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;
 311         const struct ib_send_wr *bad_wr;
 312         struct list_head *tmp;
 313         struct ib_cqe *cqe;
 314         int ret;
 315
 316         if (cc->cc_sqecount > rdma->sc_sq_depth)
 317                 return -EINVAL;
 318
 319         first_wr = NULL;
 320         cqe = &cc->cc_cqe;
 321         list_for_each(tmp, &cc->cc_rwctxts) {
 322                 struct svc_rdma_rw_ctxt *ctxt;
 323
 324                 ctxt = list_entry(tmp, struct svc_rdma_rw_ctxt, rw_list);
 325                 first_wr = rdma_rw_ctx_wrs(&ctxt->rw_ctx, rdma->sc_qp,
 326                                            rdma->sc_port_num, cqe, first_wr);
 327                 cqe = NULL;
 328         }
 329
 330         do {
 331                 if (atomic_sub_return(cc->cc_sqecount,
 332                                       &rdma->sc_sq_avail) > 0) {
 333                         ret = ib_post_send(rdma->sc_qp, first_wr, &bad_wr);
 334                         trace_svcrdma_post_rw(&cc->cc_cqe,
 335                                               cc->cc_sqecount, ret);
 336                         if (ret)
 337                                 break;
 338                         return 0;
 339                 }
 340
 341                 trace_svcrdma_sq_full(rdma);
 342                 atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
 343                 wait_event(rdma->sc_send_wait,
 344                            atomic_read(&rdma->sc_sq_avail) > cc->cc_sqecount);
 345                 trace_svcrdma_sq_retry(rdma);
 346         } while (1);
 347
 348         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 349
 350         /* If even one was posted, there will be a completion. */
 351         if (bad_wr != first_wr)
 352                 return 0;
 353
 354         atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
 355         wake_up(&rdma->sc_send_wait);
 356         return -ENOTCONN;
 357 }
 358
 359 /* Build and DMA-map an SGL that covers one kvec in an xdr_buf
 360  */
 361 static void svc_rdma_vec_to_sg(struct svc_rdma_write_info *info,
 362                                unsigned int len,
 363                                struct svc_rdma_rw_ctxt *ctxt)
 364 {
 365         struct scatterlist *sg = ctxt->rw_sg_table.sgl;
 366
 367         sg_set_buf(&sg[0], info->wi_base, len);
 368         info->wi_base += len;
 369
 370         ctxt->rw_nents = 1;
 371 }
 372
 373 /* Build and DMA-map an SGL that covers part of an xdr_buf's pagelist.
 374  */
 375 static void svc_rdma_pagelist_to_sg(struct svc_rdma_write_info *info,
 376                                     unsigned int remaining,
 377                                     struct svc_rdma_rw_ctxt *ctxt)
 378 {
 379         unsigned int sge_no, sge_bytes, page_off, page_no;
 380         struct xdr_buf *xdr = info->wi_xdr;
 381         struct scatterlist *sg;
 382         struct page **page;
 383
 384         page_off = info->wi_next_off + xdr->page_base;
 385         page_no = page_off >> PAGE_SHIFT;
 386         page_off = offset_in_page(page_off);
 387         page = xdr->pages + page_no;
 388         info->wi_next_off += remaining;
 389         sg = ctxt->rw_sg_table.sgl;
 390         sge_no = 0;
 391         do {
 392                 sge_bytes = min_t(unsigned int, remaining,
 393                                   PAGE_SIZE - page_off);
 394                 sg_set_page(sg, *page, sge_bytes, page_off);
 395
 396                 remaining -= sge_bytes;
 397                 sg = sg_next(sg);
 398                 page_off = 0;
 399                 sge_no++;
 400                 page++;
 401         } while (remaining);
 402
 403         ctxt->rw_nents = sge_no;
 404 }
 405
 406 /* Construct RDMA Write WRs to send a portion of an xdr_buf containing
 407  * an RPC Reply.
 408  */
 409 static int
 410 svc_rdma_build_writes(struct svc_rdma_write_info *info,
 411                       void (*constructor)(struct svc_rdma_write_info *info,
 412                                           unsigned int len,
 413                                           struct svc_rdma_rw_ctxt *ctxt),
 414                       unsigned int remaining)
 415 {
 416         struct svc_rdma_chunk_ctxt *cc = &info->wi_cc;
 417         struct svcxprt_rdma *rdma = cc->cc_rdma;
 418         struct svc_rdma_rw_ctxt *ctxt;
 419         __be32 *seg;
 420         int ret;
 421
 422         seg = info->wi_segs + info->wi_seg_no * rpcrdma_segment_maxsz;
 423         do {
 424                 unsigned int write_len;
 425                 u32 seg_length, seg_handle;
 426                 u64 seg_offset;
 427
 428                 if (info->wi_seg_no >= info->wi_nsegs)
 429                         goto out_overflow;
 430
 431                 seg_handle = be32_to_cpup(seg);
 432                 seg_length = be32_to_cpup(seg + 1);
 433                 xdr_decode_hyper(seg + 2, &seg_offset);
 434                 seg_offset += info->wi_seg_off;
 435
 436                 write_len = min(remaining, seg_length - info->wi_seg_off);
 437                 ctxt = svc_rdma_get_rw_ctxt(rdma,
 438                                             (write_len >> PAGE_SHIFT) + 2);
 439                 if (!ctxt)
 440                         goto out_noctx;
 441
 442                 constructor(info, write_len, ctxt);
 443                 ret = rdma_rw_ctx_init(&ctxt->rw_ctx, rdma->sc_qp,
 444                                        rdma->sc_port_num, ctxt->rw_sg_table.sgl,
 445                                        ctxt->rw_nents, 0, seg_offset,
 446                                        seg_handle, DMA_TO_DEVICE);
 447                 if (ret < 0)
 448                         goto out_initerr;
 449
 450                 trace_svcrdma_encode_wseg(seg_handle, write_len, seg_offset);
 451                 list_add(&ctxt->rw_list, &cc->cc_rwctxts);
 452                 cc->cc_sqecount += ret;
 453                 if (write_len == seg_length - info->wi_seg_off) {
 454                         seg += 4;
 455                         info->wi_seg_no++;
 456                         info->wi_seg_off = 0;
 457                 } else {
 458                         info->wi_seg_off += write_len;
 459                 }
 460                 remaining -= write_len;
 461         } while (remaining);
 462
 463         return 0;
 464
 465 out_overflow:
 466         dprintk("svcrdma: inadequate space in Write chunk (%u)\n",
 467                 info->wi_nsegs);
 468         return -E2BIG;
 469
 470 out_noctx:
 471         dprintk("svcrdma: no R/W ctxs available\n");
 472         return -ENOMEM;
 473
 474 out_initerr:
 475         svc_rdma_put_rw_ctxt(rdma, ctxt);
 476         trace_svcrdma_dma_map_rwctx(rdma, ret);
 477         return -EIO;
 478 }
 479
 480 /* Send one of an xdr_buf's kvecs by itself. To send a Reply
 481  * chunk, the whole RPC Reply is written back to the client.
 482  * This function writes either the head or tail of the xdr_buf
 483  * containing the Reply.
 484  */
 485 static int svc_rdma_send_xdr_kvec(struct svc_rdma_write_info *info,
 486                                   struct kvec *vec)
 487 {
 488         info->wi_base = vec->iov_base;
 489         return svc_rdma_build_writes(info, svc_rdma_vec_to_sg,
 490                                      vec->iov_len);
 491 }
 492
 493 /* Send an xdr_buf's page list by itself. A Write chunk is
 494  * just the page list. a Reply chunk is the head, page list,
 495  * and tail. This function is shared between the two types
 496  * of chunk.
 497  */
 498 static int svc_rdma_send_xdr_pagelist(struct svc_rdma_write_info *info,
 499                                       struct xdr_buf *xdr)
 500 {
 501         info->wi_xdr = xdr;
 502         info->wi_next_off = 0;
 503         return svc_rdma_build_writes(info, svc_rdma_pagelist_to_sg,
 504                                      xdr->page_len);
 505 }
 506
 507 /**
 508  * svc_rdma_send_write_chunk - Write all segments in a Write chunk
 509  * @rdma: controlling RDMA transport
 510  * @wr_ch: Write chunk provided by client
 511  * @xdr: xdr_buf containing the data payload
 512  *
 513  * Returns a non-negative number of bytes the chunk consumed, or
 514  *      %-E2BIG if the payload was larger than the Write chunk,
 515  *      %-EINVAL if client provided too many segments,
 516  *      %-ENOMEM if rdma_rw context pool was exhausted,
 517  *      %-ENOTCONN if posting failed (connection is lost),
 518  *      %-EIO if rdma_rw initialization failed (DMA mapping, etc).
 519  */
 520 int svc_rdma_send_write_chunk(struct svcxprt_rdma *rdma, __be32 *wr_ch,
 521                               struct xdr_buf *xdr)
 522 {
 523         struct svc_rdma_write_info *info;
 524         int ret;
 525
 526         if (!xdr->page_len)
 527                 return 0;
 528
 529         info = svc_rdma_write_info_alloc(rdma, wr_ch);
 530         if (!info)
 531                 return -ENOMEM;
 532
 533         ret = svc_rdma_send_xdr_pagelist(info, xdr);
 534         if (ret < 0)
 535                 goto out_err;
 536
 537         ret = svc_rdma_post_chunk_ctxt(&info->wi_cc);
 538         if (ret < 0)
 539                 goto out_err;
 540
 541         trace_svcrdma_encode_write(xdr->page_len);
 542         return xdr->page_len;
 543
 544 out_err:
 545         svc_rdma_write_info_free(info);
 546         return ret;
 547 }
 548
 549 /**
 550  * svc_rdma_send_reply_chunk - Write all segments in the Reply chunk
 551  * @rdma: controlling RDMA transport
 552  * @rp_ch: Reply chunk provided by client
 553  * @writelist: true if client provided a Write list
 554  * @xdr: xdr_buf containing an RPC Reply
 555  *
 556  * Returns a non-negative number of bytes the chunk consumed, or
 557  *      %-E2BIG if the payload was larger than the Reply chunk,
 558  *      %-EINVAL if client provided too many segments,
 559  *      %-ENOMEM if rdma_rw context pool was exhausted,
 560  *      %-ENOTCONN if posting failed (connection is lost),
 561  *      %-EIO if rdma_rw initialization failed (DMA mapping, etc).
 562  */
 563 int svc_rdma_send_reply_chunk(struct svcxprt_rdma *rdma, __be32 *rp_ch,
 564                               bool writelist, struct xdr_buf *xdr)
 565 {
 566         struct svc_rdma_write_info *info;
 567         int consumed, ret;
 568
 569         info = svc_rdma_write_info_alloc(rdma, rp_ch);
 570         if (!info)
 571                 return -ENOMEM;
 572
 573         ret = svc_rdma_send_xdr_kvec(info, &xdr->head[0]);
 574         if (ret < 0)
 575                 goto out_err;
 576         consumed = xdr->head[0].iov_len;
 577
 578         /* Send the page list in the Reply chunk only if the
 579          * client did not provide Write chunks.
 580          */
 581         if (!writelist && xdr->page_len) {
 582                 ret = svc_rdma_send_xdr_pagelist(info, xdr);
 583                 if (ret < 0)
 584                         goto out_err;
 585                 consumed += xdr->page_len;
 586         }
 587
 588         if (xdr->tail[0].iov_len) {
 589                 ret = svc_rdma_send_xdr_kvec(info, &xdr->tail[0]);
 590                 if (ret < 0)
 591                         goto out_err;
 592                 consumed += xdr->tail[0].iov_len;
 593         }
 594
 595         ret = svc_rdma_post_chunk_ctxt(&info->wi_cc);
 596         if (ret < 0)
 597                 goto out_err;
 598
 599         trace_svcrdma_encode_reply(consumed);
 600         return consumed;
 601
 602 out_err:
 603         svc_rdma_write_info_free(info);
 604         return ret;
 605 }
 606
 607 static int svc_rdma_build_read_segment(struct svc_rdma_read_info *info,
 608                                        struct svc_rqst *rqstp,
 609                                        u32 rkey, u32 len, u64 offset)
 610 {
 611         struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
 612         struct svc_rdma_chunk_ctxt *cc = &info->ri_cc;
 613         struct svc_rdma_rw_ctxt *ctxt;
 614         unsigned int sge_no, seg_len;
 615         struct scatterlist *sg;
 616         int ret;
 617
 618         sge_no = PAGE_ALIGN(info->ri_pageoff + len) >> PAGE_SHIFT;
 619         ctxt = svc_rdma_get_rw_ctxt(cc->cc_rdma, sge_no);
 620         if (!ctxt)
 621                 goto out_noctx;
 622         ctxt->rw_nents = sge_no;
 623
 624         sg = ctxt->rw_sg_table.sgl;
 625         for (sge_no = 0; sge_no < ctxt->rw_nents; sge_no++) {
 626                 seg_len = min_t(unsigned int, len,
 627                                 PAGE_SIZE - info->ri_pageoff);
 628
 629                 head->rc_arg.pages[info->ri_pageno] =
 630                         rqstp->rq_pages[info->ri_pageno];
 631                 if (!info->ri_pageoff)
 632                         head->rc_page_count++;
 633
 634                 sg_set_page(sg, rqstp->rq_pages[info->ri_pageno],
 635                             seg_len, info->ri_pageoff);
 636                 sg = sg_next(sg);
 637
 638                 info->ri_pageoff += seg_len;
 639                 if (info->ri_pageoff == PAGE_SIZE) {
 640                         info->ri_pageno++;
 641                         info->ri_pageoff = 0;
 642                 }
 643                 len -= seg_len;
 644
 645                 /* Safety check */
 646                 if (len &&
 647                     &rqstp->rq_pages[info->ri_pageno + 1] > rqstp->rq_page_end)
 648                         goto out_overrun;
 649         }
 650
 651         ret = rdma_rw_ctx_init(&ctxt->rw_ctx, cc->cc_rdma->sc_qp,
 652                                cc->cc_rdma->sc_port_num,
 653                                ctxt->rw_sg_table.sgl, ctxt->rw_nents,
 654                                0, offset, rkey, DMA_FROM_DEVICE);
 655         if (ret < 0)
 656                 goto out_initerr;
 657
 658         list_add(&ctxt->rw_list, &cc->cc_rwctxts);
 659         cc->cc_sqecount += ret;
 660         return 0;
 661
 662 out_noctx:
 663         dprintk("svcrdma: no R/W ctxs available\n");
 664         return -ENOMEM;
 665
 666 out_overrun:
 667         dprintk("svcrdma: request overruns rq_pages\n");
 668         return -EINVAL;
 669
 670 out_initerr:
 671         trace_svcrdma_dma_map_rwctx(cc->cc_rdma, ret);
 672         svc_rdma_put_rw_ctxt(cc->cc_rdma, ctxt);
 673         return -EIO;
 674 }
 675
 676 /* Walk the segments in the Read chunk starting at @p and construct
 677  * RDMA Read operations to pull the chunk to the server.
 678  */
 679 static int svc_rdma_build_read_chunk(struct svc_rqst *rqstp,
 680                                      struct svc_rdma_read_info *info,
 681                                      __be32 *p)
 682 {
 683         unsigned int i;
 684         int ret;
 685
 686         ret = -EINVAL;
 687         info->ri_chunklen = 0;
 688         while (*p++ != xdr_zero && be32_to_cpup(p++) == info->ri_position) {
 689                 u32 rs_handle, rs_length;
 690                 u64 rs_offset;
 691
 692                 rs_handle = be32_to_cpup(p++);
 693                 rs_length = be32_to_cpup(p++);
 694                 p = xdr_decode_hyper(p, &rs_offset);
 695
 696                 ret = svc_rdma_build_read_segment(info, rqstp,
 697                                                   rs_handle, rs_length,
 698                                                   rs_offset);
 699                 if (ret < 0)
 700                         break;
 701
 702                 trace_svcrdma_encode_rseg(rs_handle, rs_length, rs_offset);
 703                 info->ri_chunklen += rs_length;
 704         }
 705
 706         /* Pages under I/O have been copied to head->rc_pages.
 707          * Prevent their premature release by svc_xprt_release() .
 708          */
 709         for (i = 0; i < info->ri_readctxt->rc_page_count; i++)
 710                 rqstp->rq_pages[i] = NULL;
 711
 712         return ret;
 713 }
 714
 715 /* Construct RDMA Reads to pull over a normal Read chunk. The chunk
 716  * data lands in the page list of head->rc_arg.pages.
 717  *
 718  * Currently NFSD does not look at the head->rc_arg.tail[0] iovec.
 719  * Therefore, XDR round-up of the Read chunk and trailing
 720  * inline content must both be added at the end of the pagelist.
 721  */
 722 static int svc_rdma_build_normal_read_chunk(struct svc_rqst *rqstp,
 723                                             struct svc_rdma_read_info *info,
 724                                             __be32 *p)
 725 {
 726         struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
 727         int ret;
 728
 729         ret = svc_rdma_build_read_chunk(rqstp, info, p);
 730         if (ret < 0)
 731                 goto out;
 732
 733         trace_svcrdma_encode_read(info->ri_chunklen, info->ri_position);
 734
 735         head->rc_hdr_count = 0;
 736
 737         /* Split the Receive buffer between the head and tail
 738          * buffers at Read chunk's position. XDR roundup of the
 739          * chunk is not included in either the pagelist or in
 740          * the tail.
 741          */
 742         head->rc_arg.tail[0].iov_base =
 743                 head->rc_arg.head[0].iov_base + info->ri_position;
 744         head->rc_arg.tail[0].iov_len =
 745                 head->rc_arg.head[0].iov_len - info->ri_position;
 746         head->rc_arg.head[0].iov_len = info->ri_position;
 747
 748         /* Read chunk may need XDR roundup (see RFC 8166, s. 3.4.5.2).
 749          *
 750          * If the client already rounded up the chunk length, the
 751          * length does not change. Otherwise, the length of the page
 752          * list is increased to include XDR round-up.
 753          *
 754          * Currently these chunks always start at page offset 0,
 755          * thus the rounded-up length never crosses a page boundary.
 756          */
 757         info->ri_chunklen = XDR_QUADLEN(info->ri_chunklen) << 2;
 758
 759         head->rc_arg.page_len = info->ri_chunklen;
 760         head->rc_arg.len += info->ri_chunklen;
 761         head->rc_arg.buflen += info->ri_chunklen;
 762
 763 out:
 764         return ret;
 765 }
 766
 767 /* Construct RDMA Reads to pull over a Position Zero Read chunk.
 768  * The start of the data lands in the first page just after
 769  * the Transport header, and the rest lands in the page list of
 770  * head->rc_arg.pages.
 771  *
 772  * Assumptions:
 773  *      - A PZRC has an XDR-aligned length (no implicit round-up).
 774  *      - There can be no trailing inline content (IOW, we assume
 775  *        a PZRC is never sent in an RDMA_MSG message, though it's
 776  *        allowed by spec).
 777  */
 778 static int svc_rdma_build_pz_read_chunk(struct svc_rqst *rqstp,
 779                                         struct svc_rdma_read_info *info,
 780                                         __be32 *p)
 781 {
 782         struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
 783         int ret;
 784
 785         ret = svc_rdma_build_read_chunk(rqstp, info, p);
 786         if (ret < 0)
 787                 goto out;
 788
 789         trace_svcrdma_encode_pzr(info->ri_chunklen);
 790
 791         head->rc_arg.len += info->ri_chunklen;
 792         head->rc_arg.buflen += info->ri_chunklen;
 793
 794         head->rc_hdr_count = 1;
 795         head->rc_arg.head[0].iov_base = page_address(head->rc_pages[0]);
 796         head->rc_arg.head[0].iov_len = min_t(size_t, PAGE_SIZE,
 797                                              info->ri_chunklen);
 798
 799         head->rc_arg.page_len = info->ri_chunklen -
 800                                 head->rc_arg.head[0].iov_len;
 801
 802 out:
 803         return ret;
 804 }
 805
 806 /**
 807  * svc_rdma_recv_read_chunk - Pull a Read chunk from the client
 808  * @rdma: controlling RDMA transport
 809  * @rqstp: set of pages to use as Read sink buffers
 810  * @head: pages under I/O collect here
 811  * @p: pointer to start of Read chunk
 812  *
 813  * Returns:
 814  *      %0 if all needed RDMA Reads were posted successfully,
 815  *      %-EINVAL if client provided too many segments,
 816  *      %-ENOMEM if rdma_rw context pool was exhausted,
 817  *      %-ENOTCONN if posting failed (connection is lost),
 818  *      %-EIO if rdma_rw initialization failed (DMA mapping, etc).
 819  *
 820  * Assumptions:
 821  * - All Read segments in @p have the same Position value.
 822  */
 823 int svc_rdma_recv_read_chunk(struct svcxprt_rdma *rdma, struct svc_rqst *rqstp,
 824                              struct svc_rdma_recv_ctxt *head, __be32 *p)
 825 {
 826         struct svc_rdma_read_info *info;
 827         int ret;
 828
 829         /* The request (with page list) is constructed in
 830          * head->rc_arg. Pages involved with RDMA Read I/O are
 831          * transferred there.
 832          */
 833         head->rc_arg.head[0] = rqstp->rq_arg.head[0];
 834         head->rc_arg.tail[0] = rqstp->rq_arg.tail[0];
 835         head->rc_arg.pages = head->rc_pages;
 836         head->rc_arg.page_base = 0;
 837         head->rc_arg.page_len = 0;
 838         head->rc_arg.len = rqstp->rq_arg.len;
 839         head->rc_arg.buflen = rqstp->rq_arg.buflen;
 840
 841         info = svc_rdma_read_info_alloc(rdma);
 842         if (!info)
 843                 return -ENOMEM;
 844         info->ri_readctxt = head;
 845         info->ri_pageno = 0;
 846         info->ri_pageoff = 0;
 847
 848         info->ri_position = be32_to_cpup(p + 1);
 849         if (info->ri_position)
 850                 ret = svc_rdma_build_normal_read_chunk(rqstp, info, p);
 851         else
 852                 ret = svc_rdma_build_pz_read_chunk(rqstp, info, p);
 853         if (ret < 0)
 854                 goto out_err;
 855
 856         ret = svc_rdma_post_chunk_ctxt(&info->ri_cc);
 857         if (ret < 0)
 858                 goto out_err;
 859         return 0;
 860
 861 out_err:
 862         svc_rdma_read_info_free(info);
 863         return ret;
 864 }