2 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
3 * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
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
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
15 * Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
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.
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
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.
40 * Author: Tom Tucker <tom@opengridcomputing.com>
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>
51 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
53 static int map_xdr(struct svcxprt_rdma *xprt,
55 struct svc_rdma_req_map *vec)
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");
69 /* Skip the first sge, this is for the RPCRDMA header */
73 vec->sge[sge_no].iov_base = xdr->head[0].iov_base;
74 vec->sge[sge_no].iov_len = xdr->head[0].iov_len;
79 page_bytes = xdr->page_len;
80 page_off = xdr->page_base;
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;
90 page_off = 0; /* reset for next time through loop */
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;
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);
109 static dma_addr_t dma_map_xdr(struct svcxprt_rdma *xprt,
111 u32 xdr_off, size_t len, int dir)
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);
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;
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);
134 dma_addr = ib_dma_map_page(xprt->sc_cm_id->device, page, xdr_off,
135 min_t(size_t, PAGE_SIZE, len), dir);
140 * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
142 static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
144 u32 xdr_off, int write_len,
145 struct svc_rdma_req_map *vec)
147 struct ib_send_wr write_wr;
154 struct svc_rdma_op_ctxt *ctxt;
156 if (vec->count > RPCSVC_MAXPAGES) {
157 pr_err("svcrdma: Too many pages (%lu)\n", vec->count);
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);
166 ctxt = svc_rdma_get_context(xprt);
167 ctxt->direction = DMA_TO_DEVICE;
170 /* Find the SGE associated with xdr_off */
171 for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < vec->count;
173 if (vec->sge[xdr_sge_no].iov_len > bc)
175 bc -= vec->sge[xdr_sge_no].iov_len;
182 /* Copy the remaining SGE */
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;
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,
194 atomic_inc(&xprt->sc_dma_used);
195 sge[sge_no].lkey = xprt->sc_dma_lkey;
200 if (xdr_sge_no > vec->count) {
201 pr_err("svcrdma: Too many sges (%d)\n", xdr_sge_no);
205 if (sge_no == xprt->sc_max_sge)
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;
221 atomic_inc(&rdma_stat_write);
222 if (svc_rdma_send(xprt, &write_wr))
224 return write_len - bc;
226 svc_rdma_unmap_dma(ctxt);
227 svc_rdma_put_context(ctxt, 0);
228 /* Fatal error, close transport */
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)
238 u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
243 struct rpcrdma_write_array *arg_ary;
244 struct rpcrdma_write_array *res_ary;
247 arg_ary = svc_rdma_get_write_array(rdma_argp);
250 res_ary = (struct rpcrdma_write_array *)
251 &rdma_resp->rm_body.rm_chunks[1];
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;
257 struct rpcrdma_segment *arg_ch;
260 arg_ch = &arg_ary->wc_array[chunk_no].wc_target;
261 write_len = min(xfer_len, ntohl(arg_ch->rs_length));
263 /* Prepare the response chunk given the length actually
265 xdr_decode_hyper((__be32 *)&arg_ch->rs_offset, &rs_offset);
266 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
272 ret = send_write(xprt, rqstp,
273 ntohl(arg_ch->rs_handle),
274 rs_offset + chunk_off,
279 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
289 /* Update the req with the number of chunks actually used */
290 svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no);
292 return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
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)
301 u32 xfer_len = rqstp->rq_res.len;
307 struct rpcrdma_segment *ch;
308 struct rpcrdma_write_array *arg_ary;
309 struct rpcrdma_write_array *res_ary;
312 arg_ary = svc_rdma_get_reply_array(rdma_argp);
315 /* XXX: need to fix when reply lists occur with read-list and or
317 res_ary = (struct rpcrdma_write_array *)
318 &rdma_resp->rm_body.rm_chunks[2];
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;
326 ch = &arg_ary->wc_array[chunk_no].wc_target;
327 write_len = min(xfer_len, htonl(ch->rs_length));
329 /* Prepare the reply chunk given the length actually
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,
337 ret = send_write(xprt, rqstp,
338 ntohl(ch->rs_handle),
339 rs_offset + chunk_off,
344 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
354 /* Update the req with the number of chunks actually used */
355 svc_rdma_xdr_encode_reply_array(res_ary, chunk_no);
357 return rqstp->rq_res.len;
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
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.
376 static int send_reply(struct svcxprt_rdma *rdma,
377 struct svc_rqst *rqstp,
379 struct rpcrdma_msg *rdma_resp,
380 struct svc_rdma_op_ctxt *ctxt,
381 struct svc_rdma_req_map *vec,
384 struct ib_send_wr send_wr;
391 /* Post a recv buffer to handle another request. */
392 ret = svc_rdma_post_recv(rdma);
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);
402 /* Prepare the context */
403 ctxt->pages[0] = page;
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);
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))
414 atomic_inc(&rdma->sc_dma_used);
416 ctxt->direction = DMA_TO_DEVICE;
418 /* Map the payload indicated by 'byte_count' */
419 for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) {
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))
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;
434 if (byte_count != 0) {
435 pr_err("svcrdma: Could not map %d bytes\n", byte_count);
439 /* Save all respages in the ctxt and remove them from the
440 * respages array. They are our pages until the I/O
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];
447 rqstp->rq_respages[page_no] = NULL;
449 * If there are more pages than SGE, terminate SGE
450 * list so that svc_rdma_unmap_dma doesn't attempt to
453 if (page_no+1 >= sge_no)
454 ctxt->sge[page_no+1].length = 0;
456 rqstp->rq_next_page = rqstp->rq_respages + 1;
458 if (sge_no > rdma->sc_max_sge) {
459 pr_err("svcrdma: Too many sges (%d)\n", sge_no);
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;
470 ret = svc_rdma_send(rdma, &send_wr);
477 svc_rdma_unmap_dma(ctxt);
478 svc_rdma_put_context(ctxt, 1);
482 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
486 int svc_rdma_sendto(struct svc_rqst *rqstp)
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;
497 struct page *res_page;
498 struct svc_rdma_op_ctxt *ctxt;
499 struct svc_rdma_req_map *vec;
501 dprintk("svcrdma: sending response for rqstp=%p\n", rqstp);
503 /* Get the RDMA request header. The receive logic always
504 * places this at the start of page 0.
506 rdma_argp = page_address(rqstp->rq_pages[0]);
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);
515 inline_bytes = rqstp->rq_res.len;
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);
522 reply_type = RDMA_NOMSG;
524 reply_type = RDMA_MSG;
525 svc_rdma_xdr_encode_reply_header(rdma, rdma_argp,
526 rdma_resp, reply_type);
528 /* Send any write-chunk data and build resp write-list */
529 ret = send_write_chunks(rdma, rdma_argp, rdma_resp,
532 printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n",
538 /* Send any reply-list data and update resp reply-list */
539 ret = send_reply_chunks(rdma, rdma_argp, rdma_resp,
542 printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n",
548 ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec,
550 svc_rdma_put_req_map(vec);
551 dprintk("svcrdma: send_reply returns %d\n", ret);
557 svc_rdma_put_req_map(vec);
558 svc_rdma_put_context(ctxt, 0);