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 	}
276 
277 	svc_xprt_put(&rdma->sc_xprt);
278 }
279 
280 /**
281  * svc_rdma_send - Post a single Send WR
282  * @rdma: transport on which to post the WR
283  * @wr: prepared Send WR to post
284  *
285  * Returns zero the Send WR was posted successfully. Otherwise, a
286  * negative errno is returned.
287  */
288 int svc_rdma_send(struct svcxprt_rdma *rdma, struct ib_send_wr *wr)
289 {
290 	int ret;
291 
292 	might_sleep();
293 
294 	/* If the SQ is full, wait until an SQ entry is available */
295 	while (1) {
296 		if ((atomic_dec_return(&rdma->sc_sq_avail) < 0)) {
297 			atomic_inc(&rdma_stat_sq_starve);
298 			trace_svcrdma_sq_full(rdma);
299 			atomic_inc(&rdma->sc_sq_avail);
300 			wait_event(rdma->sc_send_wait,
301 				   atomic_read(&rdma->sc_sq_avail) > 1);
302 			if (test_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags))
303 				return -ENOTCONN;
304 			trace_svcrdma_sq_retry(rdma);
305 			continue;
306 		}
307 
308 		svc_xprt_get(&rdma->sc_xprt);
309 		ret = ib_post_send(rdma->sc_qp, wr, NULL);
310 		trace_svcrdma_post_send(wr, ret);
311 		if (ret) {
312 			set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
313 			svc_xprt_put(&rdma->sc_xprt);
314 			wake_up(&rdma->sc_send_wait);
315 		}
316 		break;
317 	}
318 	return ret;
319 }
320 
321 static u32 xdr_padsize(u32 len)
322 {
323 	return (len & 3) ? (4 - (len & 3)) : 0;
324 }
325 
326 /* Returns length of transport header, in bytes.
327  */
328 static unsigned int svc_rdma_reply_hdr_len(__be32 *rdma_resp)
329 {
330 	unsigned int nsegs;
331 	__be32 *p;
332 
333 	p = rdma_resp;
334 
335 	/* RPC-over-RDMA V1 replies never have a Read list. */
336 	p += rpcrdma_fixed_maxsz + 1;
337 
338 	/* Skip Write list. */
339 	while (*p++ != xdr_zero) {
340 		nsegs = be32_to_cpup(p++);
341 		p += nsegs * rpcrdma_segment_maxsz;
342 	}
343 
344 	/* Skip Reply chunk. */
345 	if (*p++ != xdr_zero) {
346 		nsegs = be32_to_cpup(p++);
347 		p += nsegs * rpcrdma_segment_maxsz;
348 	}
349 
350 	return (unsigned long)p - (unsigned long)rdma_resp;
351 }
352 
353 /* One Write chunk is copied from Call transport header to Reply
354  * transport header. Each segment's length field is updated to
355  * reflect number of bytes consumed in the segment.
356  *
357  * Returns number of segments in this chunk.
358  */
359 static unsigned int xdr_encode_write_chunk(__be32 *dst, __be32 *src,
360 					   unsigned int remaining)
361 {
362 	unsigned int i, nsegs;
363 	u32 seg_len;
364 
365 	/* Write list discriminator */
366 	*dst++ = *src++;
367 
368 	/* number of segments in this chunk */
369 	nsegs = be32_to_cpup(src);
370 	*dst++ = *src++;
371 
372 	for (i = nsegs; i; i--) {
373 		/* segment's RDMA handle */
374 		*dst++ = *src++;
375 
376 		/* bytes returned in this segment */
377 		seg_len = be32_to_cpu(*src);
378 		if (remaining >= seg_len) {
379 			/* entire segment was consumed */
380 			*dst = *src;
381 			remaining -= seg_len;
382 		} else {
383 			/* segment only partly filled */
384 			*dst = cpu_to_be32(remaining);
385 			remaining = 0;
386 		}
387 		dst++; src++;
388 
389 		/* segment's RDMA offset */
390 		*dst++ = *src++;
391 		*dst++ = *src++;
392 	}
393 
394 	return nsegs;
395 }
396 
397 /* The client provided a Write list in the Call message. Fill in
398  * the segments in the first Write chunk in the Reply's transport
399  * header with the number of bytes consumed in each segment.
400  * Remaining chunks are returned unused.
401  *
402  * Assumptions:
403  *  - Client has provided only one Write chunk
404  */
405 static void svc_rdma_xdr_encode_write_list(__be32 *rdma_resp, __be32 *wr_ch,
406 					   unsigned int consumed)
407 {
408 	unsigned int nsegs;
409 	__be32 *p, *q;
410 
411 	/* RPC-over-RDMA V1 replies never have a Read list. */
412 	p = rdma_resp + rpcrdma_fixed_maxsz + 1;
413 
414 	q = wr_ch;
415 	while (*q != xdr_zero) {
416 		nsegs = xdr_encode_write_chunk(p, q, consumed);
417 		q += 2 + nsegs * rpcrdma_segment_maxsz;
418 		p += 2 + nsegs * rpcrdma_segment_maxsz;
419 		consumed = 0;
420 	}
421 
422 	/* Terminate Write list */
423 	*p++ = xdr_zero;
424 
425 	/* Reply chunk discriminator; may be replaced later */
426 	*p = xdr_zero;
427 }
428 
429 /* The client provided a Reply chunk in the Call message. Fill in
430  * the segments in the Reply chunk in the Reply message with the
431  * number of bytes consumed in each segment.
432  *
433  * Assumptions:
434  * - Reply can always fit in the provided Reply chunk
435  */
436 static void svc_rdma_xdr_encode_reply_chunk(__be32 *rdma_resp, __be32 *rp_ch,
437 					    unsigned int consumed)
438 {
439 	__be32 *p;
440 
441 	/* Find the Reply chunk in the Reply's xprt header.
442 	 * RPC-over-RDMA V1 replies never have a Read list.
443 	 */
444 	p = rdma_resp + rpcrdma_fixed_maxsz + 1;
445 
446 	/* Skip past Write list */
447 	while (*p++ != xdr_zero)
448 		p += 1 + be32_to_cpup(p) * rpcrdma_segment_maxsz;
449 
450 	xdr_encode_write_chunk(p, rp_ch, consumed);
451 }
452 
453 /* Parse the RPC Call's transport header.
454  */
455 static void svc_rdma_get_write_arrays(__be32 *rdma_argp,
456 				      __be32 **write, __be32 **reply)
457 {
458 	__be32 *p;
459 
460 	p = rdma_argp + rpcrdma_fixed_maxsz;
461 
462 	/* Read list */
463 	while (*p++ != xdr_zero)
464 		p += 5;
465 
466 	/* Write list */
467 	if (*p != xdr_zero) {
468 		*write = p;
469 		while (*p++ != xdr_zero)
470 			p += 1 + be32_to_cpu(*p) * 4;
471 	} else {
472 		*write = NULL;
473 		p++;
474 	}
475 
476 	/* Reply chunk */
477 	if (*p != xdr_zero)
478 		*reply = p;
479 	else
480 		*reply = NULL;
481 }
482 
483 static int svc_rdma_dma_map_page(struct svcxprt_rdma *rdma,
484 				 struct svc_rdma_send_ctxt *ctxt,
485 				 struct page *page,
486 				 unsigned long offset,
487 				 unsigned int len)
488 {
489 	struct ib_device *dev = rdma->sc_cm_id->device;
490 	dma_addr_t dma_addr;
491 
492 	dma_addr = ib_dma_map_page(dev, page, offset, len, DMA_TO_DEVICE);
493 	if (ib_dma_mapping_error(dev, dma_addr))
494 		goto out_maperr;
495 
496 	ctxt->sc_sges[ctxt->sc_cur_sge_no].addr = dma_addr;
497 	ctxt->sc_sges[ctxt->sc_cur_sge_no].length = len;
498 	ctxt->sc_send_wr.num_sge++;
499 	return 0;
500 
501 out_maperr:
502 	trace_svcrdma_dma_map_page(rdma, page);
503 	return -EIO;
504 }
505 
506 /* ib_dma_map_page() is used here because svc_rdma_dma_unmap()
507  * handles DMA-unmap and it uses ib_dma_unmap_page() exclusively.
508  */
509 static int svc_rdma_dma_map_buf(struct svcxprt_rdma *rdma,
510 				struct svc_rdma_send_ctxt *ctxt,
511 				unsigned char *base,
512 				unsigned int len)
513 {
514 	return svc_rdma_dma_map_page(rdma, ctxt, virt_to_page(base),
515 				     offset_in_page(base), len);
516 }
517 
518 /**
519  * svc_rdma_sync_reply_hdr - DMA sync the transport header buffer
520  * @rdma: controlling transport
521  * @ctxt: send_ctxt for the Send WR
522  * @len: length of transport header
523  *
524  */
525 void svc_rdma_sync_reply_hdr(struct svcxprt_rdma *rdma,
526 			     struct svc_rdma_send_ctxt *ctxt,
527 			     unsigned int len)
528 {
529 	ctxt->sc_sges[0].length = len;
530 	ctxt->sc_send_wr.num_sge++;
531 	ib_dma_sync_single_for_device(rdma->sc_pd->device,
532 				      ctxt->sc_sges[0].addr, len,
533 				      DMA_TO_DEVICE);
534 }
535 
536 /* If the xdr_buf has more elements than the device can
537  * transmit in a single RDMA Send, then the reply will
538  * have to be copied into a bounce buffer.
539  */
540 static bool svc_rdma_pull_up_needed(struct svcxprt_rdma *rdma,
541 				    struct xdr_buf *xdr,
542 				    __be32 *wr_lst)
543 {
544 	int elements;
545 
546 	/* xdr->head */
547 	elements = 1;
548 
549 	/* xdr->pages */
550 	if (!wr_lst) {
551 		unsigned int remaining;
552 		unsigned long pageoff;
553 
554 		pageoff = xdr->page_base & ~PAGE_MASK;
555 		remaining = xdr->page_len;
556 		while (remaining) {
557 			++elements;
558 			remaining -= min_t(u32, PAGE_SIZE - pageoff,
559 					   remaining);
560 			pageoff = 0;
561 		}
562 	}
563 
564 	/* xdr->tail */
565 	if (xdr->tail[0].iov_len)
566 		++elements;
567 
568 	/* assume 1 SGE is needed for the transport header */
569 	return elements >= rdma->sc_max_send_sges;
570 }
571 
572 /* The device is not capable of sending the reply directly.
573  * Assemble the elements of @xdr into the transport header
574  * buffer.
575  */
576 static int svc_rdma_pull_up_reply_msg(struct svcxprt_rdma *rdma,
577 				      struct svc_rdma_send_ctxt *ctxt,
578 				      struct xdr_buf *xdr, __be32 *wr_lst)
579 {
580 	unsigned char *dst, *tailbase;
581 	unsigned int taillen;
582 
583 	dst = ctxt->sc_xprt_buf;
584 	dst += ctxt->sc_sges[0].length;
585 
586 	memcpy(dst, xdr->head[0].iov_base, xdr->head[0].iov_len);
587 	dst += xdr->head[0].iov_len;
588 
589 	tailbase = xdr->tail[0].iov_base;
590 	taillen = xdr->tail[0].iov_len;
591 	if (wr_lst) {
592 		u32 xdrpad;
593 
594 		xdrpad = xdr_padsize(xdr->page_len);
595 		if (taillen && xdrpad) {
596 			tailbase += xdrpad;
597 			taillen -= xdrpad;
598 		}
599 	} else {
600 		unsigned int len, remaining;
601 		unsigned long pageoff;
602 		struct page **ppages;
603 
604 		ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
605 		pageoff = xdr->page_base & ~PAGE_MASK;
606 		remaining = xdr->page_len;
607 		while (remaining) {
608 			len = min_t(u32, PAGE_SIZE - pageoff, remaining);
609 
610 			memcpy(dst, page_address(*ppages), len);
611 			remaining -= len;
612 			dst += len;
613 			pageoff = 0;
614 		}
615 	}
616 
617 	if (taillen)
618 		memcpy(dst, tailbase, taillen);
619 
620 	ctxt->sc_sges[0].length += xdr->len;
621 	ib_dma_sync_single_for_device(rdma->sc_pd->device,
622 				      ctxt->sc_sges[0].addr,
623 				      ctxt->sc_sges[0].length,
624 				      DMA_TO_DEVICE);
625 
626 	return 0;
627 }
628 
629 /* svc_rdma_map_reply_msg - Map the buffer holding RPC message
630  * @rdma: controlling transport
631  * @ctxt: send_ctxt for the Send WR
632  * @xdr: prepared xdr_buf containing RPC message
633  * @wr_lst: pointer to Call header's Write list, or NULL
634  *
635  * Load the xdr_buf into the ctxt's sge array, and DMA map each
636  * element as it is added.
637  *
638  * Returns zero on success, or a negative errno on failure.
639  */
640 int svc_rdma_map_reply_msg(struct svcxprt_rdma *rdma,
641 			   struct svc_rdma_send_ctxt *ctxt,
642 			   struct xdr_buf *xdr, __be32 *wr_lst)
643 {
644 	unsigned int len, remaining;
645 	unsigned long page_off;
646 	struct page **ppages;
647 	unsigned char *base;
648 	u32 xdr_pad;
649 	int ret;
650 
651 	if (svc_rdma_pull_up_needed(rdma, xdr, wr_lst))
652 		return svc_rdma_pull_up_reply_msg(rdma, ctxt, xdr, wr_lst);
653 
654 	++ctxt->sc_cur_sge_no;
655 	ret = svc_rdma_dma_map_buf(rdma, ctxt,
656 				   xdr->head[0].iov_base,
657 				   xdr->head[0].iov_len);
658 	if (ret < 0)
659 		return ret;
660 
661 	/* If a Write chunk is present, the xdr_buf's page list
662 	 * is not included inline. However the Upper Layer may
663 	 * have added XDR padding in the tail buffer, and that
664 	 * should not be included inline.
665 	 */
666 	if (wr_lst) {
667 		base = xdr->tail[0].iov_base;
668 		len = xdr->tail[0].iov_len;
669 		xdr_pad = xdr_padsize(xdr->page_len);
670 
671 		if (len && xdr_pad) {
672 			base += xdr_pad;
673 			len -= xdr_pad;
674 		}
675 
676 		goto tail;
677 	}
678 
679 	ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
680 	page_off = xdr->page_base & ~PAGE_MASK;
681 	remaining = xdr->page_len;
682 	while (remaining) {
683 		len = min_t(u32, PAGE_SIZE - page_off, remaining);
684 
685 		++ctxt->sc_cur_sge_no;
686 		ret = svc_rdma_dma_map_page(rdma, ctxt, *ppages++,
687 					    page_off, len);
688 		if (ret < 0)
689 			return ret;
690 
691 		remaining -= len;
692 		page_off = 0;
693 	}
694 
695 	base = xdr->tail[0].iov_base;
696 	len = xdr->tail[0].iov_len;
697 tail:
698 	if (len) {
699 		++ctxt->sc_cur_sge_no;
700 		ret = svc_rdma_dma_map_buf(rdma, ctxt, base, len);
701 		if (ret < 0)
702 			return ret;
703 	}
704 
705 	return 0;
706 }
707 
708 /* The svc_rqst and all resources it owns are released as soon as
709  * svc_rdma_sendto returns. Transfer pages under I/O to the ctxt
710  * so they are released by the Send completion handler.
711  */
712 static void svc_rdma_save_io_pages(struct svc_rqst *rqstp,
713 				   struct svc_rdma_send_ctxt *ctxt)
714 {
715 	int i, pages = rqstp->rq_next_page - rqstp->rq_respages;
716 
717 	ctxt->sc_page_count += pages;
718 	for (i = 0; i < pages; i++) {
719 		ctxt->sc_pages[i] = rqstp->rq_respages[i];
720 		rqstp->rq_respages[i] = NULL;
721 	}
722 
723 	/* Prevent svc_xprt_release from releasing pages in rq_pages */
724 	rqstp->rq_next_page = rqstp->rq_respages;
725 }
726 
727 /* Prepare the portion of the RPC Reply that will be transmitted
728  * via RDMA Send. The RPC-over-RDMA transport header is prepared
729  * in sc_sges[0], and the RPC xdr_buf is prepared in following sges.
730  *
731  * Depending on whether a Write list or Reply chunk is present,
732  * the server may send all, a portion of, or none of the xdr_buf.
733  * In the latter case, only the transport header (sc_sges[0]) is
734  * transmitted.
735  *
736  * RDMA Send is the last step of transmitting an RPC reply. Pages
737  * involved in the earlier RDMA Writes are here transferred out
738  * of the rqstp and into the sctxt's page array. These pages are
739  * DMA unmapped by each Write completion, but the subsequent Send
740  * completion finally releases these pages.
741  *
742  * Assumptions:
743  * - The Reply's transport header will never be larger than a page.
744  */
745 static int svc_rdma_send_reply_msg(struct svcxprt_rdma *rdma,
746 				   struct svc_rdma_send_ctxt *sctxt,
747 				   struct svc_rdma_recv_ctxt *rctxt,
748 				   struct svc_rqst *rqstp,
749 				   __be32 *wr_lst, __be32 *rp_ch)
750 {
751 	int ret;
752 
753 	if (!rp_ch) {
754 		ret = svc_rdma_map_reply_msg(rdma, sctxt,
755 					     &rqstp->rq_res, wr_lst);
756 		if (ret < 0)
757 			return ret;
758 	}
759 
760 	svc_rdma_save_io_pages(rqstp, sctxt);
761 
762 	if (rctxt->rc_inv_rkey) {
763 		sctxt->sc_send_wr.opcode = IB_WR_SEND_WITH_INV;
764 		sctxt->sc_send_wr.ex.invalidate_rkey = rctxt->rc_inv_rkey;
765 	} else {
766 		sctxt->sc_send_wr.opcode = IB_WR_SEND;
767 	}
768 	dprintk("svcrdma: posting Send WR with %u sge(s)\n",
769 		sctxt->sc_send_wr.num_sge);
770 	return svc_rdma_send(rdma, &sctxt->sc_send_wr);
771 }
772 
773 /* Given the client-provided Write and Reply chunks, the server was not
774  * able to form a complete reply. Return an RDMA_ERROR message so the
775  * client can retire this RPC transaction. As above, the Send completion
776  * routine releases payload pages that were part of a previous RDMA Write.
777  *
778  * Remote Invalidation is skipped for simplicity.
779  */
780 static int svc_rdma_send_error_msg(struct svcxprt_rdma *rdma,
781 				   struct svc_rdma_send_ctxt *ctxt,
782 				   struct svc_rqst *rqstp)
783 {
784 	__be32 *p;
785 	int ret;
786 
787 	p = ctxt->sc_xprt_buf;
788 	trace_svcrdma_err_chunk(*p);
789 	p += 3;
790 	*p++ = rdma_error;
791 	*p   = err_chunk;
792 	svc_rdma_sync_reply_hdr(rdma, ctxt, RPCRDMA_HDRLEN_ERR);
793 
794 	svc_rdma_save_io_pages(rqstp, ctxt);
795 
796 	ctxt->sc_send_wr.opcode = IB_WR_SEND;
797 	ret = svc_rdma_send(rdma, &ctxt->sc_send_wr);
798 	if (ret) {
799 		svc_rdma_send_ctxt_put(rdma, ctxt);
800 		return ret;
801 	}
802 
803 	return 0;
804 }
805 
806 /**
807  * svc_rdma_sendto - Transmit an RPC reply
808  * @rqstp: processed RPC request, reply XDR already in ::rq_res
809  *
810  * Any resources still associated with @rqstp are released upon return.
811  * If no reply message was possible, the connection is closed.
812  *
813  * Returns:
814  *	%0 if an RPC reply has been successfully posted,
815  *	%-ENOMEM if a resource shortage occurred (connection is lost),
816  *	%-ENOTCONN if posting failed (connection is lost).
817  */
818 int svc_rdma_sendto(struct svc_rqst *rqstp)
819 {
820 	struct svc_xprt *xprt = rqstp->rq_xprt;
821 	struct svcxprt_rdma *rdma =
822 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
823 	struct svc_rdma_recv_ctxt *rctxt = rqstp->rq_xprt_ctxt;
824 	__be32 *p, *rdma_argp, *rdma_resp, *wr_lst, *rp_ch;
825 	struct xdr_buf *xdr = &rqstp->rq_res;
826 	struct svc_rdma_send_ctxt *sctxt;
827 	int ret;
828 
829 	rdma_argp = rctxt->rc_recv_buf;
830 	svc_rdma_get_write_arrays(rdma_argp, &wr_lst, &rp_ch);
831 
832 	/* Create the RDMA response header. xprt->xpt_mutex,
833 	 * acquired in svc_send(), serializes RPC replies. The
834 	 * code path below that inserts the credit grant value
835 	 * into each transport header runs only inside this
836 	 * critical section.
837 	 */
838 	ret = -ENOMEM;
839 	sctxt = svc_rdma_send_ctxt_get(rdma);
840 	if (!sctxt)
841 		goto err0;
842 	rdma_resp = sctxt->sc_xprt_buf;
843 
844 	p = rdma_resp;
845 	*p++ = *rdma_argp;
846 	*p++ = *(rdma_argp + 1);
847 	*p++ = rdma->sc_fc_credits;
848 	*p++ = rp_ch ? rdma_nomsg : rdma_msg;
849 
850 	/* Start with empty chunks */
851 	*p++ = xdr_zero;
852 	*p++ = xdr_zero;
853 	*p   = xdr_zero;
854 
855 	if (wr_lst) {
856 		/* XXX: Presume the client sent only one Write chunk */
857 		ret = svc_rdma_send_write_chunk(rdma, wr_lst, xdr);
858 		if (ret < 0)
859 			goto err2;
860 		svc_rdma_xdr_encode_write_list(rdma_resp, wr_lst, ret);
861 	}
862 	if (rp_ch) {
863 		ret = svc_rdma_send_reply_chunk(rdma, rp_ch, wr_lst, xdr);
864 		if (ret < 0)
865 			goto err2;
866 		svc_rdma_xdr_encode_reply_chunk(rdma_resp, rp_ch, ret);
867 	}
868 
869 	svc_rdma_sync_reply_hdr(rdma, sctxt, svc_rdma_reply_hdr_len(rdma_resp));
870 	ret = svc_rdma_send_reply_msg(rdma, sctxt, rctxt, rqstp,
871 				      wr_lst, rp_ch);
872 	if (ret < 0)
873 		goto err1;
874 	ret = 0;
875 
876 out:
877 	rqstp->rq_xprt_ctxt = NULL;
878 	svc_rdma_recv_ctxt_put(rdma, rctxt);
879 	return ret;
880 
881  err2:
882 	if (ret != -E2BIG && ret != -EINVAL)
883 		goto err1;
884 
885 	ret = svc_rdma_send_error_msg(rdma, sctxt, rqstp);
886 	if (ret < 0)
887 		goto err1;
888 	ret = 0;
889 	goto out;
890 
891  err1:
892 	svc_rdma_send_ctxt_put(rdma, sctxt);
893  err0:
894 	trace_svcrdma_send_failed(rqstp, ret);
895 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
896 	ret = -ENOTCONN;
897 	goto out;
898 }
899