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 | ||
f895b252 | 56 | #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) |
e9601828 TT |
57 | static const char transfertypes[][12] = { |
58 | "pure inline", /* no chunks */ | |
59 | " read chunk", /* some argument via rdma read */ | |
60 | "*read chunk", /* entire request via rdma read */ | |
61 | "write chunk", /* some result via rdma write */ | |
62 | "reply chunk" /* entire reply via rdma write */ | |
63 | }; | |
64 | #endif | |
65 | ||
66 | /* | |
67 | * Chunk assembly from upper layer xdr_buf. | |
68 | * | |
69 | * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk | |
70 | * elements. Segments are then coalesced when registered, if possible | |
71 | * within the selected memreg mode. | |
c93c6223 CL |
72 | * |
73 | * Returns positive number of segments converted, or a negative errno. | |
e9601828 TT |
74 | */ |
75 | ||
76 | static int | |
2a428b2b | 77 | rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos, |
e9601828 TT |
78 | enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs) |
79 | { | |
80 | int len, n = 0, p; | |
bd7ea31b TT |
81 | int page_base; |
82 | struct page **ppages; | |
e9601828 TT |
83 | |
84 | if (pos == 0 && xdrbuf->head[0].iov_len) { | |
85 | seg[n].mr_page = NULL; | |
86 | seg[n].mr_offset = xdrbuf->head[0].iov_base; | |
87 | seg[n].mr_len = xdrbuf->head[0].iov_len; | |
e9601828 TT |
88 | ++n; |
89 | } | |
90 | ||
bd7ea31b TT |
91 | len = xdrbuf->page_len; |
92 | ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT); | |
93 | page_base = xdrbuf->page_base & ~PAGE_MASK; | |
94 | p = 0; | |
95 | while (len && n < nsegs) { | |
196c6998 SM |
96 | if (!ppages[p]) { |
97 | /* alloc the pagelist for receiving buffer */ | |
98 | ppages[p] = alloc_page(GFP_ATOMIC); | |
99 | if (!ppages[p]) | |
c93c6223 | 100 | return -ENOMEM; |
196c6998 | 101 | } |
bd7ea31b TT |
102 | seg[n].mr_page = ppages[p]; |
103 | seg[n].mr_offset = (void *)(unsigned long) page_base; | |
104 | seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len); | |
c93c6223 CL |
105 | if (seg[n].mr_len > PAGE_SIZE) |
106 | return -EIO; | |
bd7ea31b | 107 | len -= seg[n].mr_len; |
e9601828 | 108 | ++n; |
bd7ea31b TT |
109 | ++p; |
110 | page_base = 0; /* page offset only applies to first page */ | |
e9601828 TT |
111 | } |
112 | ||
bd7ea31b TT |
113 | /* Message overflows the seg array */ |
114 | if (len && n == nsegs) | |
c93c6223 | 115 | return -EIO; |
bd7ea31b | 116 | |
50e1092b | 117 | if (xdrbuf->tail[0].iov_len) { |
9191ca3b TT |
118 | /* the rpcrdma protocol allows us to omit any trailing |
119 | * xdr pad bytes, saving the server an RDMA operation. */ | |
120 | if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize) | |
121 | return n; | |
e9601828 | 122 | if (n == nsegs) |
bd7ea31b | 123 | /* Tail remains, but we're out of segments */ |
c93c6223 | 124 | return -EIO; |
e9601828 TT |
125 | seg[n].mr_page = NULL; |
126 | seg[n].mr_offset = xdrbuf->tail[0].iov_base; | |
127 | seg[n].mr_len = xdrbuf->tail[0].iov_len; | |
e9601828 TT |
128 | ++n; |
129 | } | |
130 | ||
e9601828 TT |
131 | return n; |
132 | } | |
133 | ||
134 | /* | |
135 | * Create read/write chunk lists, and reply chunks, for RDMA | |
136 | * | |
137 | * Assume check against THRESHOLD has been done, and chunks are required. | |
138 | * Assume only encoding one list entry for read|write chunks. The NFSv3 | |
139 | * protocol is simple enough to allow this as it only has a single "bulk | |
140 | * result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The | |
141 | * RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.) | |
142 | * | |
143 | * When used for a single reply chunk (which is a special write | |
144 | * chunk used for the entire reply, rather than just the data), it | |
145 | * is used primarily for READDIR and READLINK which would otherwise | |
146 | * be severely size-limited by a small rdma inline read max. The server | |
147 | * response will come back as an RDMA Write, followed by a message | |
148 | * of type RDMA_NOMSG carrying the xid and length. As a result, reply | |
149 | * chunks do not provide data alignment, however they do not require | |
150 | * "fixup" (moving the response to the upper layer buffer) either. | |
151 | * | |
152 | * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64): | |
153 | * | |
154 | * Read chunklist (a linked list): | |
155 | * N elements, position P (same P for all chunks of same arg!): | |
156 | * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0 | |
157 | * | |
158 | * Write chunklist (a list of (one) counted array): | |
159 | * N elements: | |
160 | * 1 - N - HLOO - HLOO - ... - HLOO - 0 | |
161 | * | |
162 | * Reply chunk (a counted array): | |
163 | * N elements: | |
164 | * 1 - N - HLOO - HLOO - ... - HLOO | |
c93c6223 CL |
165 | * |
166 | * Returns positive RPC/RDMA header size, or negative errno. | |
e9601828 TT |
167 | */ |
168 | ||
c93c6223 | 169 | static ssize_t |
e9601828 TT |
170 | rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target, |
171 | struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type) | |
172 | { | |
173 | struct rpcrdma_req *req = rpcr_to_rdmar(rqst); | |
a4f0835c | 174 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); |
c93c6223 | 175 | int n, nsegs, nchunks = 0; |
2a428b2b | 176 | unsigned int pos; |
e9601828 TT |
177 | struct rpcrdma_mr_seg *seg = req->rl_segments; |
178 | struct rpcrdma_read_chunk *cur_rchunk = NULL; | |
179 | struct rpcrdma_write_array *warray = NULL; | |
180 | struct rpcrdma_write_chunk *cur_wchunk = NULL; | |
2d8a9726 | 181 | __be32 *iptr = headerp->rm_body.rm_chunks; |
e9601828 TT |
182 | |
183 | if (type == rpcrdma_readch || type == rpcrdma_areadch) { | |
184 | /* a read chunk - server will RDMA Read our memory */ | |
185 | cur_rchunk = (struct rpcrdma_read_chunk *) iptr; | |
186 | } else { | |
187 | /* a write or reply chunk - server will RDMA Write our memory */ | |
188 | *iptr++ = xdr_zero; /* encode a NULL read chunk list */ | |
189 | if (type == rpcrdma_replych) | |
190 | *iptr++ = xdr_zero; /* a NULL write chunk list */ | |
191 | warray = (struct rpcrdma_write_array *) iptr; | |
192 | cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1); | |
193 | } | |
194 | ||
195 | if (type == rpcrdma_replych || type == rpcrdma_areadch) | |
196 | pos = 0; | |
197 | else | |
198 | pos = target->head[0].iov_len; | |
199 | ||
200 | nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS); | |
c93c6223 CL |
201 | if (nsegs < 0) |
202 | return nsegs; | |
e9601828 TT |
203 | |
204 | do { | |
c93c6223 | 205 | n = rpcrdma_register_external(seg, nsegs, |
e9601828 TT |
206 | cur_wchunk != NULL, r_xprt); |
207 | if (n <= 0) | |
208 | goto out; | |
209 | if (cur_rchunk) { /* read */ | |
210 | cur_rchunk->rc_discrim = xdr_one; | |
211 | /* all read chunks have the same "position" */ | |
284f4902 CL |
212 | cur_rchunk->rc_position = cpu_to_be32(pos); |
213 | cur_rchunk->rc_target.rs_handle = | |
214 | cpu_to_be32(seg->mr_rkey); | |
215 | cur_rchunk->rc_target.rs_length = | |
216 | cpu_to_be32(seg->mr_len); | |
e9601828 | 217 | xdr_encode_hyper( |
2d8a9726 | 218 | (__be32 *)&cur_rchunk->rc_target.rs_offset, |
e9601828 TT |
219 | seg->mr_base); |
220 | dprintk("RPC: %s: read chunk " | |
2a428b2b | 221 | "elem %d@0x%llx:0x%x pos %u (%s)\n", __func__, |
e08a132b SR |
222 | seg->mr_len, (unsigned long long)seg->mr_base, |
223 | seg->mr_rkey, pos, n < nsegs ? "more" : "last"); | |
e9601828 TT |
224 | cur_rchunk++; |
225 | r_xprt->rx_stats.read_chunk_count++; | |
226 | } else { /* write/reply */ | |
284f4902 CL |
227 | cur_wchunk->wc_target.rs_handle = |
228 | cpu_to_be32(seg->mr_rkey); | |
229 | cur_wchunk->wc_target.rs_length = | |
230 | cpu_to_be32(seg->mr_len); | |
e9601828 | 231 | xdr_encode_hyper( |
2d8a9726 | 232 | (__be32 *)&cur_wchunk->wc_target.rs_offset, |
e9601828 TT |
233 | seg->mr_base); |
234 | dprintk("RPC: %s: %s chunk " | |
235 | "elem %d@0x%llx:0x%x (%s)\n", __func__, | |
236 | (type == rpcrdma_replych) ? "reply" : "write", | |
e08a132b SR |
237 | seg->mr_len, (unsigned long long)seg->mr_base, |
238 | seg->mr_rkey, n < nsegs ? "more" : "last"); | |
e9601828 TT |
239 | cur_wchunk++; |
240 | if (type == rpcrdma_replych) | |
241 | r_xprt->rx_stats.reply_chunk_count++; | |
242 | else | |
243 | r_xprt->rx_stats.write_chunk_count++; | |
244 | r_xprt->rx_stats.total_rdma_request += seg->mr_len; | |
245 | } | |
246 | nchunks++; | |
247 | seg += n; | |
248 | nsegs -= n; | |
249 | } while (nsegs); | |
250 | ||
251 | /* success. all failures return above */ | |
252 | req->rl_nchunks = nchunks; | |
253 | ||
e9601828 TT |
254 | /* |
255 | * finish off header. If write, marshal discrim and nchunks. | |
256 | */ | |
257 | if (cur_rchunk) { | |
2d8a9726 | 258 | iptr = (__be32 *) cur_rchunk; |
e9601828 TT |
259 | *iptr++ = xdr_zero; /* finish the read chunk list */ |
260 | *iptr++ = xdr_zero; /* encode a NULL write chunk list */ | |
261 | *iptr++ = xdr_zero; /* encode a NULL reply chunk */ | |
262 | } else { | |
263 | warray->wc_discrim = xdr_one; | |
284f4902 | 264 | warray->wc_nchunks = cpu_to_be32(nchunks); |
2d8a9726 | 265 | iptr = (__be32 *) cur_wchunk; |
e9601828 TT |
266 | if (type == rpcrdma_writech) { |
267 | *iptr++ = xdr_zero; /* finish the write chunk list */ | |
268 | *iptr++ = xdr_zero; /* encode a NULL reply chunk */ | |
269 | } | |
270 | } | |
271 | ||
272 | /* | |
273 | * Return header size. | |
274 | */ | |
275 | return (unsigned char *)iptr - (unsigned char *)headerp; | |
276 | ||
277 | out: | |
539431a4 CL |
278 | if (r_xprt->rx_ia.ri_memreg_strategy != RPCRDMA_FRMR) { |
279 | for (pos = 0; nchunks--;) | |
280 | pos += rpcrdma_deregister_external( | |
281 | &req->rl_segments[pos], r_xprt); | |
282 | } | |
c93c6223 | 283 | return n; |
e9601828 TT |
284 | } |
285 | ||
6ab59945 CL |
286 | /* |
287 | * Marshal chunks. This routine returns the header length | |
288 | * consumed by marshaling. | |
289 | * | |
290 | * Returns positive RPC/RDMA header size, or negative errno. | |
291 | */ | |
292 | ||
293 | ssize_t | |
294 | rpcrdma_marshal_chunks(struct rpc_rqst *rqst, ssize_t result) | |
295 | { | |
296 | struct rpcrdma_req *req = rpcr_to_rdmar(rqst); | |
297 | struct rpcrdma_msg *headerp = (struct rpcrdma_msg *)req->rl_base; | |
298 | ||
299 | if (req->rl_rtype != rpcrdma_noch) | |
300 | result = rpcrdma_create_chunks(rqst, &rqst->rq_snd_buf, | |
301 | headerp, req->rl_rtype); | |
302 | else if (req->rl_wtype != rpcrdma_noch) | |
303 | result = rpcrdma_create_chunks(rqst, &rqst->rq_rcv_buf, | |
304 | headerp, req->rl_wtype); | |
305 | return result; | |
306 | } | |
307 | ||
e9601828 TT |
308 | /* |
309 | * Copy write data inline. | |
310 | * This function is used for "small" requests. Data which is passed | |
311 | * to RPC via iovecs (or page list) is copied directly into the | |
312 | * pre-registered memory buffer for this request. For small amounts | |
313 | * of data, this is efficient. The cutoff value is tunable. | |
314 | */ | |
315 | static int | |
316 | rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad) | |
317 | { | |
318 | int i, npages, curlen; | |
319 | int copy_len; | |
320 | unsigned char *srcp, *destp; | |
321 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); | |
bd7ea31b TT |
322 | int page_base; |
323 | struct page **ppages; | |
e9601828 TT |
324 | |
325 | destp = rqst->rq_svec[0].iov_base; | |
326 | curlen = rqst->rq_svec[0].iov_len; | |
327 | destp += curlen; | |
328 | /* | |
329 | * Do optional padding where it makes sense. Alignment of write | |
330 | * payload can help the server, if our setting is accurate. | |
331 | */ | |
332 | pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/); | |
333 | if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH) | |
334 | pad = 0; /* don't pad this request */ | |
335 | ||
336 | dprintk("RPC: %s: pad %d destp 0x%p len %d hdrlen %d\n", | |
337 | __func__, pad, destp, rqst->rq_slen, curlen); | |
338 | ||
339 | copy_len = rqst->rq_snd_buf.page_len; | |
b38ab40a TT |
340 | |
341 | if (rqst->rq_snd_buf.tail[0].iov_len) { | |
342 | curlen = rqst->rq_snd_buf.tail[0].iov_len; | |
343 | if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) { | |
344 | memmove(destp + copy_len, | |
345 | rqst->rq_snd_buf.tail[0].iov_base, curlen); | |
346 | r_xprt->rx_stats.pullup_copy_count += curlen; | |
347 | } | |
348 | dprintk("RPC: %s: tail destp 0x%p len %d\n", | |
349 | __func__, destp + copy_len, curlen); | |
350 | rqst->rq_svec[0].iov_len += curlen; | |
351 | } | |
e9601828 | 352 | r_xprt->rx_stats.pullup_copy_count += copy_len; |
bd7ea31b TT |
353 | |
354 | page_base = rqst->rq_snd_buf.page_base; | |
355 | ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT); | |
356 | page_base &= ~PAGE_MASK; | |
357 | npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT; | |
e9601828 | 358 | for (i = 0; copy_len && i < npages; i++) { |
bd7ea31b | 359 | curlen = PAGE_SIZE - page_base; |
e9601828 TT |
360 | if (curlen > copy_len) |
361 | curlen = copy_len; | |
362 | dprintk("RPC: %s: page %d destp 0x%p len %d curlen %d\n", | |
363 | __func__, i, destp, copy_len, curlen); | |
b8541786 | 364 | srcp = kmap_atomic(ppages[i]); |
bd7ea31b | 365 | memcpy(destp, srcp+page_base, curlen); |
b8541786 | 366 | kunmap_atomic(srcp); |
e9601828 TT |
367 | rqst->rq_svec[0].iov_len += curlen; |
368 | destp += curlen; | |
369 | copy_len -= curlen; | |
bd7ea31b | 370 | page_base = 0; |
e9601828 | 371 | } |
e9601828 TT |
372 | /* header now contains entire send message */ |
373 | return pad; | |
374 | } | |
375 | ||
376 | /* | |
377 | * Marshal a request: the primary job of this routine is to choose | |
378 | * the transfer modes. See comments below. | |
379 | * | |
380 | * Uses multiple RDMA IOVs for a request: | |
381 | * [0] -- RPC RDMA header, which uses memory from the *start* of the | |
382 | * preregistered buffer that already holds the RPC data in | |
383 | * its middle. | |
384 | * [1] -- the RPC header/data, marshaled by RPC and the NFS protocol. | |
385 | * [2] -- optional padding. | |
386 | * [3] -- if padded, header only in [1] and data here. | |
c93c6223 CL |
387 | * |
388 | * Returns zero on success, otherwise a negative errno. | |
e9601828 TT |
389 | */ |
390 | ||
391 | int | |
392 | rpcrdma_marshal_req(struct rpc_rqst *rqst) | |
393 | { | |
a4f0835c | 394 | struct rpc_xprt *xprt = rqst->rq_xprt; |
e9601828 TT |
395 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); |
396 | struct rpcrdma_req *req = rpcr_to_rdmar(rqst); | |
397 | char *base; | |
c93c6223 CL |
398 | size_t rpclen, padlen; |
399 | ssize_t hdrlen; | |
e9601828 TT |
400 | struct rpcrdma_msg *headerp; |
401 | ||
402 | /* | |
403 | * rpclen gets amount of data in first buffer, which is the | |
404 | * pre-registered buffer. | |
405 | */ | |
406 | base = rqst->rq_svec[0].iov_base; | |
407 | rpclen = rqst->rq_svec[0].iov_len; | |
408 | ||
409 | /* build RDMA header in private area at front */ | |
410 | headerp = (struct rpcrdma_msg *) req->rl_base; | |
284f4902 | 411 | /* don't byte-swap XID, it's already done in request */ |
e9601828 | 412 | headerp->rm_xid = rqst->rq_xid; |
284f4902 CL |
413 | headerp->rm_vers = rpcrdma_version; |
414 | headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_max_requests); | |
415 | headerp->rm_type = rdma_msg; | |
e9601828 TT |
416 | |
417 | /* | |
418 | * Chunks needed for results? | |
419 | * | |
420 | * o If the expected result is under the inline threshold, all ops | |
421 | * return as inline (but see later). | |
422 | * o Large non-read ops return as a single reply chunk. | |
423 | * o Large read ops return data as write chunk(s), header as inline. | |
424 | * | |
425 | * Note: the NFS code sending down multiple result segments implies | |
426 | * the op is one of read, readdir[plus], readlink or NFSv4 getacl. | |
427 | */ | |
428 | ||
429 | /* | |
430 | * This code can handle read chunks, write chunks OR reply | |
431 | * chunks -- only one type. If the request is too big to fit | |
432 | * inline, then we will choose read chunks. If the request is | |
433 | * a READ, then use write chunks to separate the file data | |
434 | * into pages; otherwise use reply chunks. | |
435 | */ | |
436 | if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst)) | |
6ab59945 | 437 | req->rl_wtype = rpcrdma_noch; |
e9601828 | 438 | else if (rqst->rq_rcv_buf.page_len == 0) |
6ab59945 | 439 | req->rl_wtype = rpcrdma_replych; |
e9601828 | 440 | else if (rqst->rq_rcv_buf.flags & XDRBUF_READ) |
6ab59945 | 441 | req->rl_wtype = rpcrdma_writech; |
e9601828 | 442 | else |
6ab59945 | 443 | req->rl_wtype = rpcrdma_replych; |
e9601828 TT |
444 | |
445 | /* | |
446 | * Chunks needed for arguments? | |
447 | * | |
448 | * o If the total request is under the inline threshold, all ops | |
449 | * are sent as inline. | |
450 | * o Large non-write ops are sent with the entire message as a | |
451 | * single read chunk (protocol 0-position special case). | |
452 | * o Large write ops transmit data as read chunk(s), header as | |
453 | * inline. | |
454 | * | |
455 | * Note: the NFS code sending down multiple argument segments | |
456 | * implies the op is a write. | |
457 | * TBD check NFSv4 setacl | |
458 | */ | |
459 | if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst)) | |
6ab59945 | 460 | req->rl_rtype = rpcrdma_noch; |
e9601828 | 461 | else if (rqst->rq_snd_buf.page_len == 0) |
6ab59945 | 462 | req->rl_rtype = rpcrdma_areadch; |
e9601828 | 463 | else |
6ab59945 | 464 | req->rl_rtype = rpcrdma_readch; |
e9601828 TT |
465 | |
466 | /* The following simplification is not true forever */ | |
6ab59945 CL |
467 | if (req->rl_rtype != rpcrdma_noch && req->rl_wtype == rpcrdma_replych) |
468 | req->rl_wtype = rpcrdma_noch; | |
469 | if (req->rl_rtype != rpcrdma_noch && req->rl_wtype != rpcrdma_noch) { | |
c93c6223 CL |
470 | dprintk("RPC: %s: cannot marshal multiple chunk lists\n", |
471 | __func__); | |
472 | return -EIO; | |
473 | } | |
e9601828 | 474 | |
e9601828 TT |
475 | hdrlen = 28; /*sizeof *headerp;*/ |
476 | padlen = 0; | |
477 | ||
478 | /* | |
479 | * Pull up any extra send data into the preregistered buffer. | |
480 | * When padding is in use and applies to the transfer, insert | |
481 | * it and change the message type. | |
482 | */ | |
6ab59945 | 483 | if (req->rl_rtype == rpcrdma_noch) { |
e9601828 TT |
484 | |
485 | padlen = rpcrdma_inline_pullup(rqst, | |
486 | RPCRDMA_INLINE_PAD_VALUE(rqst)); | |
487 | ||
488 | if (padlen) { | |
284f4902 | 489 | headerp->rm_type = rdma_msgp; |
e9601828 | 490 | headerp->rm_body.rm_padded.rm_align = |
284f4902 | 491 | cpu_to_be32(RPCRDMA_INLINE_PAD_VALUE(rqst)); |
e9601828 | 492 | headerp->rm_body.rm_padded.rm_thresh = |
284f4902 | 493 | cpu_to_be32(RPCRDMA_INLINE_PAD_THRESH); |
e9601828 TT |
494 | headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero; |
495 | headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero; | |
496 | headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero; | |
497 | hdrlen += 2 * sizeof(u32); /* extra words in padhdr */ | |
6ab59945 | 498 | if (req->rl_wtype != rpcrdma_noch) { |
c93c6223 CL |
499 | dprintk("RPC: %s: invalid chunk list\n", |
500 | __func__); | |
501 | return -EIO; | |
502 | } | |
e9601828 TT |
503 | } else { |
504 | headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero; | |
505 | headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero; | |
506 | headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero; | |
507 | /* new length after pullup */ | |
508 | rpclen = rqst->rq_svec[0].iov_len; | |
509 | /* | |
510 | * Currently we try to not actually use read inline. | |
511 | * Reply chunks have the desirable property that | |
512 | * they land, packed, directly in the target buffers | |
513 | * without headers, so they require no fixup. The | |
514 | * additional RDMA Write op sends the same amount | |
515 | * of data, streams on-the-wire and adds no overhead | |
516 | * on receive. Therefore, we request a reply chunk | |
517 | * for non-writes wherever feasible and efficient. | |
518 | */ | |
6ab59945 CL |
519 | if (req->rl_wtype == rpcrdma_noch) |
520 | req->rl_wtype = rpcrdma_replych; | |
e9601828 TT |
521 | } |
522 | } | |
523 | ||
6ab59945 | 524 | hdrlen = rpcrdma_marshal_chunks(rqst, hdrlen); |
c93c6223 CL |
525 | if (hdrlen < 0) |
526 | return hdrlen; | |
e9601828 | 527 | |
5f37d561 TT |
528 | dprintk("RPC: %s: %s: hdrlen %zd rpclen %zd padlen %zd" |
529 | " headerp 0x%p base 0x%p lkey 0x%x\n", | |
6ab59945 | 530 | __func__, transfertypes[req->rl_wtype], hdrlen, rpclen, padlen, |
e9601828 TT |
531 | headerp, base, req->rl_iov.lkey); |
532 | ||
533 | /* | |
534 | * initialize send_iov's - normally only two: rdma chunk header and | |
535 | * single preregistered RPC header buffer, but if padding is present, | |
536 | * then use a preregistered (and zeroed) pad buffer between the RPC | |
537 | * header and any write data. In all non-rdma cases, any following | |
538 | * data has been copied into the RPC header buffer. | |
539 | */ | |
540 | req->rl_send_iov[0].addr = req->rl_iov.addr; | |
541 | req->rl_send_iov[0].length = hdrlen; | |
542 | req->rl_send_iov[0].lkey = req->rl_iov.lkey; | |
543 | ||
544 | req->rl_send_iov[1].addr = req->rl_iov.addr + (base - req->rl_base); | |
545 | req->rl_send_iov[1].length = rpclen; | |
546 | req->rl_send_iov[1].lkey = req->rl_iov.lkey; | |
547 | ||
548 | req->rl_niovs = 2; | |
549 | ||
550 | if (padlen) { | |
551 | struct rpcrdma_ep *ep = &r_xprt->rx_ep; | |
552 | ||
553 | req->rl_send_iov[2].addr = ep->rep_pad.addr; | |
554 | req->rl_send_iov[2].length = padlen; | |
555 | req->rl_send_iov[2].lkey = ep->rep_pad.lkey; | |
556 | ||
557 | req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen; | |
558 | req->rl_send_iov[3].length = rqst->rq_slen - rpclen; | |
559 | req->rl_send_iov[3].lkey = req->rl_iov.lkey; | |
560 | ||
561 | req->rl_niovs = 4; | |
562 | } | |
563 | ||
564 | return 0; | |
565 | } | |
566 | ||
567 | /* | |
568 | * Chase down a received write or reply chunklist to get length | |
569 | * RDMA'd by server. See map at rpcrdma_create_chunks()! :-) | |
570 | */ | |
571 | static int | |
d4b37ff7 | 572 | rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp) |
e9601828 TT |
573 | { |
574 | unsigned int i, total_len; | |
575 | struct rpcrdma_write_chunk *cur_wchunk; | |
576 | ||
284f4902 | 577 | i = be32_to_cpu(**iptrp); |
e9601828 TT |
578 | if (i > max) |
579 | return -1; | |
580 | cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1); | |
581 | total_len = 0; | |
582 | while (i--) { | |
583 | struct rpcrdma_segment *seg = &cur_wchunk->wc_target; | |
584 | ifdebug(FACILITY) { | |
585 | u64 off; | |
2d8a9726 | 586 | xdr_decode_hyper((__be32 *)&seg->rs_offset, &off); |
e9601828 TT |
587 | dprintk("RPC: %s: chunk %d@0x%llx:0x%x\n", |
588 | __func__, | |
284f4902 | 589 | be32_to_cpu(seg->rs_length), |
e08a132b | 590 | (unsigned long long)off, |
284f4902 | 591 | be32_to_cpu(seg->rs_handle)); |
e9601828 | 592 | } |
284f4902 | 593 | total_len += be32_to_cpu(seg->rs_length); |
e9601828 TT |
594 | ++cur_wchunk; |
595 | } | |
596 | /* check and adjust for properly terminated write chunk */ | |
597 | if (wrchunk) { | |
2d8a9726 | 598 | __be32 *w = (__be32 *) cur_wchunk; |
e9601828 TT |
599 | if (*w++ != xdr_zero) |
600 | return -1; | |
601 | cur_wchunk = (struct rpcrdma_write_chunk *) w; | |
602 | } | |
603 | if ((char *) cur_wchunk > rep->rr_base + rep->rr_len) | |
604 | return -1; | |
605 | ||
2d8a9726 | 606 | *iptrp = (__be32 *) cur_wchunk; |
e9601828 TT |
607 | return total_len; |
608 | } | |
609 | ||
610 | /* | |
611 | * Scatter inline received data back into provided iov's. | |
612 | */ | |
613 | static void | |
9191ca3b | 614 | rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad) |
e9601828 TT |
615 | { |
616 | int i, npages, curlen, olen; | |
617 | char *destp; | |
bd7ea31b TT |
618 | struct page **ppages; |
619 | int page_base; | |
e9601828 TT |
620 | |
621 | curlen = rqst->rq_rcv_buf.head[0].iov_len; | |
622 | if (curlen > copy_len) { /* write chunk header fixup */ | |
623 | curlen = copy_len; | |
624 | rqst->rq_rcv_buf.head[0].iov_len = curlen; | |
625 | } | |
626 | ||
627 | dprintk("RPC: %s: srcp 0x%p len %d hdrlen %d\n", | |
628 | __func__, srcp, copy_len, curlen); | |
629 | ||
630 | /* Shift pointer for first receive segment only */ | |
631 | rqst->rq_rcv_buf.head[0].iov_base = srcp; | |
632 | srcp += curlen; | |
633 | copy_len -= curlen; | |
634 | ||
635 | olen = copy_len; | |
636 | i = 0; | |
637 | rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen; | |
bd7ea31b TT |
638 | page_base = rqst->rq_rcv_buf.page_base; |
639 | ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT); | |
640 | page_base &= ~PAGE_MASK; | |
641 | ||
e9601828 | 642 | if (copy_len && rqst->rq_rcv_buf.page_len) { |
bd7ea31b | 643 | npages = PAGE_ALIGN(page_base + |
e9601828 TT |
644 | rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT; |
645 | for (; i < npages; i++) { | |
bd7ea31b | 646 | curlen = PAGE_SIZE - page_base; |
e9601828 TT |
647 | if (curlen > copy_len) |
648 | curlen = copy_len; | |
649 | dprintk("RPC: %s: page %d" | |
650 | " srcp 0x%p len %d curlen %d\n", | |
651 | __func__, i, srcp, copy_len, curlen); | |
b8541786 | 652 | destp = kmap_atomic(ppages[i]); |
bd7ea31b TT |
653 | memcpy(destp + page_base, srcp, curlen); |
654 | flush_dcache_page(ppages[i]); | |
b8541786 | 655 | kunmap_atomic(destp); |
e9601828 TT |
656 | srcp += curlen; |
657 | copy_len -= curlen; | |
658 | if (copy_len == 0) | |
659 | break; | |
bd7ea31b | 660 | page_base = 0; |
e9601828 | 661 | } |
2b7bbc96 | 662 | } |
e9601828 TT |
663 | |
664 | if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) { | |
665 | curlen = copy_len; | |
666 | if (curlen > rqst->rq_rcv_buf.tail[0].iov_len) | |
667 | curlen = rqst->rq_rcv_buf.tail[0].iov_len; | |
668 | if (rqst->rq_rcv_buf.tail[0].iov_base != srcp) | |
b38ab40a | 669 | memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen); |
e9601828 TT |
670 | dprintk("RPC: %s: tail srcp 0x%p len %d curlen %d\n", |
671 | __func__, srcp, copy_len, curlen); | |
672 | rqst->rq_rcv_buf.tail[0].iov_len = curlen; | |
673 | copy_len -= curlen; ++i; | |
674 | } else | |
675 | rqst->rq_rcv_buf.tail[0].iov_len = 0; | |
676 | ||
9191ca3b TT |
677 | if (pad) { |
678 | /* implicit padding on terminal chunk */ | |
679 | unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base; | |
680 | while (pad--) | |
681 | p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0; | |
682 | } | |
683 | ||
e9601828 TT |
684 | if (copy_len) |
685 | dprintk("RPC: %s: %d bytes in" | |
686 | " %d extra segments (%d lost)\n", | |
687 | __func__, olen, i, copy_len); | |
688 | ||
689 | /* TBD avoid a warning from call_decode() */ | |
690 | rqst->rq_private_buf = rqst->rq_rcv_buf; | |
691 | } | |
692 | ||
e9601828 | 693 | void |
254f91e2 | 694 | rpcrdma_connect_worker(struct work_struct *work) |
e9601828 | 695 | { |
254f91e2 CL |
696 | struct rpcrdma_ep *ep = |
697 | container_of(work, struct rpcrdma_ep, rep_connect_worker.work); | |
e9601828 TT |
698 | struct rpc_xprt *xprt = ep->rep_xprt; |
699 | ||
700 | spin_lock_bh(&xprt->transport_lock); | |
575448bd TT |
701 | if (++xprt->connect_cookie == 0) /* maintain a reserved value */ |
702 | ++xprt->connect_cookie; | |
e9601828 TT |
703 | if (ep->rep_connected > 0) { |
704 | if (!xprt_test_and_set_connected(xprt)) | |
705 | xprt_wake_pending_tasks(xprt, 0); | |
706 | } else { | |
707 | if (xprt_test_and_clear_connected(xprt)) | |
926449ba | 708 | xprt_wake_pending_tasks(xprt, -ENOTCONN); |
e9601828 TT |
709 | } |
710 | spin_unlock_bh(&xprt->transport_lock); | |
711 | } | |
712 | ||
254f91e2 CL |
713 | /* |
714 | * This function is called when an async event is posted to | |
715 | * the connection which changes the connection state. All it | |
716 | * does at this point is mark the connection up/down, the rpc | |
717 | * timers do the rest. | |
718 | */ | |
719 | void | |
720 | rpcrdma_conn_func(struct rpcrdma_ep *ep) | |
721 | { | |
722 | schedule_delayed_work(&ep->rep_connect_worker, 0); | |
723 | } | |
724 | ||
e9601828 TT |
725 | /* |
726 | * Called as a tasklet to do req/reply match and complete a request | |
727 | * Errors must result in the RPC task either being awakened, or | |
728 | * allowed to timeout, to discover the errors at that time. | |
729 | */ | |
730 | void | |
731 | rpcrdma_reply_handler(struct rpcrdma_rep *rep) | |
732 | { | |
733 | struct rpcrdma_msg *headerp; | |
734 | struct rpcrdma_req *req; | |
735 | struct rpc_rqst *rqst; | |
736 | struct rpc_xprt *xprt = rep->rr_xprt; | |
737 | struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); | |
2d8a9726 | 738 | __be32 *iptr; |
b45ccfd2 | 739 | int rdmalen, status; |
e7ce710a | 740 | unsigned long cwnd; |
e9601828 TT |
741 | |
742 | /* Check status. If bad, signal disconnect and return rep to pool */ | |
743 | if (rep->rr_len == ~0U) { | |
744 | rpcrdma_recv_buffer_put(rep); | |
745 | if (r_xprt->rx_ep.rep_connected == 1) { | |
746 | r_xprt->rx_ep.rep_connected = -EIO; | |
747 | rpcrdma_conn_func(&r_xprt->rx_ep); | |
748 | } | |
749 | return; | |
750 | } | |
751 | if (rep->rr_len < 28) { | |
752 | dprintk("RPC: %s: short/invalid reply\n", __func__); | |
753 | goto repost; | |
754 | } | |
755 | headerp = (struct rpcrdma_msg *) rep->rr_base; | |
284f4902 | 756 | if (headerp->rm_vers != rpcrdma_version) { |
e9601828 | 757 | dprintk("RPC: %s: invalid version %d\n", |
284f4902 | 758 | __func__, be32_to_cpu(headerp->rm_vers)); |
e9601828 TT |
759 | goto repost; |
760 | } | |
761 | ||
762 | /* Get XID and try for a match. */ | |
763 | spin_lock(&xprt->transport_lock); | |
764 | rqst = xprt_lookup_rqst(xprt, headerp->rm_xid); | |
765 | if (rqst == NULL) { | |
766 | spin_unlock(&xprt->transport_lock); | |
767 | dprintk("RPC: %s: reply 0x%p failed " | |
768 | "to match any request xid 0x%08x len %d\n", | |
769 | __func__, rep, headerp->rm_xid, rep->rr_len); | |
770 | repost: | |
771 | r_xprt->rx_stats.bad_reply_count++; | |
772 | rep->rr_func = rpcrdma_reply_handler; | |
773 | if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep)) | |
774 | rpcrdma_recv_buffer_put(rep); | |
775 | ||
776 | return; | |
777 | } | |
778 | ||
779 | /* get request object */ | |
780 | req = rpcr_to_rdmar(rqst); | |
4a6862b3 TT |
781 | if (req->rl_reply) { |
782 | spin_unlock(&xprt->transport_lock); | |
783 | dprintk("RPC: %s: duplicate reply 0x%p to RPC " | |
784 | "request 0x%p: xid 0x%08x\n", __func__, rep, req, | |
785 | headerp->rm_xid); | |
786 | goto repost; | |
787 | } | |
e9601828 TT |
788 | |
789 | dprintk("RPC: %s: reply 0x%p completes request 0x%p\n" | |
790 | " RPC request 0x%p xid 0x%08x\n", | |
791 | __func__, rep, req, rqst, headerp->rm_xid); | |
792 | ||
e9601828 TT |
793 | /* from here on, the reply is no longer an orphan */ |
794 | req->rl_reply = rep; | |
18906972 | 795 | xprt->reestablish_timeout = 0; |
e9601828 TT |
796 | |
797 | /* check for expected message types */ | |
798 | /* The order of some of these tests is important. */ | |
799 | switch (headerp->rm_type) { | |
284f4902 | 800 | case rdma_msg: |
e9601828 TT |
801 | /* never expect read chunks */ |
802 | /* never expect reply chunks (two ways to check) */ | |
803 | /* never expect write chunks without having offered RDMA */ | |
804 | if (headerp->rm_body.rm_chunks[0] != xdr_zero || | |
805 | (headerp->rm_body.rm_chunks[1] == xdr_zero && | |
806 | headerp->rm_body.rm_chunks[2] != xdr_zero) || | |
807 | (headerp->rm_body.rm_chunks[1] != xdr_zero && | |
808 | req->rl_nchunks == 0)) | |
809 | goto badheader; | |
810 | if (headerp->rm_body.rm_chunks[1] != xdr_zero) { | |
811 | /* count any expected write chunks in read reply */ | |
812 | /* start at write chunk array count */ | |
813 | iptr = &headerp->rm_body.rm_chunks[2]; | |
814 | rdmalen = rpcrdma_count_chunks(rep, | |
815 | req->rl_nchunks, 1, &iptr); | |
816 | /* check for validity, and no reply chunk after */ | |
817 | if (rdmalen < 0 || *iptr++ != xdr_zero) | |
818 | goto badheader; | |
819 | rep->rr_len -= | |
820 | ((unsigned char *)iptr - (unsigned char *)headerp); | |
821 | status = rep->rr_len + rdmalen; | |
822 | r_xprt->rx_stats.total_rdma_reply += rdmalen; | |
9191ca3b TT |
823 | /* special case - last chunk may omit padding */ |
824 | if (rdmalen &= 3) { | |
825 | rdmalen = 4 - rdmalen; | |
826 | status += rdmalen; | |
827 | } | |
e9601828 TT |
828 | } else { |
829 | /* else ordinary inline */ | |
9191ca3b | 830 | rdmalen = 0; |
2d8a9726 | 831 | iptr = (__be32 *)((unsigned char *)headerp + 28); |
e9601828 TT |
832 | rep->rr_len -= 28; /*sizeof *headerp;*/ |
833 | status = rep->rr_len; | |
834 | } | |
835 | /* Fix up the rpc results for upper layer */ | |
9191ca3b | 836 | rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen); |
e9601828 TT |
837 | break; |
838 | ||
284f4902 | 839 | case rdma_nomsg: |
e9601828 TT |
840 | /* never expect read or write chunks, always reply chunks */ |
841 | if (headerp->rm_body.rm_chunks[0] != xdr_zero || | |
842 | headerp->rm_body.rm_chunks[1] != xdr_zero || | |
843 | headerp->rm_body.rm_chunks[2] != xdr_one || | |
844 | req->rl_nchunks == 0) | |
845 | goto badheader; | |
2d8a9726 | 846 | iptr = (__be32 *)((unsigned char *)headerp + 28); |
e9601828 TT |
847 | rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr); |
848 | if (rdmalen < 0) | |
849 | goto badheader; | |
850 | r_xprt->rx_stats.total_rdma_reply += rdmalen; | |
851 | /* Reply chunk buffer already is the reply vector - no fixup. */ | |
852 | status = rdmalen; | |
853 | break; | |
854 | ||
855 | badheader: | |
856 | default: | |
857 | dprintk("%s: invalid rpcrdma reply header (type %d):" | |
858 | " chunks[012] == %d %d %d" | |
859 | " expected chunks <= %d\n", | |
284f4902 | 860 | __func__, be32_to_cpu(headerp->rm_type), |
e9601828 TT |
861 | headerp->rm_body.rm_chunks[0], |
862 | headerp->rm_body.rm_chunks[1], | |
863 | headerp->rm_body.rm_chunks[2], | |
864 | req->rl_nchunks); | |
865 | status = -EIO; | |
866 | r_xprt->rx_stats.bad_reply_count++; | |
867 | break; | |
868 | } | |
869 | ||
e7ce710a CL |
870 | cwnd = xprt->cwnd; |
871 | xprt->cwnd = atomic_read(&r_xprt->rx_buf.rb_credits) << RPC_CWNDSHIFT; | |
872 | if (xprt->cwnd > cwnd) | |
873 | xprt_release_rqst_cong(rqst->rq_task); | |
874 | ||
e9601828 TT |
875 | dprintk("RPC: %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n", |
876 | __func__, xprt, rqst, status); | |
877 | xprt_complete_rqst(rqst->rq_task, status); | |
878 | spin_unlock(&xprt->transport_lock); | |
879 | } |