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 
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>
49 
50 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
51 
52 /*
53  * Replace the pages in the rq_argpages array with the pages from the SGE in
54  * the RDMA_RECV completion. The SGL should contain full pages up until the
55  * last one.
56  */
57 static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
58 			       struct svc_rdma_op_ctxt *ctxt,
59 			       u32 byte_count)
60 {
61 	struct page *page;
62 	u32 bc;
63 	int sge_no;
64 
65 	/* Swap the page in the SGE with the page in argpages */
66 	page = ctxt->pages[0];
67 	put_page(rqstp->rq_pages[0]);
68 	rqstp->rq_pages[0] = page;
69 
70 	/* Set up the XDR head */
71 	rqstp->rq_arg.head[0].iov_base = page_address(page);
72 	rqstp->rq_arg.head[0].iov_len = min(byte_count, ctxt->sge[0].length);
73 	rqstp->rq_arg.len = byte_count;
74 	rqstp->rq_arg.buflen = byte_count;
75 
76 	/* Compute bytes past head in the SGL */
77 	bc = byte_count - rqstp->rq_arg.head[0].iov_len;
78 
79 	/* If data remains, store it in the pagelist */
80 	rqstp->rq_arg.page_len = bc;
81 	rqstp->rq_arg.page_base = 0;
82 	rqstp->rq_arg.pages = &rqstp->rq_pages[1];
83 	sge_no = 1;
84 	while (bc && sge_no < ctxt->count) {
85 		page = ctxt->pages[sge_no];
86 		put_page(rqstp->rq_pages[sge_no]);
87 		rqstp->rq_pages[sge_no] = page;
88 		bc -= min(bc, ctxt->sge[sge_no].length);
89 		rqstp->rq_arg.buflen += ctxt->sge[sge_no].length;
90 		sge_no++;
91 	}
92 	rqstp->rq_respages = &rqstp->rq_pages[sge_no];
93 
94 	/* We should never run out of SGE because the limit is defined to
95 	 * support the max allowed RPC data length
96 	 */
97 	BUG_ON(bc && (sge_no == ctxt->count));
98 	BUG_ON((rqstp->rq_arg.head[0].iov_len + rqstp->rq_arg.page_len)
99 	       != byte_count);
100 	BUG_ON(rqstp->rq_arg.len != byte_count);
101 
102 	/* If not all pages were used from the SGL, free the remaining ones */
103 	bc = sge_no;
104 	while (sge_no < ctxt->count) {
105 		page = ctxt->pages[sge_no++];
106 		put_page(page);
107 	}
108 	ctxt->count = bc;
109 
110 	/* Set up tail */
111 	rqstp->rq_arg.tail[0].iov_base = NULL;
112 	rqstp->rq_arg.tail[0].iov_len = 0;
113 }
114 
115 /* Encode a read-chunk-list as an array of IB SGE
116  *
117  * Assumptions:
118  * - chunk[0]->position points to pages[0] at an offset of 0
119  * - pages[] is not physically or virtually contigous and consists of
120  *   PAGE_SIZE elements.
121  *
122  * Output:
123  * - sge array pointing into pages[] array.
124  * - chunk_sge array specifying sge index and count for each
125  *   chunk in the read list
126  *
127  */
128 static int rdma_rcl_to_sge(struct svcxprt_rdma *xprt,
129 			   struct svc_rqst *rqstp,
130 			   struct svc_rdma_op_ctxt *head,
131 			   struct rpcrdma_msg *rmsgp,
132 			   struct svc_rdma_req_map *rpl_map,
133 			   struct svc_rdma_req_map *chl_map,
134 			   int ch_count,
135 			   int byte_count)
136 {
137 	int sge_no;
138 	int sge_bytes;
139 	int page_off;
140 	int page_no;
141 	int ch_bytes;
142 	int ch_no;
143 	struct rpcrdma_read_chunk *ch;
144 
145 	sge_no = 0;
146 	page_no = 0;
147 	page_off = 0;
148 	ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
149 	ch_no = 0;
150 	ch_bytes = ch->rc_target.rs_length;
151 	head->arg.head[0] = rqstp->rq_arg.head[0];
152 	head->arg.tail[0] = rqstp->rq_arg.tail[0];
153 	head->arg.pages = &head->pages[head->count];
154 	head->hdr_count = head->count; /* save count of hdr pages */
155 	head->arg.page_base = 0;
156 	head->arg.page_len = ch_bytes;
157 	head->arg.len = rqstp->rq_arg.len + ch_bytes;
158 	head->arg.buflen = rqstp->rq_arg.buflen + ch_bytes;
159 	head->count++;
160 	chl_map->ch[0].start = 0;
161 	while (byte_count) {
162 		rpl_map->sge[sge_no].iov_base =
163 			page_address(rqstp->rq_arg.pages[page_no]) + page_off;
164 		sge_bytes = min_t(int, PAGE_SIZE-page_off, ch_bytes);
165 		rpl_map->sge[sge_no].iov_len = sge_bytes;
166 		/*
167 		 * Don't bump head->count here because the same page
168 		 * may be used by multiple SGE.
169 		 */
170 		head->arg.pages[page_no] = rqstp->rq_arg.pages[page_no];
171 		rqstp->rq_respages = &rqstp->rq_arg.pages[page_no+1];
172 
173 		byte_count -= sge_bytes;
174 		ch_bytes -= sge_bytes;
175 		sge_no++;
176 		/*
177 		 * If all bytes for this chunk have been mapped to an
178 		 * SGE, move to the next SGE
179 		 */
180 		if (ch_bytes == 0) {
181 			chl_map->ch[ch_no].count =
182 				sge_no - chl_map->ch[ch_no].start;
183 			ch_no++;
184 			ch++;
185 			chl_map->ch[ch_no].start = sge_no;
186 			ch_bytes = ch->rc_target.rs_length;
187 			/* If bytes remaining account for next chunk */
188 			if (byte_count) {
189 				head->arg.page_len += ch_bytes;
190 				head->arg.len += ch_bytes;
191 				head->arg.buflen += ch_bytes;
192 			}
193 		}
194 		/*
195 		 * If this SGE consumed all of the page, move to the
196 		 * next page
197 		 */
198 		if ((sge_bytes + page_off) == PAGE_SIZE) {
199 			page_no++;
200 			page_off = 0;
201 			/*
202 			 * If there are still bytes left to map, bump
203 			 * the page count
204 			 */
205 			if (byte_count)
206 				head->count++;
207 		} else
208 			page_off += sge_bytes;
209 	}
210 	BUG_ON(byte_count != 0);
211 	return sge_no;
212 }
213 
214 static void rdma_set_ctxt_sge(struct svcxprt_rdma *xprt,
215 			      struct svc_rdma_op_ctxt *ctxt,
216 			      struct kvec *vec,
217 			      u64 *sgl_offset,
218 			      int count)
219 {
220 	int i;
221 
222 	ctxt->count = count;
223 	ctxt->direction = DMA_FROM_DEVICE;
224 	for (i = 0; i < count; i++) {
225 		atomic_inc(&xprt->sc_dma_used);
226 		ctxt->sge[i].addr =
227 			ib_dma_map_single(xprt->sc_cm_id->device,
228 					  vec[i].iov_base, vec[i].iov_len,
229 					  DMA_FROM_DEVICE);
230 		ctxt->sge[i].length = vec[i].iov_len;
231 		ctxt->sge[i].lkey = xprt->sc_phys_mr->lkey;
232 		*sgl_offset = *sgl_offset + vec[i].iov_len;
233 	}
234 }
235 
236 static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count)
237 {
238 	if ((RDMA_TRANSPORT_IWARP ==
239 	     rdma_node_get_transport(xprt->sc_cm_id->
240 				     device->node_type))
241 	    && sge_count > 1)
242 		return 1;
243 	else
244 		return min_t(int, sge_count, xprt->sc_max_sge);
245 }
246 
247 /*
248  * Use RDMA_READ to read data from the advertised client buffer into the
249  * XDR stream starting at rq_arg.head[0].iov_base.
250  * Each chunk in the array
251  * contains the following fields:
252  * discrim      - '1', This isn't used for data placement
253  * position     - The xdr stream offset (the same for every chunk)
254  * handle       - RMR for client memory region
255  * length       - data transfer length
256  * offset       - 64 bit tagged offset in remote memory region
257  *
258  * On our side, we need to read into a pagelist. The first page immediately
259  * follows the RPC header.
260  *
261  * This function returns:
262  * 0 - No error and no read-list found.
263  *
264  * 1 - Successful read-list processing. The data is not yet in
265  * the pagelist and therefore the RPC request must be deferred. The
266  * I/O completion will enqueue the transport again and
267  * svc_rdma_recvfrom will complete the request.
268  *
269  * <0 - Error processing/posting read-list.
270  *
271  * NOTE: The ctxt must not be touched after the last WR has been posted
272  * because the I/O completion processing may occur on another
273  * processor and free / modify the context. Ne touche pas!
274  */
275 static int rdma_read_xdr(struct svcxprt_rdma *xprt,
276 			 struct rpcrdma_msg *rmsgp,
277 			 struct svc_rqst *rqstp,
278 			 struct svc_rdma_op_ctxt *hdr_ctxt)
279 {
280 	struct ib_send_wr read_wr;
281 	int err = 0;
282 	int ch_no;
283 	int ch_count;
284 	int byte_count;
285 	int sge_count;
286 	u64 sgl_offset;
287 	struct rpcrdma_read_chunk *ch;
288 	struct svc_rdma_op_ctxt *ctxt = NULL;
289 	struct svc_rdma_req_map *rpl_map;
290 	struct svc_rdma_req_map *chl_map;
291 
292 	/* If no read list is present, return 0 */
293 	ch = svc_rdma_get_read_chunk(rmsgp);
294 	if (!ch)
295 		return 0;
296 
297 	/* Allocate temporary reply and chunk maps */
298 	rpl_map = svc_rdma_get_req_map();
299 	chl_map = svc_rdma_get_req_map();
300 
301 	svc_rdma_rcl_chunk_counts(ch, &ch_count, &byte_count);
302 	if (ch_count > RPCSVC_MAXPAGES)
303 		return -EINVAL;
304 	sge_count = rdma_rcl_to_sge(xprt, rqstp, hdr_ctxt, rmsgp,
305 				    rpl_map, chl_map,
306 				    ch_count, byte_count);
307 	sgl_offset = 0;
308 	ch_no = 0;
309 
310 	for (ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
311 	     ch->rc_discrim != 0; ch++, ch_no++) {
312 next_sge:
313 		ctxt = svc_rdma_get_context(xprt);
314 		ctxt->direction = DMA_FROM_DEVICE;
315 		clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
316 
317 		/* Prepare READ WR */
318 		memset(&read_wr, 0, sizeof read_wr);
319 		ctxt->wr_op = IB_WR_RDMA_READ;
320 		read_wr.wr_id = (unsigned long)ctxt;
321 		read_wr.opcode = IB_WR_RDMA_READ;
322 		read_wr.send_flags = IB_SEND_SIGNALED;
323 		read_wr.wr.rdma.rkey = ch->rc_target.rs_handle;
324 		read_wr.wr.rdma.remote_addr =
325 			get_unaligned(&(ch->rc_target.rs_offset)) +
326 			sgl_offset;
327 		read_wr.sg_list = ctxt->sge;
328 		read_wr.num_sge =
329 			rdma_read_max_sge(xprt, chl_map->ch[ch_no].count);
330 		rdma_set_ctxt_sge(xprt, ctxt,
331 				  &rpl_map->sge[chl_map->ch[ch_no].start],
332 				  &sgl_offset,
333 				  read_wr.num_sge);
334 		if (((ch+1)->rc_discrim == 0) &&
335 		    (read_wr.num_sge == chl_map->ch[ch_no].count)) {
336 			/*
337 			 * Mark the last RDMA_READ with a bit to
338 			 * indicate all RPC data has been fetched from
339 			 * the client and the RPC needs to be enqueued.
340 			 */
341 			set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
342 			ctxt->read_hdr = hdr_ctxt;
343 		}
344 		/* Post the read */
345 		err = svc_rdma_send(xprt, &read_wr);
346 		if (err) {
347 			printk(KERN_ERR "svcrdma: Error %d posting RDMA_READ\n",
348 			       err);
349 			set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
350 			svc_rdma_put_context(ctxt, 0);
351 			goto out;
352 		}
353 		atomic_inc(&rdma_stat_read);
354 
355 		if (read_wr.num_sge < chl_map->ch[ch_no].count) {
356 			chl_map->ch[ch_no].count -= read_wr.num_sge;
357 			chl_map->ch[ch_no].start += read_wr.num_sge;
358 			goto next_sge;
359 		}
360 		sgl_offset = 0;
361 		err = 1;
362 	}
363 
364  out:
365 	svc_rdma_put_req_map(rpl_map);
366 	svc_rdma_put_req_map(chl_map);
367 
368 	/* Detach arg pages. svc_recv will replenish them */
369 	for (ch_no = 0; &rqstp->rq_pages[ch_no] < rqstp->rq_respages; ch_no++)
370 		rqstp->rq_pages[ch_no] = NULL;
371 
372 	/*
373 	 * Detach res pages. svc_release must see a resused count of
374 	 * zero or it will attempt to put them.
375 	 */
376 	while (rqstp->rq_resused)
377 		rqstp->rq_respages[--rqstp->rq_resused] = NULL;
378 
379 	return err;
380 }
381 
382 static int rdma_read_complete(struct svc_rqst *rqstp,
383 			      struct svc_rdma_op_ctxt *head)
384 {
385 	int page_no;
386 	int ret;
387 
388 	BUG_ON(!head);
389 
390 	/* Copy RPC pages */
391 	for (page_no = 0; page_no < head->count; page_no++) {
392 		put_page(rqstp->rq_pages[page_no]);
393 		rqstp->rq_pages[page_no] = head->pages[page_no];
394 	}
395 	/* Point rq_arg.pages past header */
396 	rqstp->rq_arg.pages = &rqstp->rq_pages[head->hdr_count];
397 	rqstp->rq_arg.page_len = head->arg.page_len;
398 	rqstp->rq_arg.page_base = head->arg.page_base;
399 
400 	/* rq_respages starts after the last arg page */
401 	rqstp->rq_respages = &rqstp->rq_arg.pages[page_no];
402 	rqstp->rq_resused = 0;
403 
404 	/* Rebuild rq_arg head and tail. */
405 	rqstp->rq_arg.head[0] = head->arg.head[0];
406 	rqstp->rq_arg.tail[0] = head->arg.tail[0];
407 	rqstp->rq_arg.len = head->arg.len;
408 	rqstp->rq_arg.buflen = head->arg.buflen;
409 
410 	/* Free the context */
411 	svc_rdma_put_context(head, 0);
412 
413 	/* XXX: What should this be? */
414 	rqstp->rq_prot = IPPROTO_MAX;
415 	svc_xprt_copy_addrs(rqstp, rqstp->rq_xprt);
416 
417 	ret = rqstp->rq_arg.head[0].iov_len
418 		+ rqstp->rq_arg.page_len
419 		+ rqstp->rq_arg.tail[0].iov_len;
420 	dprintk("svcrdma: deferred read ret=%d, rq_arg.len =%d, "
421 		"rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
422 		ret, rqstp->rq_arg.len,	rqstp->rq_arg.head[0].iov_base,
423 		rqstp->rq_arg.head[0].iov_len);
424 
425 	svc_xprt_received(rqstp->rq_xprt);
426 	return ret;
427 }
428 
429 /*
430  * Set up the rqstp thread context to point to the RQ buffer. If
431  * necessary, pull additional data from the client with an RDMA_READ
432  * request.
433  */
434 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
435 {
436 	struct svc_xprt *xprt = rqstp->rq_xprt;
437 	struct svcxprt_rdma *rdma_xprt =
438 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
439 	struct svc_rdma_op_ctxt *ctxt = NULL;
440 	struct rpcrdma_msg *rmsgp;
441 	int ret = 0;
442 	int len;
443 
444 	dprintk("svcrdma: rqstp=%p\n", rqstp);
445 
446 	spin_lock_bh(&rdma_xprt->sc_read_complete_lock);
447 	if (!list_empty(&rdma_xprt->sc_read_complete_q)) {
448 		ctxt = list_entry(rdma_xprt->sc_read_complete_q.next,
449 				  struct svc_rdma_op_ctxt,
450 				  dto_q);
451 		list_del_init(&ctxt->dto_q);
452 	}
453 	spin_unlock_bh(&rdma_xprt->sc_read_complete_lock);
454 	if (ctxt)
455 		return rdma_read_complete(rqstp, ctxt);
456 
457 	spin_lock_bh(&rdma_xprt->sc_rq_dto_lock);
458 	if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
459 		ctxt = list_entry(rdma_xprt->sc_rq_dto_q.next,
460 				  struct svc_rdma_op_ctxt,
461 				  dto_q);
462 		list_del_init(&ctxt->dto_q);
463 	} else {
464 		atomic_inc(&rdma_stat_rq_starve);
465 		clear_bit(XPT_DATA, &xprt->xpt_flags);
466 		ctxt = NULL;
467 	}
468 	spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
469 	if (!ctxt) {
470 		/* This is the EAGAIN path. The svc_recv routine will
471 		 * return -EAGAIN, the nfsd thread will go to call into
472 		 * svc_recv again and we shouldn't be on the active
473 		 * transport list
474 		 */
475 		if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
476 			goto close_out;
477 
478 		BUG_ON(ret);
479 		goto out;
480 	}
481 	dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p, status=%d\n",
482 		ctxt, rdma_xprt, rqstp, ctxt->wc_status);
483 	BUG_ON(ctxt->wc_status != IB_WC_SUCCESS);
484 	atomic_inc(&rdma_stat_recv);
485 
486 	/* Build up the XDR from the receive buffers. */
487 	rdma_build_arg_xdr(rqstp, ctxt, ctxt->byte_len);
488 
489 	/* Decode the RDMA header. */
490 	len = svc_rdma_xdr_decode_req(&rmsgp, rqstp);
491 	rqstp->rq_xprt_hlen = len;
492 
493 	/* If the request is invalid, reply with an error */
494 	if (len < 0) {
495 		if (len == -ENOSYS)
496 			svc_rdma_send_error(rdma_xprt, rmsgp, ERR_VERS);
497 		goto close_out;
498 	}
499 
500 	/* Read read-list data. */
501 	ret = rdma_read_xdr(rdma_xprt, rmsgp, rqstp, ctxt);
502 	if (ret > 0) {
503 		/* read-list posted, defer until data received from client. */
504 		svc_xprt_received(xprt);
505 		return 0;
506 	}
507 	if (ret < 0) {
508 		/* Post of read-list failed, free context. */
509 		svc_rdma_put_context(ctxt, 1);
510 		return 0;
511 	}
512 
513 	ret = rqstp->rq_arg.head[0].iov_len
514 		+ rqstp->rq_arg.page_len
515 		+ rqstp->rq_arg.tail[0].iov_len;
516 	svc_rdma_put_context(ctxt, 0);
517  out:
518 	dprintk("svcrdma: ret = %d, rq_arg.len =%d, "
519 		"rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
520 		ret, rqstp->rq_arg.len,
521 		rqstp->rq_arg.head[0].iov_base,
522 		rqstp->rq_arg.head[0].iov_len);
523 	rqstp->rq_prot = IPPROTO_MAX;
524 	svc_xprt_copy_addrs(rqstp, xprt);
525 	svc_xprt_received(xprt);
526 	return ret;
527 
528  close_out:
529 	if (ctxt)
530 		svc_rdma_put_context(ctxt, 1);
531 	dprintk("svcrdma: transport %p is closing\n", xprt);
532 	/*
533 	 * Set the close bit and enqueue it. svc_recv will see the
534 	 * close bit and call svc_xprt_delete
535 	 */
536 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
537 	svc_xprt_received(xprt);
538 	return 0;
539 }
540