Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / net / sunrpc / xprtrdma / svc_rdma_sendto.c
1 /*
2  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
3  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the BSD-type
9  * license below:
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  *
15  *      Redistributions of source code must retain the above copyright
16  *      notice, this list of conditions and the following disclaimer.
17  *
18  *      Redistributions in binary form must reproduce the above
19  *      copyright notice, this list of conditions and the following
20  *      disclaimer in the documentation and/or other materials provided
21  *      with the distribution.
22  *
23  *      Neither the name of the Network Appliance, Inc. nor the names of
24  *      its contributors may be used to endorse or promote products
25  *      derived from this software without specific prior written
26  *      permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39  *
40  * Author: Tom Tucker <tom@opengridcomputing.com>
41  */
42
43 #include <linux/sunrpc/debug.h>
44 #include <linux/sunrpc/rpc_rdma.h>
45 #include <linux/spinlock.h>
46 #include <asm/unaligned.h>
47 #include <rdma/ib_verbs.h>
48 #include <rdma/rdma_cm.h>
49 #include <linux/sunrpc/svc_rdma.h>
50
51 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
52
53 static int map_xdr(struct svcxprt_rdma *xprt,
54                    struct xdr_buf *xdr,
55                    struct svc_rdma_req_map *vec)
56 {
57         int sge_no;
58         u32 sge_bytes;
59         u32 page_bytes;
60         u32 page_off;
61         int page_no;
62
63         if (xdr->len !=
64             (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len)) {
65                 pr_err("svcrdma: map_xdr: XDR buffer length error\n");
66                 return -EIO;
67         }
68
69         /* Skip the first sge, this is for the RPCRDMA header */
70         sge_no = 1;
71
72         /* Head SGE */
73         vec->sge[sge_no].iov_base = xdr->head[0].iov_base;
74         vec->sge[sge_no].iov_len = xdr->head[0].iov_len;
75         sge_no++;
76
77         /* pages SGE */
78         page_no = 0;
79         page_bytes = xdr->page_len;
80         page_off = xdr->page_base;
81         while (page_bytes) {
82                 vec->sge[sge_no].iov_base =
83                         page_address(xdr->pages[page_no]) + page_off;
84                 sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
85                 page_bytes -= sge_bytes;
86                 vec->sge[sge_no].iov_len = sge_bytes;
87
88                 sge_no++;
89                 page_no++;
90                 page_off = 0; /* reset for next time through loop */
91         }
92
93         /* Tail SGE */
94         if (xdr->tail[0].iov_len) {
95                 vec->sge[sge_no].iov_base = xdr->tail[0].iov_base;
96                 vec->sge[sge_no].iov_len = xdr->tail[0].iov_len;
97                 sge_no++;
98         }
99
100         dprintk("svcrdma: map_xdr: sge_no %d page_no %d "
101                 "page_base %u page_len %u head_len %zu tail_len %zu\n",
102                 sge_no, page_no, xdr->page_base, xdr->page_len,
103                 xdr->head[0].iov_len, xdr->tail[0].iov_len);
104
105         vec->count = sge_no;
106         return 0;
107 }
108
109 static dma_addr_t dma_map_xdr(struct svcxprt_rdma *xprt,
110                               struct xdr_buf *xdr,
111                               u32 xdr_off, size_t len, int dir)
112 {
113         struct page *page;
114         dma_addr_t dma_addr;
115         if (xdr_off < xdr->head[0].iov_len) {
116                 /* This offset is in the head */
117                 xdr_off += (unsigned long)xdr->head[0].iov_base & ~PAGE_MASK;
118                 page = virt_to_page(xdr->head[0].iov_base);
119         } else {
120                 xdr_off -= xdr->head[0].iov_len;
121                 if (xdr_off < xdr->page_len) {
122                         /* This offset is in the page list */
123                         xdr_off += xdr->page_base;
124                         page = xdr->pages[xdr_off >> PAGE_SHIFT];
125                         xdr_off &= ~PAGE_MASK;
126                 } else {
127                         /* This offset is in the tail */
128                         xdr_off -= xdr->page_len;
129                         xdr_off += (unsigned long)
130                                 xdr->tail[0].iov_base & ~PAGE_MASK;
131                         page = virt_to_page(xdr->tail[0].iov_base);
132                 }
133         }
134         dma_addr = ib_dma_map_page(xprt->sc_cm_id->device, page, xdr_off,
135                                    min_t(size_t, PAGE_SIZE, len), dir);
136         return dma_addr;
137 }
138
139 /* Assumptions:
140  * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
141  */
142 static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
143                       u32 rmr, u64 to,
144                       u32 xdr_off, int write_len,
145                       struct svc_rdma_req_map *vec)
146 {
147         struct ib_send_wr write_wr;
148         struct ib_sge *sge;
149         int xdr_sge_no;
150         int sge_no;
151         int sge_bytes;
152         int sge_off;
153         int bc;
154         struct svc_rdma_op_ctxt *ctxt;
155
156         if (vec->count > RPCSVC_MAXPAGES) {
157                 pr_err("svcrdma: Too many pages (%lu)\n", vec->count);
158                 return -EIO;
159         }
160
161         dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, "
162                 "write_len=%d, vec->sge=%p, vec->count=%lu\n",
163                 rmr, (unsigned long long)to, xdr_off,
164                 write_len, vec->sge, vec->count);
165
166         ctxt = svc_rdma_get_context(xprt);
167         ctxt->direction = DMA_TO_DEVICE;
168         sge = ctxt->sge;
169
170         /* Find the SGE associated with xdr_off */
171         for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < vec->count;
172              xdr_sge_no++) {
173                 if (vec->sge[xdr_sge_no].iov_len > bc)
174                         break;
175                 bc -= vec->sge[xdr_sge_no].iov_len;
176         }
177
178         sge_off = bc;
179         bc = write_len;
180         sge_no = 0;
181
182         /* Copy the remaining SGE */
183         while (bc != 0) {
184                 sge_bytes = min_t(size_t,
185                           bc, vec->sge[xdr_sge_no].iov_len-sge_off);
186                 sge[sge_no].length = sge_bytes;
187                 sge[sge_no].addr =
188                         dma_map_xdr(xprt, &rqstp->rq_res, xdr_off,
189                                     sge_bytes, DMA_TO_DEVICE);
190                 xdr_off += sge_bytes;
191                 if (ib_dma_mapping_error(xprt->sc_cm_id->device,
192                                          sge[sge_no].addr))
193                         goto err;
194                 atomic_inc(&xprt->sc_dma_used);
195                 sge[sge_no].lkey = xprt->sc_dma_lkey;
196                 ctxt->count++;
197                 sge_off = 0;
198                 sge_no++;
199                 xdr_sge_no++;
200                 if (xdr_sge_no > vec->count) {
201                         pr_err("svcrdma: Too many sges (%d)\n", xdr_sge_no);
202                         goto err;
203                 }
204                 bc -= sge_bytes;
205                 if (sge_no == xprt->sc_max_sge)
206                         break;
207         }
208
209         /* Prepare WRITE WR */
210         memset(&write_wr, 0, sizeof write_wr);
211         ctxt->wr_op = IB_WR_RDMA_WRITE;
212         write_wr.wr_id = (unsigned long)ctxt;
213         write_wr.sg_list = &sge[0];
214         write_wr.num_sge = sge_no;
215         write_wr.opcode = IB_WR_RDMA_WRITE;
216         write_wr.send_flags = IB_SEND_SIGNALED;
217         write_wr.wr.rdma.rkey = rmr;
218         write_wr.wr.rdma.remote_addr = to;
219
220         /* Post It */
221         atomic_inc(&rdma_stat_write);
222         if (svc_rdma_send(xprt, &write_wr))
223                 goto err;
224         return write_len - bc;
225  err:
226         svc_rdma_unmap_dma(ctxt);
227         svc_rdma_put_context(ctxt, 0);
228         /* Fatal error, close transport */
229         return -EIO;
230 }
231
232 static int send_write_chunks(struct svcxprt_rdma *xprt,
233                              struct rpcrdma_msg *rdma_argp,
234                              struct rpcrdma_msg *rdma_resp,
235                              struct svc_rqst *rqstp,
236                              struct svc_rdma_req_map *vec)
237 {
238         u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
239         int write_len;
240         u32 xdr_off;
241         int chunk_off;
242         int chunk_no;
243         struct rpcrdma_write_array *arg_ary;
244         struct rpcrdma_write_array *res_ary;
245         int ret;
246
247         arg_ary = svc_rdma_get_write_array(rdma_argp);
248         if (!arg_ary)
249                 return 0;
250         res_ary = (struct rpcrdma_write_array *)
251                 &rdma_resp->rm_body.rm_chunks[1];
252
253         /* Write chunks start at the pagelist */
254         for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0;
255              xfer_len && chunk_no < arg_ary->wc_nchunks;
256              chunk_no++) {
257                 struct rpcrdma_segment *arg_ch;
258                 u64 rs_offset;
259
260                 arg_ch = &arg_ary->wc_array[chunk_no].wc_target;
261                 write_len = min(xfer_len, ntohl(arg_ch->rs_length));
262
263                 /* Prepare the response chunk given the length actually
264                  * written */
265                 xdr_decode_hyper((__be32 *)&arg_ch->rs_offset, &rs_offset);
266                 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
267                                                 arg_ch->rs_handle,
268                                                 arg_ch->rs_offset,
269                                                 write_len);
270                 chunk_off = 0;
271                 while (write_len) {
272                         ret = send_write(xprt, rqstp,
273                                          ntohl(arg_ch->rs_handle),
274                                          rs_offset + chunk_off,
275                                          xdr_off,
276                                          write_len,
277                                          vec);
278                         if (ret <= 0) {
279                                 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
280                                         ret);
281                                 return -EIO;
282                         }
283                         chunk_off += ret;
284                         xdr_off += ret;
285                         xfer_len -= ret;
286                         write_len -= ret;
287                 }
288         }
289         /* Update the req with the number of chunks actually used */
290         svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no);
291
292         return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
293 }
294
295 static int send_reply_chunks(struct svcxprt_rdma *xprt,
296                              struct rpcrdma_msg *rdma_argp,
297                              struct rpcrdma_msg *rdma_resp,
298                              struct svc_rqst *rqstp,
299                              struct svc_rdma_req_map *vec)
300 {
301         u32 xfer_len = rqstp->rq_res.len;
302         int write_len;
303         u32 xdr_off;
304         int chunk_no;
305         int chunk_off;
306         int nchunks;
307         struct rpcrdma_segment *ch;
308         struct rpcrdma_write_array *arg_ary;
309         struct rpcrdma_write_array *res_ary;
310         int ret;
311
312         arg_ary = svc_rdma_get_reply_array(rdma_argp);
313         if (!arg_ary)
314                 return 0;
315         /* XXX: need to fix when reply lists occur with read-list and or
316          * write-list */
317         res_ary = (struct rpcrdma_write_array *)
318                 &rdma_resp->rm_body.rm_chunks[2];
319
320         /* xdr offset starts at RPC message */
321         nchunks = ntohl(arg_ary->wc_nchunks);
322         for (xdr_off = 0, chunk_no = 0;
323              xfer_len && chunk_no < nchunks;
324              chunk_no++) {
325                 u64 rs_offset;
326                 ch = &arg_ary->wc_array[chunk_no].wc_target;
327                 write_len = min(xfer_len, htonl(ch->rs_length));
328
329                 /* Prepare the reply chunk given the length actually
330                  * written */
331                 xdr_decode_hyper((__be32 *)&ch->rs_offset, &rs_offset);
332                 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
333                                                 ch->rs_handle, ch->rs_offset,
334                                                 write_len);
335                 chunk_off = 0;
336                 while (write_len) {
337                         ret = send_write(xprt, rqstp,
338                                          ntohl(ch->rs_handle),
339                                          rs_offset + chunk_off,
340                                          xdr_off,
341                                          write_len,
342                                          vec);
343                         if (ret <= 0) {
344                                 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
345                                         ret);
346                                 return -EIO;
347                         }
348                         chunk_off += ret;
349                         xdr_off += ret;
350                         xfer_len -= ret;
351                         write_len -= ret;
352                 }
353         }
354         /* Update the req with the number of chunks actually used */
355         svc_rdma_xdr_encode_reply_array(res_ary, chunk_no);
356
357         return rqstp->rq_res.len;
358 }
359
360 /* This function prepares the portion of the RPCRDMA message to be
361  * sent in the RDMA_SEND. This function is called after data sent via
362  * RDMA has already been transmitted. There are three cases:
363  * - The RPCRDMA header, RPC header, and payload are all sent in a
364  *   single RDMA_SEND. This is the "inline" case.
365  * - The RPCRDMA header and some portion of the RPC header and data
366  *   are sent via this RDMA_SEND and another portion of the data is
367  *   sent via RDMA.
368  * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
369  *   header and data are all transmitted via RDMA.
370  * In all three cases, this function prepares the RPCRDMA header in
371  * sge[0], the 'type' parameter indicates the type to place in the
372  * RPCRDMA header, and the 'byte_count' field indicates how much of
373  * the XDR to include in this RDMA_SEND. NB: The offset of the payload
374  * to send is zero in the XDR.
375  */
376 static int send_reply(struct svcxprt_rdma *rdma,
377                       struct svc_rqst *rqstp,
378                       struct page *page,
379                       struct rpcrdma_msg *rdma_resp,
380                       struct svc_rdma_op_ctxt *ctxt,
381                       struct svc_rdma_req_map *vec,
382                       int byte_count)
383 {
384         struct ib_send_wr send_wr;
385         int sge_no;
386         int sge_bytes;
387         int page_no;
388         int pages;
389         int ret;
390
391         /* Post a recv buffer to handle another request. */
392         ret = svc_rdma_post_recv(rdma);
393         if (ret) {
394                 printk(KERN_INFO
395                        "svcrdma: could not post a receive buffer, err=%d."
396                        "Closing transport %p.\n", ret, rdma);
397                 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
398                 svc_rdma_put_context(ctxt, 0);
399                 return -ENOTCONN;
400         }
401
402         /* Prepare the context */
403         ctxt->pages[0] = page;
404         ctxt->count = 1;
405
406         /* Prepare the SGE for the RPCRDMA Header */
407         ctxt->sge[0].lkey = rdma->sc_dma_lkey;
408         ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
409         ctxt->sge[0].addr =
410             ib_dma_map_page(rdma->sc_cm_id->device, page, 0,
411                             ctxt->sge[0].length, DMA_TO_DEVICE);
412         if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr))
413                 goto err;
414         atomic_inc(&rdma->sc_dma_used);
415
416         ctxt->direction = DMA_TO_DEVICE;
417
418         /* Map the payload indicated by 'byte_count' */
419         for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) {
420                 int xdr_off = 0;
421                 sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count);
422                 byte_count -= sge_bytes;
423                 ctxt->sge[sge_no].addr =
424                         dma_map_xdr(rdma, &rqstp->rq_res, xdr_off,
425                                     sge_bytes, DMA_TO_DEVICE);
426                 xdr_off += sge_bytes;
427                 if (ib_dma_mapping_error(rdma->sc_cm_id->device,
428                                          ctxt->sge[sge_no].addr))
429                         goto err;
430                 atomic_inc(&rdma->sc_dma_used);
431                 ctxt->sge[sge_no].lkey = rdma->sc_dma_lkey;
432                 ctxt->sge[sge_no].length = sge_bytes;
433         }
434         if (byte_count != 0) {
435                 pr_err("svcrdma: Could not map %d bytes\n", byte_count);
436                 goto err;
437         }
438
439         /* Save all respages in the ctxt and remove them from the
440          * respages array. They are our pages until the I/O
441          * completes.
442          */
443         pages = rqstp->rq_next_page - rqstp->rq_respages;
444         for (page_no = 0; page_no < pages; page_no++) {
445                 ctxt->pages[page_no+1] = rqstp->rq_respages[page_no];
446                 ctxt->count++;
447                 rqstp->rq_respages[page_no] = NULL;
448                 /*
449                  * If there are more pages than SGE, terminate SGE
450                  * list so that svc_rdma_unmap_dma doesn't attempt to
451                  * unmap garbage.
452                  */
453                 if (page_no+1 >= sge_no)
454                         ctxt->sge[page_no+1].length = 0;
455         }
456         rqstp->rq_next_page = rqstp->rq_respages + 1;
457
458         if (sge_no > rdma->sc_max_sge) {
459                 pr_err("svcrdma: Too many sges (%d)\n", sge_no);
460                 goto err;
461         }
462         memset(&send_wr, 0, sizeof send_wr);
463         ctxt->wr_op = IB_WR_SEND;
464         send_wr.wr_id = (unsigned long)ctxt;
465         send_wr.sg_list = ctxt->sge;
466         send_wr.num_sge = sge_no;
467         send_wr.opcode = IB_WR_SEND;
468         send_wr.send_flags =  IB_SEND_SIGNALED;
469
470         ret = svc_rdma_send(rdma, &send_wr);
471         if (ret)
472                 goto err;
473
474         return 0;
475
476  err:
477         svc_rdma_unmap_dma(ctxt);
478         svc_rdma_put_context(ctxt, 1);
479         return -EIO;
480 }
481
482 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
483 {
484 }
485
486 int svc_rdma_sendto(struct svc_rqst *rqstp)
487 {
488         struct svc_xprt *xprt = rqstp->rq_xprt;
489         struct svcxprt_rdma *rdma =
490                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
491         struct rpcrdma_msg *rdma_argp;
492         struct rpcrdma_msg *rdma_resp;
493         struct rpcrdma_write_array *reply_ary;
494         enum rpcrdma_proc reply_type;
495         int ret;
496         int inline_bytes;
497         struct page *res_page;
498         struct svc_rdma_op_ctxt *ctxt;
499         struct svc_rdma_req_map *vec;
500
501         dprintk("svcrdma: sending response for rqstp=%p\n", rqstp);
502
503         /* Get the RDMA request header. The receive logic always
504          * places this at the start of page 0.
505          */
506         rdma_argp = page_address(rqstp->rq_pages[0]);
507
508         /* Build an req vec for the XDR */
509         ctxt = svc_rdma_get_context(rdma);
510         ctxt->direction = DMA_TO_DEVICE;
511         vec = svc_rdma_get_req_map();
512         ret = map_xdr(rdma, &rqstp->rq_res, vec);
513         if (ret)
514                 goto err0;
515         inline_bytes = rqstp->rq_res.len;
516
517         /* Create the RDMA response header */
518         res_page = svc_rdma_get_page();
519         rdma_resp = page_address(res_page);
520         reply_ary = svc_rdma_get_reply_array(rdma_argp);
521         if (reply_ary)
522                 reply_type = RDMA_NOMSG;
523         else
524                 reply_type = RDMA_MSG;
525         svc_rdma_xdr_encode_reply_header(rdma, rdma_argp,
526                                          rdma_resp, reply_type);
527
528         /* Send any write-chunk data and build resp write-list */
529         ret = send_write_chunks(rdma, rdma_argp, rdma_resp,
530                                 rqstp, vec);
531         if (ret < 0) {
532                 printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n",
533                        ret);
534                 goto err1;
535         }
536         inline_bytes -= ret;
537
538         /* Send any reply-list data and update resp reply-list */
539         ret = send_reply_chunks(rdma, rdma_argp, rdma_resp,
540                                 rqstp, vec);
541         if (ret < 0) {
542                 printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n",
543                        ret);
544                 goto err1;
545         }
546         inline_bytes -= ret;
547
548         ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec,
549                          inline_bytes);
550         svc_rdma_put_req_map(vec);
551         dprintk("svcrdma: send_reply returns %d\n", ret);
552         return ret;
553
554  err1:
555         put_page(res_page);
556  err0:
557         svc_rdma_put_req_map(vec);
558         svc_rdma_put_context(ctxt, 0);
559         return ret;
560 }