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