a1ac9997da98df63b1eaa547cf6b569f55d16381
[muen/linux.git] / net / sunrpc / xprtrdma / svc_rdma_rw.c
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[0];
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(sizeof(*ctxt) +
68                                SG_CHUNK_SIZE * sizeof(struct scatterlist),
69                                GFP_KERNEL);
70                 if (!ctxt)
71                         goto out;
72                 INIT_LIST_HEAD(&ctxt->rw_list);
73         }
74
75         ctxt->rw_sg_table.sgl = ctxt->rw_first_sgl;
76         if (sg_alloc_table_chained(&ctxt->rw_sg_table, sges,
77                                    ctxt->rw_sg_table.sgl)) {
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, true);
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                 if (wc->status != IB_WC_WR_FLUSH_ERR)
219                         pr_err("svcrdma: write ctx: %s (%u/0x%x)\n",
220                                ib_wc_status_msg(wc->status),
221                                wc->status, wc->vendor_err);
222         }
223
224         svc_rdma_write_info_free(info);
225 }
226
227 /* State for pulling a Read chunk.
228  */
229 struct svc_rdma_read_info {
230         struct svc_rdma_recv_ctxt       *ri_readctxt;
231         unsigned int                    ri_position;
232         unsigned int                    ri_pageno;
233         unsigned int                    ri_pageoff;
234         unsigned int                    ri_chunklen;
235
236         struct svc_rdma_chunk_ctxt      ri_cc;
237 };
238
239 static struct svc_rdma_read_info *
240 svc_rdma_read_info_alloc(struct svcxprt_rdma *rdma)
241 {
242         struct svc_rdma_read_info *info;
243
244         info = kmalloc(sizeof(*info), GFP_KERNEL);
245         if (!info)
246                 return info;
247
248         svc_rdma_cc_init(rdma, &info->ri_cc);
249         info->ri_cc.cc_cqe.done = svc_rdma_wc_read_done;
250         return info;
251 }
252
253 static void svc_rdma_read_info_free(struct svc_rdma_read_info *info)
254 {
255         svc_rdma_cc_release(&info->ri_cc, DMA_FROM_DEVICE);
256         kfree(info);
257 }
258
259 /**
260  * svc_rdma_wc_read_done - Handle completion of an RDMA Read ctx
261  * @cq: controlling Completion Queue
262  * @wc: Work Completion
263  *
264  */
265 static void svc_rdma_wc_read_done(struct ib_cq *cq, struct ib_wc *wc)
266 {
267         struct ib_cqe *cqe = wc->wr_cqe;
268         struct svc_rdma_chunk_ctxt *cc =
269                         container_of(cqe, struct svc_rdma_chunk_ctxt, cc_cqe);
270         struct svcxprt_rdma *rdma = cc->cc_rdma;
271         struct svc_rdma_read_info *info =
272                         container_of(cc, struct svc_rdma_read_info, ri_cc);
273
274         trace_svcrdma_wc_read(wc);
275
276         atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
277         wake_up(&rdma->sc_send_wait);
278
279         if (unlikely(wc->status != IB_WC_SUCCESS)) {
280                 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
281                 if (wc->status != IB_WC_WR_FLUSH_ERR)
282                         pr_err("svcrdma: read ctx: %s (%u/0x%x)\n",
283                                ib_wc_status_msg(wc->status),
284                                wc->status, wc->vendor_err);
285                 svc_rdma_recv_ctxt_put(rdma, info->ri_readctxt);
286         } else {
287                 spin_lock(&rdma->sc_rq_dto_lock);
288                 list_add_tail(&info->ri_readctxt->rc_list,
289                               &rdma->sc_read_complete_q);
290                 spin_unlock(&rdma->sc_rq_dto_lock);
291
292                 set_bit(XPT_DATA, &rdma->sc_xprt.xpt_flags);
293                 svc_xprt_enqueue(&rdma->sc_xprt);
294         }
295
296         svc_rdma_read_info_free(info);
297 }
298
299 /* This function sleeps when the transport's Send Queue is congested.
300  *
301  * Assumptions:
302  * - If ib_post_send() succeeds, only one completion is expected,
303  *   even if one or more WRs are flushed. This is true when posting
304  *   an rdma_rw_ctx or when posting a single signaled WR.
305  */
306 static int svc_rdma_post_chunk_ctxt(struct svc_rdma_chunk_ctxt *cc)
307 {
308         struct svcxprt_rdma *rdma = cc->cc_rdma;
309         struct svc_xprt *xprt = &rdma->sc_xprt;
310         struct ib_send_wr *first_wr, *bad_wr;
311         struct list_head *tmp;
312         struct ib_cqe *cqe;
313         int ret;
314
315         if (cc->cc_sqecount > rdma->sc_sq_depth)
316                 return -EINVAL;
317
318         first_wr = NULL;
319         cqe = &cc->cc_cqe;
320         list_for_each(tmp, &cc->cc_rwctxts) {
321                 struct svc_rdma_rw_ctxt *ctxt;
322
323                 ctxt = list_entry(tmp, struct svc_rdma_rw_ctxt, rw_list);
324                 first_wr = rdma_rw_ctx_wrs(&ctxt->rw_ctx, rdma->sc_qp,
325                                            rdma->sc_port_num, cqe, first_wr);
326                 cqe = NULL;
327         }
328
329         do {
330                 if (atomic_sub_return(cc->cc_sqecount,
331                                       &rdma->sc_sq_avail) > 0) {
332                         ret = ib_post_send(rdma->sc_qp, first_wr, &bad_wr);
333                         trace_svcrdma_post_rw(&cc->cc_cqe,
334                                               cc->cc_sqecount, ret);
335                         if (ret)
336                                 break;
337                         return 0;
338                 }
339
340                 trace_svcrdma_sq_full(rdma);
341                 atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
342                 wait_event(rdma->sc_send_wait,
343                            atomic_read(&rdma->sc_sq_avail) > cc->cc_sqecount);
344                 trace_svcrdma_sq_retry(rdma);
345         } while (1);
346
347         set_bit(XPT_CLOSE, &xprt->xpt_flags);
348
349         /* If even one was posted, there will be a completion. */
350         if (bad_wr != first_wr)
351                 return 0;
352
353         atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
354         wake_up(&rdma->sc_send_wait);
355         return -ENOTCONN;
356 }
357
358 /* Build and DMA-map an SGL that covers one kvec in an xdr_buf
359  */
360 static void svc_rdma_vec_to_sg(struct svc_rdma_write_info *info,
361                                unsigned int len,
362                                struct svc_rdma_rw_ctxt *ctxt)
363 {
364         struct scatterlist *sg = ctxt->rw_sg_table.sgl;
365
366         sg_set_buf(&sg[0], info->wi_base, len);
367         info->wi_base += len;
368
369         ctxt->rw_nents = 1;
370 }
371
372 /* Build and DMA-map an SGL that covers part of an xdr_buf's pagelist.
373  */
374 static void svc_rdma_pagelist_to_sg(struct svc_rdma_write_info *info,
375                                     unsigned int remaining,
376                                     struct svc_rdma_rw_ctxt *ctxt)
377 {
378         unsigned int sge_no, sge_bytes, page_off, page_no;
379         struct xdr_buf *xdr = info->wi_xdr;
380         struct scatterlist *sg;
381         struct page **page;
382
383         page_off = info->wi_next_off + xdr->page_base;
384         page_no = page_off >> PAGE_SHIFT;
385         page_off = offset_in_page(page_off);
386         page = xdr->pages + page_no;
387         info->wi_next_off += remaining;
388         sg = ctxt->rw_sg_table.sgl;
389         sge_no = 0;
390         do {
391                 sge_bytes = min_t(unsigned int, remaining,
392                                   PAGE_SIZE - page_off);
393                 sg_set_page(sg, *page, sge_bytes, page_off);
394
395                 remaining -= sge_bytes;
396                 sg = sg_next(sg);
397                 page_off = 0;
398                 sge_no++;
399                 page++;
400         } while (remaining);
401
402         ctxt->rw_nents = sge_no;
403 }
404
405 /* Construct RDMA Write WRs to send a portion of an xdr_buf containing
406  * an RPC Reply.
407  */
408 static int
409 svc_rdma_build_writes(struct svc_rdma_write_info *info,
410                       void (*constructor)(struct svc_rdma_write_info *info,
411                                           unsigned int len,
412                                           struct svc_rdma_rw_ctxt *ctxt),
413                       unsigned int remaining)
414 {
415         struct svc_rdma_chunk_ctxt *cc = &info->wi_cc;
416         struct svcxprt_rdma *rdma = cc->cc_rdma;
417         struct svc_rdma_rw_ctxt *ctxt;
418         __be32 *seg;
419         int ret;
420
421         seg = info->wi_segs + info->wi_seg_no * rpcrdma_segment_maxsz;
422         do {
423                 unsigned int write_len;
424                 u32 seg_length, seg_handle;
425                 u64 seg_offset;
426
427                 if (info->wi_seg_no >= info->wi_nsegs)
428                         goto out_overflow;
429
430                 seg_handle = be32_to_cpup(seg);
431                 seg_length = be32_to_cpup(seg + 1);
432                 xdr_decode_hyper(seg + 2, &seg_offset);
433                 seg_offset += info->wi_seg_off;
434
435                 write_len = min(remaining, seg_length - info->wi_seg_off);
436                 ctxt = svc_rdma_get_rw_ctxt(rdma,
437                                             (write_len >> PAGE_SHIFT) + 2);
438                 if (!ctxt)
439                         goto out_noctx;
440
441                 constructor(info, write_len, ctxt);
442                 ret = rdma_rw_ctx_init(&ctxt->rw_ctx, rdma->sc_qp,
443                                        rdma->sc_port_num, ctxt->rw_sg_table.sgl,
444                                        ctxt->rw_nents, 0, seg_offset,
445                                        seg_handle, DMA_TO_DEVICE);
446                 if (ret < 0)
447                         goto out_initerr;
448
449                 trace_svcrdma_encode_wseg(seg_handle, write_len, seg_offset);
450                 list_add(&ctxt->rw_list, &cc->cc_rwctxts);
451                 cc->cc_sqecount += ret;
452                 if (write_len == seg_length - info->wi_seg_off) {
453                         seg += 4;
454                         info->wi_seg_no++;
455                         info->wi_seg_off = 0;
456                 } else {
457                         info->wi_seg_off += write_len;
458                 }
459                 remaining -= write_len;
460         } while (remaining);
461
462         return 0;
463
464 out_overflow:
465         dprintk("svcrdma: inadequate space in Write chunk (%u)\n",
466                 info->wi_nsegs);
467         return -E2BIG;
468
469 out_noctx:
470         dprintk("svcrdma: no R/W ctxs available\n");
471         return -ENOMEM;
472
473 out_initerr:
474         svc_rdma_put_rw_ctxt(rdma, ctxt);
475         trace_svcrdma_dma_map_rwctx(rdma, ret);
476         return -EIO;
477 }
478
479 /* Send one of an xdr_buf's kvecs by itself. To send a Reply
480  * chunk, the whole RPC Reply is written back to the client.
481  * This function writes either the head or tail of the xdr_buf
482  * containing the Reply.
483  */
484 static int svc_rdma_send_xdr_kvec(struct svc_rdma_write_info *info,
485                                   struct kvec *vec)
486 {
487         info->wi_base = vec->iov_base;
488         return svc_rdma_build_writes(info, svc_rdma_vec_to_sg,
489                                      vec->iov_len);
490 }
491
492 /* Send an xdr_buf's page list by itself. A Write chunk is
493  * just the page list. a Reply chunk is the head, page list,
494  * and tail. This function is shared between the two types
495  * of chunk.
496  */
497 static int svc_rdma_send_xdr_pagelist(struct svc_rdma_write_info *info,
498                                       struct xdr_buf *xdr)
499 {
500         info->wi_xdr = xdr;
501         info->wi_next_off = 0;
502         return svc_rdma_build_writes(info, svc_rdma_pagelist_to_sg,
503                                      xdr->page_len);
504 }
505
506 /**
507  * svc_rdma_send_write_chunk - Write all segments in a Write chunk
508  * @rdma: controlling RDMA transport
509  * @wr_ch: Write chunk provided by client
510  * @xdr: xdr_buf containing the data payload
511  *
512  * Returns a non-negative number of bytes the chunk consumed, or
513  *      %-E2BIG if the payload was larger than the Write chunk,
514  *      %-EINVAL if client provided too many segments,
515  *      %-ENOMEM if rdma_rw context pool was exhausted,
516  *      %-ENOTCONN if posting failed (connection is lost),
517  *      %-EIO if rdma_rw initialization failed (DMA mapping, etc).
518  */
519 int svc_rdma_send_write_chunk(struct svcxprt_rdma *rdma, __be32 *wr_ch,
520                               struct xdr_buf *xdr)
521 {
522         struct svc_rdma_write_info *info;
523         int ret;
524
525         if (!xdr->page_len)
526                 return 0;
527
528         info = svc_rdma_write_info_alloc(rdma, wr_ch);
529         if (!info)
530                 return -ENOMEM;
531
532         ret = svc_rdma_send_xdr_pagelist(info, xdr);
533         if (ret < 0)
534                 goto out_err;
535
536         ret = svc_rdma_post_chunk_ctxt(&info->wi_cc);
537         if (ret < 0)
538                 goto out_err;
539
540         trace_svcrdma_encode_write(xdr->page_len);
541         return xdr->page_len;
542
543 out_err:
544         svc_rdma_write_info_free(info);
545         return ret;
546 }
547
548 /**
549  * svc_rdma_send_reply_chunk - Write all segments in the Reply chunk
550  * @rdma: controlling RDMA transport
551  * @rp_ch: Reply chunk provided by client
552  * @writelist: true if client provided a Write list
553  * @xdr: xdr_buf containing an RPC Reply
554  *
555  * Returns a non-negative number of bytes the chunk consumed, or
556  *      %-E2BIG if the payload was larger than the Reply chunk,
557  *      %-EINVAL if client provided too many segments,
558  *      %-ENOMEM if rdma_rw context pool was exhausted,
559  *      %-ENOTCONN if posting failed (connection is lost),
560  *      %-EIO if rdma_rw initialization failed (DMA mapping, etc).
561  */
562 int svc_rdma_send_reply_chunk(struct svcxprt_rdma *rdma, __be32 *rp_ch,
563                               bool writelist, struct xdr_buf *xdr)
564 {
565         struct svc_rdma_write_info *info;
566         int consumed, ret;
567
568         info = svc_rdma_write_info_alloc(rdma, rp_ch);
569         if (!info)
570                 return -ENOMEM;
571
572         ret = svc_rdma_send_xdr_kvec(info, &xdr->head[0]);
573         if (ret < 0)
574                 goto out_err;
575         consumed = xdr->head[0].iov_len;
576
577         /* Send the page list in the Reply chunk only if the
578          * client did not provide Write chunks.
579          */
580         if (!writelist && xdr->page_len) {
581                 ret = svc_rdma_send_xdr_pagelist(info, xdr);
582                 if (ret < 0)
583                         goto out_err;
584                 consumed += xdr->page_len;
585         }
586
587         if (xdr->tail[0].iov_len) {
588                 ret = svc_rdma_send_xdr_kvec(info, &xdr->tail[0]);
589                 if (ret < 0)
590                         goto out_err;
591                 consumed += xdr->tail[0].iov_len;
592         }
593
594         ret = svc_rdma_post_chunk_ctxt(&info->wi_cc);
595         if (ret < 0)
596                 goto out_err;
597
598         trace_svcrdma_encode_reply(consumed);
599         return consumed;
600
601 out_err:
602         svc_rdma_write_info_free(info);
603         return ret;
604 }
605
606 static int svc_rdma_build_read_segment(struct svc_rdma_read_info *info,
607                                        struct svc_rqst *rqstp,
608                                        u32 rkey, u32 len, u64 offset)
609 {
610         struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
611         struct svc_rdma_chunk_ctxt *cc = &info->ri_cc;
612         struct svc_rdma_rw_ctxt *ctxt;
613         unsigned int sge_no, seg_len;
614         struct scatterlist *sg;
615         int ret;
616
617         sge_no = PAGE_ALIGN(info->ri_pageoff + len) >> PAGE_SHIFT;
618         ctxt = svc_rdma_get_rw_ctxt(cc->cc_rdma, sge_no);
619         if (!ctxt)
620                 goto out_noctx;
621         ctxt->rw_nents = sge_no;
622
623         sg = ctxt->rw_sg_table.sgl;
624         for (sge_no = 0; sge_no < ctxt->rw_nents; sge_no++) {
625                 seg_len = min_t(unsigned int, len,
626                                 PAGE_SIZE - info->ri_pageoff);
627
628                 head->rc_arg.pages[info->ri_pageno] =
629                         rqstp->rq_pages[info->ri_pageno];
630                 if (!info->ri_pageoff)
631                         head->rc_page_count++;
632
633                 sg_set_page(sg, rqstp->rq_pages[info->ri_pageno],
634                             seg_len, info->ri_pageoff);
635                 sg = sg_next(sg);
636
637                 info->ri_pageoff += seg_len;
638                 if (info->ri_pageoff == PAGE_SIZE) {
639                         info->ri_pageno++;
640                         info->ri_pageoff = 0;
641                 }
642                 len -= seg_len;
643
644                 /* Safety check */
645                 if (len &&
646                     &rqstp->rq_pages[info->ri_pageno + 1] > rqstp->rq_page_end)
647                         goto out_overrun;
648         }
649
650         ret = rdma_rw_ctx_init(&ctxt->rw_ctx, cc->cc_rdma->sc_qp,
651                                cc->cc_rdma->sc_port_num,
652                                ctxt->rw_sg_table.sgl, ctxt->rw_nents,
653                                0, offset, rkey, DMA_FROM_DEVICE);
654         if (ret < 0)
655                 goto out_initerr;
656
657         list_add(&ctxt->rw_list, &cc->cc_rwctxts);
658         cc->cc_sqecount += ret;
659         return 0;
660
661 out_noctx:
662         dprintk("svcrdma: no R/W ctxs available\n");
663         return -ENOMEM;
664
665 out_overrun:
666         dprintk("svcrdma: request overruns rq_pages\n");
667         return -EINVAL;
668
669 out_initerr:
670         trace_svcrdma_dma_map_rwctx(cc->cc_rdma, ret);
671         svc_rdma_put_rw_ctxt(cc->cc_rdma, ctxt);
672         return -EIO;
673 }
674
675 /* Walk the segments in the Read chunk starting at @p and construct
676  * RDMA Read operations to pull the chunk to the server.
677  */
678 static int svc_rdma_build_read_chunk(struct svc_rqst *rqstp,
679                                      struct svc_rdma_read_info *info,
680                                      __be32 *p)
681 {
682         unsigned int i;
683         int ret;
684
685         ret = -EINVAL;
686         info->ri_chunklen = 0;
687         while (*p++ != xdr_zero && be32_to_cpup(p++) == info->ri_position) {
688                 u32 rs_handle, rs_length;
689                 u64 rs_offset;
690
691                 rs_handle = be32_to_cpup(p++);
692                 rs_length = be32_to_cpup(p++);
693                 p = xdr_decode_hyper(p, &rs_offset);
694
695                 ret = svc_rdma_build_read_segment(info, rqstp,
696                                                   rs_handle, rs_length,
697                                                   rs_offset);
698                 if (ret < 0)
699                         break;
700
701                 trace_svcrdma_encode_rseg(rs_handle, rs_length, rs_offset);
702                 info->ri_chunklen += rs_length;
703         }
704
705         /* Pages under I/O have been copied to head->rc_pages.
706          * Prevent their premature release by svc_xprt_release() .
707          */
708         for (i = 0; i < info->ri_readctxt->rc_page_count; i++)
709                 rqstp->rq_pages[i] = NULL;
710
711         return ret;
712 }
713
714 /* Construct RDMA Reads to pull over a normal Read chunk. The chunk
715  * data lands in the page list of head->rc_arg.pages.
716  *
717  * Currently NFSD does not look at the head->rc_arg.tail[0] iovec.
718  * Therefore, XDR round-up of the Read chunk and trailing
719  * inline content must both be added at the end of the pagelist.
720  */
721 static int svc_rdma_build_normal_read_chunk(struct svc_rqst *rqstp,
722                                             struct svc_rdma_read_info *info,
723                                             __be32 *p)
724 {
725         struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
726         int ret;
727
728         ret = svc_rdma_build_read_chunk(rqstp, info, p);
729         if (ret < 0)
730                 goto out;
731
732         trace_svcrdma_encode_read(info->ri_chunklen, info->ri_position);
733
734         head->rc_hdr_count = 0;
735
736         /* Split the Receive buffer between the head and tail
737          * buffers at Read chunk's position. XDR roundup of the
738          * chunk is not included in either the pagelist or in
739          * the tail.
740          */
741         head->rc_arg.tail[0].iov_base =
742                 head->rc_arg.head[0].iov_base + info->ri_position;
743         head->rc_arg.tail[0].iov_len =
744                 head->rc_arg.head[0].iov_len - info->ri_position;
745         head->rc_arg.head[0].iov_len = info->ri_position;
746
747         /* Read chunk may need XDR roundup (see RFC 8166, s. 3.4.5.2).
748          *
749          * If the client already rounded up the chunk length, the
750          * length does not change. Otherwise, the length of the page
751          * list is increased to include XDR round-up.
752          *
753          * Currently these chunks always start at page offset 0,
754          * thus the rounded-up length never crosses a page boundary.
755          */
756         info->ri_chunklen = XDR_QUADLEN(info->ri_chunklen) << 2;
757
758         head->rc_arg.page_len = info->ri_chunklen;
759         head->rc_arg.len += info->ri_chunklen;
760         head->rc_arg.buflen += info->ri_chunklen;
761
762 out:
763         return ret;
764 }
765
766 /* Construct RDMA Reads to pull over a Position Zero Read chunk.
767  * The start of the data lands in the first page just after
768  * the Transport header, and the rest lands in the page list of
769  * head->rc_arg.pages.
770  *
771  * Assumptions:
772  *      - A PZRC has an XDR-aligned length (no implicit round-up).
773  *      - There can be no trailing inline content (IOW, we assume
774  *        a PZRC is never sent in an RDMA_MSG message, though it's
775  *        allowed by spec).
776  */
777 static int svc_rdma_build_pz_read_chunk(struct svc_rqst *rqstp,
778                                         struct svc_rdma_read_info *info,
779                                         __be32 *p)
780 {
781         struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
782         int ret;
783
784         ret = svc_rdma_build_read_chunk(rqstp, info, p);
785         if (ret < 0)
786                 goto out;
787
788         trace_svcrdma_encode_pzr(info->ri_chunklen);
789
790         head->rc_arg.len += info->ri_chunklen;
791         head->rc_arg.buflen += info->ri_chunklen;
792
793         head->rc_hdr_count = 1;
794         head->rc_arg.head[0].iov_base = page_address(head->rc_pages[0]);
795         head->rc_arg.head[0].iov_len = min_t(size_t, PAGE_SIZE,
796                                              info->ri_chunklen);
797
798         head->rc_arg.page_len = info->ri_chunklen -
799                                 head->rc_arg.head[0].iov_len;
800
801 out:
802         return ret;
803 }
804
805 /**
806  * svc_rdma_recv_read_chunk - Pull a Read chunk from the client
807  * @rdma: controlling RDMA transport
808  * @rqstp: set of pages to use as Read sink buffers
809  * @head: pages under I/O collect here
810  * @p: pointer to start of Read chunk
811  *
812  * Returns:
813  *      %0 if all needed RDMA Reads were posted successfully,
814  *      %-EINVAL if client provided too many segments,
815  *      %-ENOMEM if rdma_rw context pool was exhausted,
816  *      %-ENOTCONN if posting failed (connection is lost),
817  *      %-EIO if rdma_rw initialization failed (DMA mapping, etc).
818  *
819  * Assumptions:
820  * - All Read segments in @p have the same Position value.
821  */
822 int svc_rdma_recv_read_chunk(struct svcxprt_rdma *rdma, struct svc_rqst *rqstp,
823                              struct svc_rdma_recv_ctxt *head, __be32 *p)
824 {
825         struct svc_rdma_read_info *info;
826         int ret;
827
828         /* The request (with page list) is constructed in
829          * head->rc_arg. Pages involved with RDMA Read I/O are
830          * transferred there.
831          */
832         head->rc_arg.head[0] = rqstp->rq_arg.head[0];
833         head->rc_arg.tail[0] = rqstp->rq_arg.tail[0];
834         head->rc_arg.pages = head->rc_pages;
835         head->rc_arg.page_base = 0;
836         head->rc_arg.page_len = 0;
837         head->rc_arg.len = rqstp->rq_arg.len;
838         head->rc_arg.buflen = rqstp->rq_arg.buflen;
839
840         info = svc_rdma_read_info_alloc(rdma);
841         if (!info)
842                 return -ENOMEM;
843         info->ri_readctxt = head;
844         info->ri_pageno = 0;
845         info->ri_pageoff = 0;
846
847         info->ri_position = be32_to_cpup(p + 1);
848         if (info->ri_position)
849                 ret = svc_rdma_build_normal_read_chunk(rqstp, info, p);
850         else
851                 ret = svc_rdma_build_pz_read_chunk(rqstp, info, p);
852         if (ret < 0)
853                 goto out_err;
854
855         ret = svc_rdma_post_chunk_ctxt(&info->ri_cc);
856         if (ret < 0)
857                 goto out_err;
858         return 0;
859
860 out_err:
861         svc_rdma_read_info_free(info);
862         return ret;
863 }