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 	int nchunks;
244 	struct rpcrdma_write_array *arg_ary;
245 	struct rpcrdma_write_array *res_ary;
246 	int ret;
247 
248 	arg_ary = svc_rdma_get_write_array(rdma_argp);
249 	if (!arg_ary)
250 		return 0;
251 	res_ary = (struct rpcrdma_write_array *)
252 		&rdma_resp->rm_body.rm_chunks[1];
253 
254 	/* Write chunks start at the pagelist */
255 	nchunks = be32_to_cpu(arg_ary->wc_nchunks);
256 	for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0;
257 	     xfer_len && chunk_no < nchunks;
258 	     chunk_no++) {
259 		struct rpcrdma_segment *arg_ch;
260 		u64 rs_offset;
261 
262 		arg_ch = &arg_ary->wc_array[chunk_no].wc_target;
263 		write_len = min(xfer_len, be32_to_cpu(arg_ch->rs_length));
264 
265 		/* Prepare the response chunk given the length actually
266 		 * written */
267 		xdr_decode_hyper((__be32 *)&arg_ch->rs_offset, &rs_offset);
268 		svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
269 						arg_ch->rs_handle,
270 						arg_ch->rs_offset,
271 						write_len);
272 		chunk_off = 0;
273 		while (write_len) {
274 			ret = send_write(xprt, rqstp,
275 					 be32_to_cpu(arg_ch->rs_handle),
276 					 rs_offset + chunk_off,
277 					 xdr_off,
278 					 write_len,
279 					 vec);
280 			if (ret <= 0) {
281 				dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
282 					ret);
283 				return -EIO;
284 			}
285 			chunk_off += ret;
286 			xdr_off += ret;
287 			xfer_len -= ret;
288 			write_len -= ret;
289 		}
290 	}
291 	/* Update the req with the number of chunks actually used */
292 	svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no);
293 
294 	return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len;
295 }
296 
297 static int send_reply_chunks(struct svcxprt_rdma *xprt,
298 			     struct rpcrdma_msg *rdma_argp,
299 			     struct rpcrdma_msg *rdma_resp,
300 			     struct svc_rqst *rqstp,
301 			     struct svc_rdma_req_map *vec)
302 {
303 	u32 xfer_len = rqstp->rq_res.len;
304 	int write_len;
305 	u32 xdr_off;
306 	int chunk_no;
307 	int chunk_off;
308 	int nchunks;
309 	struct rpcrdma_segment *ch;
310 	struct rpcrdma_write_array *arg_ary;
311 	struct rpcrdma_write_array *res_ary;
312 	int ret;
313 
314 	arg_ary = svc_rdma_get_reply_array(rdma_argp);
315 	if (!arg_ary)
316 		return 0;
317 	/* XXX: need to fix when reply lists occur with read-list and or
318 	 * write-list */
319 	res_ary = (struct rpcrdma_write_array *)
320 		&rdma_resp->rm_body.rm_chunks[2];
321 
322 	/* xdr offset starts at RPC message */
323 	nchunks = be32_to_cpu(arg_ary->wc_nchunks);
324 	for (xdr_off = 0, chunk_no = 0;
325 	     xfer_len && chunk_no < nchunks;
326 	     chunk_no++) {
327 		u64 rs_offset;
328 		ch = &arg_ary->wc_array[chunk_no].wc_target;
329 		write_len = min(xfer_len, be32_to_cpu(ch->rs_length));
330 
331 		/* Prepare the reply chunk given the length actually
332 		 * written */
333 		xdr_decode_hyper((__be32 *)&ch->rs_offset, &rs_offset);
334 		svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no,
335 						ch->rs_handle, ch->rs_offset,
336 						write_len);
337 		chunk_off = 0;
338 		while (write_len) {
339 			ret = send_write(xprt, rqstp,
340 					 be32_to_cpu(ch->rs_handle),
341 					 rs_offset + chunk_off,
342 					 xdr_off,
343 					 write_len,
344 					 vec);
345 			if (ret <= 0) {
346 				dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n",
347 					ret);
348 				return -EIO;
349 			}
350 			chunk_off += ret;
351 			xdr_off += ret;
352 			xfer_len -= ret;
353 			write_len -= ret;
354 		}
355 	}
356 	/* Update the req with the number of chunks actually used */
357 	svc_rdma_xdr_encode_reply_array(res_ary, chunk_no);
358 
359 	return rqstp->rq_res.len;
360 }
361 
362 /* This function prepares the portion of the RPCRDMA message to be
363  * sent in the RDMA_SEND. This function is called after data sent via
364  * RDMA has already been transmitted. There are three cases:
365  * - The RPCRDMA header, RPC header, and payload are all sent in a
366  *   single RDMA_SEND. This is the "inline" case.
367  * - The RPCRDMA header and some portion of the RPC header and data
368  *   are sent via this RDMA_SEND and another portion of the data is
369  *   sent via RDMA.
370  * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC
371  *   header and data are all transmitted via RDMA.
372  * In all three cases, this function prepares the RPCRDMA header in
373  * sge[0], the 'type' parameter indicates the type to place in the
374  * RPCRDMA header, and the 'byte_count' field indicates how much of
375  * the XDR to include in this RDMA_SEND. NB: The offset of the payload
376  * to send is zero in the XDR.
377  */
378 static int send_reply(struct svcxprt_rdma *rdma,
379 		      struct svc_rqst *rqstp,
380 		      struct page *page,
381 		      struct rpcrdma_msg *rdma_resp,
382 		      struct svc_rdma_op_ctxt *ctxt,
383 		      struct svc_rdma_req_map *vec,
384 		      int byte_count)
385 {
386 	struct ib_send_wr send_wr;
387 	int sge_no;
388 	int sge_bytes;
389 	int page_no;
390 	int pages;
391 	int ret;
392 
393 	/* Post a recv buffer to handle another request. */
394 	ret = svc_rdma_post_recv(rdma);
395 	if (ret) {
396 		printk(KERN_INFO
397 		       "svcrdma: could not post a receive buffer, err=%d."
398 		       "Closing transport %p.\n", ret, rdma);
399 		set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
400 		svc_rdma_put_context(ctxt, 0);
401 		return -ENOTCONN;
402 	}
403 
404 	/* Prepare the context */
405 	ctxt->pages[0] = page;
406 	ctxt->count = 1;
407 
408 	/* Prepare the SGE for the RPCRDMA Header */
409 	ctxt->sge[0].lkey = rdma->sc_dma_lkey;
410 	ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp);
411 	ctxt->sge[0].addr =
412 	    ib_dma_map_page(rdma->sc_cm_id->device, page, 0,
413 			    ctxt->sge[0].length, DMA_TO_DEVICE);
414 	if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr))
415 		goto err;
416 	atomic_inc(&rdma->sc_dma_used);
417 
418 	ctxt->direction = DMA_TO_DEVICE;
419 
420 	/* Map the payload indicated by 'byte_count' */
421 	for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) {
422 		int xdr_off = 0;
423 		sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count);
424 		byte_count -= sge_bytes;
425 		ctxt->sge[sge_no].addr =
426 			dma_map_xdr(rdma, &rqstp->rq_res, xdr_off,
427 				    sge_bytes, DMA_TO_DEVICE);
428 		xdr_off += sge_bytes;
429 		if (ib_dma_mapping_error(rdma->sc_cm_id->device,
430 					 ctxt->sge[sge_no].addr))
431 			goto err;
432 		atomic_inc(&rdma->sc_dma_used);
433 		ctxt->sge[sge_no].lkey = rdma->sc_dma_lkey;
434 		ctxt->sge[sge_no].length = sge_bytes;
435 	}
436 	if (byte_count != 0) {
437 		pr_err("svcrdma: Could not map %d bytes\n", byte_count);
438 		goto err;
439 	}
440 
441 	/* Save all respages in the ctxt and remove them from the
442 	 * respages array. They are our pages until the I/O
443 	 * completes.
444 	 */
445 	pages = rqstp->rq_next_page - rqstp->rq_respages;
446 	for (page_no = 0; page_no < pages; page_no++) {
447 		ctxt->pages[page_no+1] = rqstp->rq_respages[page_no];
448 		ctxt->count++;
449 		rqstp->rq_respages[page_no] = NULL;
450 		/*
451 		 * If there are more pages than SGE, terminate SGE
452 		 * list so that svc_rdma_unmap_dma doesn't attempt to
453 		 * unmap garbage.
454 		 */
455 		if (page_no+1 >= sge_no)
456 			ctxt->sge[page_no+1].length = 0;
457 	}
458 	rqstp->rq_next_page = rqstp->rq_respages + 1;
459 
460 	if (sge_no > rdma->sc_max_sge) {
461 		pr_err("svcrdma: Too many sges (%d)\n", sge_no);
462 		goto err;
463 	}
464 	memset(&send_wr, 0, sizeof send_wr);
465 	ctxt->wr_op = IB_WR_SEND;
466 	send_wr.wr_id = (unsigned long)ctxt;
467 	send_wr.sg_list = ctxt->sge;
468 	send_wr.num_sge = sge_no;
469 	send_wr.opcode = IB_WR_SEND;
470 	send_wr.send_flags =  IB_SEND_SIGNALED;
471 
472 	ret = svc_rdma_send(rdma, &send_wr);
473 	if (ret)
474 		goto err;
475 
476 	return 0;
477 
478  err:
479 	svc_rdma_unmap_dma(ctxt);
480 	svc_rdma_put_context(ctxt, 1);
481 	return -EIO;
482 }
483 
484 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
485 {
486 }
487 
488 int svc_rdma_sendto(struct svc_rqst *rqstp)
489 {
490 	struct svc_xprt *xprt = rqstp->rq_xprt;
491 	struct svcxprt_rdma *rdma =
492 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
493 	struct rpcrdma_msg *rdma_argp;
494 	struct rpcrdma_msg *rdma_resp;
495 	struct rpcrdma_write_array *reply_ary;
496 	enum rpcrdma_proc reply_type;
497 	int ret;
498 	int inline_bytes;
499 	struct page *res_page;
500 	struct svc_rdma_op_ctxt *ctxt;
501 	struct svc_rdma_req_map *vec;
502 
503 	dprintk("svcrdma: sending response for rqstp=%p\n", rqstp);
504 
505 	/* Get the RDMA request header. The receive logic always
506 	 * places this at the start of page 0.
507 	 */
508 	rdma_argp = page_address(rqstp->rq_pages[0]);
509 
510 	/* Build an req vec for the XDR */
511 	ctxt = svc_rdma_get_context(rdma);
512 	ctxt->direction = DMA_TO_DEVICE;
513 	vec = svc_rdma_get_req_map();
514 	ret = map_xdr(rdma, &rqstp->rq_res, vec);
515 	if (ret)
516 		goto err0;
517 	inline_bytes = rqstp->rq_res.len;
518 
519 	/* Create the RDMA response header */
520 	res_page = alloc_page(GFP_KERNEL | __GFP_NOFAIL);
521 	rdma_resp = page_address(res_page);
522 	reply_ary = svc_rdma_get_reply_array(rdma_argp);
523 	if (reply_ary)
524 		reply_type = RDMA_NOMSG;
525 	else
526 		reply_type = RDMA_MSG;
527 	svc_rdma_xdr_encode_reply_header(rdma, rdma_argp,
528 					 rdma_resp, reply_type);
529 
530 	/* Send any write-chunk data and build resp write-list */
531 	ret = send_write_chunks(rdma, rdma_argp, rdma_resp,
532 				rqstp, vec);
533 	if (ret < 0) {
534 		printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n",
535 		       ret);
536 		goto err1;
537 	}
538 	inline_bytes -= ret;
539 
540 	/* Send any reply-list data and update resp reply-list */
541 	ret = send_reply_chunks(rdma, rdma_argp, rdma_resp,
542 				rqstp, vec);
543 	if (ret < 0) {
544 		printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n",
545 		       ret);
546 		goto err1;
547 	}
548 	inline_bytes -= ret;
549 
550 	ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec,
551 			 inline_bytes);
552 	svc_rdma_put_req_map(vec);
553 	dprintk("svcrdma: send_reply returns %d\n", ret);
554 	return ret;
555 
556  err1:
557 	put_page(res_page);
558  err0:
559 	svc_rdma_put_req_map(vec);
560 	svc_rdma_put_context(ctxt, 0);
561 	return ret;
562 }
563