xref: /linux/net/sunrpc/xprtrdma/svc_rdma_sendto.c (revision b9b77222d4ff6b5bb8f5d87fca20de0910618bb9)
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 	struct ib_send_wr *bad_wr;
295 	int ret;
296 
297 	might_sleep();
298 
299 	/* If the SQ is full, wait until an SQ entry is available */
300 	while (1) {
301 		if ((atomic_dec_return(&rdma->sc_sq_avail) < 0)) {
302 			atomic_inc(&rdma_stat_sq_starve);
303 			trace_svcrdma_sq_full(rdma);
304 			atomic_inc(&rdma->sc_sq_avail);
305 			wait_event(rdma->sc_send_wait,
306 				   atomic_read(&rdma->sc_sq_avail) > 1);
307 			if (test_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags))
308 				return -ENOTCONN;
309 			trace_svcrdma_sq_retry(rdma);
310 			continue;
311 		}
312 
313 		svc_xprt_get(&rdma->sc_xprt);
314 		ret = ib_post_send(rdma->sc_qp, wr, &bad_wr);
315 		trace_svcrdma_post_send(wr, ret);
316 		if (ret) {
317 			set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
318 			svc_xprt_put(&rdma->sc_xprt);
319 			wake_up(&rdma->sc_send_wait);
320 		}
321 		break;
322 	}
323 	return ret;
324 }
325 
326 static u32 xdr_padsize(u32 len)
327 {
328 	return (len & 3) ? (4 - (len & 3)) : 0;
329 }
330 
331 /* Returns length of transport header, in bytes.
332  */
333 static unsigned int svc_rdma_reply_hdr_len(__be32 *rdma_resp)
334 {
335 	unsigned int nsegs;
336 	__be32 *p;
337 
338 	p = rdma_resp;
339 
340 	/* RPC-over-RDMA V1 replies never have a Read list. */
341 	p += rpcrdma_fixed_maxsz + 1;
342 
343 	/* Skip Write list. */
344 	while (*p++ != xdr_zero) {
345 		nsegs = be32_to_cpup(p++);
346 		p += nsegs * rpcrdma_segment_maxsz;
347 	}
348 
349 	/* Skip Reply chunk. */
350 	if (*p++ != xdr_zero) {
351 		nsegs = be32_to_cpup(p++);
352 		p += nsegs * rpcrdma_segment_maxsz;
353 	}
354 
355 	return (unsigned long)p - (unsigned long)rdma_resp;
356 }
357 
358 /* One Write chunk is copied from Call transport header to Reply
359  * transport header. Each segment's length field is updated to
360  * reflect number of bytes consumed in the segment.
361  *
362  * Returns number of segments in this chunk.
363  */
364 static unsigned int xdr_encode_write_chunk(__be32 *dst, __be32 *src,
365 					   unsigned int remaining)
366 {
367 	unsigned int i, nsegs;
368 	u32 seg_len;
369 
370 	/* Write list discriminator */
371 	*dst++ = *src++;
372 
373 	/* number of segments in this chunk */
374 	nsegs = be32_to_cpup(src);
375 	*dst++ = *src++;
376 
377 	for (i = nsegs; i; i--) {
378 		/* segment's RDMA handle */
379 		*dst++ = *src++;
380 
381 		/* bytes returned in this segment */
382 		seg_len = be32_to_cpu(*src);
383 		if (remaining >= seg_len) {
384 			/* entire segment was consumed */
385 			*dst = *src;
386 			remaining -= seg_len;
387 		} else {
388 			/* segment only partly filled */
389 			*dst = cpu_to_be32(remaining);
390 			remaining = 0;
391 		}
392 		dst++; src++;
393 
394 		/* segment's RDMA offset */
395 		*dst++ = *src++;
396 		*dst++ = *src++;
397 	}
398 
399 	return nsegs;
400 }
401 
402 /* The client provided a Write list in the Call message. Fill in
403  * the segments in the first Write chunk in the Reply's transport
404  * header with the number of bytes consumed in each segment.
405  * Remaining chunks are returned unused.
406  *
407  * Assumptions:
408  *  - Client has provided only one Write chunk
409  */
410 static void svc_rdma_xdr_encode_write_list(__be32 *rdma_resp, __be32 *wr_ch,
411 					   unsigned int consumed)
412 {
413 	unsigned int nsegs;
414 	__be32 *p, *q;
415 
416 	/* RPC-over-RDMA V1 replies never have a Read list. */
417 	p = rdma_resp + rpcrdma_fixed_maxsz + 1;
418 
419 	q = wr_ch;
420 	while (*q != xdr_zero) {
421 		nsegs = xdr_encode_write_chunk(p, q, consumed);
422 		q += 2 + nsegs * rpcrdma_segment_maxsz;
423 		p += 2 + nsegs * rpcrdma_segment_maxsz;
424 		consumed = 0;
425 	}
426 
427 	/* Terminate Write list */
428 	*p++ = xdr_zero;
429 
430 	/* Reply chunk discriminator; may be replaced later */
431 	*p = xdr_zero;
432 }
433 
434 /* The client provided a Reply chunk in the Call message. Fill in
435  * the segments in the Reply chunk in the Reply message with the
436  * number of bytes consumed in each segment.
437  *
438  * Assumptions:
439  * - Reply can always fit in the provided Reply chunk
440  */
441 static void svc_rdma_xdr_encode_reply_chunk(__be32 *rdma_resp, __be32 *rp_ch,
442 					    unsigned int consumed)
443 {
444 	__be32 *p;
445 
446 	/* Find the Reply chunk in the Reply's xprt header.
447 	 * RPC-over-RDMA V1 replies never have a Read list.
448 	 */
449 	p = rdma_resp + rpcrdma_fixed_maxsz + 1;
450 
451 	/* Skip past Write list */
452 	while (*p++ != xdr_zero)
453 		p += 1 + be32_to_cpup(p) * rpcrdma_segment_maxsz;
454 
455 	xdr_encode_write_chunk(p, rp_ch, consumed);
456 }
457 
458 /* Parse the RPC Call's transport header.
459  */
460 static void svc_rdma_get_write_arrays(__be32 *rdma_argp,
461 				      __be32 **write, __be32 **reply)
462 {
463 	__be32 *p;
464 
465 	p = rdma_argp + rpcrdma_fixed_maxsz;
466 
467 	/* Read list */
468 	while (*p++ != xdr_zero)
469 		p += 5;
470 
471 	/* Write list */
472 	if (*p != xdr_zero) {
473 		*write = p;
474 		while (*p++ != xdr_zero)
475 			p += 1 + be32_to_cpu(*p) * 4;
476 	} else {
477 		*write = NULL;
478 		p++;
479 	}
480 
481 	/* Reply chunk */
482 	if (*p != xdr_zero)
483 		*reply = p;
484 	else
485 		*reply = NULL;
486 }
487 
488 /* RPC-over-RDMA Version One private extension: Remote Invalidation.
489  * Responder's choice: requester signals it can handle Send With
490  * Invalidate, and responder chooses one rkey to invalidate.
491  *
492  * Find a candidate rkey to invalidate when sending a reply.  Picks the
493  * first R_key it finds in the chunk lists.
494  *
495  * Returns zero if RPC's chunk lists are empty.
496  */
497 static u32 svc_rdma_get_inv_rkey(__be32 *rdma_argp,
498 				 __be32 *wr_lst, __be32 *rp_ch)
499 {
500 	__be32 *p;
501 
502 	p = rdma_argp + rpcrdma_fixed_maxsz;
503 	if (*p != xdr_zero)
504 		p += 2;
505 	else if (wr_lst && be32_to_cpup(wr_lst + 1))
506 		p = wr_lst + 2;
507 	else if (rp_ch && be32_to_cpup(rp_ch + 1))
508 		p = rp_ch + 2;
509 	else
510 		return 0;
511 	return be32_to_cpup(p);
512 }
513 
514 static int svc_rdma_dma_map_page(struct svcxprt_rdma *rdma,
515 				 struct svc_rdma_send_ctxt *ctxt,
516 				 struct page *page,
517 				 unsigned long offset,
518 				 unsigned int len)
519 {
520 	struct ib_device *dev = rdma->sc_cm_id->device;
521 	dma_addr_t dma_addr;
522 
523 	dma_addr = ib_dma_map_page(dev, page, offset, len, DMA_TO_DEVICE);
524 	if (ib_dma_mapping_error(dev, dma_addr))
525 		goto out_maperr;
526 
527 	ctxt->sc_sges[ctxt->sc_cur_sge_no].addr = dma_addr;
528 	ctxt->sc_sges[ctxt->sc_cur_sge_no].length = len;
529 	ctxt->sc_send_wr.num_sge++;
530 	return 0;
531 
532 out_maperr:
533 	trace_svcrdma_dma_map_page(rdma, page);
534 	return -EIO;
535 }
536 
537 /* ib_dma_map_page() is used here because svc_rdma_dma_unmap()
538  * handles DMA-unmap and it uses ib_dma_unmap_page() exclusively.
539  */
540 static int svc_rdma_dma_map_buf(struct svcxprt_rdma *rdma,
541 				struct svc_rdma_send_ctxt *ctxt,
542 				unsigned char *base,
543 				unsigned int len)
544 {
545 	return svc_rdma_dma_map_page(rdma, ctxt, virt_to_page(base),
546 				     offset_in_page(base), len);
547 }
548 
549 /**
550  * svc_rdma_sync_reply_hdr - DMA sync the transport header buffer
551  * @rdma: controlling transport
552  * @ctxt: send_ctxt for the Send WR
553  * @len: length of transport header
554  *
555  */
556 void svc_rdma_sync_reply_hdr(struct svcxprt_rdma *rdma,
557 			     struct svc_rdma_send_ctxt *ctxt,
558 			     unsigned int len)
559 {
560 	ctxt->sc_sges[0].length = len;
561 	ctxt->sc_send_wr.num_sge++;
562 	ib_dma_sync_single_for_device(rdma->sc_pd->device,
563 				      ctxt->sc_sges[0].addr, len,
564 				      DMA_TO_DEVICE);
565 }
566 
567 /* svc_rdma_map_reply_msg - Map the buffer holding RPC message
568  * @rdma: controlling transport
569  * @ctxt: send_ctxt for the Send WR
570  * @xdr: prepared xdr_buf containing RPC message
571  * @wr_lst: pointer to Call header's Write list, or NULL
572  *
573  * Load the xdr_buf into the ctxt's sge array, and DMA map each
574  * element as it is added.
575  *
576  * Returns zero on success, or a negative errno on failure.
577  */
578 int svc_rdma_map_reply_msg(struct svcxprt_rdma *rdma,
579 			   struct svc_rdma_send_ctxt *ctxt,
580 			   struct xdr_buf *xdr, __be32 *wr_lst)
581 {
582 	unsigned int len, remaining;
583 	unsigned long page_off;
584 	struct page **ppages;
585 	unsigned char *base;
586 	u32 xdr_pad;
587 	int ret;
588 
589 	if (++ctxt->sc_cur_sge_no >= rdma->sc_max_send_sges)
590 		return -EIO;
591 	ret = svc_rdma_dma_map_buf(rdma, ctxt,
592 				   xdr->head[0].iov_base,
593 				   xdr->head[0].iov_len);
594 	if (ret < 0)
595 		return ret;
596 
597 	/* If a Write chunk is present, the xdr_buf's page list
598 	 * is not included inline. However the Upper Layer may
599 	 * have added XDR padding in the tail buffer, and that
600 	 * should not be included inline.
601 	 */
602 	if (wr_lst) {
603 		base = xdr->tail[0].iov_base;
604 		len = xdr->tail[0].iov_len;
605 		xdr_pad = xdr_padsize(xdr->page_len);
606 
607 		if (len && xdr_pad) {
608 			base += xdr_pad;
609 			len -= xdr_pad;
610 		}
611 
612 		goto tail;
613 	}
614 
615 	ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
616 	page_off = xdr->page_base & ~PAGE_MASK;
617 	remaining = xdr->page_len;
618 	while (remaining) {
619 		len = min_t(u32, PAGE_SIZE - page_off, remaining);
620 
621 		if (++ctxt->sc_cur_sge_no >= rdma->sc_max_send_sges)
622 			return -EIO;
623 		ret = svc_rdma_dma_map_page(rdma, ctxt, *ppages++,
624 					    page_off, len);
625 		if (ret < 0)
626 			return ret;
627 
628 		remaining -= len;
629 		page_off = 0;
630 	}
631 
632 	base = xdr->tail[0].iov_base;
633 	len = xdr->tail[0].iov_len;
634 tail:
635 	if (len) {
636 		if (++ctxt->sc_cur_sge_no >= rdma->sc_max_send_sges)
637 			return -EIO;
638 		ret = svc_rdma_dma_map_buf(rdma, ctxt, base, len);
639 		if (ret < 0)
640 			return ret;
641 	}
642 
643 	return 0;
644 }
645 
646 /* The svc_rqst and all resources it owns are released as soon as
647  * svc_rdma_sendto returns. Transfer pages under I/O to the ctxt
648  * so they are released by the Send completion handler.
649  */
650 static void svc_rdma_save_io_pages(struct svc_rqst *rqstp,
651 				   struct svc_rdma_send_ctxt *ctxt)
652 {
653 	int i, pages = rqstp->rq_next_page - rqstp->rq_respages;
654 
655 	ctxt->sc_page_count += pages;
656 	for (i = 0; i < pages; i++) {
657 		ctxt->sc_pages[i] = rqstp->rq_respages[i];
658 		rqstp->rq_respages[i] = NULL;
659 	}
660 	rqstp->rq_next_page = rqstp->rq_respages + 1;
661 }
662 
663 /* Prepare the portion of the RPC Reply that will be transmitted
664  * via RDMA Send. The RPC-over-RDMA transport header is prepared
665  * in sc_sges[0], and the RPC xdr_buf is prepared in following sges.
666  *
667  * Depending on whether a Write list or Reply chunk is present,
668  * the server may send all, a portion of, or none of the xdr_buf.
669  * In the latter case, only the transport header (sc_sges[0]) is
670  * transmitted.
671  *
672  * RDMA Send is the last step of transmitting an RPC reply. Pages
673  * involved in the earlier RDMA Writes are here transferred out
674  * of the rqstp and into the ctxt's page array. These pages are
675  * DMA unmapped by each Write completion, but the subsequent Send
676  * completion finally releases these pages.
677  *
678  * Assumptions:
679  * - The Reply's transport header will never be larger than a page.
680  */
681 static int svc_rdma_send_reply_msg(struct svcxprt_rdma *rdma,
682 				   struct svc_rdma_send_ctxt *ctxt,
683 				   __be32 *rdma_argp,
684 				   struct svc_rqst *rqstp,
685 				   __be32 *wr_lst, __be32 *rp_ch)
686 {
687 	int ret;
688 
689 	if (!rp_ch) {
690 		ret = svc_rdma_map_reply_msg(rdma, ctxt,
691 					     &rqstp->rq_res, wr_lst);
692 		if (ret < 0)
693 			return ret;
694 	}
695 
696 	svc_rdma_save_io_pages(rqstp, ctxt);
697 
698 	ctxt->sc_send_wr.opcode = IB_WR_SEND;
699 	if (rdma->sc_snd_w_inv) {
700 		ctxt->sc_send_wr.ex.invalidate_rkey =
701 			svc_rdma_get_inv_rkey(rdma_argp, wr_lst, rp_ch);
702 		if (ctxt->sc_send_wr.ex.invalidate_rkey)
703 			ctxt->sc_send_wr.opcode = IB_WR_SEND_WITH_INV;
704 	}
705 	dprintk("svcrdma: posting Send WR with %u sge(s)\n",
706 		ctxt->sc_send_wr.num_sge);
707 	return svc_rdma_send(rdma, &ctxt->sc_send_wr);
708 }
709 
710 /* Given the client-provided Write and Reply chunks, the server was not
711  * able to form a complete reply. Return an RDMA_ERROR message so the
712  * client can retire this RPC transaction. As above, the Send completion
713  * routine releases payload pages that were part of a previous RDMA Write.
714  *
715  * Remote Invalidation is skipped for simplicity.
716  */
717 static int svc_rdma_send_error_msg(struct svcxprt_rdma *rdma,
718 				   struct svc_rdma_send_ctxt *ctxt,
719 				   struct svc_rqst *rqstp)
720 {
721 	__be32 *p;
722 	int ret;
723 
724 	p = ctxt->sc_xprt_buf;
725 	trace_svcrdma_err_chunk(*p);
726 	p += 3;
727 	*p++ = rdma_error;
728 	*p   = err_chunk;
729 	svc_rdma_sync_reply_hdr(rdma, ctxt, RPCRDMA_HDRLEN_ERR);
730 
731 	svc_rdma_save_io_pages(rqstp, ctxt);
732 
733 	ctxt->sc_send_wr.opcode = IB_WR_SEND;
734 	ret = svc_rdma_send(rdma, &ctxt->sc_send_wr);
735 	if (ret) {
736 		svc_rdma_send_ctxt_put(rdma, ctxt);
737 		return ret;
738 	}
739 
740 	return 0;
741 }
742 
743 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp)
744 {
745 }
746 
747 /**
748  * svc_rdma_sendto - Transmit an RPC reply
749  * @rqstp: processed RPC request, reply XDR already in ::rq_res
750  *
751  * Any resources still associated with @rqstp are released upon return.
752  * If no reply message was possible, the connection is closed.
753  *
754  * Returns:
755  *	%0 if an RPC reply has been successfully posted,
756  *	%-ENOMEM if a resource shortage occurred (connection is lost),
757  *	%-ENOTCONN if posting failed (connection is lost).
758  */
759 int svc_rdma_sendto(struct svc_rqst *rqstp)
760 {
761 	struct svc_xprt *xprt = rqstp->rq_xprt;
762 	struct svcxprt_rdma *rdma =
763 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
764 	struct svc_rdma_recv_ctxt *rctxt = rqstp->rq_xprt_ctxt;
765 	__be32 *p, *rdma_argp, *rdma_resp, *wr_lst, *rp_ch;
766 	struct xdr_buf *xdr = &rqstp->rq_res;
767 	struct svc_rdma_send_ctxt *sctxt;
768 	int ret;
769 
770 	rdma_argp = rctxt->rc_recv_buf;
771 	svc_rdma_get_write_arrays(rdma_argp, &wr_lst, &rp_ch);
772 
773 	/* Create the RDMA response header. xprt->xpt_mutex,
774 	 * acquired in svc_send(), serializes RPC replies. The
775 	 * code path below that inserts the credit grant value
776 	 * into each transport header runs only inside this
777 	 * critical section.
778 	 */
779 	ret = -ENOMEM;
780 	sctxt = svc_rdma_send_ctxt_get(rdma);
781 	if (!sctxt)
782 		goto err0;
783 	rdma_resp = sctxt->sc_xprt_buf;
784 
785 	p = rdma_resp;
786 	*p++ = *rdma_argp;
787 	*p++ = *(rdma_argp + 1);
788 	*p++ = rdma->sc_fc_credits;
789 	*p++ = rp_ch ? rdma_nomsg : rdma_msg;
790 
791 	/* Start with empty chunks */
792 	*p++ = xdr_zero;
793 	*p++ = xdr_zero;
794 	*p   = xdr_zero;
795 
796 	if (wr_lst) {
797 		/* XXX: Presume the client sent only one Write chunk */
798 		ret = svc_rdma_send_write_chunk(rdma, wr_lst, xdr);
799 		if (ret < 0)
800 			goto err2;
801 		svc_rdma_xdr_encode_write_list(rdma_resp, wr_lst, ret);
802 	}
803 	if (rp_ch) {
804 		ret = svc_rdma_send_reply_chunk(rdma, rp_ch, wr_lst, xdr);
805 		if (ret < 0)
806 			goto err2;
807 		svc_rdma_xdr_encode_reply_chunk(rdma_resp, rp_ch, ret);
808 	}
809 
810 	svc_rdma_sync_reply_hdr(rdma, sctxt, svc_rdma_reply_hdr_len(rdma_resp));
811 	ret = svc_rdma_send_reply_msg(rdma, sctxt, rdma_argp, rqstp,
812 				      wr_lst, rp_ch);
813 	if (ret < 0)
814 		goto err1;
815 	ret = 0;
816 
817 out:
818 	rqstp->rq_xprt_ctxt = NULL;
819 	svc_rdma_recv_ctxt_put(rdma, rctxt);
820 	return ret;
821 
822  err2:
823 	if (ret != -E2BIG && ret != -EINVAL)
824 		goto err1;
825 
826 	ret = svc_rdma_send_error_msg(rdma, sctxt, rqstp);
827 	if (ret < 0)
828 		goto err1;
829 	ret = 0;
830 	goto out;
831 
832  err1:
833 	svc_rdma_send_ctxt_put(rdma, sctxt);
834  err0:
835 	trace_svcrdma_send_failed(rqstp, ret);
836 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
837 	ret = -ENOTCONN;
838 	goto out;
839 }
840