl_nchunks = nchunks;
349 
350 	/*
351 	 * finish off header. If write, marshal discrim and nchunks.
352 	 */
353 	if (cur_rchunk) {
354 		iptr = (__be32 *) cur_rchunk;
355 		*iptr++ = xdr_zero;	/* finish the read chunk list */
356 		*iptr++ = xdr_zero;	/* encode a NULL write chunk list */
357 		*iptr++ = xdr_zero;	/* encode a NULL reply chunk */
358 	} else {
359 		warray->wc_discrim = xdr_one;
360 		warray->wc_nchunks = cpu_to_be32(nchunks);
361 		iptr = (__be32 *) cur_wchunk;
362 		if (type == rpcrdma_writech) {
363 			*iptr++ = xdr_zero; /* finish the write chunk list */
364 			*iptr++ = xdr_zero; /* encode a NULL reply chunk */
365 		}
366 	}
367 
368 	/*
369 	 * Return header size.
370 	 */
371 	return (unsigned char *)iptr - (unsigned char *)headerp;
372 
373 out:
374 	for (pos = 0; nchunks--;)
375 		pos += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt,
376 						      &req->rl_segments[pos]);
377 	return n;
378 }
379 
380 /*
381  * Copy write data inline.
382  * This function is used for "small" requests. Data which is passed
383  * to RPC via iovecs (or page list) is copied directly into the
384  * pre-registered memory buffer for this request. For small amounts
385  * of data, this is efficient. The cutoff value is tunable.
386  */
387 static void rpcrdma_inline_pullup(struct rpc_rqst *rqst)
388 {
389 	int i, npages, curlen;
390 	int copy_len;
391 	unsigned char *srcp, *destp;
392 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
393 	int page_base;
394 	struct page **ppages;
395 
396 	destp = rqst->rq_svec[0].iov_base;
397 	curlen = rqst->rq_svec[0].iov_len;
398 	destp += curlen;
399 
400 	dprintk("RPC:       %s: destp 0x%p len %d hdrlen %d\n",
401 		__func__, destp, rqst->rq_slen, curlen);
402 
403 	copy_len = rqst->rq_snd_buf.page_len;
404 
405 	if (rqst->rq_snd_buf.tail[0].iov_len) {
406 		curlen = rqst->rq_snd_buf.tail[0].iov_len;
407 		if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) {
408 			memmove(destp + copy_len,
409 				rqst->rq_snd_buf.tail[0].iov_base, curlen);
410 			r_xprt->rx_stats.pullup_copy_count += curlen;
411 		}
412 		dprintk("RPC:       %s: tail destp 0x%p len %d\n",
413 			__func__, destp + copy_len, curlen);
414 		rqst->rq_svec[0].iov_len += curlen;
415 	}
416 	r_xprt->rx_stats.pullup_copy_count += copy_len;
417 
418 	page_base = rqst->rq_snd_buf.page_base;
419 	ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT);
420 	page_base &= ~PAGE_MASK;
421 	npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT;
422 	for (i = 0; copy_len && i < npages; i++) {
423 		curlen = PAGE_SIZE - page_base;
424 		if (curlen > copy_len)
425 			curlen = copy_len;
426 		dprintk("RPC:       %s: page %d destp 0x%p len %d curlen %d\n",
427 			__func__, i, destp, copy_len, curlen);
428 		srcp = kmap_atomic(ppages[i]);
429 		memcpy(destp, srcp+page_base, curlen);
430 		kunmap_atomic(srcp);
431 		rqst->rq_svec[0].iov_len += curlen;
432 		destp += curlen;
433 		copy_len -= curlen;
434 		page_base = 0;
435 	}
436 	/* header now contains entire send message */
437 }
438 
439 /*
440  * Marshal a request: the primary job of this routine is to choose
441  * the transfer modes. See comments below.
442  *
443  * Uses multiple RDMA IOVs for a request:
444  *  [0] -- RPC RDMA header, which uses memory from the *start* of the
445  *         preregistered buffer that already holds the RPC data in
446  *         its middle.
447  *  [1] -- the RPC header/data, marshaled by RPC and the NFS protocol.
448  *  [2] -- optional padding.
449  *  [3] -- if padded, header only in [1] and data here.
450  *
451  * Returns zero on success, otherwise a negative errno.
452  */
453 
454 int
455 rpcrdma_marshal_req(struct rpc_rqst *rqst)
456 {
457 	struct rpc_xprt *xprt = rqst->rq_xprt;
458 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
459 	struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
460 	char *base;
461 	size_t rpclen;
462 	ssize_t hdrlen;
463 	enum rpcrdma_chunktype rtype, wtype;
464 	struct rpcrdma_msg *headerp;
465 
466 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
467 	if (test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state))
468 		return rpcrdma_bc_marshal_reply(rqst);
469 #endif
470 
471 	/*
472 	 * rpclen gets amount of data in first buffer, which is the
473 	 * pre-registered buffer.
474 	 */
475 	base = rqst->rq_svec[0].iov_base;
476 	rpclen = rqst->rq_svec[0].iov_len;
477 
478 	headerp = rdmab_to_msg(req->rl_rdmabuf);
479 	/* don't byte-swap XID, it's already done in request */
480 	headerp->rm_xid = rqst->rq_xid;
481 	headerp->rm_vers = rpcrdma_version;
482 	headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_max_requests);
483 	headerp->rm_type = rdma_msg;
484 
485 	/*
486 	 * Chunks needed for results?
487 	 *
488 	 * o Read ops return data as write chunk(s), header as inline.
489 	 * o If the expected result is under the inline threshold, all ops
490 	 *   return as inline.
491 	 * o Large non-read ops return as a single reply chunk.
492 	 */
493 	if (rqst->rq_rcv_buf.flags & XDRBUF_READ)
494 		wtype = rpcrdma_writech;
495 	else if (rpcrdma_results_inline(rqst))
496 		wtype = rpcrdma_noch;
497 	else
498 		wtype = rpcrdma_replych;
499 
500 	/*
501 	 * Chunks needed for arguments?
502 	 *
503 	 * o If the total request is under the inline threshold, all ops
504 	 *   are sent as inline.
505 	 * o Large write ops transmit data as read chunk(s), header as
506 	 *   inline.
507 	 * o Large non-write ops are sent with the entire message as a
508 	 *   single read chunk (protocol 0-position special case).
509 	 *
510 	 * This assumes that the upper layer does not present a request
511 	 * that both has a data payload, and whose non-data arguments
512 	 * by themselves are larger than the inline threshold.
513 	 */
514 	if (rpcrdma_args_inline(rqst)) {
515 		rtype = rpcrdma_noch;
516 	} else if (rqst->rq_snd_buf.flags & XDRBUF_WRITE) {
517 		rtype = rpcrdma_readch;
518 	} else {
519 		r_xprt->rx_stats.nomsg_call_count++;
520 		headerp->rm_type = htonl(RDMA_NOMSG);
521 		rtype = rpcrdma_areadch;
522 		rpclen = 0;
523 	}
524 
525 	/* The following simplification is not true forever */
526 	if (rtype != rpcrdma_noch && wtype == rpcrdma_replych)
527 		wtype = rpcrdma_noch;
528 	if (rtype != rpcrdma_noch && wtype != rpcrdma_noch) {
529 		dprintk("RPC:       %s: cannot marshal multiple chunk lists\n",
530 			__func__);
531 		return -EIO;
532 	}
533 
534 	hdrlen = RPCRDMA_HDRLEN_MIN;
535 
536 	/*
537 	 * Pull up any extra send data into the preregistered buffer.
538 	 * When padding is in use and applies to the transfer, insert
539 	 * it and change the message type.
540 	 */
541 	if (rtype == rpcrdma_noch) {
542 
543 		rpcrdma_inline_pullup(rqst);
544 
545 		headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero;
546 		headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero;
547 		headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero;
548 		/* new length after pullup */
549 		rpclen = rqst->rq_svec[0].iov_len;
550 	} else if (rtype == rpcrdma_readch)
551 		rpclen += rpcrdma_tail_pullup(&rqst->rq_snd_buf);
552 	if (rtype != rpcrdma_noch) {
553 		hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_snd_buf,
554 					       headerp, rtype);
555 		wtype = rtype;	/* simplify dprintk */
556 
557 	} else if (wtype != rpcrdma_noch) {
558 		hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_rcv_buf,
559 					       headerp, wtype);
560 	}
561 	if (hdrlen < 0)
562 		return hdrlen;
563 
564 	dprintk("RPC:       %s: %s: hdrlen %zd rpclen %zd"
565 		" headerp 0x%p base 0x%p lkey 0x%x\n",
566 		__func__, transfertypes[wtype], hdrlen, rpclen,
567 		headerp, base, rdmab_lkey(req->rl_rdmabuf));
568 
569 	/*
570 	 * initialize send_iov's - normally only two: rdma chunk header and
571 	 * single preregistered RPC header buffer, but if padding is present,
572 	 * then use a preregistered (and zeroed) pad buffer between the RPC
573 	 * header and any write data. In all non-rdma cases, any following
574 	 * data has been copied into the RPC header buffer.
575 	 */
576 	req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf);
577 	req->rl_send_iov[0].length = hdrlen;
578 	req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf);
579 
580 	req->rl_niovs = 1;
581 	if (rtype == rpcrdma_areadch)
582 		return 0;
583 
584 	req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf);
585 	req->rl_send_iov[1].length = rpclen;
586 	req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf);
587 
588 	req->rl_niovs = 2;
589 	return 0;
590 }
591 
592 /*
593  * Chase down a received write or reply chunklist to get length
594  * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
595  */
596 static int
597 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
598 {
599 	unsigned int i, total_len;
600 	struct rpcrdma_write_chunk *cur_wchunk;
601 	char *base = (char *)rdmab_to_msg(rep->rr_rdmabuf);
602 
603 	i = be32_to_cpu(**iptrp);
604 	if (i > max)
605 		return -1;
606 	cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
607 	total_len = 0;
608 	while (i--) {
609 		struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
610 		ifdebug(FACILITY) {
611 			u64 off;
612 			xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
613 			dprintk("RPC:       %s: chunk %d@0x%llx:0x%x\n",
614 				__func__,
615 				be32_to_cpu(seg->rs_length),
616 				(unsigned long long)off,
617 				be32_to_cpu(seg->rs_handle));
618 		}
619 		total_len += be32_to_cpu(seg->rs_length);
620 		++cur_wchunk;
621 	}
622 	/* check and adjust for properly terminated write chunk */
623 	if (wrchunk) {
624 		__be32 *w = (__be32 *) cur_wchunk;
625 		if (*w++ != xdr_zero)
626 			return -1;
627 		cur_wchunk = (struct rpcrdma_write_chunk *) w;
628 	}
629 	if ((char *)cur_wchunk > base + rep->rr_len)
630 		return -1;
631 
632 	*iptrp = (__be32 *) cur_wchunk;
633 	return total_len;
634 }
635 
636 /*
637  * Scatter inline received data back into provided iov's.
638  */
639 static void
640 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad)
641 {
642 	int i, npages, curlen, olen;
643 	char *destp;
644 	struct page **ppages;
645 	int page_base;
646 
647 	curlen = rqst->rq_rcv_buf.head[0].iov_len;
648 	if (curlen > copy_len) {	/* write chunk header fixup */
649 		curlen = copy_len;
650 		rqst->rq_rcv_buf.head[0].iov_len = curlen;
651 	}
652 
653 	dprintk("RPC:       %s: srcp 0x%p len %d hdrlen %d\n",
654 		__func__, srcp, copy_len, curlen);
655 
656 	/* Shift pointer for first receive segment only */
657 	rqst->rq_rcv_buf.head[0].iov_base = srcp;
658 	srcp += curlen;
659 	copy_len -= curlen;
660 
661 	olen = copy_len;
662 	i = 0;
663 	rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen;
664 	page_base = rqst->rq_rcv_buf.page_base;
665 	ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT);
666 	page_base &= ~PAGE_MASK;
667 
668 	if (copy_len && rqst->rq_rcv_buf.page_len) {
669 		npages = PAGE_ALIGN(page_base +
670 			rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT;
671 		for (; i < npages; i++) {
672 			curlen = PAGE_SIZE - page_base;
673 			if (curlen > copy_len)
674 				curlen = copy_len;
675 			dprintk("RPC:       %s: page %d"
676 				" srcp 0x%p len %d curlen %d\n",
677 				__func__, i, srcp, copy_len, curlen);
678 			destp = kmap_atomic(ppages[i]);
679 			memcpy(destp + page_base, srcp, curlen);
680 			flush_dcache_page(ppages[i]);
681 			kunmap_atomic(destp);
682 			srcp += curlen;
683 			copy_len -= curlen;
684 			if (copy_len == 0)
685 				break;
686 			page_base = 0;
687 		}
688 	}
689 
690 	if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) {
691 		curlen = copy_len;
692 		if (curlen > rqst->rq_rcv_buf.tail[0].iov_len)
693 			curlen = rqst->rq_rcv_buf.tail[0].iov_len;
694 		if (rqst->rq_rcv_buf.tail[0].iov_base != srcp)
695 			memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen);
696 		dprintk("RPC:       %s: tail srcp 0x%p len %d curlen %d\n",
697 			__func__, srcp, copy_len, curlen);
698 		rqst->rq_rcv_buf.tail[0].iov_len = curlen;
699 		copy_len -= curlen; ++i;
700 	} else
701 		rqst->rq_rcv_buf.tail[0].iov_len = 0;
702 
703 	if (pad) {
704 		/* implicit padding on terminal chunk */
705 		unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base;
706 		while (pad--)
707 			p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0;
708 	}
709 
710 	if (copy_len)
711 		dprintk("RPC:       %s: %d bytes in"
712 			" %d extra segments (%d lost)\n",
713 			__func__, olen, i, copy_len);
714 
715 	/* TBD avoid a warning from call_decode() */
716 	rqst->rq_private_buf = rqst->rq_rcv_buf;
717 }
718 
719 void
720 rpcrdma_connect_worker(struct work_struct *work)
721 {
722 	struct rpcrdma_ep *ep =
723 		container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
724 	struct rpcrdma_xprt *r_xprt =
725 		container_of(ep, struct rpcrdma_xprt, rx_ep);
726 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
727 
728 	spin_lock_bh(&xprt->transport_lock);
729 	if (++xprt->connect_cookie == 0)	/* maintain a reserved value */
730 		++xprt->connect_cookie;
731 	if (ep->rep_connected > 0) {
732 		if (!xprt_test_and_set_connected(xprt))
733 			xprt_wake_pending_tasks(xprt, 0);
734 	} else {
735 		if (xprt_test_and_clear_connected(xprt))
736 			xprt_wake_pending_tasks(xprt, -ENOTCONN);
737 	}
738 	spin_unlock_bh(&xprt->transport_lock);
739 }
740 
741 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
742 /* By convention, backchannel calls arrive via rdma_msg type
743  * messages, and never populate the chunk lists. This makes
744  * the RPC/RDMA header small and fixed in size, so it is
745  * straightforward to check the RPC header's direction field.
746  */
747 static bool
748 rpcrdma_is_bcall(struct rpcrdma_msg *headerp)
749 {
750 	__be32 *p = (__be32 *)headerp;
751 
752 	if (headerp->rm_type != rdma_msg)
753 		return false;
754 	if (headerp->rm_body.rm_chunks[0] != xdr_zero)
755 		return false;
756 	if (headerp->rm_body.rm_chunks[1] != xdr_zero)
757 		return false;
758 	if (headerp->rm_body.rm_chunks[2] != xdr_zero)
759 		return false;
760 
761 	/* sanity */
762 	if (p[7] != headerp->rm_xid)
763 		return false;
764 	/* call direction */
765 	if (p[8] != cpu_to_be32(RPC_CALL))
766 		return false;
767 
768 	return true;
769 }
770 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
771 
772 /*
773  * This function is called when an async event is posted to
774  * the connection which changes the connection state. All it
775  * does at this point is mark the connection up/down, the rpc
776  * timers do the rest.
777  */
778 void
779 rpcrdma_conn_func(struct rpcrdma_ep *ep)
780 {
781 	schedule_delayed_work(&ep->rep_connect_worker, 0);
782 }
783 
784 /* Process received RPC/RDMA messages.
785  *
786  * Errors must result in the RPC task either being awakened, or
787  * allowed to timeout, to discover the errors at that time.
788  */
789 void
790 rpcrdma_reply_handler(struct rpcrdma_rep *rep)
791 {
792 	struct rpcrdma_msg *headerp;
793 	struct rpcrdma_req *req;
794 	struct rpc_rqst *rqst;
795 	struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
796 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
797 	__be32 *iptr;
798 	int rdmalen, status, rmerr;
799 	unsigned long cwnd;
800 
801 	dprintk("RPC:       %s: incoming rep %p\n", __func__, rep);
802 
803 	if (rep->rr_len == RPCRDMA_BAD_LEN)
804 		goto out_badstatus;
805 	if (rep->rr_len < RPCRDMA_HDRLEN_ERR)
806 		goto out_shortreply;
807 
808 	headerp = rdmab_to_msg(rep->rr_rdmabuf);
809 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
810 	if (rpcrdma_is_bcall(headerp))
811 		goto out_bcall;
812 #endif
813 
814 	/* Match incoming rpcrdma_rep to an rpcrdma_req to
815 	 * get context for handling any incoming chunks.
816 	 */
817 	spin_lock_bh(&xprt->transport_lock);
818 	rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
819 	if (!rqst)
820 		goto out_nomatch;
821 
822 	req = rpcr_to_rdmar(rqst);
823 	if (req->rl_reply)
824 		goto out_duplicate;
825 
826 	/* Sanity checking has passed. We are now committed
827 	 * to complete this transaction.
828 	 */
829 	list_del_init(&rqst->rq_list);
830 	spin_unlock_bh(&xprt->transport_lock);
831 	dprintk("RPC:       %s: reply %p completes request %p (xid 0x%08x)\n",
832 		__func__, rep, req, be32_to_cpu(headerp->rm_xid));
833 
834 	/* from here on, the reply is no longer an orphan */
835 	req->rl_reply = rep;
836 	xprt->reestablish_timeout = 0;
837 
838 	if (headerp->rm_vers != rpcrdma_version)
839 		goto out_badversion;
840 
841 	/* check for expected message types */
842 	/* The order of some of these tests is important. */
843 	switch (headerp->rm_type) {
844 	case rdma_msg:
845 		/* never expect read chunks */
846 		/* never expect reply chunks (two ways to check) */
847 		/* never expect write chunks without having offered RDMA */
848 		if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
849 		    (headerp->rm_body.rm_chunks[1] == xdr_zero &&
850 		     headerp->rm_body.rm_chunks[2] != xdr_zero) ||
851 		    (headerp->rm_body.rm_chunks[1] != xdr_zero &&
852 		     req->rl_nchunks == 0))
853 			goto badheader;
854 		if (headerp->rm_body.rm_chunks[1] != xdr_zero) {
855 			/* count any expected write chunks in read reply */
856 			/* start at write chunk array count */
857 			iptr = &headerp->rm_body.rm_chunks[2];
858 			rdmalen = rpcrdma_count_chunks(rep,
859 						req->rl_nchunks, 1, &iptr);
860 			/* check for validity, and no reply chunk after */
861 			if (rdmalen < 0 || *iptr++ != xdr_zero)
862 				goto badheader;
863 			rep->rr_len -=
864 			    ((unsigned char *)iptr - (unsigned char *)headerp);
865 			status = rep->rr_len + rdmalen;
866 			r_xprt->rx_stats.total_rdma_reply += rdmalen;
867 			/* special case - last chunk may omit padding */
868 			if (rdmalen &= 3) {
869 				rdmalen = 4 - rdmalen;
870 				status += rdmalen;
871 			}
872 		} else {
873 			/* else ordinary inline */
874 			rdmalen = 0;
875 			iptr = (__be32 *)((unsigned char *)headerp +
876 							RPCRDMA_HDRLEN_MIN);
877 			rep->rr_len -= RPCRDMA_HDRLEN_MIN;
878 			status = rep->rr_len;
879 		}
880 		/* Fix up the rpc results for upper layer */
881 		rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen);
882 		break;
883 
884 	case rdma_nomsg:
885 		/* never expect read or write chunks, always reply chunks */
886 		if (headerp->rm_body.rm_chunks[0] != xdr_zero ||
887 		    headerp->rm_body.rm_chunks[1] != xdr_zero ||
888 		    headerp->rm_body.rm_chunks[2] != xdr_one ||
889 		    req->rl_nchunks == 0)
890 			goto badheader;
891 		iptr = (__be32 *)((unsigned char *)headerp +
892 							RPCRDMA_HDRLEN_MIN);
893 		rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
894 		if (rdmalen < 0)
895 			goto badheader;
896 		r_xprt->rx_stats.total_rdma_reply += rdmalen;
897 		/* Reply chunk buffer already is the reply vector - no fixup. */
898 		status = rdmalen;
899 		break;
900 
901 	case rdma_error:
902 		goto out_rdmaerr;
903 
904 badheader:
905 	default:
906 		dprintk("%s: invalid rpcrdma reply header (type %d):"
907 				" chunks[012] == %d %d %d"
908 				" expected chunks <= %d\n",
909 				__func__, be32_to_cpu(headerp->rm_type),
910 				headerp->rm_body.rm_chunks[0],
911 				headerp->rm_body.rm_chunks[1],
912 				headerp->rm_body.rm_chunks[2],
913 				req->rl_nchunks);
914 		status = -EIO;
915 		r_xprt->rx_stats.bad_reply_count++;
916 		break;
917 	}
918 
919 out:
920 	/* Invalidate and flush the data payloads before waking the
921 	 * waiting application. This guarantees the memory region is
922 	 * properly fenced from the server before the application
923 	 * accesses the data. It also ensures proper send flow
924 	 * control: waking the next RPC waits until this RPC has
925 	 * relinquished all its Send Queue entries.
926 	 */
927 	if (req->rl_nchunks)
928 		r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt, req);
929 
930 	spin_lock_bh(&xprt->transport_lock);
931 	cwnd = xprt->cwnd;
932 	xprt->cwnd = atomic_read(&r_xprt->rx_buf.rb_credits) << RPC_CWNDSHIFT;
933 	if (xprt->cwnd > cwnd)
934 		xprt_release_rqst_cong(rqst->rq_task);
935 
936 	xprt_complete_rqst(rqst->rq_task, status);
937 	spin_unlock_bh(&xprt->transport_lock);
938 	dprintk("RPC:       %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n",
939 			__func__, xprt, rqst, status);
940 	return;
941 
942 out_badstatus:
943 	rpcrdma_recv_buffer_put(rep);
944 	if (r_xprt->rx_ep.rep_connected == 1) {
945 		r_xprt->rx_ep.rep_connected = -EIO;
946 		rpcrdma_conn_func(&r_xprt->rx_ep);
947 	}
948 	return;
949 
950 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
951 out_bcall:
952 	rpcrdma_bc_receive_call(r_xprt, rep);
953 	return;
954 #endif
955 
956 /* If the incoming reply terminated a pending RPC, the next
957  * RPC call will post a replacement receive buffer as it is
958  * being marshaled.
959  */
960 out_badversion:
961 	dprintk("RPC:       %s: invalid version %d\n",
962 		__func__, be32_to_cpu(headerp->rm_vers));
963 	status = -EIO;
964 	r_xprt->rx_stats.bad_reply_count++;
965 	goto out;
966 
967 out_rdmaerr:
968 	rmerr = be32_to_cpu(headerp->rm_body.rm_error.rm_err);
969 	switch (rmerr) {
970 	case ERR_VERS:
971 		pr_err("%s: server reports header version error (%u-%u)\n",
972 		       __func__,
973 		       be32_to_cpu(headerp->rm_body.rm_error.rm_vers_low),
974 		       be32_to_cpu(headerp->rm_body.rm_error.rm_vers_high));
975 		break;
976 	case ERR_CHUNK:
977 		pr_err("%s: server reports header decoding error\n",
978 		       __func__);
979 		break;
980 	default:
981 		pr_err("%s: server reports unknown error %d\n",
982 		       __func__, rmerr);
983 	}
984 	status = -EREMOTEIO;
985 	r_xprt->rx_stats.bad_reply_count++;
986 	goto out;
987 
988 /* If no pending RPC transaction was matched, post a replacement
989  * receive buffer before returning.
990  */
991 out_shortreply:
992 	dprintk("RPC:       %s: short/invalid reply\n", __func__);
993 	goto repost;
994 
995 out_nomatch:
996 	spin_unlock_bh(&xprt->transport_lock);
997 	dprintk("RPC:       %s: no match for incoming xid 0x%08x len %d\n",
998 		__func__, be32_to_cpu(headerp->rm_xid),
999 		rep->rr_len);
1000 	goto repost;
1001 
1002 out_duplicate:
1003 	spin_unlock_bh(&xprt->transport_lock);
1004 	dprintk("RPC:       %s: "
1005 		"duplicate reply %p to RPC request %p: xid 0x%08x\n",
1006 		__func__, rep, req, be32_to_cpu(headerp->rm_xid));
1007 
1008 repost:
1009 	r_xprt->rx_stats.bad_reply_count++;
1010 	if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep))
1011 		rpcrdma_recv_buffer_put(rep);
1012 }
1013