svc_rdma_sendto.c 19 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702
  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. #include <linux/sunrpc/debug.h>
  42. #include <linux/sunrpc/rpc_rdma.h>
  43. #include <linux/spinlock.h>
  44. #include <asm/unaligned.h>
  45. #include <rdma/ib_verbs.h>
  46. #include <rdma/rdma_cm.h>
  47. #include <linux/sunrpc/svc_rdma.h>
  48. #define RPCDBG_FACILITY RPCDBG_SVCXPRT
  49. /* Encode an XDR as an array of IB SGE
  50. *
  51. * Assumptions:
  52. * - head[0] is physically contiguous.
  53. * - tail[0] is physically contiguous.
  54. * - pages[] is not physically or virtually contigous and consists of
  55. * PAGE_SIZE elements.
  56. *
  57. * Output:
  58. * SGE[0] reserved for RCPRDMA header
  59. * SGE[1] data from xdr->head[]
  60. * SGE[2..sge_count-2] data from xdr->pages[]
  61. * SGE[sge_count-1] data from xdr->tail.
  62. *
  63. * The max SGE we need is the length of the XDR / pagesize + one for
  64. * head + one for tail + one for RPCRDMA header. Since RPCSVC_MAXPAGES
  65. * reserves a page for both the request and the reply header, and this
  66. * array is only concerned with the reply we are assured that we have
  67. * on extra page for the RPCRMDA header.
  68. */
  69. static int fast_reg_xdr(struct svcxprt_rdma *xprt,
  70. struct xdr_buf *xdr,
  71. struct svc_rdma_req_map *vec)
  72. {
  73. int sge_no;
  74. u32 sge_bytes;
  75. u32 page_bytes;
  76. u32 page_off;
  77. int page_no = 0;
  78. u8 *frva;
  79. struct svc_rdma_fastreg_mr *frmr;
  80. frmr = svc_rdma_get_frmr(xprt);
  81. if (IS_ERR(frmr))
  82. return -ENOMEM;
  83. vec->frmr = frmr;
  84. /* Skip the RPCRDMA header */
  85. sge_no = 1;
  86. /* Map the head. */
  87. frva = (void *)((unsigned long)(xdr->head[0].iov_base) & PAGE_MASK);
  88. vec->sge[sge_no].iov_base = xdr->head[0].iov_base;
  89. vec->sge[sge_no].iov_len = xdr->head[0].iov_len;
  90. vec->count = 2;
  91. sge_no++;
  92. /* Build the FRMR */
  93. frmr->kva = frva;
  94. frmr->direction = DMA_TO_DEVICE;
  95. frmr->access_flags = 0;
  96. frmr->map_len = PAGE_SIZE;
  97. frmr->page_list_len = 1;
  98. frmr->page_list->page_list[page_no] =
  99. ib_dma_map_single(xprt->sc_cm_id->device,
  100. (void *)xdr->head[0].iov_base,
  101. PAGE_SIZE, DMA_TO_DEVICE);
  102. if (ib_dma_mapping_error(xprt->sc_cm_id->device,
  103. frmr->page_list->page_list[page_no]))
  104. goto fatal_err;
  105. atomic_inc(&xprt->sc_dma_used);
  106. page_off = xdr->page_base;
  107. page_bytes = xdr->page_len + page_off;
  108. if (!page_bytes)
  109. goto encode_tail;
  110. /* Map the pages */
  111. vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off;
  112. vec->sge[sge_no].iov_len = page_bytes;
  113. sge_no++;
  114. while (page_bytes) {
  115. struct page *page;
  116. page = xdr->pages[page_no++];
  117. sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
  118. page_bytes -= sge_bytes;
  119. frmr->page_list->page_list[page_no] =
  120. ib_dma_map_page(xprt->sc_cm_id->device, page, 0,
  121. PAGE_SIZE, DMA_TO_DEVICE);
  122. if (ib_dma_mapping_error(xprt->sc_cm_id->device,
  123. frmr->page_list->page_list[page_no]))
  124. goto fatal_err;
  125. atomic_inc(&xprt->sc_dma_used);
  126. page_off = 0; /* reset for next time through loop */
  127. frmr->map_len += PAGE_SIZE;
  128. frmr->page_list_len++;
  129. }
  130. vec->count++;
  131. encode_tail:
  132. /* Map tail */
  133. if (0 == xdr->tail[0].iov_len)
  134. goto done;
  135. vec->count++;
  136. vec->sge[sge_no].iov_len = xdr->tail[0].iov_len;
  137. if (((unsigned long)xdr->tail[0].iov_base & PAGE_MASK) ==
  138. ((unsigned long)xdr->head[0].iov_base & PAGE_MASK)) {
  139. /*
  140. * If head and tail use the same page, we don't need
  141. * to map it again.
  142. */
  143. vec->sge[sge_no].iov_base = xdr->tail[0].iov_base;
  144. } else {
  145. void *va;
  146. /* Map another page for the tail */
  147. page_off = (unsigned long)xdr->tail[0].iov_base & ~PAGE_MASK;
  148. va = (void *)((unsigned long)xdr->tail[0].iov_base & PAGE_MASK);
  149. vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off;
  150. frmr->page_list->page_list[page_no] =
  151. ib_dma_map_single(xprt->sc_cm_id->device, va, PAGE_SIZE,
  152. DMA_TO_DEVICE);
  153. if (ib_dma_mapping_error(xprt->sc_cm_id->device,
  154. frmr->page_list->page_list[page_no]))
  155. goto fatal_err;
  156. atomic_inc(&xprt->sc_dma_used);
  157. frmr->map_len += PAGE_SIZE;
  158. frmr->page_list_len++;
  159. }
  160. done:
  161. if (svc_rdma_fastreg(xprt, frmr))
  162. goto fatal_err;
  163. return 0;
  164. fatal_err:
  165. printk("svcrdma: Error fast registering memory for xprt %p\n", xprt);
  166. svc_rdma_put_frmr(xprt, frmr);
  167. return -EIO;
  168. }
  169. static int map_xdr(struct svcxprt_rdma *xprt,
  170. struct xdr_buf *xdr,
  171. struct svc_rdma_req_map *vec)
  172. {
  173. int sge_no;
  174. u32 sge_bytes;
  175. u32 page_bytes;
  176. u32 page_off;
  177. int page_no;
  178. BUG_ON(xdr->len !=
  179. (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len));
  180. if (xprt->sc_frmr_pg_list_len)
  181. return fast_reg_xdr(xprt, xdr, vec);
  182. /* Skip the first sge, this is for the RPCRDMA header */
  183. sge_no = 1;
  184. /* Head SGE */
  185. vec->sge[sge_no].iov_base = xdr->head[0].iov_base;
  186. vec->sge[sge_no].iov_len = xdr->head[0].iov_len;
  187. sge_no++;
  188. /* pages SGE */
  189. page_no = 0;
  190. page_bytes = xdr->page_len;
  191. page_off = xdr->page_base;
  192. while (page_bytes) {
  193. vec->sge[sge_no].iov_base =
  194. page_address(xdr->pages[page_no]) + page_off;
  195. sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
  196. page_bytes -= sge_bytes;
  197. vec->sge[sge_no].iov_len = sge_bytes;
  198. sge_no++;
  199. page_no++;
  200. page_off = 0; /* reset for next time through loop */
  201. }
  202. /* Tail SGE */
  203. if (xdr->tail[0].iov_len) {
  204. vec->sge[sge_no].iov_base = xdr->tail[0].iov_base;
  205. vec->sge[sge_no].iov_len = xdr->tail[0].iov_len;
  206. sge_no++;
  207. }
  208. dprintk("svcrdma: map_xdr: sge_no %d page_no %d "
  209. "page_base %u page_len %u head_len %zu tail_len %zu\n",
  210. sge_no, page_no, xdr->page_base, xdr->page_len,
  211. xdr->head[0].iov_len, xdr->tail[0].iov_len);
  212. vec->count = sge_no;
  213. return 0;
  214. }
  215. /* Assumptions:
  216. * - We are using FRMR
  217. * - or -
  218. * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
  219. */
  220. static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
  221. u32 rmr, u64 to,
  222. u32 xdr_off, int write_len,
  223. struct svc_rdma_req_map *vec)
  224. {
  225. struct ib_send_wr write_wr;
  226. struct ib_sge *sge;
  227. int xdr_sge_no;
  228. int sge_no;
  229. int sge_bytes;
  230. int sge_off;
  231. int bc;
  232. struct svc_rdma_op_ctxt *ctxt;
  233. BUG_ON(vec->count > RPCSVC_MAXPAGES);
  234. dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, "
  235. "write_len=%d, vec->sge=%p, vec->count=%lu\n",
  236. rmr, (unsigned long long)to, xdr_off,
  237. write_len, vec->sge, vec->count);
  238. ctxt = svc_rdma_get_context(xprt);
  239. ctxt->direction = DMA_TO_DEVICE;
  240. sge = ctxt->sge;
  241. /* Find the SGE associated with xdr_off */
  242. for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < vec->count;
  243. xdr_sge_no++) {
  244. if (vec->sge[xdr_sge_no].iov_len > bc)
  245. break;
  246. bc -= vec->sge[xdr_sge_no].iov_len;
  247. }
  248. sge_off = bc;
  249. bc = write_len;
  250. sge_no = 0;
  251. /* Copy the remaining SGE */
  252. while (bc != 0) {
  253. sge_bytes = min_t(size_t,
  254. bc, vec->sge[xdr_sge_no].iov_len-sge_off);
  255. sge[sge_no].length = sge_bytes;
  256. if (!vec->frmr) {
  257. sge[sge_no].addr =
  258. ib_dma_map_single(xprt->sc_cm_id->device,
  259. (void *)
  260. vec->sge[xdr_sge_no].iov_base + sge_off,
  261. sge_bytes, DMA_TO_DEVICE);
  262. if (ib_dma_mapping_error(xprt->sc_cm_id->device,
  263. sge[sge_no].addr))
  264. goto err;
  265. atomic_inc(&xprt->sc_dma_used);
  266. sge[sge_no].lkey = xprt->sc_dma_lkey;
  267. } else {
  268. sge[sge_no].addr = (unsigned long)
  269. vec->sge[xdr_sge_no].iov_base + sge_off;
  270. sge[sge_no].lkey = vec->frmr->mr->lkey;
  271. }
  272. ctxt->count++;
  273. ctxt->frmr = vec->frmr;
  274. sge_off = 0;
  275. sge_no++;
  276. xdr_sge_no++;
  277. BUG_ON(xdr_sge_no > vec->count);
  278. bc -= sge_bytes;
  279. }
  280. /* Prepare WRITE WR */
  281. memset(&write_wr, 0, sizeof write_wr);
  282. ctxt->wr_op = IB_WR_RDMA_WRITE;
  283. write_wr.wr_id = (unsigned long)ctxt;
  284. write_wr.sg_list = &sge[0];
  285. write_wr.num_sge = sge_no;
  286. write_wr.opcode = IB_WR_RDMA_WRITE;
  287. write_wr.send_flags = IB_SEND_SIGNALED;
  288. write_wr.wr.rdma.rkey = rmr;
  289. write_wr.wr.rdma.remote_addr = to;
  290. /* Post It */
  291. atomic_inc(&rdma_stat_write);
  292. if (svc_rdma_send(xprt, &write_wr))
  293. goto err;
  294. return 0;
  295. err:
  296. svc_rdma_put_context(ctxt, 0);
  297. /* Fatal error, close transport */
  298. return -EIO;
  299. }
  300. static int send_write_chunks(struct svcxprt_rdma *xprt,
  301. struct rpcrdma_msg *rdma_argp,
  302. struct rpcrdma_msg *rdma_resp,
  303. struct svc_rqst *rqstp,
  304. struct svc_rdma_req_map *vec)
  305. {
  306. u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
  307. int write_len;
  308. int max_write;
  309. u32 xdr_off;
  310. int chunk_off;
  311. int chunk_no;
  312. struct rpcrdma_write_array *arg_ary;
  313. struct rpcrdma_write_array *res_ary;
  314. int ret;
  315. arg_ary = svc_rdma_get_write_array(rdma_argp);
  316. if (!arg_ary)
  317. return 0;
  318. res_ary = (struct rpcrdma_write_array *)
  319. &rdma_resp->rm_body.rm_chunks[1];
  320. if (vec->frmr)
  321. max_write = vec->frmr->map_len;
  322. else
  323. max_write = xprt->sc_max_sge * PAGE_SIZE;
  324. /* Write chunks start at the pagelist */
  325. for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0;
  326. xfer_len && chunk_no < arg_ary->wc_nchunks;
  327. chunk_no++) {
  328. struct rpcrdma_segment *arg_ch;
  329. u64 rs_offset;
  330. arg_ch = &arg_ary->wc_array[chunk_no].wc_target;
  331. write_len = min(xfer_len, arg_ch->rs_length);
  332. /* Prepare the response chunk given the length actually
  333. * written */
  334. rs_offset = get_unaligned(&(arg_ch->rs_offset));
  335. svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
  336. arg_ch->rs_handle,
  337. rs_offset,
  338. write_len);
  339. chunk_off = 0;
  340. while (write_len) {
  341. int this_write;
  342. this_write = min(write_len, max_write);
  343. ret = send_write(xprt, rqstp,
  344. arg_ch->rs_handle,
  345. rs_offset + chunk_off,
  346. xdr_off,
  347. this_write,
  348. vec);
  349. if (ret) {
  350. dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
  351. ret);
  352. return -EIO;
  353. }
  354. chunk_off += this_write;
  355. xdr_off += this_write;
  356. xfer_len -= this_write;
  357. write_len -= this_write;
  358. }
  359. }
  360. /* Update the req with the number of chunks actually used */
  361. svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no);
  362. return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
  363. }
  364. static int send_reply_chunks(struct svcxprt_rdma *xprt,
  365. struct rpcrdma_msg *rdma_argp,
  366. struct rpcrdma_msg *rdma_resp,
  367. struct svc_rqst *rqstp,
  368. struct svc_rdma_req_map *vec)
  369. {
  370. u32 xfer_len = rqstp->rq_res.len;
  371. int write_len;
  372. int max_write;
  373. u32 xdr_off;
  374. int chunk_no;
  375. int chunk_off;
  376. struct rpcrdma_segment *ch;
  377. struct rpcrdma_write_array *arg_ary;
  378. struct rpcrdma_write_array *res_ary;
  379. int ret;
  380. arg_ary = svc_rdma_get_reply_array(rdma_argp);
  381. if (!arg_ary)
  382. return 0;
  383. /* XXX: need to fix when reply lists occur with read-list and or
  384. * write-list */
  385. res_ary = (struct rpcrdma_write_array *)
  386. &rdma_resp->rm_body.rm_chunks[2];
  387. if (vec->frmr)
  388. max_write = vec->frmr->map_len;
  389. else
  390. max_write = xprt->sc_max_sge * PAGE_SIZE;
  391. /* xdr offset starts at RPC message */
  392. for (xdr_off = 0, chunk_no = 0;
  393. xfer_len && chunk_no < arg_ary->wc_nchunks;
  394. chunk_no++) {
  395. u64 rs_offset;
  396. ch = &arg_ary->wc_array[chunk_no].wc_target;
  397. write_len = min(xfer_len, ch->rs_length);
  398. /* Prepare the reply chunk given the length actually
  399. * written */
  400. rs_offset = get_unaligned(&(ch->rs_offset));
  401. svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
  402. ch->rs_handle, rs_offset,
  403. write_len);
  404. chunk_off = 0;
  405. while (write_len) {
  406. int this_write;
  407. this_write = min(write_len, max_write);
  408. ret = send_write(xprt, rqstp,
  409. ch->rs_handle,
  410. rs_offset + chunk_off,
  411. xdr_off,
  412. this_write,
  413. vec);
  414. if (ret) {
  415. dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
  416. ret);
  417. return -EIO;
  418. }
  419. chunk_off += this_write;
  420. xdr_off += this_write;
  421. xfer_len -= this_write;
  422. write_len -= this_write;
  423. }
  424. }
  425. /* Update the req with the number of chunks actually used */
  426. svc_rdma_xdr_encode_reply_array(res_ary, chunk_no);
  427. return rqstp->rq_res.len;
  428. }
  429. /* This function prepares the portion of the RPCRDMA message to be
  430. * sent in the RDMA_SEND. This function is called after data sent via
  431. * RDMA has already been transmitted. There are three cases:
  432. * - The RPCRDMA header, RPC header, and payload are all sent in a
  433. * single RDMA_SEND. This is the "inline" case.
  434. * - The RPCRDMA header and some portion of the RPC header and data
  435. * are sent via this RDMA_SEND and another portion of the data is
  436. * sent via RDMA.
  437. * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
  438. * header and data are all transmitted via RDMA.
  439. * In all three cases, this function prepares the RPCRDMA header in
  440. * sge[0], the 'type' parameter indicates the type to place in the
  441. * RPCRDMA header, and the 'byte_count' field indicates how much of
  442. * the XDR to include in this RDMA_SEND.
  443. */
  444. static int send_reply(struct svcxprt_rdma *rdma,
  445. struct svc_rqst *rqstp,
  446. struct page *page,
  447. struct rpcrdma_msg *rdma_resp,
  448. struct svc_rdma_op_ctxt *ctxt,
  449. struct svc_rdma_req_map *vec,
  450. int byte_count)
  451. {
  452. struct ib_send_wr send_wr;
  453. struct ib_send_wr inv_wr;
  454. int sge_no;
  455. int sge_bytes;
  456. int page_no;
  457. int ret;
  458. /* Post a recv buffer to handle another request. */
  459. ret = svc_rdma_post_recv(rdma);
  460. if (ret) {
  461. printk(KERN_INFO
  462. "svcrdma: could not post a receive buffer, err=%d."
  463. "Closing transport %p.\n", ret, rdma);
  464. set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
  465. svc_rdma_put_context(ctxt, 0);
  466. return -ENOTCONN;
  467. }
  468. /* Prepare the context */
  469. ctxt->pages[0] = page;
  470. ctxt->count = 1;
  471. ctxt->frmr = vec->frmr;
  472. if (vec->frmr)
  473. set_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
  474. else
  475. clear_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
  476. /* Prepare the SGE for the RPCRDMA Header */
  477. ctxt->sge[0].addr =
  478. ib_dma_map_page(rdma->sc_cm_id->device,
  479. page, 0, PAGE_SIZE, DMA_TO_DEVICE);
  480. if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr))
  481. goto err;
  482. atomic_inc(&rdma->sc_dma_used);
  483. ctxt->direction = DMA_TO_DEVICE;
  484. ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
  485. ctxt->sge[0].lkey = rdma->sc_dma_lkey;
  486. /* Determine how many of our SGE are to be transmitted */
  487. for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) {
  488. sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count);
  489. byte_count -= sge_bytes;
  490. if (!vec->frmr) {
  491. ctxt->sge[sge_no].addr =
  492. ib_dma_map_single(rdma->sc_cm_id->device,
  493. vec->sge[sge_no].iov_base,
  494. sge_bytes, DMA_TO_DEVICE);
  495. if (ib_dma_mapping_error(rdma->sc_cm_id->device,
  496. ctxt->sge[sge_no].addr))
  497. goto err;
  498. atomic_inc(&rdma->sc_dma_used);
  499. ctxt->sge[sge_no].lkey = rdma->sc_dma_lkey;
  500. } else {
  501. ctxt->sge[sge_no].addr = (unsigned long)
  502. vec->sge[sge_no].iov_base;
  503. ctxt->sge[sge_no].lkey = vec->frmr->mr->lkey;
  504. }
  505. ctxt->sge[sge_no].length = sge_bytes;
  506. }
  507. BUG_ON(byte_count != 0);
  508. /* Save all respages in the ctxt and remove them from the
  509. * respages array. They are our pages until the I/O
  510. * completes.
  511. */
  512. for (page_no = 0; page_no < rqstp->rq_resused; page_no++) {
  513. ctxt->pages[page_no+1] = rqstp->rq_respages[page_no];
  514. ctxt->count++;
  515. rqstp->rq_respages[page_no] = NULL;
  516. /*
  517. * If there are more pages than SGE, terminate SGE
  518. * list so that svc_rdma_unmap_dma doesn't attempt to
  519. * unmap garbage.
  520. */
  521. if (page_no+1 >= sge_no)
  522. ctxt->sge[page_no+1].length = 0;
  523. }
  524. BUG_ON(sge_no > rdma->sc_max_sge);
  525. memset(&send_wr, 0, sizeof send_wr);
  526. ctxt->wr_op = IB_WR_SEND;
  527. send_wr.wr_id = (unsigned long)ctxt;
  528. send_wr.sg_list = ctxt->sge;
  529. send_wr.num_sge = sge_no;
  530. send_wr.opcode = IB_WR_SEND;
  531. send_wr.send_flags = IB_SEND_SIGNALED;
  532. if (vec->frmr) {
  533. /* Prepare INVALIDATE WR */
  534. memset(&inv_wr, 0, sizeof inv_wr);
  535. inv_wr.opcode = IB_WR_LOCAL_INV;
  536. inv_wr.send_flags = IB_SEND_SIGNALED;
  537. inv_wr.ex.invalidate_rkey =
  538. vec->frmr->mr->lkey;
  539. send_wr.next = &inv_wr;
  540. }
  541. ret = svc_rdma_send(rdma, &send_wr);
  542. if (ret)
  543. goto err;
  544. return 0;
  545. err:
  546. svc_rdma_put_frmr(rdma, vec->frmr);
  547. svc_rdma_put_context(ctxt, 1);
  548. return -EIO;
  549. }
  550. void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
  551. {
  552. }
  553. /*
  554. * Return the start of an xdr buffer.
  555. */
  556. static void *xdr_start(struct xdr_buf *xdr)
  557. {
  558. return xdr->head[0].iov_base -
  559. (xdr->len -
  560. xdr->page_len -
  561. xdr->tail[0].iov_len -
  562. xdr->head[0].iov_len);
  563. }
  564. int svc_rdma_sendto(struct svc_rqst *rqstp)
  565. {
  566. struct svc_xprt *xprt = rqstp->rq_xprt;
  567. struct svcxprt_rdma *rdma =
  568. container_of(xprt, struct svcxprt_rdma, sc_xprt);
  569. struct rpcrdma_msg *rdma_argp;
  570. struct rpcrdma_msg *rdma_resp;
  571. struct rpcrdma_write_array *reply_ary;
  572. enum rpcrdma_proc reply_type;
  573. int ret;
  574. int inline_bytes;
  575. struct page *res_page;
  576. struct svc_rdma_op_ctxt *ctxt;
  577. struct svc_rdma_req_map *vec;
  578. dprintk("svcrdma: sending response for rqstp=%p\n", rqstp);
  579. /* Get the RDMA request header. */
  580. rdma_argp = xdr_start(&rqstp->rq_arg);
  581. /* Build an req vec for the XDR */
  582. ctxt = svc_rdma_get_context(rdma);
  583. ctxt->direction = DMA_TO_DEVICE;
  584. vec = svc_rdma_get_req_map();
  585. ret = map_xdr(rdma, &rqstp->rq_res, vec);
  586. if (ret)
  587. goto err0;
  588. inline_bytes = rqstp->rq_res.len;
  589. /* Create the RDMA response header */
  590. res_page = svc_rdma_get_page();
  591. rdma_resp = page_address(res_page);
  592. reply_ary = svc_rdma_get_reply_array(rdma_argp);
  593. if (reply_ary)
  594. reply_type = RDMA_NOMSG;
  595. else
  596. reply_type = RDMA_MSG;
  597. svc_rdma_xdr_encode_reply_header(rdma, rdma_argp,
  598. rdma_resp, reply_type);
  599. /* Send any write-chunk data and build resp write-list */
  600. ret = send_write_chunks(rdma, rdma_argp, rdma_resp,
  601. rqstp, vec);
  602. if (ret < 0) {
  603. printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n",
  604. ret);
  605. goto err1;
  606. }
  607. inline_bytes -= ret;
  608. /* Send any reply-list data and update resp reply-list */
  609. ret = send_reply_chunks(rdma, rdma_argp, rdma_resp,
  610. rqstp, vec);
  611. if (ret < 0) {
  612. printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n",
  613. ret);
  614. goto err1;
  615. }
  616. inline_bytes -= ret;
  617. ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec,
  618. inline_bytes);
  619. svc_rdma_put_req_map(vec);
  620. dprintk("svcrdma: send_reply returns %d\n", ret);
  621. return ret;
  622. err1:
  623. put_page(res_page);
  624. err0:
  625. svc_rdma_put_req_map(vec);
  626. svc_rdma_put_context(ctxt, 0);
  627. return ret;
  628. }