Commit | Line | Data |
---|---|---|
c06b540a TT |
1 | /* |
2 | * Copyright (c) 2005-2006 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 | * Author: Tom Tucker <tom@opengridcomputing.com> | |
40 | */ | |
41 | ||
42 | #include <linux/sunrpc/debug.h> | |
43 | #include <linux/sunrpc/rpc_rdma.h> | |
44 | #include <linux/spinlock.h> | |
45 | #include <asm/unaligned.h> | |
46 | #include <rdma/ib_verbs.h> | |
47 | #include <rdma/rdma_cm.h> | |
48 | #include <linux/sunrpc/svc_rdma.h> | |
49 | ||
50 | #define RPCDBG_FACILITY RPCDBG_SVCXPRT | |
51 | ||
52 | /* Encode an XDR as an array of IB SGE | |
53 | * | |
54 | * Assumptions: | |
55 | * - head[0] is physically contiguous. | |
56 | * - tail[0] is physically contiguous. | |
57 | * - pages[] is not physically or virtually contigous and consists of | |
58 | * PAGE_SIZE elements. | |
59 | * | |
60 | * Output: | |
61 | * SGE[0] reserved for RCPRDMA header | |
62 | * SGE[1] data from xdr->head[] | |
63 | * SGE[2..sge_count-2] data from xdr->pages[] | |
64 | * SGE[sge_count-1] data from xdr->tail. | |
65 | * | |
34d16e42 TT |
66 | * The max SGE we need is the length of the XDR / pagesize + one for |
67 | * head + one for tail + one for RPCRDMA header. Since RPCSVC_MAXPAGES | |
68 | * reserves a page for both the request and the reply header, and this | |
69 | * array is only concerned with the reply we are assured that we have | |
70 | * on extra page for the RPCRMDA header. | |
c06b540a | 71 | */ |
5eaa65b2 | 72 | static int fast_reg_xdr(struct svcxprt_rdma *xprt, |
afd566ea TT |
73 | struct xdr_buf *xdr, |
74 | struct svc_rdma_req_map *vec) | |
75 | { | |
76 | int sge_no; | |
77 | u32 sge_bytes; | |
78 | u32 page_bytes; | |
79 | u32 page_off; | |
80 | int page_no = 0; | |
81 | u8 *frva; | |
82 | struct svc_rdma_fastreg_mr *frmr; | |
83 | ||
84 | frmr = svc_rdma_get_frmr(xprt); | |
85 | if (IS_ERR(frmr)) | |
86 | return -ENOMEM; | |
87 | vec->frmr = frmr; | |
88 | ||
89 | /* Skip the RPCRDMA header */ | |
90 | sge_no = 1; | |
91 | ||
92 | /* Map the head. */ | |
93 | frva = (void *)((unsigned long)(xdr->head[0].iov_base) & PAGE_MASK); | |
94 | vec->sge[sge_no].iov_base = xdr->head[0].iov_base; | |
95 | vec->sge[sge_no].iov_len = xdr->head[0].iov_len; | |
96 | vec->count = 2; | |
97 | sge_no++; | |
98 | ||
99 | /* Build the FRMR */ | |
100 | frmr->kva = frva; | |
101 | frmr->direction = DMA_TO_DEVICE; | |
102 | frmr->access_flags = 0; | |
103 | frmr->map_len = PAGE_SIZE; | |
104 | frmr->page_list_len = 1; | |
105 | frmr->page_list->page_list[page_no] = | |
106 | ib_dma_map_single(xprt->sc_cm_id->device, | |
107 | (void *)xdr->head[0].iov_base, | |
108 | PAGE_SIZE, DMA_TO_DEVICE); | |
109 | if (ib_dma_mapping_error(xprt->sc_cm_id->device, | |
110 | frmr->page_list->page_list[page_no])) | |
111 | goto fatal_err; | |
112 | atomic_inc(&xprt->sc_dma_used); | |
113 | ||
114 | page_off = xdr->page_base; | |
115 | page_bytes = xdr->page_len + page_off; | |
116 | if (!page_bytes) | |
117 | goto encode_tail; | |
118 | ||
119 | /* Map the pages */ | |
120 | vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off; | |
121 | vec->sge[sge_no].iov_len = page_bytes; | |
122 | sge_no++; | |
123 | while (page_bytes) { | |
124 | struct page *page; | |
125 | ||
126 | page = xdr->pages[page_no++]; | |
127 | sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off)); | |
128 | page_bytes -= sge_bytes; | |
129 | ||
130 | frmr->page_list->page_list[page_no] = | |
131 | ib_dma_map_page(xprt->sc_cm_id->device, page, 0, | |
132 | PAGE_SIZE, DMA_TO_DEVICE); | |
133 | if (ib_dma_mapping_error(xprt->sc_cm_id->device, | |
134 | frmr->page_list->page_list[page_no])) | |
135 | goto fatal_err; | |
136 | ||
137 | atomic_inc(&xprt->sc_dma_used); | |
138 | page_off = 0; /* reset for next time through loop */ | |
139 | frmr->map_len += PAGE_SIZE; | |
140 | frmr->page_list_len++; | |
141 | } | |
142 | vec->count++; | |
143 | ||
144 | encode_tail: | |
145 | /* Map tail */ | |
146 | if (0 == xdr->tail[0].iov_len) | |
147 | goto done; | |
148 | ||
149 | vec->count++; | |
150 | vec->sge[sge_no].iov_len = xdr->tail[0].iov_len; | |
151 | ||
152 | if (((unsigned long)xdr->tail[0].iov_base & PAGE_MASK) == | |
153 | ((unsigned long)xdr->head[0].iov_base & PAGE_MASK)) { | |
154 | /* | |
155 | * If head and tail use the same page, we don't need | |
156 | * to map it again. | |
157 | */ | |
158 | vec->sge[sge_no].iov_base = xdr->tail[0].iov_base; | |
159 | } else { | |
160 | void *va; | |
161 | ||
162 | /* Map another page for the tail */ | |
163 | page_off = (unsigned long)xdr->tail[0].iov_base & ~PAGE_MASK; | |
164 | va = (void *)((unsigned long)xdr->tail[0].iov_base & PAGE_MASK); | |
165 | vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off; | |
166 | ||
167 | frmr->page_list->page_list[page_no] = | |
168 | ib_dma_map_single(xprt->sc_cm_id->device, va, PAGE_SIZE, | |
169 | DMA_TO_DEVICE); | |
170 | if (ib_dma_mapping_error(xprt->sc_cm_id->device, | |
171 | frmr->page_list->page_list[page_no])) | |
172 | goto fatal_err; | |
173 | atomic_inc(&xprt->sc_dma_used); | |
174 | frmr->map_len += PAGE_SIZE; | |
175 | frmr->page_list_len++; | |
176 | } | |
177 | ||
178 | done: | |
179 | if (svc_rdma_fastreg(xprt, frmr)) | |
180 | goto fatal_err; | |
181 | ||
182 | return 0; | |
183 | ||
184 | fatal_err: | |
185 | printk("svcrdma: Error fast registering memory for xprt %p\n", xprt); | |
186 | svc_rdma_put_frmr(xprt, frmr); | |
187 | return -EIO; | |
188 | } | |
189 | ||
190 | static int map_xdr(struct svcxprt_rdma *xprt, | |
191 | struct xdr_buf *xdr, | |
192 | struct svc_rdma_req_map *vec) | |
c06b540a | 193 | { |
c06b540a | 194 | int sge_no; |
c06b540a TT |
195 | u32 sge_bytes; |
196 | u32 page_bytes; | |
34d16e42 | 197 | u32 page_off; |
c06b540a TT |
198 | int page_no; |
199 | ||
34d16e42 TT |
200 | BUG_ON(xdr->len != |
201 | (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len)); | |
202 | ||
afd566ea TT |
203 | if (xprt->sc_frmr_pg_list_len) |
204 | return fast_reg_xdr(xprt, xdr, vec); | |
205 | ||
c06b540a TT |
206 | /* Skip the first sge, this is for the RPCRDMA header */ |
207 | sge_no = 1; | |
208 | ||
209 | /* Head SGE */ | |
34d16e42 TT |
210 | vec->sge[sge_no].iov_base = xdr->head[0].iov_base; |
211 | vec->sge[sge_no].iov_len = xdr->head[0].iov_len; | |
c06b540a TT |
212 | sge_no++; |
213 | ||
214 | /* pages SGE */ | |
215 | page_no = 0; | |
216 | page_bytes = xdr->page_len; | |
217 | page_off = xdr->page_base; | |
34d16e42 TT |
218 | while (page_bytes) { |
219 | vec->sge[sge_no].iov_base = | |
220 | page_address(xdr->pages[page_no]) + page_off; | |
221 | sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off)); | |
c06b540a | 222 | page_bytes -= sge_bytes; |
34d16e42 | 223 | vec->sge[sge_no].iov_len = sge_bytes; |
c06b540a TT |
224 | |
225 | sge_no++; | |
226 | page_no++; | |
227 | page_off = 0; /* reset for next time through loop */ | |
228 | } | |
229 | ||
230 | /* Tail SGE */ | |
34d16e42 TT |
231 | if (xdr->tail[0].iov_len) { |
232 | vec->sge[sge_no].iov_base = xdr->tail[0].iov_base; | |
233 | vec->sge[sge_no].iov_len = xdr->tail[0].iov_len; | |
c06b540a TT |
234 | sge_no++; |
235 | } | |
236 | ||
b1e1e158 | 237 | dprintk("svcrdma: map_xdr: sge_no %d page_no %d " |
2e3c230b | 238 | "page_base %u page_len %u head_len %zu tail_len %zu\n", |
b1e1e158 TT |
239 | sge_no, page_no, xdr->page_base, xdr->page_len, |
240 | xdr->head[0].iov_len, xdr->tail[0].iov_len); | |
241 | ||
34d16e42 | 242 | vec->count = sge_no; |
afd566ea | 243 | return 0; |
c06b540a TT |
244 | } |
245 | ||
c06b540a | 246 | /* Assumptions: |
afd566ea TT |
247 | * - We are using FRMR |
248 | * - or - | |
c06b540a TT |
249 | * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE |
250 | */ | |
251 | static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp, | |
252 | u32 rmr, u64 to, | |
253 | u32 xdr_off, int write_len, | |
34d16e42 | 254 | struct svc_rdma_req_map *vec) |
c06b540a | 255 | { |
c06b540a TT |
256 | struct ib_send_wr write_wr; |
257 | struct ib_sge *sge; | |
258 | int xdr_sge_no; | |
259 | int sge_no; | |
260 | int sge_bytes; | |
261 | int sge_off; | |
262 | int bc; | |
263 | struct svc_rdma_op_ctxt *ctxt; | |
c06b540a | 264 | |
34d16e42 | 265 | BUG_ON(vec->count > RPCSVC_MAXPAGES); |
c06b540a | 266 | dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, " |
34d16e42 | 267 | "write_len=%d, vec->sge=%p, vec->count=%lu\n", |
bb50c801 | 268 | rmr, (unsigned long long)to, xdr_off, |
34d16e42 | 269 | write_len, vec->sge, vec->count); |
c06b540a TT |
270 | |
271 | ctxt = svc_rdma_get_context(xprt); | |
34d16e42 TT |
272 | ctxt->direction = DMA_TO_DEVICE; |
273 | sge = ctxt->sge; | |
c06b540a TT |
274 | |
275 | /* Find the SGE associated with xdr_off */ | |
34d16e42 | 276 | for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < vec->count; |
c06b540a | 277 | xdr_sge_no++) { |
34d16e42 | 278 | if (vec->sge[xdr_sge_no].iov_len > bc) |
c06b540a | 279 | break; |
34d16e42 | 280 | bc -= vec->sge[xdr_sge_no].iov_len; |
c06b540a TT |
281 | } |
282 | ||
283 | sge_off = bc; | |
284 | bc = write_len; | |
285 | sge_no = 0; | |
286 | ||
287 | /* Copy the remaining SGE */ | |
afd566ea TT |
288 | while (bc != 0) { |
289 | sge_bytes = min_t(size_t, | |
290 | bc, vec->sge[xdr_sge_no].iov_len-sge_off); | |
c06b540a | 291 | sge[sge_no].length = sge_bytes; |
afd566ea TT |
292 | if (!vec->frmr) { |
293 | sge[sge_no].addr = | |
294 | ib_dma_map_single(xprt->sc_cm_id->device, | |
295 | (void *) | |
296 | vec->sge[xdr_sge_no].iov_base + sge_off, | |
297 | sge_bytes, DMA_TO_DEVICE); | |
298 | if (ib_dma_mapping_error(xprt->sc_cm_id->device, | |
299 | sge[sge_no].addr)) | |
300 | goto err; | |
301 | atomic_inc(&xprt->sc_dma_used); | |
302 | sge[sge_no].lkey = xprt->sc_dma_lkey; | |
303 | } else { | |
304 | sge[sge_no].addr = (unsigned long) | |
305 | vec->sge[xdr_sge_no].iov_base + sge_off; | |
306 | sge[sge_no].lkey = vec->frmr->mr->lkey; | |
307 | } | |
308 | ctxt->count++; | |
309 | ctxt->frmr = vec->frmr; | |
c06b540a TT |
310 | sge_off = 0; |
311 | sge_no++; | |
312 | xdr_sge_no++; | |
afd566ea | 313 | BUG_ON(xdr_sge_no > vec->count); |
c06b540a TT |
314 | bc -= sge_bytes; |
315 | } | |
316 | ||
c06b540a TT |
317 | /* Prepare WRITE WR */ |
318 | memset(&write_wr, 0, sizeof write_wr); | |
319 | ctxt->wr_op = IB_WR_RDMA_WRITE; | |
320 | write_wr.wr_id = (unsigned long)ctxt; | |
321 | write_wr.sg_list = &sge[0]; | |
322 | write_wr.num_sge = sge_no; | |
323 | write_wr.opcode = IB_WR_RDMA_WRITE; | |
324 | write_wr.send_flags = IB_SEND_SIGNALED; | |
325 | write_wr.wr.rdma.rkey = rmr; | |
326 | write_wr.wr.rdma.remote_addr = to; | |
327 | ||
328 | /* Post It */ | |
329 | atomic_inc(&rdma_stat_write); | |
34d16e42 TT |
330 | if (svc_rdma_send(xprt, &write_wr)) |
331 | goto err; | |
332 | return 0; | |
333 | err: | |
334 | svc_rdma_put_context(ctxt, 0); | |
335 | /* Fatal error, close transport */ | |
336 | return -EIO; | |
c06b540a TT |
337 | } |
338 | ||
339 | static int send_write_chunks(struct svcxprt_rdma *xprt, | |
340 | struct rpcrdma_msg *rdma_argp, | |
341 | struct rpcrdma_msg *rdma_resp, | |
342 | struct svc_rqst *rqstp, | |
34d16e42 | 343 | struct svc_rdma_req_map *vec) |
c06b540a TT |
344 | { |
345 | u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len; | |
346 | int write_len; | |
347 | int max_write; | |
348 | u32 xdr_off; | |
349 | int chunk_off; | |
350 | int chunk_no; | |
351 | struct rpcrdma_write_array *arg_ary; | |
352 | struct rpcrdma_write_array *res_ary; | |
353 | int ret; | |
354 | ||
355 | arg_ary = svc_rdma_get_write_array(rdma_argp); | |
356 | if (!arg_ary) | |
357 | return 0; | |
358 | res_ary = (struct rpcrdma_write_array *) | |
359 | &rdma_resp->rm_body.rm_chunks[1]; | |
360 | ||
afd566ea TT |
361 | if (vec->frmr) |
362 | max_write = vec->frmr->map_len; | |
363 | else | |
364 | max_write = xprt->sc_max_sge * PAGE_SIZE; | |
c06b540a TT |
365 | |
366 | /* Write chunks start at the pagelist */ | |
367 | for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0; | |
368 | xfer_len && chunk_no < arg_ary->wc_nchunks; | |
369 | chunk_no++) { | |
370 | struct rpcrdma_segment *arg_ch; | |
371 | u64 rs_offset; | |
372 | ||
373 | arg_ch = &arg_ary->wc_array[chunk_no].wc_target; | |
374 | write_len = min(xfer_len, arg_ch->rs_length); | |
375 | ||
376 | /* Prepare the response chunk given the length actually | |
377 | * written */ | |
378 | rs_offset = get_unaligned(&(arg_ch->rs_offset)); | |
379 | svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no, | |
380 | arg_ch->rs_handle, | |
381 | rs_offset, | |
382 | write_len); | |
383 | chunk_off = 0; | |
384 | while (write_len) { | |
385 | int this_write; | |
386 | this_write = min(write_len, max_write); | |
387 | ret = send_write(xprt, rqstp, | |
388 | arg_ch->rs_handle, | |
389 | rs_offset + chunk_off, | |
390 | xdr_off, | |
391 | this_write, | |
34d16e42 | 392 | vec); |
c06b540a TT |
393 | if (ret) { |
394 | dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n", | |
395 | ret); | |
396 | return -EIO; | |
397 | } | |
398 | chunk_off += this_write; | |
399 | xdr_off += this_write; | |
400 | xfer_len -= this_write; | |
401 | write_len -= this_write; | |
402 | } | |
403 | } | |
404 | /* Update the req with the number of chunks actually used */ | |
405 | svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no); | |
406 | ||
407 | return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len; | |
408 | } | |
409 | ||
410 | static int send_reply_chunks(struct svcxprt_rdma *xprt, | |
411 | struct rpcrdma_msg *rdma_argp, | |
412 | struct rpcrdma_msg *rdma_resp, | |
413 | struct svc_rqst *rqstp, | |
34d16e42 | 414 | struct svc_rdma_req_map *vec) |
c06b540a TT |
415 | { |
416 | u32 xfer_len = rqstp->rq_res.len; | |
417 | int write_len; | |
418 | int max_write; | |
419 | u32 xdr_off; | |
420 | int chunk_no; | |
421 | int chunk_off; | |
422 | struct rpcrdma_segment *ch; | |
423 | struct rpcrdma_write_array *arg_ary; | |
424 | struct rpcrdma_write_array *res_ary; | |
425 | int ret; | |
426 | ||
427 | arg_ary = svc_rdma_get_reply_array(rdma_argp); | |
428 | if (!arg_ary) | |
429 | return 0; | |
430 | /* XXX: need to fix when reply lists occur with read-list and or | |
431 | * write-list */ | |
432 | res_ary = (struct rpcrdma_write_array *) | |
433 | &rdma_resp->rm_body.rm_chunks[2]; | |
434 | ||
afd566ea TT |
435 | if (vec->frmr) |
436 | max_write = vec->frmr->map_len; | |
437 | else | |
438 | max_write = xprt->sc_max_sge * PAGE_SIZE; | |
c06b540a TT |
439 | |
440 | /* xdr offset starts at RPC message */ | |
441 | for (xdr_off = 0, chunk_no = 0; | |
442 | xfer_len && chunk_no < arg_ary->wc_nchunks; | |
443 | chunk_no++) { | |
444 | u64 rs_offset; | |
445 | ch = &arg_ary->wc_array[chunk_no].wc_target; | |
446 | write_len = min(xfer_len, ch->rs_length); | |
447 | ||
c06b540a TT |
448 | /* Prepare the reply chunk given the length actually |
449 | * written */ | |
450 | rs_offset = get_unaligned(&(ch->rs_offset)); | |
451 | svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no, | |
452 | ch->rs_handle, rs_offset, | |
453 | write_len); | |
454 | chunk_off = 0; | |
455 | while (write_len) { | |
456 | int this_write; | |
457 | ||
458 | this_write = min(write_len, max_write); | |
459 | ret = send_write(xprt, rqstp, | |
460 | ch->rs_handle, | |
461 | rs_offset + chunk_off, | |
462 | xdr_off, | |
463 | this_write, | |
34d16e42 | 464 | vec); |
c06b540a TT |
465 | if (ret) { |
466 | dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n", | |
467 | ret); | |
468 | return -EIO; | |
469 | } | |
470 | chunk_off += this_write; | |
471 | xdr_off += this_write; | |
472 | xfer_len -= this_write; | |
473 | write_len -= this_write; | |
474 | } | |
475 | } | |
476 | /* Update the req with the number of chunks actually used */ | |
477 | svc_rdma_xdr_encode_reply_array(res_ary, chunk_no); | |
478 | ||
479 | return rqstp->rq_res.len; | |
480 | } | |
481 | ||
482 | /* This function prepares the portion of the RPCRDMA message to be | |
483 | * sent in the RDMA_SEND. This function is called after data sent via | |
484 | * RDMA has already been transmitted. There are three cases: | |
485 | * - The RPCRDMA header, RPC header, and payload are all sent in a | |
486 | * single RDMA_SEND. This is the "inline" case. | |
487 | * - The RPCRDMA header and some portion of the RPC header and data | |
488 | * are sent via this RDMA_SEND and another portion of the data is | |
489 | * sent via RDMA. | |
490 | * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC | |
491 | * header and data are all transmitted via RDMA. | |
492 | * In all three cases, this function prepares the RPCRDMA header in | |
493 | * sge[0], the 'type' parameter indicates the type to place in the | |
494 | * RPCRDMA header, and the 'byte_count' field indicates how much of | |
495 | * the XDR to include in this RDMA_SEND. | |
496 | */ | |
497 | static int send_reply(struct svcxprt_rdma *rdma, | |
498 | struct svc_rqst *rqstp, | |
499 | struct page *page, | |
500 | struct rpcrdma_msg *rdma_resp, | |
501 | struct svc_rdma_op_ctxt *ctxt, | |
34d16e42 | 502 | struct svc_rdma_req_map *vec, |
c06b540a TT |
503 | int byte_count) |
504 | { | |
505 | struct ib_send_wr send_wr; | |
afd566ea | 506 | struct ib_send_wr inv_wr; |
c06b540a TT |
507 | int sge_no; |
508 | int sge_bytes; | |
509 | int page_no; | |
510 | int ret; | |
511 | ||
0e7f011a TT |
512 | /* Post a recv buffer to handle another request. */ |
513 | ret = svc_rdma_post_recv(rdma); | |
514 | if (ret) { | |
515 | printk(KERN_INFO | |
516 | "svcrdma: could not post a receive buffer, err=%d." | |
517 | "Closing transport %p.\n", ret, rdma); | |
518 | set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags); | |
5ac461a6 TT |
519 | svc_rdma_put_context(ctxt, 0); |
520 | return -ENOTCONN; | |
0e7f011a TT |
521 | } |
522 | ||
c06b540a TT |
523 | /* Prepare the context */ |
524 | ctxt->pages[0] = page; | |
525 | ctxt->count = 1; | |
afd566ea TT |
526 | ctxt->frmr = vec->frmr; |
527 | if (vec->frmr) | |
528 | set_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags); | |
529 | else | |
530 | clear_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags); | |
c06b540a TT |
531 | |
532 | /* Prepare the SGE for the RPCRDMA Header */ | |
533 | ctxt->sge[0].addr = | |
534 | ib_dma_map_page(rdma->sc_cm_id->device, | |
535 | page, 0, PAGE_SIZE, DMA_TO_DEVICE); | |
afd566ea TT |
536 | if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr)) |
537 | goto err; | |
538 | atomic_inc(&rdma->sc_dma_used); | |
539 | ||
c06b540a | 540 | ctxt->direction = DMA_TO_DEVICE; |
afd566ea | 541 | |
c06b540a | 542 | ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp); |
afd566ea | 543 | ctxt->sge[0].lkey = rdma->sc_dma_lkey; |
c06b540a TT |
544 | |
545 | /* Determine how many of our SGE are to be transmitted */ | |
34d16e42 TT |
546 | for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) { |
547 | sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count); | |
c06b540a | 548 | byte_count -= sge_bytes; |
afd566ea TT |
549 | if (!vec->frmr) { |
550 | ctxt->sge[sge_no].addr = | |
551 | ib_dma_map_single(rdma->sc_cm_id->device, | |
552 | vec->sge[sge_no].iov_base, | |
553 | sge_bytes, DMA_TO_DEVICE); | |
554 | if (ib_dma_mapping_error(rdma->sc_cm_id->device, | |
555 | ctxt->sge[sge_no].addr)) | |
556 | goto err; | |
557 | atomic_inc(&rdma->sc_dma_used); | |
558 | ctxt->sge[sge_no].lkey = rdma->sc_dma_lkey; | |
559 | } else { | |
560 | ctxt->sge[sge_no].addr = (unsigned long) | |
561 | vec->sge[sge_no].iov_base; | |
562 | ctxt->sge[sge_no].lkey = vec->frmr->mr->lkey; | |
563 | } | |
34d16e42 | 564 | ctxt->sge[sge_no].length = sge_bytes; |
c06b540a TT |
565 | } |
566 | BUG_ON(byte_count != 0); | |
567 | ||
568 | /* Save all respages in the ctxt and remove them from the | |
569 | * respages array. They are our pages until the I/O | |
570 | * completes. | |
571 | */ | |
572 | for (page_no = 0; page_no < rqstp->rq_resused; page_no++) { | |
573 | ctxt->pages[page_no+1] = rqstp->rq_respages[page_no]; | |
574 | ctxt->count++; | |
575 | rqstp->rq_respages[page_no] = NULL; | |
afd566ea TT |
576 | /* |
577 | * If there are more pages than SGE, terminate SGE | |
578 | * list so that svc_rdma_unmap_dma doesn't attempt to | |
579 | * unmap garbage. | |
580 | */ | |
34d16e42 TT |
581 | if (page_no+1 >= sge_no) |
582 | ctxt->sge[page_no+1].length = 0; | |
c06b540a | 583 | } |
c06b540a TT |
584 | BUG_ON(sge_no > rdma->sc_max_sge); |
585 | memset(&send_wr, 0, sizeof send_wr); | |
586 | ctxt->wr_op = IB_WR_SEND; | |
587 | send_wr.wr_id = (unsigned long)ctxt; | |
588 | send_wr.sg_list = ctxt->sge; | |
589 | send_wr.num_sge = sge_no; | |
590 | send_wr.opcode = IB_WR_SEND; | |
591 | send_wr.send_flags = IB_SEND_SIGNALED; | |
afd566ea TT |
592 | if (vec->frmr) { |
593 | /* Prepare INVALIDATE WR */ | |
594 | memset(&inv_wr, 0, sizeof inv_wr); | |
595 | inv_wr.opcode = IB_WR_LOCAL_INV; | |
596 | inv_wr.send_flags = IB_SEND_SIGNALED; | |
597 | inv_wr.ex.invalidate_rkey = | |
598 | vec->frmr->mr->lkey; | |
599 | send_wr.next = &inv_wr; | |
600 | } | |
c06b540a TT |
601 | |
602 | ret = svc_rdma_send(rdma, &send_wr); | |
603 | if (ret) | |
afd566ea | 604 | goto err; |
c06b540a | 605 | |
afd566ea TT |
606 | return 0; |
607 | ||
608 | err: | |
609 | svc_rdma_put_frmr(rdma, vec->frmr); | |
610 | svc_rdma_put_context(ctxt, 1); | |
611 | return -EIO; | |
c06b540a TT |
612 | } |
613 | ||
614 | void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp) | |
615 | { | |
616 | } | |
617 | ||
618 | /* | |
619 | * Return the start of an xdr buffer. | |
620 | */ | |
621 | static void *xdr_start(struct xdr_buf *xdr) | |
622 | { | |
623 | return xdr->head[0].iov_base - | |
624 | (xdr->len - | |
625 | xdr->page_len - | |
626 | xdr->tail[0].iov_len - | |
627 | xdr->head[0].iov_len); | |
628 | } | |
629 | ||
630 | int svc_rdma_sendto(struct svc_rqst *rqstp) | |
631 | { | |
632 | struct svc_xprt *xprt = rqstp->rq_xprt; | |
633 | struct svcxprt_rdma *rdma = | |
634 | container_of(xprt, struct svcxprt_rdma, sc_xprt); | |
635 | struct rpcrdma_msg *rdma_argp; | |
636 | struct rpcrdma_msg *rdma_resp; | |
637 | struct rpcrdma_write_array *reply_ary; | |
638 | enum rpcrdma_proc reply_type; | |
639 | int ret; | |
640 | int inline_bytes; | |
c06b540a TT |
641 | struct page *res_page; |
642 | struct svc_rdma_op_ctxt *ctxt; | |
34d16e42 | 643 | struct svc_rdma_req_map *vec; |
c06b540a TT |
644 | |
645 | dprintk("svcrdma: sending response for rqstp=%p\n", rqstp); | |
646 | ||
647 | /* Get the RDMA request header. */ | |
648 | rdma_argp = xdr_start(&rqstp->rq_arg); | |
649 | ||
34d16e42 | 650 | /* Build an req vec for the XDR */ |
c06b540a TT |
651 | ctxt = svc_rdma_get_context(rdma); |
652 | ctxt->direction = DMA_TO_DEVICE; | |
34d16e42 | 653 | vec = svc_rdma_get_req_map(); |
afd566ea TT |
654 | ret = map_xdr(rdma, &rqstp->rq_res, vec); |
655 | if (ret) | |
656 | goto err0; | |
c06b540a TT |
657 | inline_bytes = rqstp->rq_res.len; |
658 | ||
659 | /* Create the RDMA response header */ | |
660 | res_page = svc_rdma_get_page(); | |
661 | rdma_resp = page_address(res_page); | |
662 | reply_ary = svc_rdma_get_reply_array(rdma_argp); | |
663 | if (reply_ary) | |
664 | reply_type = RDMA_NOMSG; | |
665 | else | |
666 | reply_type = RDMA_MSG; | |
667 | svc_rdma_xdr_encode_reply_header(rdma, rdma_argp, | |
668 | rdma_resp, reply_type); | |
669 | ||
670 | /* Send any write-chunk data and build resp write-list */ | |
671 | ret = send_write_chunks(rdma, rdma_argp, rdma_resp, | |
34d16e42 | 672 | rqstp, vec); |
c06b540a TT |
673 | if (ret < 0) { |
674 | printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n", | |
675 | ret); | |
afd566ea | 676 | goto err1; |
c06b540a TT |
677 | } |
678 | inline_bytes -= ret; | |
679 | ||
680 | /* Send any reply-list data and update resp reply-list */ | |
681 | ret = send_reply_chunks(rdma, rdma_argp, rdma_resp, | |
34d16e42 | 682 | rqstp, vec); |
c06b540a TT |
683 | if (ret < 0) { |
684 | printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n", | |
685 | ret); | |
afd566ea | 686 | goto err1; |
c06b540a TT |
687 | } |
688 | inline_bytes -= ret; | |
689 | ||
34d16e42 | 690 | ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec, |
c06b540a | 691 | inline_bytes); |
34d16e42 | 692 | svc_rdma_put_req_map(vec); |
c06b540a TT |
693 | dprintk("svcrdma: send_reply returns %d\n", ret); |
694 | return ret; | |
afd566ea TT |
695 | |
696 | err1: | |
697 | put_page(res_page); | |
698 | err0: | |
34d16e42 | 699 | svc_rdma_put_req_map(vec); |
c06b540a | 700 | svc_rdma_put_context(ctxt, 0); |
c06b540a TT |
701 | return ret; |
702 | } |