2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
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
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
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.
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
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.
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.
48 #include "xprt_rdma.h"
50 #include <linux/highmem.h>
53 # define RPCDBG_FACILITY RPCDBG_TRANS
56 enum rpcrdma_chunktype {
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 */
75 * Chunk assembly from upper layer xdr_buf.
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.
81 * Note, this routine is never called if the connection's memory
82 * registration strategy is 0 (bounce buffers).
86 rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos,
87 enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs)
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;
98 if (xdrbuf->page_len && (xdrbuf->pages[0] != NULL)) {
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;
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;
120 if (xdrbuf->tail[0].iov_len) {
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;
133 * Create read/write chunk lists, and reply chunks, for RDMA
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.)
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.
150 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
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
156 * Write chunklist (a list of (one) counted array):
158 * 1 - N - HLOO - HLOO - ... - HLOO - 0
160 * Reply chunk (a counted array):
162 * 1 - N - HLOO - HLOO - ... - HLOO
166 rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target,
167 struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type)
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;
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;
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;
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);
191 if (type == rpcrdma_replych || type == rpcrdma_areadch)
194 pos = target->head[0].iov_len;
196 nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS);
201 /* bind/register the memory, then build chunk from result. */
202 int n = rpcrdma_register_external(seg, nsegs,
203 cur_wchunk != NULL, r_xprt);
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);
213 (__be32 *)&cur_rchunk->rc_target.rs_offset,
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");
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);
225 (__be32 *)&cur_wchunk->wc_target.rs_offset,
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");
233 if (type == rpcrdma_replych)
234 r_xprt->rx_stats.reply_chunk_count++;
236 r_xprt->rx_stats.write_chunk_count++;
237 r_xprt->rx_stats.total_rdma_request += seg->mr_len;
244 /* success. all failures return above */
245 req->rl_nchunks = nchunks;
247 BUG_ON(nchunks == 0);
250 * finish off header. If write, marshal discrim and nchunks.
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 */
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 */
268 * Return header size.
270 return (unsigned char *)iptr - (unsigned char *)headerp;
273 for (pos = 0; nchunks--;)
274 pos += rpcrdma_deregister_external(
275 &req->rl_segments[pos], r_xprt, NULL);
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.
287 rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad)
289 int i, npages, curlen;
291 unsigned char *srcp, *destp;
292 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
294 destp = rqst->rq_svec[0].iov_base;
295 curlen = rqst->rq_svec[0].iov_len;
298 * Do optional padding where it makes sense. Alignment of write
299 * payload can help the server, if our setting is accurate.
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 */
305 dprintk("RPC: %s: pad %d destp 0x%p len %d hdrlen %d\n",
306 __func__, pad, destp, rqst->rq_slen, curlen);
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++) {
313 curlen = PAGE_SIZE - rqst->rq_snd_buf.page_base;
316 if (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],
323 memcpy(destp, srcp+rqst->rq_snd_buf.page_base, curlen);
325 memcpy(destp, srcp, curlen);
326 kunmap_atomic(srcp, KM_SKB_SUNRPC_DATA);
327 rqst->rq_svec[0].iov_len += curlen;
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) {
335 rqst->rq_snd_buf.tail[0].iov_base, curlen);
336 r_xprt->rx_stats.pullup_copy_count += curlen;
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;
342 /* header now contains entire send message */
347 * Marshal a request: the primary job of this routine is to choose
348 * the transfer modes. See comments below.
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
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.
360 rpcrdma_marshal_req(struct rpc_rqst *rqst)
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);
366 size_t hdrlen, rpclen, padlen;
367 enum rpcrdma_chunktype rtype, wtype;
368 struct rpcrdma_msg *headerp;
371 * rpclen gets amount of data in first buffer, which is the
372 * pre-registered buffer.
374 base = rqst->rq_svec[0].iov_base;
375 rpclen = rqst->rq_svec[0].iov_len;
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);
386 * Chunks needed for results?
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.
393 * Note: the NFS code sending down multiple result segments implies
394 * the op is one of read, readdir[plus], readlink or NFSv4 getacl.
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.
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;
411 wtype = rpcrdma_replych;
414 * Chunks needed for arguments?
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
423 * Note: the NFS code sending down multiple argument segments
424 * implies the op is a write.
425 * TBD check NFSv4 setacl
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;
432 rtype = rpcrdma_readch;
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);
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);
447 hdrlen = 28; /*sizeof *headerp;*/
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.
455 if (rtype == rpcrdma_noch) {
457 padlen = rpcrdma_inline_pullup(rqst,
458 RPCRDMA_INLINE_PAD_VALUE(rqst));
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);
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;
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.
488 if (wtype == rpcrdma_noch &&
489 r_xprt->rx_ia.ri_memreg_strategy > RPCRDMA_REGISTER)
490 wtype = rpcrdma_replych;
495 * Marshal chunks. This routine will return the header length
496 * consumed by marshaling.
498 if (rtype != rpcrdma_noch) {
499 hdrlen = rpcrdma_create_chunks(rqst,
500 &rqst->rq_snd_buf, headerp, rtype);
501 wtype = rtype; /* simplify dprintk */
503 } else if (wtype != rpcrdma_noch) {
504 hdrlen = rpcrdma_create_chunks(rqst,
505 &rqst->rq_rcv_buf, headerp, wtype);
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);
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.
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;
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;
534 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
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;
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;
551 * Chase down a received write or reply chunklist to get length
552 * RDMA'd by server. See map at rpcrdma_create_chunks()! :-)
555 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp)
557 unsigned int i, total_len;
558 struct rpcrdma_write_chunk *cur_wchunk;
560 i = ntohl(**iptrp); /* get array count */
563 cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1);
566 struct rpcrdma_segment *seg = &cur_wchunk->wc_target;
569 xdr_decode_hyper((__be32 *)&seg->rs_offset, &off);
570 dprintk("RPC: %s: chunk %d@0x%llx:0x%x\n",
572 ntohl(seg->rs_length),
573 (unsigned long long)off,
574 ntohl(seg->rs_handle));
576 total_len += ntohl(seg->rs_length);
579 /* check and adjust for properly terminated write chunk */
581 __be32 *w = (__be32 *) cur_wchunk;
582 if (*w++ != xdr_zero)
584 cur_wchunk = (struct rpcrdma_write_chunk *) w;
586 if ((char *) cur_wchunk > rep->rr_base + rep->rr_len)
589 *iptrp = (__be32 *) cur_wchunk;
594 * Scatter inline received data back into provided iov's.
597 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len)
599 int i, npages, curlen, olen;
602 curlen = rqst->rq_rcv_buf.head[0].iov_len;
603 if (curlen > copy_len) { /* write chunk header fixup */
605 rqst->rq_rcv_buf.head[0].iov_len = curlen;
608 dprintk("RPC: %s: srcp 0x%p len %d hdrlen %d\n",
609 __func__, srcp, copy_len, curlen);
611 /* Shift pointer for first receive segment only */
612 rqst->rq_rcv_buf.head[0].iov_base = srcp;
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++) {
624 curlen = PAGE_SIZE - rqst->rq_rcv_buf.page_base;
627 if (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],
635 memcpy(destp + rqst->rq_rcv_buf.page_base,
638 memcpy(destp, srcp, curlen);
639 flush_dcache_page(rqst->rq_rcv_buf.pages[i]);
640 kunmap_atomic(destp, KM_SKB_SUNRPC_DATA);
646 rqst->rq_rcv_buf.page_len = olen - copy_len;
648 rqst->rq_rcv_buf.page_len = 0;
650 if (copy_len && rqst->rq_rcv_buf.tail[0].iov_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;
661 rqst->rq_rcv_buf.tail[0].iov_len = 0;
664 dprintk("RPC: %s: %d bytes in"
665 " %d extra segments (%d lost)\n",
666 __func__, olen, i, copy_len);
668 /* TBD avoid a warning from call_decode() */
669 rqst->rq_private_buf = rqst->rq_rcv_buf;
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.
679 rpcrdma_conn_func(struct rpcrdma_ep *ep)
681 struct rpc_xprt *xprt = ep->rep_xprt;
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);
688 if (xprt_test_and_clear_connected(xprt))
689 xprt_wake_pending_tasks(xprt, ep->rep_connected);
691 spin_unlock_bh(&xprt->transport_lock);
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.
699 rpcrdma_unbind_func(struct rpcrdma_rep *rep)
701 wake_up(&rep->rr_unbind);
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.
710 rpcrdma_reply_handler(struct rpcrdma_rep *rep)
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);
718 int i, rdmalen, status;
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);
729 if (rep->rr_len < 28) {
730 dprintk("RPC: %s: short/invalid reply\n", __func__);
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));
740 /* Get XID and try for a match. */
741 spin_lock(&xprt->transport_lock);
742 rqst = xprt_lookup_rqst(xprt, headerp->rm_xid);
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);
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);
757 /* get request object */
758 req = rpcr_to_rdmar(rqst);
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);
764 BUG_ON(!req || req->rl_reply);
766 /* from here on, the reply is no longer an orphan */
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))
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)
792 ((unsigned char *)iptr - (unsigned char *)headerp);
793 status = rep->rr_len + rdmalen;
794 r_xprt->rx_stats.total_rdma_reply += rdmalen;
796 /* else ordinary inline */
797 iptr = (__be32 *)((unsigned char *)headerp + 28);
798 rep->rr_len -= 28; /*sizeof *headerp;*/
799 status = rep->rr_len;
801 /* Fix up the rpc results for upper layer */
802 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len);
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)
812 iptr = (__be32 *)((unsigned char *)headerp + 28);
813 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr);
816 r_xprt->rx_stats.total_rdma_reply += rdmalen;
817 /* Reply chunk buffer already is the reply vector - no fixup. */
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],
832 r_xprt->rx_stats.bad_reply_count++;
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],
848 case RPCRDMA_MEMWINDOWS_ASYNC:
849 for (i = 0; req->rl_nchunks--;)
850 i += rpcrdma_deregister_external(&req->rl_segments[i],
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);