Commit | Line | Data |
---|---|---|
f58851e6 | 1 | /* |
e9601828 TT |
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. | |
f58851e6 TT |
46 | */ |
47 | ||
48 | #include "xprt_rdma.h" | |
49 | ||
e9601828 TT |
50 | #include <linux/highmem.h> |
51 | ||
f895b252 | 52 | #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) |
e9601828 TT |
53 | # define RPCDBG_FACILITY RPCDBG_TRANS |
54 | #endif | |
55 | ||
e2377945 CL |
56 | enum rpcrdma_chunktype { |
57 | rpcrdma_noch = 0, | |
58 | rpcrdma_readch, | |
59 | rpcrdma_areadch, | |
60 | rpcrdma_writech, | |
61 | rpcrdma_replych | |
62 | }; | |
63 | ||
f895b252 | 64 | #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) |
e9601828 TT |
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. | |
c93c6223 CL |
80 | * |
81 | * Returns positive number of segments converted, or a negative errno. | |
e9601828 TT |
82 | */ |
83 | ||
84 | static int | |
2a428b2b | 85 | rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos, |
e9601828 TT |
86 | enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs) |
87 | { | |
88 | int len, n = 0, p; | |
bd7ea31b TT |
89 | int page_base; |
90 | struct page **ppages; | |
e9601828 TT |
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; | |
e9601828 TT |
96 | ++n; |
97 | } | |
98 | ||
bd7ea31b TT |
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) { | |
196c6998 SM |
104 | if (!ppages[p]) { |
105 | /* alloc the pagelist for receiving buffer */ | |
106 | ppages[p] = alloc_page(GFP_ATOMIC); | |
107 | if (!ppages[p]) | |
c93c6223 | 108 | return -ENOMEM; |
196c6998 | 109 | } |
bd7ea31b TT |
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); | |
c93c6223 CL |
113 | if (seg[n].mr_len > PAGE_SIZE) |
114 | return -EIO; | |
bd7ea31b | 115 | len -= seg[n].mr_len; |
e9601828 | 116 | ++n; |
bd7ea31b TT |
117 | ++p; |
118 | page_base = 0; /* page offset only applies to first page */ | |
e9601828 TT |
119 | } |
120 | ||
bd7ea31b TT |
121 | /* Message overflows the seg array */ |
122 | if (len && n == nsegs) | |
c93c6223 | 123 | return -EIO; |
bd7ea31b | 124 | |
50e1092b | 125 | if (xdrbuf->tail[0].iov_len) { |
9191ca3b TT |
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; | |
e9601828 | 130 | if (n == nsegs) |
bd7ea31b | 131 | /* Tail remains, but we're out of segments */ |
c93c6223 | 132 | return -EIO; |
e9601828 TT |
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; | |
e9601828 TT |
136 | ++n; |
137 | } | |
138 | ||
e9601828 TT |
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 | |
c93c6223 CL |
173 | * |
174 | * Returns positive RPC/RDMA header size, or negative errno. | |
e9601828 TT |
175 | */ |
176 | ||
c93c6223 | 177 | static ssize_t |
e9601828 TT |
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); | |
a4f0835c | 182 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); |
c93c6223 | 183 | int n, nsegs, nchunks = 0; |
2a428b2b | 184 | unsigned int pos; |
e9601828 TT |
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; | |
2d8a9726 | 189 | __be32 *iptr = headerp->rm_body.rm_chunks; |
9c1b4d77 | 190 | int (*map)(struct rpcrdma_xprt *, struct rpcrdma_mr_seg *, int, bool); |
e9601828 TT |
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); | |
c93c6223 CL |
210 | if (nsegs < 0) |
211 | return nsegs; | |
e9601828 | 212 | |
9c1b4d77 | 213 | map = r_xprt->rx_ia.ri_ops->ro_map; |
e9601828 | 214 | do { |
9c1b4d77 | 215 | n = map(r_xprt, seg, nsegs, cur_wchunk != NULL); |
e9601828 TT |
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" */ | |
284f4902 CL |
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); | |
e9601828 | 226 | xdr_encode_hyper( |
2d8a9726 | 227 | (__be32 *)&cur_rchunk->rc_target.rs_offset, |
e9601828 TT |
228 | seg->mr_base); |
229 | dprintk("RPC: %s: read chunk " | |
2a428b2b | 230 | "elem %d@0x%llx:0x%x pos %u (%s)\n", __func__, |
e08a132b SR |
231 | seg->mr_len, (unsigned long long)seg->mr_base, |
232 | seg->mr_rkey, pos, n < nsegs ? "more" : "last"); | |
e9601828 TT |
233 | cur_rchunk++; |
234 | r_xprt->rx_stats.read_chunk_count++; | |
235 | } else { /* write/reply */ | |
284f4902 CL |
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); | |
e9601828 | 240 | xdr_encode_hyper( |
2d8a9726 | 241 | (__be32 *)&cur_wchunk->wc_target.rs_offset, |
e9601828 TT |
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", | |
e08a132b SR |
246 | seg->mr_len, (unsigned long long)seg->mr_base, |
247 | seg->mr_rkey, n < nsegs ? "more" : "last"); | |
e9601828 TT |
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 | ||
e9601828 TT |
263 | /* |
264 | * finish off header. If write, marshal discrim and nchunks. | |
265 | */ | |
266 | if (cur_rchunk) { | |
2d8a9726 | 267 | iptr = (__be32 *) cur_rchunk; |
e9601828 TT |
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; | |
284f4902 | 273 | warray->wc_nchunks = cpu_to_be32(nchunks); |
2d8a9726 | 274 | iptr = (__be32 *) cur_wchunk; |
e9601828 TT |
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: | |
6814baea CL |
287 | for (pos = 0; nchunks--;) |
288 | pos += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt, | |
289 | &req->rl_segments[pos]); | |
c93c6223 | 290 | return n; |
e9601828 TT |
291 | } |
292 | ||
293 | /* | |
294 | * Copy write data inline. | |
295 | * This function is used for "small" requests. Data which is passed | |
296 | * to RPC via iovecs (or page list) is copied directly into the | |
297 | * pre-registered memory buffer for this request. For small amounts | |
298 | * of data, this is efficient. The cutoff value is tunable. | |
299 | */ | |
b3221d6a | 300 | static void rpcrdma_inline_pullup(struct rpc_rqst *rqst) |
e9601828 TT |
301 | { |
302 | int i, npages, curlen; | |
303 | int copy_len; | |
304 | unsigned char *srcp, *destp; | |
305 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); | |
bd7ea31b TT |
306 | int page_base; |
307 | struct page **ppages; | |
e9601828 TT |
308 | |
309 | destp = rqst->rq_svec[0].iov_base; | |
310 | curlen = rqst->rq_svec[0].iov_len; | |
311 | destp += curlen; | |
e9601828 | 312 | |
b3221d6a CL |
313 | dprintk("RPC: %s: destp 0x%p len %d hdrlen %d\n", |
314 | __func__, destp, rqst->rq_slen, curlen); | |
e9601828 TT |
315 | |
316 | copy_len = rqst->rq_snd_buf.page_len; | |
b38ab40a TT |
317 | |
318 | if (rqst->rq_snd_buf.tail[0].iov_len) { | |
319 | curlen = rqst->rq_snd_buf.tail[0].iov_len; | |
320 | if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) { | |
321 | memmove(destp + copy_len, | |
322 | rqst->rq_snd_buf.tail[0].iov_base, curlen); | |
323 | r_xprt->rx_stats.pullup_copy_count += curlen; | |
324 | } | |
325 | dprintk("RPC: %s: tail destp 0x%p len %d\n", | |
326 | __func__, destp + copy_len, curlen); | |
327 | rqst->rq_svec[0].iov_len += curlen; | |
328 | } | |
e9601828 | 329 | r_xprt->rx_stats.pullup_copy_count += copy_len; |
bd7ea31b TT |
330 | |
331 | page_base = rqst->rq_snd_buf.page_base; | |
332 | ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT); | |
333 | page_base &= ~PAGE_MASK; | |
334 | npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT; | |
e9601828 | 335 | for (i = 0; copy_len && i < npages; i++) { |
bd7ea31b | 336 | curlen = PAGE_SIZE - page_base; |
e9601828 TT |
337 | if (curlen > copy_len) |
338 | curlen = copy_len; | |
339 | dprintk("RPC: %s: page %d destp 0x%p len %d curlen %d\n", | |
340 | __func__, i, destp, copy_len, curlen); | |
b8541786 | 341 | srcp = kmap_atomic(ppages[i]); |
bd7ea31b | 342 | memcpy(destp, srcp+page_base, curlen); |
b8541786 | 343 | kunmap_atomic(srcp); |
e9601828 TT |
344 | rqst->rq_svec[0].iov_len += curlen; |
345 | destp += curlen; | |
346 | copy_len -= curlen; | |
bd7ea31b | 347 | page_base = 0; |
e9601828 | 348 | } |
e9601828 | 349 | /* header now contains entire send message */ |
e9601828 TT |
350 | } |
351 | ||
352 | /* | |
353 | * Marshal a request: the primary job of this routine is to choose | |
354 | * the transfer modes. See comments below. | |
355 | * | |
356 | * Uses multiple RDMA IOVs for a request: | |
357 | * [0] -- RPC RDMA header, which uses memory from the *start* of the | |
358 | * preregistered buffer that already holds the RPC data in | |
359 | * its middle. | |
360 | * [1] -- the RPC header/data, marshaled by RPC and the NFS protocol. | |
361 | * [2] -- optional padding. | |
362 | * [3] -- if padded, header only in [1] and data here. | |
c93c6223 CL |
363 | * |
364 | * Returns zero on success, otherwise a negative errno. | |
e9601828 TT |
365 | */ |
366 | ||
367 | int | |
368 | rpcrdma_marshal_req(struct rpc_rqst *rqst) | |
369 | { | |
a4f0835c | 370 | struct rpc_xprt *xprt = rqst->rq_xprt; |
e9601828 TT |
371 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); |
372 | struct rpcrdma_req *req = rpcr_to_rdmar(rqst); | |
373 | char *base; | |
b3221d6a | 374 | size_t rpclen; |
c93c6223 | 375 | ssize_t hdrlen; |
e2377945 | 376 | enum rpcrdma_chunktype rtype, wtype; |
e9601828 TT |
377 | struct rpcrdma_msg *headerp; |
378 | ||
379 | /* | |
380 | * rpclen gets amount of data in first buffer, which is the | |
381 | * pre-registered buffer. | |
382 | */ | |
383 | base = rqst->rq_svec[0].iov_base; | |
384 | rpclen = rqst->rq_svec[0].iov_len; | |
385 | ||
85275c87 | 386 | headerp = rdmab_to_msg(req->rl_rdmabuf); |
284f4902 | 387 | /* don't byte-swap XID, it's already done in request */ |
e9601828 | 388 | headerp->rm_xid = rqst->rq_xid; |
284f4902 CL |
389 | headerp->rm_vers = rpcrdma_version; |
390 | headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_max_requests); | |
391 | headerp->rm_type = rdma_msg; | |
e9601828 TT |
392 | |
393 | /* | |
394 | * Chunks needed for results? | |
395 | * | |
396 | * o If the expected result is under the inline threshold, all ops | |
397 | * return as inline (but see later). | |
398 | * o Large non-read ops return as a single reply chunk. | |
399 | * o Large read ops return data as write chunk(s), header as inline. | |
400 | * | |
401 | * Note: the NFS code sending down multiple result segments implies | |
402 | * the op is one of read, readdir[plus], readlink or NFSv4 getacl. | |
403 | */ | |
404 | ||
405 | /* | |
406 | * This code can handle read chunks, write chunks OR reply | |
407 | * chunks -- only one type. If the request is too big to fit | |
408 | * inline, then we will choose read chunks. If the request is | |
409 | * a READ, then use write chunks to separate the file data | |
410 | * into pages; otherwise use reply chunks. | |
411 | */ | |
412 | if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst)) | |
e2377945 | 413 | wtype = rpcrdma_noch; |
e9601828 | 414 | else if (rqst->rq_rcv_buf.page_len == 0) |
e2377945 | 415 | wtype = rpcrdma_replych; |
e9601828 | 416 | else if (rqst->rq_rcv_buf.flags & XDRBUF_READ) |
e2377945 | 417 | wtype = rpcrdma_writech; |
e9601828 | 418 | else |
e2377945 | 419 | wtype = rpcrdma_replych; |
e9601828 TT |
420 | |
421 | /* | |
422 | * Chunks needed for arguments? | |
423 | * | |
424 | * o If the total request is under the inline threshold, all ops | |
425 | * are sent as inline. | |
426 | * o Large non-write ops are sent with the entire message as a | |
427 | * single read chunk (protocol 0-position special case). | |
428 | * o Large write ops transmit data as read chunk(s), header as | |
429 | * inline. | |
430 | * | |
431 | * Note: the NFS code sending down multiple argument segments | |
432 | * implies the op is a write. | |
433 | * TBD check NFSv4 setacl | |
434 | */ | |
435 | if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst)) | |
e2377945 | 436 | rtype = rpcrdma_noch; |
e9601828 | 437 | else if (rqst->rq_snd_buf.page_len == 0) |
e2377945 | 438 | rtype = rpcrdma_areadch; |
e9601828 | 439 | else |
e2377945 | 440 | rtype = rpcrdma_readch; |
e9601828 TT |
441 | |
442 | /* The following simplification is not true forever */ | |
e2377945 CL |
443 | if (rtype != rpcrdma_noch && wtype == rpcrdma_replych) |
444 | wtype = rpcrdma_noch; | |
445 | if (rtype != rpcrdma_noch && wtype != rpcrdma_noch) { | |
c93c6223 CL |
446 | dprintk("RPC: %s: cannot marshal multiple chunk lists\n", |
447 | __func__); | |
448 | return -EIO; | |
449 | } | |
e9601828 | 450 | |
f2846481 | 451 | hdrlen = RPCRDMA_HDRLEN_MIN; |
e9601828 TT |
452 | |
453 | /* | |
454 | * Pull up any extra send data into the preregistered buffer. | |
455 | * When padding is in use and applies to the transfer, insert | |
456 | * it and change the message type. | |
457 | */ | |
e2377945 | 458 | if (rtype == rpcrdma_noch) { |
e9601828 | 459 | |
b3221d6a CL |
460 | rpcrdma_inline_pullup(rqst); |
461 | ||
462 | headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero; | |
463 | headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero; | |
464 | headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero; | |
465 | /* new length after pullup */ | |
466 | rpclen = rqst->rq_svec[0].iov_len; | |
467 | /* Currently we try to not actually use read inline. | |
468 | * Reply chunks have the desirable property that | |
469 | * they land, packed, directly in the target buffers | |
470 | * without headers, so they require no fixup. The | |
471 | * additional RDMA Write op sends the same amount | |
472 | * of data, streams on-the-wire and adds no overhead | |
473 | * on receive. Therefore, we request a reply chunk | |
474 | * for non-writes wherever feasible and efficient. | |
475 | */ | |
476 | if (wtype == rpcrdma_noch) | |
477 | wtype = rpcrdma_replych; | |
e9601828 TT |
478 | } |
479 | ||
e2377945 CL |
480 | if (rtype != rpcrdma_noch) { |
481 | hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_snd_buf, | |
482 | headerp, rtype); | |
483 | wtype = rtype; /* simplify dprintk */ | |
484 | ||
485 | } else if (wtype != rpcrdma_noch) { | |
486 | hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_rcv_buf, | |
487 | headerp, wtype); | |
488 | } | |
c93c6223 CL |
489 | if (hdrlen < 0) |
490 | return hdrlen; | |
e9601828 | 491 | |
b3221d6a | 492 | dprintk("RPC: %s: %s: hdrlen %zd rpclen %zd" |
5f37d561 | 493 | " headerp 0x%p base 0x%p lkey 0x%x\n", |
b3221d6a | 494 | __func__, transfertypes[wtype], hdrlen, rpclen, |
85275c87 | 495 | headerp, base, rdmab_lkey(req->rl_rdmabuf)); |
e9601828 TT |
496 | |
497 | /* | |
498 | * initialize send_iov's - normally only two: rdma chunk header and | |
499 | * single preregistered RPC header buffer, but if padding is present, | |
500 | * then use a preregistered (and zeroed) pad buffer between the RPC | |
501 | * header and any write data. In all non-rdma cases, any following | |
502 | * data has been copied into the RPC header buffer. | |
503 | */ | |
85275c87 | 504 | req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf); |
e9601828 | 505 | req->rl_send_iov[0].length = hdrlen; |
85275c87 | 506 | req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf); |
e9601828 | 507 | |
0ca77dc3 | 508 | req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf); |
e9601828 | 509 | req->rl_send_iov[1].length = rpclen; |
0ca77dc3 | 510 | req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf); |
e9601828 TT |
511 | |
512 | req->rl_niovs = 2; | |
e9601828 TT |
513 | return 0; |
514 | } | |
515 | ||
516 | /* | |
517 | * Chase down a received write or reply chunklist to get length | |
518 | * RDMA'd by server. See map at rpcrdma_create_chunks()! :-) | |
519 | */ | |
520 | static int | |
d4b37ff7 | 521 | rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp) |
e9601828 TT |
522 | { |
523 | unsigned int i, total_len; | |
524 | struct rpcrdma_write_chunk *cur_wchunk; | |
6b1184cd | 525 | char *base = (char *)rdmab_to_msg(rep->rr_rdmabuf); |
e9601828 | 526 | |
284f4902 | 527 | i = be32_to_cpu(**iptrp); |
e9601828 TT |
528 | if (i > max) |
529 | return -1; | |
530 | cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1); | |
531 | total_len = 0; | |
532 | while (i--) { | |
533 | struct rpcrdma_segment *seg = &cur_wchunk->wc_target; | |
534 | ifdebug(FACILITY) { | |
535 | u64 off; | |
2d8a9726 | 536 | xdr_decode_hyper((__be32 *)&seg->rs_offset, &off); |
e9601828 TT |
537 | dprintk("RPC: %s: chunk %d@0x%llx:0x%x\n", |
538 | __func__, | |
284f4902 | 539 | be32_to_cpu(seg->rs_length), |
e08a132b | 540 | (unsigned long long)off, |
284f4902 | 541 | be32_to_cpu(seg->rs_handle)); |
e9601828 | 542 | } |
284f4902 | 543 | total_len += be32_to_cpu(seg->rs_length); |
e9601828 TT |
544 | ++cur_wchunk; |
545 | } | |
546 | /* check and adjust for properly terminated write chunk */ | |
547 | if (wrchunk) { | |
2d8a9726 | 548 | __be32 *w = (__be32 *) cur_wchunk; |
e9601828 TT |
549 | if (*w++ != xdr_zero) |
550 | return -1; | |
551 | cur_wchunk = (struct rpcrdma_write_chunk *) w; | |
552 | } | |
6b1184cd | 553 | if ((char *)cur_wchunk > base + rep->rr_len) |
e9601828 TT |
554 | return -1; |
555 | ||
2d8a9726 | 556 | *iptrp = (__be32 *) cur_wchunk; |
e9601828 TT |
557 | return total_len; |
558 | } | |
559 | ||
560 | /* | |
561 | * Scatter inline received data back into provided iov's. | |
562 | */ | |
563 | static void | |
9191ca3b | 564 | rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad) |
e9601828 TT |
565 | { |
566 | int i, npages, curlen, olen; | |
567 | char *destp; | |
bd7ea31b TT |
568 | struct page **ppages; |
569 | int page_base; | |
e9601828 TT |
570 | |
571 | curlen = rqst->rq_rcv_buf.head[0].iov_len; | |
572 | if (curlen > copy_len) { /* write chunk header fixup */ | |
573 | curlen = copy_len; | |
574 | rqst->rq_rcv_buf.head[0].iov_len = curlen; | |
575 | } | |
576 | ||
577 | dprintk("RPC: %s: srcp 0x%p len %d hdrlen %d\n", | |
578 | __func__, srcp, copy_len, curlen); | |
579 | ||
580 | /* Shift pointer for first receive segment only */ | |
581 | rqst->rq_rcv_buf.head[0].iov_base = srcp; | |
582 | srcp += curlen; | |
583 | copy_len -= curlen; | |
584 | ||
585 | olen = copy_len; | |
586 | i = 0; | |
587 | rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen; | |
bd7ea31b TT |
588 | page_base = rqst->rq_rcv_buf.page_base; |
589 | ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT); | |
590 | page_base &= ~PAGE_MASK; | |
591 | ||
e9601828 | 592 | if (copy_len && rqst->rq_rcv_buf.page_len) { |
bd7ea31b | 593 | npages = PAGE_ALIGN(page_base + |
e9601828 TT |
594 | rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT; |
595 | for (; i < npages; i++) { | |
bd7ea31b | 596 | curlen = PAGE_SIZE - page_base; |
e9601828 TT |
597 | if (curlen > copy_len) |
598 | curlen = copy_len; | |
599 | dprintk("RPC: %s: page %d" | |
600 | " srcp 0x%p len %d curlen %d\n", | |
601 | __func__, i, srcp, copy_len, curlen); | |
b8541786 | 602 | destp = kmap_atomic(ppages[i]); |
bd7ea31b TT |
603 | memcpy(destp + page_base, srcp, curlen); |
604 | flush_dcache_page(ppages[i]); | |
b8541786 | 605 | kunmap_atomic(destp); |
e9601828 TT |
606 | srcp += curlen; |
607 | copy_len -= curlen; | |
608 | if (copy_len == 0) | |
609 | break; | |
bd7ea31b | 610 | page_base = 0; |
e9601828 | 611 | } |
2b7bbc96 | 612 | } |
e9601828 TT |
613 | |
614 | if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) { | |
615 | curlen = copy_len; | |
616 | if (curlen > rqst->rq_rcv_buf.tail[0].iov_len) | |
617 | curlen = rqst->rq_rcv_buf.tail[0].iov_len; | |
618 | if (rqst->rq_rcv_buf.tail[0].iov_base != srcp) | |
b38ab40a | 619 | memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen); |
e9601828 TT |
620 | dprintk("RPC: %s: tail srcp 0x%p len %d curlen %d\n", |
621 | __func__, srcp, copy_len, curlen); | |
622 | rqst->rq_rcv_buf.tail[0].iov_len = curlen; | |
623 | copy_len -= curlen; ++i; | |
624 | } else | |
625 | rqst->rq_rcv_buf.tail[0].iov_len = 0; | |
626 | ||
9191ca3b TT |
627 | if (pad) { |
628 | /* implicit padding on terminal chunk */ | |
629 | unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base; | |
630 | while (pad--) | |
631 | p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0; | |
632 | } | |
633 | ||
e9601828 TT |
634 | if (copy_len) |
635 | dprintk("RPC: %s: %d bytes in" | |
636 | " %d extra segments (%d lost)\n", | |
637 | __func__, olen, i, copy_len); | |
638 | ||
639 | /* TBD avoid a warning from call_decode() */ | |
640 | rqst->rq_private_buf = rqst->rq_rcv_buf; | |
641 | } | |
642 | ||
e9601828 | 643 | void |
254f91e2 | 644 | rpcrdma_connect_worker(struct work_struct *work) |
e9601828 | 645 | { |
254f91e2 CL |
646 | struct rpcrdma_ep *ep = |
647 | container_of(work, struct rpcrdma_ep, rep_connect_worker.work); | |
afadc468 CL |
648 | struct rpcrdma_xprt *r_xprt = |
649 | container_of(ep, struct rpcrdma_xprt, rx_ep); | |
650 | struct rpc_xprt *xprt = &r_xprt->rx_xprt; | |
e9601828 TT |
651 | |
652 | spin_lock_bh(&xprt->transport_lock); | |
575448bd TT |
653 | if (++xprt->connect_cookie == 0) /* maintain a reserved value */ |
654 | ++xprt->connect_cookie; | |
e9601828 TT |
655 | if (ep->rep_connected > 0) { |
656 | if (!xprt_test_and_set_connected(xprt)) | |
657 | xprt_wake_pending_tasks(xprt, 0); | |
658 | } else { | |
659 | if (xprt_test_and_clear_connected(xprt)) | |
926449ba | 660 | xprt_wake_pending_tasks(xprt, -ENOTCONN); |
e9601828 TT |
661 | } |
662 | spin_unlock_bh(&xprt->transport_lock); | |
663 | } | |
664 | ||
254f91e2 CL |
665 | /* |
666 | * This function is called when an async event is posted to | |
667 | * the connection which changes the connection state. All it | |
668 | * does at this point is mark the connection up/down, the rpc | |
669 | * timers do the rest. | |
670 | */ | |
671 | void | |
672 | rpcrdma_conn_func(struct rpcrdma_ep *ep) | |
673 | { | |
674 | schedule_delayed_work(&ep->rep_connect_worker, 0); | |
675 | } | |
676 | ||
e9601828 TT |
677 | /* |
678 | * Called as a tasklet to do req/reply match and complete a request | |
679 | * Errors must result in the RPC task either being awakened, or | |
680 | * allowed to timeout, to discover the errors at that time. | |
681 | */ | |
682 | void | |
683 | rpcrdma_reply_handler(struct rpcrdma_rep *rep) | |
684 | { | |
685 | struct rpcrdma_msg *headerp; | |
686 | struct rpcrdma_req *req; | |
687 | struct rpc_rqst *rqst; | |
fed171b3 CL |
688 | struct rpcrdma_xprt *r_xprt = rep->rr_rxprt; |
689 | struct rpc_xprt *xprt = &r_xprt->rx_xprt; | |
2d8a9726 | 690 | __be32 *iptr; |
9b1dcbc8 | 691 | int rdmalen, status; |
e7ce710a | 692 | unsigned long cwnd; |
9b1dcbc8 | 693 | u32 credits; |
e9601828 TT |
694 | |
695 | /* Check status. If bad, signal disconnect and return rep to pool */ | |
696 | if (rep->rr_len == ~0U) { | |
697 | rpcrdma_recv_buffer_put(rep); | |
698 | if (r_xprt->rx_ep.rep_connected == 1) { | |
699 | r_xprt->rx_ep.rep_connected = -EIO; | |
700 | rpcrdma_conn_func(&r_xprt->rx_ep); | |
701 | } | |
702 | return; | |
703 | } | |
f2846481 | 704 | if (rep->rr_len < RPCRDMA_HDRLEN_MIN) { |
e9601828 TT |
705 | dprintk("RPC: %s: short/invalid reply\n", __func__); |
706 | goto repost; | |
707 | } | |
6b1184cd | 708 | headerp = rdmab_to_msg(rep->rr_rdmabuf); |
284f4902 | 709 | if (headerp->rm_vers != rpcrdma_version) { |
e9601828 | 710 | dprintk("RPC: %s: invalid version %d\n", |
284f4902 | 711 | __func__, be32_to_cpu(headerp->rm_vers)); |
e9601828 TT |
712 | goto repost; |
713 | } | |
714 | ||
715 | /* Get XID and try for a match. */ | |
716 | spin_lock(&xprt->transport_lock); | |
717 | rqst = xprt_lookup_rqst(xprt, headerp->rm_xid); | |
718 | if (rqst == NULL) { | |
719 | spin_unlock(&xprt->transport_lock); | |
720 | dprintk("RPC: %s: reply 0x%p failed " | |
721 | "to match any request xid 0x%08x len %d\n", | |
052151a9 CL |
722 | __func__, rep, be32_to_cpu(headerp->rm_xid), |
723 | rep->rr_len); | |
e9601828 TT |
724 | repost: |
725 | r_xprt->rx_stats.bad_reply_count++; | |
e9601828 TT |
726 | if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep)) |
727 | rpcrdma_recv_buffer_put(rep); | |
728 | ||
729 | return; | |
730 | } | |
731 | ||
732 | /* get request object */ | |
733 | req = rpcr_to_rdmar(rqst); | |
4a6862b3 TT |
734 | if (req->rl_reply) { |
735 | spin_unlock(&xprt->transport_lock); | |
736 | dprintk("RPC: %s: duplicate reply 0x%p to RPC " | |
737 | "request 0x%p: xid 0x%08x\n", __func__, rep, req, | |
052151a9 | 738 | be32_to_cpu(headerp->rm_xid)); |
4a6862b3 TT |
739 | goto repost; |
740 | } | |
e9601828 TT |
741 | |
742 | dprintk("RPC: %s: reply 0x%p completes request 0x%p\n" | |
743 | " RPC request 0x%p xid 0x%08x\n", | |
052151a9 CL |
744 | __func__, rep, req, rqst, |
745 | be32_to_cpu(headerp->rm_xid)); | |
e9601828 | 746 | |
e9601828 TT |
747 | /* from here on, the reply is no longer an orphan */ |
748 | req->rl_reply = rep; | |
18906972 | 749 | xprt->reestablish_timeout = 0; |
e9601828 TT |
750 | |
751 | /* check for expected message types */ | |
752 | /* The order of some of these tests is important. */ | |
753 | switch (headerp->rm_type) { | |
284f4902 | 754 | case rdma_msg: |
e9601828 TT |
755 | /* never expect read chunks */ |
756 | /* never expect reply chunks (two ways to check) */ | |
757 | /* never expect write chunks without having offered RDMA */ | |
758 | if (headerp->rm_body.rm_chunks[0] != xdr_zero || | |
759 | (headerp->rm_body.rm_chunks[1] == xdr_zero && | |
760 | headerp->rm_body.rm_chunks[2] != xdr_zero) || | |
761 | (headerp->rm_body.rm_chunks[1] != xdr_zero && | |
762 | req->rl_nchunks == 0)) | |
763 | goto badheader; | |
764 | if (headerp->rm_body.rm_chunks[1] != xdr_zero) { | |
765 | /* count any expected write chunks in read reply */ | |
766 | /* start at write chunk array count */ | |
767 | iptr = &headerp->rm_body.rm_chunks[2]; | |
768 | rdmalen = rpcrdma_count_chunks(rep, | |
769 | req->rl_nchunks, 1, &iptr); | |
770 | /* check for validity, and no reply chunk after */ | |
771 | if (rdmalen < 0 || *iptr++ != xdr_zero) | |
772 | goto badheader; | |
773 | rep->rr_len -= | |
774 | ((unsigned char *)iptr - (unsigned char *)headerp); | |
775 | status = rep->rr_len + rdmalen; | |
776 | r_xprt->rx_stats.total_rdma_reply += rdmalen; | |
9191ca3b TT |
777 | /* special case - last chunk may omit padding */ |
778 | if (rdmalen &= 3) { | |
779 | rdmalen = 4 - rdmalen; | |
780 | status += rdmalen; | |
781 | } | |
e9601828 TT |
782 | } else { |
783 | /* else ordinary inline */ | |
9191ca3b | 784 | rdmalen = 0; |
f2846481 CL |
785 | iptr = (__be32 *)((unsigned char *)headerp + |
786 | RPCRDMA_HDRLEN_MIN); | |
787 | rep->rr_len -= RPCRDMA_HDRLEN_MIN; | |
e9601828 TT |
788 | status = rep->rr_len; |
789 | } | |
790 | /* Fix up the rpc results for upper layer */ | |
9191ca3b | 791 | rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen); |
e9601828 TT |
792 | break; |
793 | ||
284f4902 | 794 | case rdma_nomsg: |
e9601828 TT |
795 | /* never expect read or write chunks, always reply chunks */ |
796 | if (headerp->rm_body.rm_chunks[0] != xdr_zero || | |
797 | headerp->rm_body.rm_chunks[1] != xdr_zero || | |
798 | headerp->rm_body.rm_chunks[2] != xdr_one || | |
799 | req->rl_nchunks == 0) | |
800 | goto badheader; | |
f2846481 CL |
801 | iptr = (__be32 *)((unsigned char *)headerp + |
802 | RPCRDMA_HDRLEN_MIN); | |
e9601828 TT |
803 | rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr); |
804 | if (rdmalen < 0) | |
805 | goto badheader; | |
806 | r_xprt->rx_stats.total_rdma_reply += rdmalen; | |
807 | /* Reply chunk buffer already is the reply vector - no fixup. */ | |
808 | status = rdmalen; | |
809 | break; | |
810 | ||
811 | badheader: | |
812 | default: | |
813 | dprintk("%s: invalid rpcrdma reply header (type %d):" | |
814 | " chunks[012] == %d %d %d" | |
815 | " expected chunks <= %d\n", | |
284f4902 | 816 | __func__, be32_to_cpu(headerp->rm_type), |
e9601828 TT |
817 | headerp->rm_body.rm_chunks[0], |
818 | headerp->rm_body.rm_chunks[1], | |
819 | headerp->rm_body.rm_chunks[2], | |
820 | req->rl_nchunks); | |
821 | status = -EIO; | |
822 | r_xprt->rx_stats.bad_reply_count++; | |
823 | break; | |
824 | } | |
825 | ||
eba8ff66 CL |
826 | credits = be32_to_cpu(headerp->rm_credit); |
827 | if (credits == 0) | |
828 | credits = 1; /* don't deadlock */ | |
829 | else if (credits > r_xprt->rx_buf.rb_max_requests) | |
830 | credits = r_xprt->rx_buf.rb_max_requests; | |
831 | ||
e7ce710a | 832 | cwnd = xprt->cwnd; |
eba8ff66 | 833 | xprt->cwnd = credits << RPC_CWNDSHIFT; |
e7ce710a CL |
834 | if (xprt->cwnd > cwnd) |
835 | xprt_release_rqst_cong(rqst->rq_task); | |
836 | ||
e9601828 TT |
837 | dprintk("RPC: %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n", |
838 | __func__, xprt, rqst, status); | |
839 | xprt_complete_rqst(rqst->rq_task, status); | |
840 | spin_unlock(&xprt->transport_lock); | |
841 | } |