1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2016-2018 Oracle.  All rights reserved.
4  *
5  * Use the core R/W API to move RPC-over-RDMA Read and Write chunks.
6  */
7 
8 #include <rdma/rw.h>
9 
10 #include <linux/sunrpc/rpc_rdma.h>
11 #include <linux/sunrpc/svc_rdma.h>
12 #include <linux/sunrpc/debug.h>
13 
14 #include "xprt_rdma.h"
15 #include <trace/events/rpcrdma.h>
16 
17 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
18 
19 static void svc_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc);
20 static void svc_rdma_wc_read_done(struct ib_cq *cq, struct ib_wc *wc);
21 
22 /* Each R/W context contains state for one chain of RDMA Read or
23  * Write Work Requests.
24  *
25  * Each WR chain handles a single contiguous server-side buffer,
26  * because scatterlist entries after the first have to start on
27  * page alignment. xdr_buf iovecs cannot guarantee alignment.
28  *
29  * Each WR chain handles only one R_key. Each RPC-over-RDMA segment
30  * from a client may contain a unique R_key, so each WR chain moves
31  * up to one segment at a time.
32  *
33  * The scatterlist makes this data structure over 4KB in size. To
34  * make it less likely to fail, and to handle the allocation for
35  * smaller I/O requests without disabling bottom-halves, these
36  * contexts are created on demand, but cached and reused until the
37  * controlling svcxprt_rdma is destroyed.
38  */
39 struct svc_rdma_rw_ctxt {
40 	struct list_head	rw_list;
41 	struct rdma_rw_ctx	rw_ctx;
42 	int			rw_nents;
43 	struct sg_table		rw_sg_table;
44 	struct scatterlist	rw_first_sgl[];
45 };
46 
47 static inline struct svc_rdma_rw_ctxt *
48 svc_rdma_next_ctxt(struct list_head *list)
49 {
50 	return list_first_entry_or_null(list, struct svc_rdma_rw_ctxt,
51 					rw_list);
52 }
53 
54 static struct svc_rdma_rw_ctxt *
55 svc_rdma_get_rw_ctxt(struct svcxprt_rdma *rdma, unsigned int sges)
56 {
57 	struct svc_rdma_rw_ctxt *ctxt;
58 
59 	spin_lock(&rdma->sc_rw_ctxt_lock);
60 
61 	ctxt = svc_rdma_next_ctxt(&rdma->sc_rw_ctxts);
62 	if (ctxt) {
63 		list_del(&ctxt->rw_list);
64 		spin_unlock(&rdma->sc_rw_ctxt_lock);
65 	} else {
66 		spin_unlock(&rdma->sc_rw_ctxt_lock);
67 		ctxt = kmalloc(struct_size(ctxt, rw_first_sgl, SG_CHUNK_SIZE),
68 			       GFP_KERNEL);
69 		if (!ctxt)
70 			goto out;
71 		INIT_LIST_HEAD(&ctxt->rw_list);
72 	}
73 
74 	ctxt->rw_sg_table.sgl = ctxt->rw_first_sgl;
75 	if (sg_alloc_table_chained(&ctxt->rw_sg_table, sges,
76 				   ctxt->rw_sg_table.sgl,
77 				   SG_CHUNK_SIZE)) {
78 		kfree(ctxt);
79 		ctxt = NULL;
80 	}
81 out:
82 	return ctxt;
83 }
84 
85 static void svc_rdma_put_rw_ctxt(struct svcxprt_rdma *rdma,
86 				 struct svc_rdma_rw_ctxt *ctxt)
87 {
88 	sg_free_table_chained(&ctxt->rw_sg_table, SG_CHUNK_SIZE);
89 
90 	spin_lock(&rdma->sc_rw_ctxt_lock);
91 	list_add(&ctxt->rw_list, &rdma->sc_rw_ctxts);
92 	spin_unlock(&rdma->sc_rw_ctxt_lock);
93 }
94 
95 /**
96  * svc_rdma_destroy_rw_ctxts - Free accumulated R/W contexts
97  * @rdma: transport about to be destroyed
98  *
99  */
100 void svc_rdma_destroy_rw_ctxts(struct svcxprt_rdma *rdma)
101 {
102 	struct svc_rdma_rw_ctxt *ctxt;
103 
104 	while ((ctxt = svc_rdma_next_ctxt(&rdma->sc_rw_ctxts)) != NULL) {
105 		list_del(&ctxt->rw_list);
106 		kfree(ctxt);
107 	}
108 }
109 
110 /* A chunk context tracks all I/O for moving one Read or Write
111  * chunk. This is a a set of rdma_rw's that handle data movement
112  * for all segments of one chunk.
113  *
114  * These are small, acquired with a single allocator call, and
115  * no more than one is needed per chunk. They are allocated on
116  * demand, and not cached.
117  */
118 struct svc_rdma_chunk_ctxt {
119 	struct ib_cqe		cc_cqe;
120 	struct svcxprt_rdma	*cc_rdma;
121 	struct list_head	cc_rwctxts;
122 	int			cc_sqecount;
123 };
124 
125 static void svc_rdma_cc_init(struct svcxprt_rdma *rdma,
126 			     struct svc_rdma_chunk_ctxt *cc)
127 {
128 	cc->cc_rdma = rdma;
129 	svc_xprt_get(&rdma->sc_xprt);
130 
131 	INIT_LIST_HEAD(&cc->cc_rwctxts);
132 	cc->cc_sqecount = 0;
133 }
134 
135 static void svc_rdma_cc_release(struct svc_rdma_chunk_ctxt *cc,
136 				enum dma_data_direction dir)
137 {
138 	struct svcxprt_rdma *rdma = cc->cc_rdma;
139 	struct svc_rdma_rw_ctxt *ctxt;
140 
141 	while ((ctxt = svc_rdma_next_ctxt(&cc->cc_rwctxts)) != NULL) {
142 		list_del(&ctxt->rw_list);
143 
144 		rdma_rw_ctx_destroy(&ctxt->rw_ctx, rdma->sc_qp,
145 				    rdma->sc_port_num, ctxt->rw_sg_table.sgl,
146 				    ctxt->rw_nents, dir);
147 		svc_rdma_put_rw_ctxt(rdma, ctxt);
148 	}
149 	svc_xprt_put(&rdma->sc_xprt);
150 }
151 
152 /* State for sending a Write or Reply chunk.
153  *  - Tracks progress of writing one chunk over all its segments
154  *  - Stores arguments for the SGL constructor functions
155  */
156 struct svc_rdma_write_info {
157 	/* write state of this chunk */
158 	unsigned int		wi_seg_off;
159 	unsigned int		wi_seg_no;
160 	unsigned int		wi_nsegs;
161 	__be32			*wi_segs;
162 
163 	/* SGL constructor arguments */
164 	struct xdr_buf		*wi_xdr;
165 	unsigned char		*wi_base;
166 	unsigned int		wi_next_off;
167 
168 	struct svc_rdma_chunk_ctxt	wi_cc;
169 };
170 
171 static struct svc_rdma_write_info *
172 svc_rdma_write_info_alloc(struct svcxprt_rdma *rdma, __be32 *chunk)
173 {
174 	struct svc_rdma_write_info *info;
175 
176 	info = kmalloc(sizeof(*info), GFP_KERNEL);
177 	if (!info)
178 		return info;
179 
180 	info->wi_seg_off = 0;
181 	info->wi_seg_no = 0;
182 	info->wi_nsegs = be32_to_cpup(++chunk);
183 	info->wi_segs = ++chunk;
184 	svc_rdma_cc_init(rdma, &info->wi_cc);
185 	info->wi_cc.cc_cqe.done = svc_rdma_write_done;
186 	return info;
187 }
188 
189 static void svc_rdma_write_info_free(struct svc_rdma_write_info *info)
190 {
191 	svc_rdma_cc_release(&info->wi_cc, DMA_TO_DEVICE);
192 	kfree(info);
193 }
194 
195 /**
196  * svc_rdma_write_done - Write chunk completion
197  * @cq: controlling Completion Queue
198  * @wc: Work Completion
199  *
200  * Pages under I/O are freed by a subsequent Send completion.
201  */
202 static void svc_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc)
203 {
204 	struct ib_cqe *cqe = wc->wr_cqe;
205 	struct svc_rdma_chunk_ctxt *cc =
206 			container_of(cqe, struct svc_rdma_chunk_ctxt, cc_cqe);
207 	struct svcxprt_rdma *rdma = cc->cc_rdma;
208 	struct svc_rdma_write_info *info =
209 			container_of(cc, struct svc_rdma_write_info, wi_cc);
210 
211 	trace_svcrdma_wc_write(wc);
212 
213 	atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
214 	wake_up(&rdma->sc_send_wait);
215 
216 	if (unlikely(wc->status != IB_WC_SUCCESS))
217 		set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
218 
219 	svc_rdma_write_info_free(info);
220 }
221 
222 /* State for pulling a Read chunk.
223  */
224 struct svc_rdma_read_info {
225 	struct svc_rdma_recv_ctxt	*ri_readctxt;
226 	unsigned int			ri_position;
227 	unsigned int			ri_pageno;
228 	unsigned int			ri_pageoff;
229 	unsigned int			ri_chunklen;
230 
231 	struct svc_rdma_chunk_ctxt	ri_cc;
232 };
233 
234 static struct svc_rdma_read_info *
235 svc_rdma_read_info_alloc(struct svcxprt_rdma *rdma)
236 {
237 	struct svc_rdma_read_info *info;
238 
239 	info = kmalloc(sizeof(*info), GFP_KERNEL);
240 	if (!info)
241 		return info;
242 
243 	svc_rdma_cc_init(rdma, &info->ri_cc);
244 	info->ri_cc.cc_cqe.done = svc_rdma_wc_read_done;
245 	return info;
246 }
247 
248 static void svc_rdma_read_info_free(struct svc_rdma_read_info *info)
249 {
250 	svc_rdma_cc_release(&info->ri_cc, DMA_FROM_DEVICE);
251 	kfree(info);
252 }
253 
254 /**
255  * svc_rdma_wc_read_done - Handle completion of an RDMA Read ctx
256  * @cq: controlling Completion Queue
257  * @wc: Work Completion
258  *
259  */
260 static void svc_rdma_wc_read_done(struct ib_cq *cq, struct ib_wc *wc)
261 {
262 	struct ib_cqe *cqe = wc->wr_cqe;
263 	struct svc_rdma_chunk_ctxt *cc =
264 			container_of(cqe, struct svc_rdma_chunk_ctxt, cc_cqe);
265 	struct svcxprt_rdma *rdma = cc->cc_rdma;
266 	struct svc_rdma_read_info *info =
267 			container_of(cc, struct svc_rdma_read_info, ri_cc);
268 
269 	trace_svcrdma_wc_read(wc);
270 
271 	atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
272 	wake_up(&rdma->sc_send_wait);
273 
274 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
275 		set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
276 		svc_rdma_recv_ctxt_put(rdma, info->ri_readctxt);
277 	} else {
278 		spin_lock(&rdma->sc_rq_dto_lock);
279 		list_add_tail(&info->ri_readctxt->rc_list,
280 			      &rdma->sc_read_complete_q);
281 		/* Note the unlock pairs with the smp_rmb in svc_xprt_ready: */
282 		set_bit(XPT_DATA, &rdma->sc_xprt.xpt_flags);
283 		spin_unlock(&rdma->sc_rq_dto_lock);
284 
285 		svc_xprt_enqueue(&rdma->sc_xprt);
286 	}
287 
288 	svc_rdma_read_info_free(info);
289 }
290 
291 /* This function sleeps when the transport's Send Queue is congested.
292  *
293  * Assumptions:
294  * - If ib_post_send() succeeds, only one completion is expected,
295  *   even if one or more WRs are flushed. This is true when posting
296  *   an rdma_rw_ctx or when posting a single signaled WR.
297  */
298 static int svc_rdma_post_chunk_ctxt(struct svc_rdma_chunk_ctxt *cc)
299 {
300 	struct svcxprt_rdma *rdma = cc->cc_rdma;
301 	struct svc_xprt *xprt = &rdma->sc_xprt;
302 	struct ib_send_wr *first_wr;
303 	const struct ib_send_wr *bad_wr;
304 	struct list_head *tmp;
305 	struct ib_cqe *cqe;
306 	int ret;
307 
308 	if (cc->cc_sqecount > rdma->sc_sq_depth)
309 		return -EINVAL;
310 
311 	first_wr = NULL;
312 	cqe = &cc->cc_cqe;
313 	list_for_each(tmp, &cc->cc_rwctxts) {
314 		struct svc_rdma_rw_ctxt *ctxt;
315 
316 		ctxt = list_entry(tmp, struct svc_rdma_rw_ctxt, rw_list);
317 		first_wr = rdma_rw_ctx_wrs(&ctxt->rw_ctx, rdma->sc_qp,
318 					   rdma->sc_port_num, cqe, first_wr);
319 		cqe = NULL;
320 	}
321 
322 	do {
323 		if (atomic_sub_return(cc->cc_sqecount,
324 				      &rdma->sc_sq_avail) > 0) {
325 			ret = ib_post_send(rdma->sc_qp, first_wr, &bad_wr);
326 			if (ret)
327 				break;
328 			return 0;
329 		}
330 
331 		trace_svcrdma_sq_full(rdma);
332 		atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
333 		wait_event(rdma->sc_send_wait,
334 			   atomic_read(&rdma->sc_sq_avail) > cc->cc_sqecount);
335 		trace_svcrdma_sq_retry(rdma);
336 	} while (1);
337 
338 	trace_svcrdma_sq_post_err(rdma, ret);
339 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
340 
341 	/* If even one was posted, there will be a completion. */
342 	if (bad_wr != first_wr)
343 		return 0;
344 
345 	atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
346 	wake_up(&rdma->sc_send_wait);
347 	return -ENOTCONN;
348 }
349 
350 /* Build and DMA-map an SGL that covers one kvec in an xdr_buf
351  */
352 static void svc_rdma_vec_to_sg(struct svc_rdma_write_info *info,
353 			       unsigned int len,
354 			       struct svc_rdma_rw_ctxt *ctxt)
355 {
356 	struct scatterlist *sg = ctxt->rw_sg_table.sgl;
357 
358 	sg_set_buf(&sg[0], info->wi_base, len);
359 	info->wi_base += len;
360 
361 	ctxt->rw_nents = 1;
362 }
363 
364 /* Build and DMA-map an SGL that covers part of an xdr_buf's pagelist.
365  */
366 static void svc_rdma_pagelist_to_sg(struct svc_rdma_write_info *info,
367 				    unsigned int remaining,
368 				    struct svc_rdma_rw_ctxt *ctxt)
369 {
370 	unsigned int sge_no, sge_bytes, page_off, page_no;
371 	struct xdr_buf *xdr = info->wi_xdr;
372 	struct scatterlist *sg;
373 	struct page **page;
374 
375 	page_off = info->wi_next_off + xdr->page_base;
376 	page_no = page_off >> PAGE_SHIFT;
377 	page_off = offset_in_page(page_off);
378 	page = xdr->pages + page_no;
379 	info->wi_next_off += remaining;
380 	sg = ctxt->rw_sg_table.sgl;
381 	sge_no = 0;
382 	do {
383 		sge_bytes = min_t(unsigned int, remaining,
384 				  PAGE_SIZE - page_off);
385 		sg_set_page(sg, *page, sge_bytes, page_off);
386 
387 		remaining -= sge_bytes;
388 		sg = sg_next(sg);
389 		page_off = 0;
390 		sge_no++;
391 		page++;
392 	} while (remaining);
393 
394 	ctxt->rw_nents = sge_no;
395 }
396 
397 /* Construct RDMA Write WRs to send a portion of an xdr_buf containing
398  * an RPC Reply.
399  */
400 static int
401 svc_rdma_build_writes(struct svc_rdma_write_info *info,
402 		      void (*constructor)(struct svc_rdma_write_info *info,
403 					  unsigned int len,
404 					  struct svc_rdma_rw_ctxt *ctxt),
405 		      unsigned int remaining)
406 {
407 	struct svc_rdma_chunk_ctxt *cc = &info->wi_cc;
408 	struct svcxprt_rdma *rdma = cc->cc_rdma;
409 	struct svc_rdma_rw_ctxt *ctxt;
410 	__be32 *seg;
411 	int ret;
412 
413 	seg = info->wi_segs + info->wi_seg_no * rpcrdma_segment_maxsz;
414 	do {
415 		unsigned int write_len;
416 		u32 seg_length, seg_handle;
417 		u64 seg_offset;
418 
419 		if (info->wi_seg_no >= info->wi_nsegs)
420 			goto out_overflow;
421 
422 		seg_handle = be32_to_cpup(seg);
423 		seg_length = be32_to_cpup(seg + 1);
424 		xdr_decode_hyper(seg + 2, &seg_offset);
425 		seg_offset += info->wi_seg_off;
426 
427 		write_len = min(remaining, seg_length - info->wi_seg_off);
428 		ctxt = svc_rdma_get_rw_ctxt(rdma,
429 					    (write_len >> PAGE_SHIFT) + 2);
430 		if (!ctxt)
431 			goto out_noctx;
432 
433 		constructor(info, write_len, ctxt);
434 		ret = rdma_rw_ctx_init(&ctxt->rw_ctx, rdma->sc_qp,
435 				       rdma->sc_port_num, ctxt->rw_sg_table.sgl,
436 				       ctxt->rw_nents, 0, seg_offset,
437 				       seg_handle, DMA_TO_DEVICE);
438 		if (ret < 0)
439 			goto out_initerr;
440 
441 		trace_svcrdma_send_wseg(seg_handle, write_len, seg_offset);
442 
443 		list_add(&ctxt->rw_list, &cc->cc_rwctxts);
444 		cc->cc_sqecount += ret;
445 		if (write_len == seg_length - info->wi_seg_off) {
446 			seg += 4;
447 			info->wi_seg_no++;
448 			info->wi_seg_off = 0;
449 		} else {
450 			info->wi_seg_off += write_len;
451 		}
452 		remaining -= write_len;
453 	} while (remaining);
454 
455 	return 0;
456 
457 out_overflow:
458 	dprintk("svcrdma: inadequate space in Write chunk (%u)\n",
459 		info->wi_nsegs);
460 	return -E2BIG;
461 
462 out_noctx:
463 	dprintk("svcrdma: no R/W ctxs available\n");
464 	return -ENOMEM;
465 
466 out_initerr:
467 	svc_rdma_put_rw_ctxt(rdma, ctxt);
468 	trace_svcrdma_dma_map_rwctx(rdma, ret);
469 	return -EIO;
470 }
471 
472 /* Send one of an xdr_buf's kvecs by itself. To send a Reply
473  * chunk, the whole RPC Reply is written back to the client.
474  * This function writes either the head or tail of the xdr_buf
475  * containing the Reply.
476  */
477 static int svc_rdma_send_xdr_kvec(struct svc_rdma_write_info *info,
478 				  struct kvec *vec)
479 {
480 	info->wi_base = vec->iov_base;
481 	return svc_rdma_build_writes(info, svc_rdma_vec_to_sg,
482 				     vec->iov_len);
483 }
484 
485 /* Send an xdr_buf's page list by itself. A Write chunk is just
486  * the page list. A Reply chunk is @xdr's head, page list, and
487  * tail. This function is shared between the two types of chunk.
488  */
489 static int svc_rdma_send_xdr_pagelist(struct svc_rdma_write_info *info,
490 				      struct xdr_buf *xdr,
491 				      unsigned int offset,
492 				      unsigned long length)
493 {
494 	info->wi_xdr = xdr;
495 	info->wi_next_off = offset - xdr->head[0].iov_len;
496 	return svc_rdma_build_writes(info, svc_rdma_pagelist_to_sg,
497 				     length);
498 }
499 
500 /**
501  * svc_rdma_send_write_chunk - Write all segments in a Write chunk
502  * @rdma: controlling RDMA transport
503  * @wr_ch: Write chunk provided by client
504  * @xdr: xdr_buf containing the data payload
505  * @offset: payload's byte offset in @xdr
506  * @length: size of payload, in bytes
507  *
508  * Returns a non-negative number of bytes the chunk consumed, or
509  *	%-E2BIG if the payload was larger than the Write chunk,
510  *	%-EINVAL if client provided too many segments,
511  *	%-ENOMEM if rdma_rw context pool was exhausted,
512  *	%-ENOTCONN if posting failed (connection is lost),
513  *	%-EIO if rdma_rw initialization failed (DMA mapping, etc).
514  */
515 int svc_rdma_send_write_chunk(struct svcxprt_rdma *rdma, __be32 *wr_ch,
516 			      struct xdr_buf *xdr,
517 			      unsigned int offset, unsigned long length)
518 {
519 	struct svc_rdma_write_info *info;
520 	int ret;
521 
522 	if (!length)
523 		return 0;
524 
525 	info = svc_rdma_write_info_alloc(rdma, wr_ch);
526 	if (!info)
527 		return -ENOMEM;
528 
529 	ret = svc_rdma_send_xdr_pagelist(info, xdr, offset, length);
530 	if (ret < 0)
531 		goto out_err;
532 
533 	ret = svc_rdma_post_chunk_ctxt(&info->wi_cc);
534 	if (ret < 0)
535 		goto out_err;
536 
537 	trace_svcrdma_send_write_chunk(xdr->page_len);
538 	return length;
539 
540 out_err:
541 	svc_rdma_write_info_free(info);
542 	return ret;
543 }
544 
545 /**
546  * svc_rdma_send_reply_chunk - Write all segments in the Reply chunk
547  * @rdma: controlling RDMA transport
548  * @rctxt: Write and Reply chunks from client
549  * @xdr: xdr_buf containing an RPC Reply
550  *
551  * Returns a non-negative number of bytes the chunk consumed, or
552  *	%-E2BIG if the payload was larger than the Reply chunk,
553  *	%-EINVAL if client provided too many segments,
554  *	%-ENOMEM if rdma_rw context pool was exhausted,
555  *	%-ENOTCONN if posting failed (connection is lost),
556  *	%-EIO if rdma_rw initialization failed (DMA mapping, etc).
557  */
558 int svc_rdma_send_reply_chunk(struct svcxprt_rdma *rdma,
559 			      const struct svc_rdma_recv_ctxt *rctxt,
560 			      struct xdr_buf *xdr)
561 {
562 	struct svc_rdma_write_info *info;
563 	int consumed, ret;
564 
565 	info = svc_rdma_write_info_alloc(rdma, rctxt->rc_reply_chunk);
566 	if (!info)
567 		return -ENOMEM;
568 
569 	ret = svc_rdma_send_xdr_kvec(info, &xdr->head[0]);
570 	if (ret < 0)
571 		goto out_err;
572 	consumed = xdr->head[0].iov_len;
573 
574 	/* Send the page list in the Reply chunk only if the
575 	 * client did not provide Write chunks.
576 	 */
577 	if (!rctxt->rc_write_list && xdr->page_len) {
578 		ret = svc_rdma_send_xdr_pagelist(info, xdr,
579 						 xdr->head[0].iov_len,
580 						 xdr->page_len);
581 		if (ret < 0)
582 			goto out_err;
583 		consumed += xdr->page_len;
584 	}
585 
586 	if (xdr->tail[0].iov_len) {
587 		ret = svc_rdma_send_xdr_kvec(info, &xdr->tail[0]);
588 		if (ret < 0)
589 			goto out_err;
590 		consumed += xdr->tail[0].iov_len;
591 	}
592 
593 	ret = svc_rdma_post_chunk_ctxt(&info->wi_cc);
594 	if (ret < 0)
595 		goto out_err;
596 
597 	trace_svcrdma_send_reply_chunk(consumed);
598 	return consumed;
599 
600 out_err:
601 	svc_rdma_write_info_free(info);
602 	return ret;
603 }
604 
605 static int svc_rdma_build_read_segment(struct svc_rdma_read_info *info,
606 				       struct svc_rqst *rqstp,
607 				       u32 rkey, u32 len, u64 offset)
608 {
609 	struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
610 	struct svc_rdma_chunk_ctxt *cc = &info->ri_cc;
611 	struct svc_rdma_rw_ctxt *ctxt;
612 	unsigned int sge_no, seg_len;
613 	struct scatterlist *sg;
614 	int ret;
615 
616 	sge_no = PAGE_ALIGN(info->ri_pageoff + len) >> PAGE_SHIFT;
617 	ctxt = svc_rdma_get_rw_ctxt(cc->cc_rdma, sge_no);
618 	if (!ctxt)
619 		goto out_noctx;
620 	ctxt->rw_nents = sge_no;
621 
622 	sg = ctxt->rw_sg_table.sgl;
623 	for (sge_no = 0; sge_no < ctxt->rw_nents; sge_no++) {
624 		seg_len = min_t(unsigned int, len,
625 				PAGE_SIZE - info->ri_pageoff);
626 
627 		head->rc_arg.pages[info->ri_pageno] =
628 			rqstp->rq_pages[info->ri_pageno];
629 		if (!info->ri_pageoff)
630 			head->rc_page_count++;
631 
632 		sg_set_page(sg, rqstp->rq_pages[info->ri_pageno],
633 			    seg_len, info->ri_pageoff);
634 		sg = sg_next(sg);
635 
636 		info->ri_pageoff += seg_len;
637 		if (info->ri_pageoff == PAGE_SIZE) {
638 			info->ri_pageno++;
639 			info->ri_pageoff = 0;
640 		}
641 		len -= seg_len;
642 
643 		/* Safety check */
644 		if (len &&
645 		    &rqstp->rq_pages[info->ri_pageno + 1] > rqstp->rq_page_end)
646 			goto out_overrun;
647 	}
648 
649 	ret = rdma_rw_ctx_init(&ctxt->rw_ctx, cc->cc_rdma->sc_qp,
650 			       cc->cc_rdma->sc_port_num,
651 			       ctxt->rw_sg_table.sgl, ctxt->rw_nents,
652 			       0, offset, rkey, DMA_FROM_DEVICE);
653 	if (ret < 0)
654 		goto out_initerr;
655 
656 	list_add(&ctxt->rw_list, &cc->cc_rwctxts);
657 	cc->cc_sqecount += ret;
658 	return 0;
659 
660 out_noctx:
661 	dprintk("svcrdma: no R/W ctxs available\n");
662 	return -ENOMEM;
663 
664 out_overrun:
665 	dprintk("svcrdma: request overruns rq_pages\n");
666 	return -EINVAL;
667 
668 out_initerr:
669 	trace_svcrdma_dma_map_rwctx(cc->cc_rdma, ret);
670 	svc_rdma_put_rw_ctxt(cc->cc_rdma, ctxt);
671 	return -EIO;
672 }
673 
674 /* Walk the segments in the Read chunk starting at @p and construct
675  * RDMA Read operations to pull the chunk to the server.
676  */
677 static int svc_rdma_build_read_chunk(struct svc_rqst *rqstp,
678 				     struct svc_rdma_read_info *info,
679 				     __be32 *p)
680 {
681 	unsigned int i;
682 	int ret;
683 
684 	ret = -EINVAL;
685 	info->ri_chunklen = 0;
686 	while (*p++ != xdr_zero && be32_to_cpup(p++) == info->ri_position) {
687 		u32 rs_handle, rs_length;
688 		u64 rs_offset;
689 
690 		rs_handle = be32_to_cpup(p++);
691 		rs_length = be32_to_cpup(p++);
692 		p = xdr_decode_hyper(p, &rs_offset);
693 
694 		ret = svc_rdma_build_read_segment(info, rqstp,
695 						  rs_handle, rs_length,
696 						  rs_offset);
697 		if (ret < 0)
698 			break;
699 
700 		trace_svcrdma_send_rseg(rs_handle, rs_length, rs_offset);
701 		info->ri_chunklen += rs_length;
702 	}
703 
704 	/* Pages under I/O have been copied to head->rc_pages.
705 	 * Prevent their premature release by svc_xprt_release() .
706 	 */
707 	for (i = 0; i < info->ri_readctxt->rc_page_count; i++)
708 		rqstp->rq_pages[i] = NULL;
709 
710 	return ret;
711 }
712 
713 /* Construct RDMA Reads to pull over a normal Read chunk. The chunk
714  * data lands in the page list of head->rc_arg.pages.
715  *
716  * Currently NFSD does not look at the head->rc_arg.tail[0] iovec.
717  * Therefore, XDR round-up of the Read chunk and trailing
718  * inline content must both be added at the end of the pagelist.
719  */
720 static int svc_rdma_build_normal_read_chunk(struct svc_rqst *rqstp,
721 					    struct svc_rdma_read_info *info,
722 					    __be32 *p)
723 {
724 	struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
725 	int ret;
726 
727 	ret = svc_rdma_build_read_chunk(rqstp, info, p);
728 	if (ret < 0)
729 		goto out;
730 
731 	trace_svcrdma_send_read_chunk(info->ri_chunklen, info->ri_position);
732 
733 	head->rc_hdr_count = 0;
734 
735 	/* Split the Receive buffer between the head and tail
736 	 * buffers at Read chunk's position. XDR roundup of the
737 	 * chunk is not included in either the pagelist or in
738 	 * the tail.
739 	 */
740 	head->rc_arg.tail[0].iov_base =
741 		head->rc_arg.head[0].iov_base + info->ri_position;
742 	head->rc_arg.tail[0].iov_len =
743 		head->rc_arg.head[0].iov_len - info->ri_position;
744 	head->rc_arg.head[0].iov_len = info->ri_position;
745 
746 	/* Read chunk may need XDR roundup (see RFC 8166, s. 3.4.5.2).
747 	 *
748 	 * If the client already rounded up the chunk length, the
749 	 * length does not change. Otherwise, the length of the page
750 	 * list is increased to include XDR round-up.
751 	 *
752 	 * Currently these chunks always start at page offset 0,
753 	 * thus the rounded-up length never crosses a page boundary.
754 	 */
755 	info->ri_chunklen = XDR_QUADLEN(info->ri_chunklen) << 2;
756 
757 	head->rc_arg.page_len = info->ri_chunklen;
758 	head->rc_arg.len += info->ri_chunklen;
759 	head->rc_arg.buflen += info->ri_chunklen;
760 
761 out:
762 	return ret;
763 }
764 
765 /* Construct RDMA Reads to pull over a Position Zero Read chunk.
766  * The start of the data lands in the first page just after
767  * the Transport header, and the rest lands in the page list of
768  * head->rc_arg.pages.
769  *
770  * Assumptions:
771  *	- A PZRC has an XDR-aligned length (no implicit round-up).
772  *	- There can be no trailing inline content (IOW, we assume
773  *	  a PZRC is never sent in an RDMA_MSG message, though it's
774  *	  allowed by spec).
775  */
776 static int svc_rdma_build_pz_read_chunk(struct svc_rqst *rqstp,
777 					struct svc_rdma_read_info *info,
778 					__be32 *p)
779 {
780 	struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
781 	int ret;
782 
783 	ret = svc_rdma_build_read_chunk(rqstp, info, p);
784 	if (ret < 0)
785 		goto out;
786 
787 	trace_svcrdma_send_pzr(info->ri_chunklen);
788 
789 	head->rc_arg.len += info->ri_chunklen;
790 	head->rc_arg.buflen += info->ri_chunklen;
791 
792 	head->rc_hdr_count = 1;
793 	head->rc_arg.head[0].iov_base = page_address(head->rc_pages[0]);
794 	head->rc_arg.head[0].iov_len = min_t(size_t, PAGE_SIZE,
795 					     info->ri_chunklen);
796 
797 	head->rc_arg.page_len = info->ri_chunklen -
798 				head->rc_arg.head[0].iov_len;
799 
800 out:
801 	return ret;
802 }
803 
804 /**
805  * svc_rdma_recv_read_chunk - Pull a Read chunk from the client
806  * @rdma: controlling RDMA transport
807  * @rqstp: set of pages to use as Read sink buffers
808  * @head: pages under I/O collect here
809  * @p: pointer to start of Read chunk
810  *
811  * Returns:
812  *	%0 if all needed RDMA Reads were posted successfully,
813  *	%-EINVAL if client provided too many segments,
814  *	%-ENOMEM if rdma_rw context pool was exhausted,
815  *	%-ENOTCONN if posting failed (connection is lost),
816  *	%-EIO if rdma_rw initialization failed (DMA mapping, etc).
817  *
818  * Assumptions:
819  * - All Read segments in @p have the same Position value.
820  */
821 int svc_rdma_recv_read_chunk(struct svcxprt_rdma *rdma, struct svc_rqst *rqstp,
822 			     struct svc_rdma_recv_ctxt *head, __be32 *p)
823 {
824 	struct svc_rdma_read_info *info;
825 	int ret;
826 
827 	/* The request (with page list) is constructed in
828 	 * head->rc_arg. Pages involved with RDMA Read I/O are
829 	 * transferred there.
830 	 */
831 	head->rc_arg.head[0] = rqstp->rq_arg.head[0];
832 	head->rc_arg.tail[0] = rqstp->rq_arg.tail[0];
833 	head->rc_arg.pages = head->rc_pages;
834 	head->rc_arg.page_base = 0;
835 	head->rc_arg.page_len = 0;
836 	head->rc_arg.len = rqstp->rq_arg.len;
837 	head->rc_arg.buflen = rqstp->rq_arg.buflen;
838 
839 	info = svc_rdma_read_info_alloc(rdma);
840 	if (!info)
841 		return -ENOMEM;
842 	info->ri_readctxt = head;
843 	info->ri_pageno = 0;
844 	info->ri_pageoff = 0;
845 
846 	info->ri_position = be32_to_cpup(p + 1);
847 	if (info->ri_position)
848 		ret = svc_rdma_build_normal_read_chunk(rqstp, info, p);
849 	else
850 		ret = svc_rdma_build_pz_read_chunk(rqstp, info, p);
851 	if (ret < 0)
852 		goto out_err;
853 
854 	ret = svc_rdma_post_chunk_ctxt(&info->ri_cc);
855 	if (ret < 0)
856 		goto out_err;
857 	return 0;
858 
859 out_err:
860 	svc_rdma_read_info_free(info);
861 	return ret;
862 }
863