2 * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
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
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
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.
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
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.
39 * Author: Tom Tucker <tom@opengridcomputing.com>
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>
50 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
52 /* Encode an XDR as an array of IB SGE
55 * - head[0] is physically contiguous.
56 * - tail[0] is physically contiguous.
57 * - pages[] is not physically or virtually contiguous and consists of
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.
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.
72 static int fast_reg_xdr(struct svcxprt_rdma *xprt,
74 struct svc_rdma_req_map *vec)
82 struct svc_rdma_fastreg_mr *frmr;
84 frmr = svc_rdma_get_frmr(xprt);
89 /* Skip the RPCRDMA header */
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;
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]))
112 atomic_inc(&xprt->sc_dma_used);
114 page_off = xdr->page_base;
115 page_bytes = xdr->page_len + page_off;
120 vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off;
121 vec->sge[sge_no].iov_len = page_bytes;
126 page = xdr->pages[page_no++];
127 sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
128 page_bytes -= sge_bytes;
130 frmr->page_list->page_list[page_no] =
131 ib_dma_map_single(xprt->sc_cm_id->device,
133 PAGE_SIZE, DMA_TO_DEVICE);
134 if (ib_dma_mapping_error(xprt->sc_cm_id->device,
135 frmr->page_list->page_list[page_no]))
138 atomic_inc(&xprt->sc_dma_used);
139 page_off = 0; /* reset for next time through loop */
140 frmr->map_len += PAGE_SIZE;
141 frmr->page_list_len++;
147 if (0 == xdr->tail[0].iov_len)
151 vec->sge[sge_no].iov_len = xdr->tail[0].iov_len;
153 if (((unsigned long)xdr->tail[0].iov_base & PAGE_MASK) ==
154 ((unsigned long)xdr->head[0].iov_base & PAGE_MASK)) {
156 * If head and tail use the same page, we don't need
159 vec->sge[sge_no].iov_base = xdr->tail[0].iov_base;
163 /* Map another page for the tail */
164 page_off = (unsigned long)xdr->tail[0].iov_base & ~PAGE_MASK;
165 va = (void *)((unsigned long)xdr->tail[0].iov_base & PAGE_MASK);
166 vec->sge[sge_no].iov_base = frva + frmr->map_len + page_off;
168 frmr->page_list->page_list[page_no] =
169 ib_dma_map_single(xprt->sc_cm_id->device, va, PAGE_SIZE,
171 if (ib_dma_mapping_error(xprt->sc_cm_id->device,
172 frmr->page_list->page_list[page_no]))
174 atomic_inc(&xprt->sc_dma_used);
175 frmr->map_len += PAGE_SIZE;
176 frmr->page_list_len++;
180 if (svc_rdma_fastreg(xprt, frmr))
186 printk("svcrdma: Error fast registering memory for xprt %p\n", xprt);
188 svc_rdma_put_frmr(xprt, frmr);
192 static int map_xdr(struct svcxprt_rdma *xprt,
194 struct svc_rdma_req_map *vec)
203 (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len));
205 if (xprt->sc_frmr_pg_list_len)
206 return fast_reg_xdr(xprt, xdr, vec);
208 /* Skip the first sge, this is for the RPCRDMA header */
212 vec->sge[sge_no].iov_base = xdr->head[0].iov_base;
213 vec->sge[sge_no].iov_len = xdr->head[0].iov_len;
218 page_bytes = xdr->page_len;
219 page_off = xdr->page_base;
221 vec->sge[sge_no].iov_base =
222 page_address(xdr->pages[page_no]) + page_off;
223 sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off));
224 page_bytes -= sge_bytes;
225 vec->sge[sge_no].iov_len = sge_bytes;
229 page_off = 0; /* reset for next time through loop */
233 if (xdr->tail[0].iov_len) {
234 vec->sge[sge_no].iov_base = xdr->tail[0].iov_base;
235 vec->sge[sge_no].iov_len = xdr->tail[0].iov_len;
239 dprintk("svcrdma: map_xdr: sge_no %d page_no %d "
240 "page_base %u page_len %u head_len %zu tail_len %zu\n",
241 sge_no, page_no, xdr->page_base, xdr->page_len,
242 xdr->head[0].iov_len, xdr->tail[0].iov_len);
249 * - We are using FRMR
251 * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE
253 static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp,
255 u32 xdr_off, int write_len,
256 struct svc_rdma_req_map *vec)
258 struct ib_send_wr write_wr;
265 struct svc_rdma_op_ctxt *ctxt;
267 BUG_ON(vec->count > RPCSVC_MAXPAGES);
268 dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, "
269 "write_len=%d, vec->sge=%p, vec->count=%lu\n",
270 rmr, (unsigned long long)to, xdr_off,
271 write_len, vec->sge, vec->count);
273 ctxt = svc_rdma_get_context(xprt);
274 ctxt->direction = DMA_TO_DEVICE;
277 /* Find the SGE associated with xdr_off */
278 for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < vec->count;
280 if (vec->sge[xdr_sge_no].iov_len > bc)
282 bc -= vec->sge[xdr_sge_no].iov_len;
289 /* Copy the remaining SGE */
291 sge_bytes = min_t(size_t,
292 bc, vec->sge[xdr_sge_no].iov_len-sge_off);
293 sge[sge_no].length = sge_bytes;
296 ib_dma_map_single(xprt->sc_cm_id->device,
298 vec->sge[xdr_sge_no].iov_base + sge_off,
299 sge_bytes, DMA_TO_DEVICE);
300 if (ib_dma_mapping_error(xprt->sc_cm_id->device,
303 atomic_inc(&xprt->sc_dma_used);
304 sge[sge_no].lkey = xprt->sc_dma_lkey;
306 sge[sge_no].addr = (unsigned long)
307 vec->sge[xdr_sge_no].iov_base + sge_off;
308 sge[sge_no].lkey = vec->frmr->mr->lkey;
311 ctxt->frmr = vec->frmr;
315 BUG_ON(xdr_sge_no > vec->count);
319 /* Prepare WRITE WR */
320 memset(&write_wr, 0, sizeof write_wr);
321 ctxt->wr_op = IB_WR_RDMA_WRITE;
322 write_wr.wr_id = (unsigned long)ctxt;
323 write_wr.sg_list = &sge[0];
324 write_wr.num_sge = sge_no;
325 write_wr.opcode = IB_WR_RDMA_WRITE;
326 write_wr.send_flags = IB_SEND_SIGNALED;
327 write_wr.wr.rdma.rkey = rmr;
328 write_wr.wr.rdma.remote_addr = to;
331 atomic_inc(&rdma_stat_write);
332 if (svc_rdma_send(xprt, &write_wr))
336 svc_rdma_put_context(ctxt, 0);
337 /* Fatal error, close transport */
341 static int send_write_chunks(struct svcxprt_rdma *xprt,
342 struct rpcrdma_msg *rdma_argp,
343 struct rpcrdma_msg *rdma_resp,
344 struct svc_rqst *rqstp,
345 struct svc_rdma_req_map *vec)
347 u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
353 struct rpcrdma_write_array *arg_ary;
354 struct rpcrdma_write_array *res_ary;
357 arg_ary = svc_rdma_get_write_array(rdma_argp);
360 res_ary = (struct rpcrdma_write_array *)
361 &rdma_resp->rm_body.rm_chunks[1];
364 max_write = vec->frmr->map_len;
366 max_write = xprt->sc_max_sge * PAGE_SIZE;
368 /* Write chunks start at the pagelist */
369 for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0;
370 xfer_len && chunk_no < arg_ary->wc_nchunks;
372 struct rpcrdma_segment *arg_ch;
375 arg_ch = &arg_ary->wc_array[chunk_no].wc_target;
376 write_len = min(xfer_len, arg_ch->rs_length);
378 /* Prepare the response chunk given the length actually
380 rs_offset = get_unaligned(&(arg_ch->rs_offset));
381 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
388 this_write = min(write_len, max_write);
389 ret = send_write(xprt, rqstp,
391 rs_offset + chunk_off,
396 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
400 chunk_off += this_write;
401 xdr_off += this_write;
402 xfer_len -= this_write;
403 write_len -= this_write;
406 /* Update the req with the number of chunks actually used */
407 svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no);
409 return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
412 static int send_reply_chunks(struct svcxprt_rdma *xprt,
413 struct rpcrdma_msg *rdma_argp,
414 struct rpcrdma_msg *rdma_resp,
415 struct svc_rqst *rqstp,
416 struct svc_rdma_req_map *vec)
418 u32 xfer_len = rqstp->rq_res.len;
424 struct rpcrdma_segment *ch;
425 struct rpcrdma_write_array *arg_ary;
426 struct rpcrdma_write_array *res_ary;
429 arg_ary = svc_rdma_get_reply_array(rdma_argp);
432 /* XXX: need to fix when reply lists occur with read-list and or
434 res_ary = (struct rpcrdma_write_array *)
435 &rdma_resp->rm_body.rm_chunks[2];
438 max_write = vec->frmr->map_len;
440 max_write = xprt->sc_max_sge * PAGE_SIZE;
442 /* xdr offset starts at RPC message */
443 for (xdr_off = 0, chunk_no = 0;
444 xfer_len && chunk_no < arg_ary->wc_nchunks;
447 ch = &arg_ary->wc_array[chunk_no].wc_target;
448 write_len = min(xfer_len, ch->rs_length);
450 /* Prepare the reply chunk given the length actually
452 rs_offset = get_unaligned(&(ch->rs_offset));
453 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
454 ch->rs_handle, rs_offset,
460 this_write = min(write_len, max_write);
461 ret = send_write(xprt, rqstp,
463 rs_offset + chunk_off,
468 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
472 chunk_off += this_write;
473 xdr_off += this_write;
474 xfer_len -= this_write;
475 write_len -= this_write;
478 /* Update the req with the number of chunks actually used */
479 svc_rdma_xdr_encode_reply_array(res_ary, chunk_no);
481 return rqstp->rq_res.len;
484 /* This function prepares the portion of the RPCRDMA message to be
485 * sent in the RDMA_SEND. This function is called after data sent via
486 * RDMA has already been transmitted. There are three cases:
487 * - The RPCRDMA header, RPC header, and payload are all sent in a
488 * single RDMA_SEND. This is the "inline" case.
489 * - The RPCRDMA header and some portion of the RPC header and data
490 * are sent via this RDMA_SEND and another portion of the data is
492 * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
493 * header and data are all transmitted via RDMA.
494 * In all three cases, this function prepares the RPCRDMA header in
495 * sge[0], the 'type' parameter indicates the type to place in the
496 * RPCRDMA header, and the 'byte_count' field indicates how much of
497 * the XDR to include in this RDMA_SEND.
499 static int send_reply(struct svcxprt_rdma *rdma,
500 struct svc_rqst *rqstp,
502 struct rpcrdma_msg *rdma_resp,
503 struct svc_rdma_op_ctxt *ctxt,
504 struct svc_rdma_req_map *vec,
507 struct ib_send_wr send_wr;
508 struct ib_send_wr inv_wr;
514 /* Post a recv buffer to handle another request. */
515 ret = svc_rdma_post_recv(rdma);
518 "svcrdma: could not post a receive buffer, err=%d."
519 "Closing transport %p.\n", ret, rdma);
520 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
521 svc_rdma_put_frmr(rdma, vec->frmr);
522 svc_rdma_put_context(ctxt, 0);
526 /* Prepare the context */
527 ctxt->pages[0] = page;
529 ctxt->frmr = vec->frmr;
531 set_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
533 clear_bit(RDMACTXT_F_FAST_UNREG, &ctxt->flags);
535 /* Prepare the SGE for the RPCRDMA Header */
536 ctxt->sge[0].lkey = rdma->sc_dma_lkey;
537 ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
539 ib_dma_map_single(rdma->sc_cm_id->device, page_address(page),
540 ctxt->sge[0].length, DMA_TO_DEVICE);
541 if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr))
543 atomic_inc(&rdma->sc_dma_used);
545 ctxt->direction = DMA_TO_DEVICE;
547 /* Determine how many of our SGE are to be transmitted */
548 for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) {
549 sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count);
550 byte_count -= sge_bytes;
552 ctxt->sge[sge_no].addr =
553 ib_dma_map_single(rdma->sc_cm_id->device,
554 vec->sge[sge_no].iov_base,
555 sge_bytes, DMA_TO_DEVICE);
556 if (ib_dma_mapping_error(rdma->sc_cm_id->device,
557 ctxt->sge[sge_no].addr))
559 atomic_inc(&rdma->sc_dma_used);
560 ctxt->sge[sge_no].lkey = rdma->sc_dma_lkey;
562 ctxt->sge[sge_no].addr = (unsigned long)
563 vec->sge[sge_no].iov_base;
564 ctxt->sge[sge_no].lkey = vec->frmr->mr->lkey;
566 ctxt->sge[sge_no].length = sge_bytes;
568 BUG_ON(byte_count != 0);
570 /* Save all respages in the ctxt and remove them from the
571 * respages array. They are our pages until the I/O
574 for (page_no = 0; page_no < rqstp->rq_resused; page_no++) {
575 ctxt->pages[page_no+1] = rqstp->rq_respages[page_no];
577 rqstp->rq_respages[page_no] = NULL;
579 * If there are more pages than SGE, terminate SGE
580 * list so that svc_rdma_unmap_dma doesn't attempt to
583 if (page_no+1 >= sge_no)
584 ctxt->sge[page_no+1].length = 0;
586 BUG_ON(sge_no > rdma->sc_max_sge);
587 memset(&send_wr, 0, sizeof send_wr);
588 ctxt->wr_op = IB_WR_SEND;
589 send_wr.wr_id = (unsigned long)ctxt;
590 send_wr.sg_list = ctxt->sge;
591 send_wr.num_sge = sge_no;
592 send_wr.opcode = IB_WR_SEND;
593 send_wr.send_flags = IB_SEND_SIGNALED;
595 /* Prepare INVALIDATE WR */
596 memset(&inv_wr, 0, sizeof inv_wr);
597 inv_wr.opcode = IB_WR_LOCAL_INV;
598 inv_wr.send_flags = IB_SEND_SIGNALED;
599 inv_wr.ex.invalidate_rkey =
601 send_wr.next = &inv_wr;
604 ret = svc_rdma_send(rdma, &send_wr);
611 svc_rdma_unmap_dma(ctxt);
612 svc_rdma_put_frmr(rdma, vec->frmr);
613 svc_rdma_put_context(ctxt, 1);
617 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
622 * Return the start of an xdr buffer.
624 static void *xdr_start(struct xdr_buf *xdr)
626 return xdr->head[0].iov_base -
629 xdr->tail[0].iov_len -
630 xdr->head[0].iov_len);
633 int svc_rdma_sendto(struct svc_rqst *rqstp)
635 struct svc_xprt *xprt = rqstp->rq_xprt;
636 struct svcxprt_rdma *rdma =
637 container_of(xprt, struct svcxprt_rdma, sc_xprt);
638 struct rpcrdma_msg *rdma_argp;
639 struct rpcrdma_msg *rdma_resp;
640 struct rpcrdma_write_array *reply_ary;
641 enum rpcrdma_proc reply_type;
644 struct page *res_page;
645 struct svc_rdma_op_ctxt *ctxt;
646 struct svc_rdma_req_map *vec;
648 dprintk("svcrdma: sending response for rqstp=%p\n", rqstp);
650 /* Get the RDMA request header. */
651 rdma_argp = xdr_start(&rqstp->rq_arg);
653 /* Build an req vec for the XDR */
654 ctxt = svc_rdma_get_context(rdma);
655 ctxt->direction = DMA_TO_DEVICE;
656 vec = svc_rdma_get_req_map();
657 ret = map_xdr(rdma, &rqstp->rq_res, vec);
660 inline_bytes = rqstp->rq_res.len;
662 /* Create the RDMA response header */
663 res_page = svc_rdma_get_page();
664 rdma_resp = page_address(res_page);
665 reply_ary = svc_rdma_get_reply_array(rdma_argp);
667 reply_type = RDMA_NOMSG;
669 reply_type = RDMA_MSG;
670 svc_rdma_xdr_encode_reply_header(rdma, rdma_argp,
671 rdma_resp, reply_type);
673 /* Send any write-chunk data and build resp write-list */
674 ret = send_write_chunks(rdma, rdma_argp, rdma_resp,
677 printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n",
683 /* Send any reply-list data and update resp reply-list */
684 ret = send_reply_chunks(rdma, rdma_argp, rdma_resp,
687 printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n",
693 ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec,
695 svc_rdma_put_req_map(vec);
696 dprintk("svcrdma: send_reply returns %d\n", ret);
702 svc_rdma_put_req_map(vec);
703 svc_rdma_put_context(ctxt, 0);