1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright (c) 2016-2018 Oracle. All rights reserved.
4  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
5  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the BSD-type
11  * license below:
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  *
17  *      Redistributions of source code must retain the above copyright
18  *      notice, this list of conditions and the following disclaimer.
19  *
20  *      Redistributions in binary form must reproduce the above
21  *      copyright notice, this list of conditions and the following
22  *      disclaimer in the documentation and/or other materials provided
23  *      with the distribution.
24  *
25  *      Neither the name of the Network Appliance, Inc. nor the names of
26  *      its contributors may be used to endorse or promote products
27  *      derived from this software without specific prior written
28  *      permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
31  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
32  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
33  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
34  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
35  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
36  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
37  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
38  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
39  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
40  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
41  *
42  * Author: Tom Tucker <tom@opengridcomputing.com>
43  */
44 
45 /* Operation
46  *
47  * The main entry point is svc_rdma_sendto. This is called by the
48  * RPC server when an RPC Reply is ready to be transmitted to a client.
49  *
50  * The passed-in svc_rqst contains a struct xdr_buf which holds an
51  * XDR-encoded RPC Reply message. sendto must construct the RPC-over-RDMA
52  * transport header, post all Write WRs needed for this Reply, then post
53  * a Send WR conveying the transport header and the RPC message itself to
54  * the client.
55  *
56  * svc_rdma_sendto must fully transmit the Reply before returning, as
57  * the svc_rqst will be recycled as soon as sendto returns. Remaining
58  * resources referred to by the svc_rqst are also recycled at that time.
59  * Therefore any resources that must remain longer must be detached
60  * from the svc_rqst and released later.
61  *
62  * Page Management
63  *
64  * The I/O that performs Reply transmission is asynchronous, and may
65  * complete well after sendto returns. Thus pages under I/O must be
66  * removed from the svc_rqst before sendto returns.
67  *
68  * The logic here depends on Send Queue and completion ordering. Since
69  * the Send WR is always posted last, it will always complete last. Thus
70  * when it completes, it is guaranteed that all previous Write WRs have
71  * also completed.
72  *
73  * Write WRs are constructed and posted. Each Write segment gets its own
74  * svc_rdma_rw_ctxt, allowing the Write completion handler to find and
75  * DMA-unmap the pages under I/O for that Write segment. The Write
76  * completion handler does not release any pages.
77  *
78  * When the Send WR is constructed, it also gets its own svc_rdma_send_ctxt.
79  * The ownership of all of the Reply's pages are transferred into that
80  * ctxt, the Send WR is posted, and sendto returns.
81  *
82  * The svc_rdma_send_ctxt is presented when the Send WR completes. The
83  * Send completion handler finally releases the Reply's pages.
84  *
85  * This mechanism also assumes that completions on the transport's Send
86  * Completion Queue do not run in parallel. Otherwise a Write completion
87  * and Send completion running at the same time could release pages that
88  * are still DMA-mapped.
89  *
90  * Error Handling
91  *
92  * - If the Send WR is posted successfully, it will either complete
93  *   successfully, or get flushed. Either way, the Send completion
94  *   handler releases the Reply's pages.
95  * - If the Send WR cannot be not posted, the forward path releases
96  *   the Reply's pages.
97  *
98  * This handles the case, without the use of page reference counting,
99  * where two different Write segments send portions of the same page.
100  */
101 
102 #include <linux/spinlock.h>
103 #include <asm/unaligned.h>
104 
105 #include <rdma/ib_verbs.h>
106 #include <rdma/rdma_cm.h>
107 
108 #include <linux/sunrpc/debug.h>
109 #include <linux/sunrpc/rpc_rdma.h>
110 #include <linux/sunrpc/svc_rdma.h>
111 
112 #include "xprt_rdma.h"
113 #include <trace/events/rpcrdma.h>
114 
115 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
116 
117 static void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc);
118 
119 static inline struct svc_rdma_send_ctxt *
120 svc_rdma_next_send_ctxt(struct list_head *list)
121 {
122 	return list_first_entry_or_null(list, struct svc_rdma_send_ctxt,
123 					sc_list);
124 }
125 
126 static struct svc_rdma_send_ctxt *
127 svc_rdma_send_ctxt_alloc(struct svcxprt_rdma *rdma)
128 {
129 	struct svc_rdma_send_ctxt *ctxt;
130 	dma_addr_t addr;
131 	void *buffer;
132 	size_t size;
133 	int i;
134 
135 	size = sizeof(*ctxt);
136 	size += rdma->sc_max_send_sges * sizeof(struct ib_sge);
137 	ctxt = kmalloc(size, GFP_KERNEL);
138 	if (!ctxt)
139 		goto fail0;
140 	buffer = kmalloc(rdma->sc_max_req_size, GFP_KERNEL);
141 	if (!buffer)
142 		goto fail1;
143 	addr = ib_dma_map_single(rdma->sc_pd->device, buffer,
144 				 rdma->sc_max_req_size, DMA_TO_DEVICE);
145 	if (ib_dma_mapping_error(rdma->sc_pd->device, addr))
146 		goto fail2;
147 
148 	ctxt->sc_send_wr.next = NULL;
149 	ctxt->sc_send_wr.wr_cqe = &ctxt->sc_cqe;
150 	ctxt->sc_send_wr.sg_list = ctxt->sc_sges;
151 	ctxt->sc_send_wr.send_flags = IB_SEND_SIGNALED;
152 	ctxt->sc_cqe.done = svc_rdma_wc_send;
153 	ctxt->sc_xprt_buf = buffer;
154 	ctxt->sc_sges[0].addr = addr;
155 
156 	for (i = 0; i < rdma->sc_max_send_sges; i++)
157 		ctxt->sc_sges[i].lkey = rdma->sc_pd->local_dma_lkey;
158 	return ctxt;
159 
160 fail2:
161 	kfree(buffer);
162 fail1:
163 	kfree(ctxt);
164 fail0:
165 	return NULL;
166 }
167 
168 /**
169  * svc_rdma_send_ctxts_destroy - Release all send_ctxt's for an xprt
170  * @rdma: svcxprt_rdma being torn down
171  *
172  */
173 void svc_rdma_send_ctxts_destroy(struct svcxprt_rdma *rdma)
174 {
175 	struct svc_rdma_send_ctxt *ctxt;
176 
177 	while ((ctxt = svc_rdma_next_send_ctxt(&rdma->sc_send_ctxts))) {
178 		list_del(&ctxt->sc_list);
179 		ib_dma_unmap_single(rdma->sc_pd->device,
180 				    ctxt->sc_sges[0].addr,
181 				    rdma->sc_max_req_size,
182 				    DMA_TO_DEVICE);
183 		kfree(ctxt->sc_xprt_buf);
184 		kfree(ctxt);
185 	}
186 }
187 
188 /**
189  * svc_rdma_send_ctxt_get - Get a free send_ctxt
190  * @rdma: controlling svcxprt_rdma
191  *
192  * Returns a ready-to-use send_ctxt, or NULL if none are
193  * available and a fresh one cannot be allocated.
194  */
195 struct svc_rdma_send_ctxt *svc_rdma_send_ctxt_get(struct svcxprt_rdma *rdma)
196 {
197 	struct svc_rdma_send_ctxt *ctxt;
198 
199 	spin_lock(&rdma->sc_send_lock);
200 	ctxt = svc_rdma_next_send_ctxt(&rdma->sc_send_ctxts);
201 	if (!ctxt)
202 		goto out_empty;
203 	list_del(&ctxt->sc_list);
204 	spin_unlock(&rdma->sc_send_lock);
205 
206 out:
207 	ctxt->sc_send_wr.num_sge = 0;
208 	ctxt->sc_cur_sge_no = 0;
209 	ctxt->sc_page_count = 0;
210 	return ctxt;
211 
212 out_empty:
213 	spin_unlock(&rdma->sc_send_lock);
214 	ctxt = svc_rdma_send_ctxt_alloc(rdma);
215 	if (!ctxt)
216 		return NULL;
217 	goto out;
218 }
219 
220 /**
221  * svc_rdma_send_ctxt_put - Return send_ctxt to free list
222  * @rdma: controlling svcxprt_rdma
223  * @ctxt: object to return to the free list
224  *
225  * Pages left in sc_pages are DMA unmapped and released.
226  */
227 void svc_rdma_send_ctxt_put(struct svcxprt_rdma *rdma,
228 			    struct svc_rdma_send_ctxt *ctxt)
229 {
230 	struct ib_device *device = rdma->sc_cm_id->device;
231 	unsigned int i;
232 
233 	/* The first SGE contains the transport header, which
234 	 * remains mapped until @ctxt is destroyed.
235 	 */
236 	for (i = 1; i < ctxt->sc_send_wr.num_sge; i++)
237 		ib_dma_unmap_page(device,
238 				  ctxt->sc_sges[i].addr,
239 				  ctxt->sc_sges[i].length,
240 				  DMA_TO_DEVICE);
241 
242 	for (i = 0; i < ctxt->sc_page_count; ++i)
243 		put_page(ctxt->sc_pages[i]);
244 
245 	spin_lock(&rdma->sc_send_lock);
246 	list_add(&ctxt->sc_list, &rdma->sc_send_ctxts);
247 	spin_unlock(&rdma->sc_send_lock);
248 }
249 
250 /**
251  * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC
252  * @cq: Completion Queue context
253  * @wc: Work Completion object
254  *
255  * NB: The svc_xprt/svcxprt_rdma is pinned whenever it's possible that
256  * the Send completion handler could be running.
257  */
258 static void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
259 {
260 	struct svcxprt_rdma *rdma = cq->cq_context;
261 	struct ib_cqe *cqe = wc->wr_cqe;
262 	struct svc_rdma_send_ctxt *ctxt;
263 
264 	trace_svcrdma_wc_send(wc);
265 
266 	atomic_inc(&rdma->sc_sq_avail);
267 	wake_up(&rdma->sc_send_wait);
268 
269 	ctxt = container_of(cqe, struct svc_rdma_send_ctxt, sc_cqe);
270 	svc_rdma_send_ctxt_put(rdma, ctxt);
271 
272 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
273 		set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
274 		svc_xprt_enqueue(&rdma->sc_xprt);
275 		if (wc->status != IB_WC_WR_FLUSH_ERR)
276 			pr_err("svcrdma: Send: %s (%u/0x%x)\n",
277 			       ib_wc_status_msg(wc->status),
278 			       wc->status, wc->vendor_err);
279 	}
280 
281 	svc_xprt_put(&rdma->sc_xprt);
282 }
283 
284 /**
285  * svc_rdma_send - Post a single Send WR
286  * @rdma: transport on which to post the WR
287  * @wr: prepared Send WR to post
288  *
289  * Returns zero the Send WR was posted successfully. Otherwise, a
290  * negative errno is returned.
291  */
292 int svc_rdma_send(struct svcxprt_rdma *rdma, struct ib_send_wr *wr)
293 {
294 	int ret;
295 
296 	might_sleep();
297 
298 	/* If the SQ is full, wait until an SQ entry is available */
299 	while (1) {
300 		if ((atomic_dec_return(&rdma->sc_sq_avail) < 0)) {
301 			atomic_inc(&rdma_stat_sq_starve);
302 			trace_svcrdma_sq_full(rdma);
303 			atomic_inc(&rdma->sc_sq_avail);
304 			wait_event(rdma->sc_send_wait,
305 				   atomic_read(&rdma->sc_sq_avail) > 1);
306 			if (test_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags))
307 				return -ENOTCONN;
308 			trace_svcrdma_sq_retry(rdma);
309 			continue;
310 		}
311 
312 		svc_xprt_get(&rdma->sc_xprt);
313 		ret = ib_post_send(rdma->sc_qp, wr, NULL);
314 		trace_svcrdma_post_send(wr, ret);
315 		if (ret) {
316 			set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
317 			svc_xprt_put(&rdma->sc_xprt);
318 			wake_up(&rdma->sc_send_wait);
319 		}
320 		break;
321 	}
322 	return ret;
323 }
324 
325 static u32 xdr_padsize(u32 len)
326 {
327 	return (len & 3) ? (4 - (len & 3)) : 0;
328 }
329 
330 /* Returns length of transport header, in bytes.
331  */
332 static unsigned int svc_rdma_reply_hdr_len(__be32 *rdma_resp)
333 {
334 	unsigned int nsegs;
335 	__be32 *p;
336 
337 	p = rdma_resp;
338 
339 	/* RPC-over-RDMA V1 replies never have a Read list. */
340 	p += rpcrdma_fixed_maxsz + 1;
341 
342 	/* Skip Write list. */
343 	while (*p++ != xdr_zero) {
344 		nsegs = be32_to_cpup(p++);
345 		p += nsegs * rpcrdma_segment_maxsz;
346 	}
347 
348 	/* Skip Reply chunk. */
349 	if (*p++ != xdr_zero) {
350 		nsegs = be32_to_cpup(p++);
351 		p += nsegs * rpcrdma_segment_maxsz;
352 	}
353 
354 	return (unsigned long)p - (unsigned long)rdma_resp;
355 }
356 
357 /* One Write chunk is copied from Call transport header to Reply
358  * transport header. Each segment's length field is updated to
359  * reflect number of bytes consumed in the segment.
360  *
361  * Returns number of segments in this chunk.
362  */
363 static unsigned int xdr_encode_write_chunk(__be32 *dst, __be32 *src,
364 					   unsigned int remaining)
365 {
366 	unsigned int i, nsegs;
367 	u32 seg_len;
368 
369 	/* Write list discriminator */
370 	*dst++ = *src++;
371 
372 	/* number of segments in this chunk */
373 	nsegs = be32_to_cpup(src);
374 	*dst++ = *src++;
375 
376 	for (i = nsegs; i; i--) {
377 		/* segment's RDMA handle */
378 		*dst++ = *src++;
379 
380 		/* bytes returned in this segment */
381 		seg_len = be32_to_cpu(*src);
382 		if (remaining >= seg_len) {
383 			/* entire segment was consumed */
384 			*dst = *src;
385 			remaining -= seg_len;
386 		} else {
387 			/* segment only partly filled */
388 			*dst = cpu_to_be32(remaining);
389 			remaining = 0;
390 		}
391 		dst++; src++;
392 
393 		/* segment's RDMA offset */
394 		*dst++ = *src++;
395 		*dst++ = *src++;
396 	}
397 
398 	return nsegs;
399 }
400 
401 /* The client provided a Write list in the Call message. Fill in
402  * the segments in the first Write chunk in the Reply's transport
403  * header with the number of bytes consumed in each segment.
404  * Remaining chunks are returned unused.
405  *
406  * Assumptions:
407  *  - Client has provided only one Write chunk
408  */
409 static void svc_rdma_xdr_encode_write_list(__be32 *rdma_resp, __be32 *wr_ch,
410 					   unsigned int consumed)
411 {
412 	unsigned int nsegs;
413 	__be32 *p, *q;
414 
415 	/* RPC-over-RDMA V1 replies never have a Read list. */
416 	p = rdma_resp + rpcrdma_fixed_maxsz + 1;
417 
418 	q = wr_ch;
419 	while (*q != xdr_zero) {
420 		nsegs = xdr_encode_write_chunk(p, q, consumed);
421 		q += 2 + nsegs * rpcrdma_segment_maxsz;
422 		p += 2 + nsegs * rpcrdma_segment_maxsz;
423 		consumed = 0;
424 	}
425 
426 	/* Terminate Write list */
427 	*p++ = xdr_zero;
428 
429 	/* Reply chunk discriminator; may be replaced later */
430 	*p = xdr_zero;
431 }
432 
433 /* The client provided a Reply chunk in the Call message. Fill in
434  * the segments in the Reply chunk in the Reply message with the
435  * number of bytes consumed in each segment.
436  *
437  * Assumptions:
438  * - Reply can always fit in the provided Reply chunk
439  */
440 static void svc_rdma_xdr_encode_reply_chunk(__be32 *rdma_resp, __be32 *rp_ch,
441 					    unsigned int consumed)
442 {
443 	__be32 *p;
444 
445 	/* Find the Reply chunk in the Reply's xprt header.
446 	 * RPC-over-RDMA V1 replies never have a Read list.
447 	 */
448 	p = rdma_resp + rpcrdma_fixed_maxsz + 1;
449 
450 	/* Skip past Write list */
451 	while (*p++ != xdr_zero)
452 		p += 1 + be32_to_cpup(p) * rpcrdma_segment_maxsz;
453 
454 	xdr_encode_write_chunk(p, rp_ch, consumed);
455 }
456 
457 /* Parse the RPC Call's transport header.
458  */
459 static void svc_rdma_get_write_arrays(__be32 *rdma_argp,
460 				      __be32 **write, __be32 **reply)
461 {
462 	__be32 *p;
463 
464 	p = rdma_argp + rpcrdma_fixed_maxsz;
465 
466 	/* Read list */
467 	while (*p++ != xdr_zero)
468 		p += 5;
469 
470 	/* Write list */
471 	if (*p != xdr_zero) {
472 		*write = p;
473 		while (*p++ != xdr_zero)
474 			p += 1 + be32_to_cpu(*p) * 4;
475 	} else {
476 		*write = NULL;
477 		p++;
478 	}
479 
480 	/* Reply chunk */
481 	if (*p != xdr_zero)
482 		*reply = p;
483 	else
484 		*reply = NULL;
485 }
486 
487 /* RPC-over-RDMA Version One private extension: Remote Invalidation.
488  * Responder's choice: requester signals it can handle Send With
489  * Invalidate, and responder chooses one rkey to invalidate.
490  *
491  * Find a candidate rkey to invalidate when sending a reply.  Picks the
492  * first R_key it finds in the chunk lists.
493  *
494  * Returns zero if RPC's chunk lists are empty.
495  */
496 static u32 svc_rdma_get_inv_rkey(__be32 *rdma_argp,
497 				 __be32 *wr_lst, __be32 *rp_ch)
498 {
499 	__be32 *p;
500 
501 	p = rdma_argp + rpcrdma_fixed_maxsz;
502 	if (*p != xdr_zero)
503 		p += 2;
504 	else if (wr_lst && be32_to_cpup(wr_lst + 1))
505 		p = wr_lst + 2;
506 	else if (rp_ch && be32_to_cpup(rp_ch + 1))
507 		p = rp_ch + 2;
508 	else
509 		return 0;
510 	return be32_to_cpup(p);
511 }
512 
513 static int svc_rdma_dma_map_page(struct svcxprt_rdma *rdma,
514 				 struct svc_rdma_send_ctxt *ctxt,
515 				 struct page *page,
516 				 unsigned long offset,
517 				 unsigned int len)
518 {
519 	struct ib_device *dev = rdma->sc_cm_id->device;
520 	dma_addr_t dma_addr;
521 
522 	dma_addr = ib_dma_map_page(dev, page, offset, len, DMA_TO_DEVICE);
523 	if (ib_dma_mapping_error(dev, dma_addr))
524 		goto out_maperr;
525 
526 	ctxt->sc_sges[ctxt->sc_cur_sge_no].addr = dma_addr;
527 	ctxt->sc_sges[ctxt->sc_cur_sge_no].length = len;
528 	ctxt->sc_send_wr.num_sge++;
529 	return 0;
530 
531 out_maperr:
532 	trace_svcrdma_dma_map_page(rdma, page);
533 	return -EIO;
534 }
535 
536 /* ib_dma_map_page() is used here because svc_rdma_dma_unmap()
537  * handles DMA-unmap and it uses ib_dma_unmap_page() exclusively.
538  */
539 static int svc_rdma_dma_map_buf(struct svcxprt_rdma *rdma,
540 				struct svc_rdma_send_ctxt *ctxt,
541 				unsigned char *base,
542 				unsigned int len)
543 {
544 	return svc_rdma_dma_map_page(rdma, ctxt, virt_to_page(base),
545 				     offset_in_page(base), len);
546 }
547 
548 /**
549  * svc_rdma_sync_reply_hdr - DMA sync the transport header buffer
550  * @rdma: controlling transport
551  * @ctxt: send_ctxt for the Send WR
552  * @len: length of transport header
553  *
554  */
555 void svc_rdma_sync_reply_hdr(struct svcxprt_rdma *rdma,
556 			     struct svc_rdma_send_ctxt *ctxt,
557 			     unsigned int len)
558 {
559 	ctxt->sc_sges[0].length = len;
560 	ctxt->sc_send_wr.num_sge++;
561 	ib_dma_sync_single_for_device(rdma->sc_pd->device,
562 				      ctxt->sc_sges[0].addr, len,
563 				      DMA_TO_DEVICE);
564 }
565 
566 /* svc_rdma_map_reply_msg - Map the buffer holding RPC message
567  * @rdma: controlling transport
568  * @ctxt: send_ctxt for the Send WR
569  * @xdr: prepared xdr_buf containing RPC message
570  * @wr_lst: pointer to Call header's Write list, or NULL
571  *
572  * Load the xdr_buf into the ctxt's sge array, and DMA map each
573  * element as it is added.
574  *
575  * Returns zero on success, or a negative errno on failure.
576  */
577 int svc_rdma_map_reply_msg(struct svcxprt_rdma *rdma,
578 			   struct svc_rdma_send_ctxt *ctxt,
579 			   struct xdr_buf *xdr, __be32 *wr_lst)
580 {
581 	unsigned int len, remaining;
582 	unsigned long page_off;
583 	struct page **ppages;
584 	unsigned char *base;
585 	u32 xdr_pad;
586 	int ret;
587 
588 	if (++ctxt->sc_cur_sge_no >= rdma->sc_max_send_sges)
589 		return -EIO;
590 	ret = svc_rdma_dma_map_buf(rdma, ctxt,
591 				   xdr->head[0].iov_base,
592 				   xdr->head[0].iov_len);
593 	if (ret < 0)
594 		return ret;
595 
596 	/* If a Write chunk is present, the xdr_buf's page list
597 	 * is not included inline. However the Upper Layer may
598 	 * have added XDR padding in the tail buffer, and that
599 	 * should not be included inline.
600 	 */
601 	if (wr_lst) {
602 		base = xdr->tail[0].iov_base;
603 		len = xdr->tail[0].iov_len;
604 		xdr_pad = xdr_padsize(xdr->page_len);
605 
606 		if (len && xdr_pad) {
607 			base += xdr_pad;
608 			len -= xdr_pad;
609 		}
610 
611 		goto tail;
612 	}
613 
614 	ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
615 	page_off = xdr->page_base & ~PAGE_MASK;
616 	remaining = xdr->page_len;
617 	while (remaining) {
618 		len = min_t(u32, PAGE_SIZE - page_off, remaining);
619 
620 		if (++ctxt->sc_cur_sge_no >= rdma->sc_max_send_sges)
621 			return -EIO;
622 		ret = svc_rdma_dma_map_page(rdma, ctxt, *ppages++,
623 					    page_off, len);
624 		if (ret < 0)
625 			return ret;
626 
627 		remaining -= len;
628 		page_off = 0;
629 	}
630 
631 	base = xdr->tail[0].iov_base;
632 	len = xdr->tail[0].iov_len;
633 tail:
634 	if (len) {
635 		if (++ctxt->sc_cur_sge_no >= rdma->sc_max_send_sges)
636 			return -EIO;
637 		ret = svc_rdma_dma_map_buf(rdma, ctxt, base, len);
638 		if (ret < 0)
639 			return ret;
640 	}
641 
642 	return 0;
643 }
644 
645 /* The svc_rqst and all resources it owns are released as soon as
646  * svc_rdma_sendto returns. Transfer pages under I/O to the ctxt
647  * so they are released by the Send completion handler.
648  */
649 static void svc_rdma_save_io_pages(struct svc_rqst *rqstp,
650 				   struct svc_rdma_send_ctxt *ctxt)
651 {
652 	int i, pages = rqstp->rq_next_page - rqstp->rq_respages;
653 
654 	ctxt->sc_page_count += pages;
655 	for (i = 0; i < pages; i++) {
656 		ctxt->sc_pages[i] = rqstp->rq_respages[i];
657 		rqstp->rq_respages[i] = NULL;
658 	}
659 
660 	/* Prevent svc_xprt_release from releasing pages in rq_pages */
661 	rqstp->rq_next_page = rqstp->rq_respages;
662 }
663 
664 /* Prepare the portion of the RPC Reply that will be transmitted
665  * via RDMA Send. The RPC-over-RDMA transport header is prepared
666  * in sc_sges[0], and the RPC xdr_buf is prepared in following sges.
667  *
668  * Depending on whether a Write list or Reply chunk is present,
669  * the server may send all, a portion of, or none of the xdr_buf.
670  * In the latter case, only the transport header (sc_sges[0]) is
671  * transmitted.
672  *
673  * RDMA Send is the last step of transmitting an RPC reply. Pages
674  * involved in the earlier RDMA Writes are here transferred out
675  * of the rqstp and into the ctxt's page array. These pages are
676  * DMA unmapped by each Write completion, but the subsequent Send
677  * completion finally releases these pages.
678  *
679  * Assumptions:
680  * - The Reply's transport header will never be larger than a page.
681  */
682 static int svc_rdma_send_reply_msg(struct svcxprt_rdma *rdma,
683 				   struct svc_rdma_send_ctxt *ctxt,
684 				   __be32 *rdma_argp,
685 				   struct svc_rqst *rqstp,
686 				   __be32 *wr_lst, __be32 *rp_ch)
687 {
688 	int ret;
689 
690 	if (!rp_ch) {
691 		ret = svc_rdma_map_reply_msg(rdma, ctxt,
692 					     &rqstp->rq_res, wr_lst);
693 		if (ret < 0)
694 			return ret;
695 	}
696 
697 	svc_rdma_save_io_pages(rqstp, ctxt);
698 
699 	ctxt->sc_send_wr.opcode = IB_WR_SEND;
700 	if (rdma->sc_snd_w_inv) {
701 		ctxt->sc_send_wr.ex.invalidate_rkey =
702 			svc_rdma_get_inv_rkey(rdma_argp, wr_lst, rp_ch);
703 		if (ctxt->sc_send_wr.ex.invalidate_rkey)
704 			ctxt->sc_send_wr.opcode = IB_WR_SEND_WITH_INV;
705 	}
706 	dprintk("svcrdma: posting Send WR with %u sge(s)\n",
707 		ctxt->sc_send_wr.num_sge);
708 	return svc_rdma_send(rdma, &ctxt->sc_send_wr);
709 }
710 
711 /* Given the client-provided Write and Reply chunks, the server was not
712  * able to form a complete reply. Return an RDMA_ERROR message so the
713  * client can retire this RPC transaction. As above, the Send completion
714  * routine releases payload pages that were part of a previous RDMA Write.
715  *
716  * Remote Invalidation is skipped for simplicity.
717  */
718 static int svc_rdma_send_error_msg(struct svcxprt_rdma *rdma,
719 				   struct svc_rdma_send_ctxt *ctxt,
720 				   struct svc_rqst *rqstp)
721 {
722 	__be32 *p;
723 	int ret;
724 
725 	p = ctxt->sc_xprt_buf;
726 	trace_svcrdma_err_chunk(*p);
727 	p += 3;
728 	*p++ = rdma_error;
729 	*p   = err_chunk;
730 	svc_rdma_sync_reply_hdr(rdma, ctxt, RPCRDMA_HDRLEN_ERR);
731 
732 	svc_rdma_save_io_pages(rqstp, ctxt);
733 
734 	ctxt->sc_send_wr.opcode = IB_WR_SEND;
735 	ret = svc_rdma_send(rdma, &ctxt->sc_send_wr);
736 	if (ret) {
737 		svc_rdma_send_ctxt_put(rdma, ctxt);
738 		return ret;
739 	}
740 
741 	return 0;
742 }
743 
744 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
745 {
746 }
747 
748 /**
749  * svc_rdma_sendto - Transmit an RPC reply
750  * @rqstp: processed RPC request, reply XDR already in ::rq_res
751  *
752  * Any resources still associated with @rqstp are released upon return.
753  * If no reply message was possible, the connection is closed.
754  *
755  * Returns:
756  *	%0 if an RPC reply has been successfully posted,
757  *	%-ENOMEM if a resource shortage occurred (connection is lost),
758  *	%-ENOTCONN if posting failed (connection is lost).
759  */
760 int svc_rdma_sendto(struct svc_rqst *rqstp)
761 {
762 	struct svc_xprt *xprt = rqstp->rq_xprt;
763 	struct svcxprt_rdma *rdma =
764 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
765 	struct svc_rdma_recv_ctxt *rctxt = rqstp->rq_xprt_ctxt;
766 	__be32 *p, *rdma_argp, *rdma_resp, *wr_lst, *rp_ch;
767 	struct xdr_buf *xdr = &rqstp->rq_res;
768 	struct svc_rdma_send_ctxt *sctxt;
769 	int ret;
770 
771 	rdma_argp = rctxt->rc_recv_buf;
772 	svc_rdma_get_write_arrays(rdma_argp, &wr_lst, &rp_ch);
773 
774 	/* Create the RDMA response header. xprt->xpt_mutex,
775 	 * acquired in svc_send(), serializes RPC replies. The
776 	 * code path below that inserts the credit grant value
777 	 * into each transport header runs only inside this
778 	 * critical section.
779 	 */
780 	ret = -ENOMEM;
781 	sctxt = svc_rdma_send_ctxt_get(rdma);
782 	if (!sctxt)
783 		goto err0;
784 	rdma_resp = sctxt->sc_xprt_buf;
785 
786 	p = rdma_resp;
787 	*p++ = *rdma_argp;
788 	*p++ = *(rdma_argp + 1);
789 	*p++ = rdma->sc_fc_credits;
790 	*p++ = rp_ch ? rdma_nomsg : rdma_msg;
791 
792 	/* Start with empty chunks */
793 	*p++ = xdr_zero;
794 	*p++ = xdr_zero;
795 	*p   = xdr_zero;
796 
797 	if (wr_lst) {
798 		/* XXX: Presume the client sent only one Write chunk */
799 		ret = svc_rdma_send_write_chunk(rdma, wr_lst, xdr);
800 		if (ret < 0)
801 			goto err2;
802 		svc_rdma_xdr_encode_write_list(rdma_resp, wr_lst, ret);
803 	}
804 	if (rp_ch) {
805 		ret = svc_rdma_send_reply_chunk(rdma, rp_ch, wr_lst, xdr);
806 		if (ret < 0)
807 			goto err2;
808 		svc_rdma_xdr_encode_reply_chunk(rdma_resp, rp_ch, ret);
809 	}
810 
811 	svc_rdma_sync_reply_hdr(rdma, sctxt, svc_rdma_reply_hdr_len(rdma_resp));
812 	ret = svc_rdma_send_reply_msg(rdma, sctxt, rdma_argp, rqstp,
813 				      wr_lst, rp_ch);
814 	if (ret < 0)
815 		goto err1;
816 	ret = 0;
817 
818 out:
819 	rqstp->rq_xprt_ctxt = NULL;
820 	svc_rdma_recv_ctxt_put(rdma, rctxt);
821 	return ret;
822 
823  err2:
824 	if (ret != -E2BIG && ret != -EINVAL)
825 		goto err1;
826 
827 	ret = svc_rdma_send_error_msg(rdma, sctxt, rqstp);
828 	if (ret < 0)
829 		goto err1;
830 	ret = 0;
831 	goto out;
832 
833  err1:
834 	svc_rdma_send_ctxt_put(rdma, sctxt);
835  err0:
836 	trace_svcrdma_send_failed(rqstp, ret);
837 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
838 	ret = -ENOTCONN;
839 	goto out;
840 }
841