xref: /linux/net/sunrpc/xprtrdma/svc_rdma_sendto.c (revision e0bf6c5ca2d3281f231c5f0c9bf145e9513644de)
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 }
561