fs/nfs/nfs4filelayout.c

   1 /*
   2  *  Module for the pnfs nfs4 file layout driver.
   3  *  Defines all I/O and Policy interface operations, plus code
   4  *  to register itself with the pNFS client.
   5  *
   6  *  Copyright (c) 2002
   7  *  The Regents of the University of Michigan
   8  *  All Rights Reserved
   9  *
  10  *  Dean Hildebrand <dhildebz@umich.edu>
  11  *
  12  *  Permission is granted to use, copy, create derivative works, and
  13  *  redistribute this software and such derivative works for any purpose,
  14  *  so long as the name of the University of Michigan is not used in
  15  *  any advertising or publicity pertaining to the use or distribution
  16  *  of this software without specific, written prior authorization. If
  17  *  the above copyright notice or any other identification of the
  18  *  University of Michigan is included in any copy of any portion of
  19  *  this software, then the disclaimer below must also be included.
  20  *
  21  *  This software is provided as is, without representation or warranty
  22  *  of any kind either express or implied, including without limitation
  23  *  the implied warranties of merchantability, fitness for a particular
  24  *  purpose, or noninfringement.  The Regents of the University of
  25  *  Michigan shall not be liable for any damages, including special,
  26  *  indirect, incidental, or consequential damages, with respect to any
  27  *  claim arising out of or in connection with the use of the software,
  28  *  even if it has been or is hereafter advised of the possibility of
  29  *  such damages.
  30  */
  31
  32 #include <linux/nfs_fs.h>
  33 #include <linux/nfs_page.h>
  34 #include <linux/module.h>
  35
  36 #include <linux/sunrpc/metrics.h>
  37
  38 #include "internal.h"
  39 #include "nfs4filelayout.h"
  40
  41 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  42
  43 MODULE_LICENSE("GPL");
  44 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
  45 MODULE_DESCRIPTION("The NFSv4 file layout driver");
  46
  47 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
  48
  49 static loff_t
  50 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
  51                             loff_t offset)
  52 {
  53         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
  54         u64 stripe_no;
  55         u32 rem;
  56
  57         offset -= flseg->pattern_offset;
  58         stripe_no = div_u64(offset, stripe_width);
  59         div_u64_rem(offset, flseg->stripe_unit, &rem);
  60
  61         return stripe_no * flseg->stripe_unit + rem;
  62 }
  63
  64 /* This function is used by the layout driver to calculate the
  65  * offset of the file on the dserver based on whether the
  66  * layout type is STRIPE_DENSE or STRIPE_SPARSE
  67  */
  68 static loff_t
  69 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
  70 {
  71         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
  72
  73         switch (flseg->stripe_type) {
  74         case STRIPE_SPARSE:
  75                 return offset;
  76
  77         case STRIPE_DENSE:
  78                 return filelayout_get_dense_offset(flseg, offset);
  79         }
  80
  81         BUG();
  82 }
  83
  84 static int filelayout_async_handle_error(struct rpc_task *task,
  85                                          struct nfs4_state *state,
  86                                          struct nfs_client *clp,
  87                                          int *reset)
  88 {
  89         if (task->tk_status >= 0)
  90                 return 0;
  91
  92         *reset = 0;
  93
  94         switch (task->tk_status) {
  95         case -NFS4ERR_BADSESSION:
  96         case -NFS4ERR_BADSLOT:
  97         case -NFS4ERR_BAD_HIGH_SLOT:
  98         case -NFS4ERR_DEADSESSION:
  99         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
 100         case -NFS4ERR_SEQ_FALSE_RETRY:
 101         case -NFS4ERR_SEQ_MISORDERED:
 102                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 103                         "flags 0x%x\n", __func__, task->tk_status,
 104                         clp->cl_exchange_flags);
 105                 nfs4_schedule_session_recovery(clp->cl_session);
 106                 break;
 107         case -NFS4ERR_DELAY:
 108         case -NFS4ERR_GRACE:
 109         case -EKEYEXPIRED:
 110                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 111                 break;
 112         case -NFS4ERR_RETRY_UNCACHED_REP:
 113                 break;
 114         default:
 115                 dprintk("%s DS error. Retry through MDS %d\n", __func__,
 116                         task->tk_status);
 117                 *reset = 1;
 118                 break;
 119         }
 120         task->tk_status = 0;
 121         return -EAGAIN;
 122 }
 123
 124 /* NFS_PROTO call done callback routines */
 125
 126 static int filelayout_read_done_cb(struct rpc_task *task,
 127                                 struct nfs_read_data *data)
 128 {
 129         int reset = 0;
 130
 131         dprintk("%s DS read\n", __func__);
 132
 133         if (filelayout_async_handle_error(task, data->args.context->state,
 134                                           data->ds_clp, &reset) == -EAGAIN) {
 135                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 136                         __func__, data->ds_clp, data->ds_clp->cl_session);
 137                 if (reset) {
 138                         pnfs_set_lo_fail(data->lseg);
 139                         nfs4_reset_read(task, data);
 140                 }
 141                 rpc_restart_call_prepare(task);
 142                 return -EAGAIN;
 143         }
 144
 145         return 0;
 146 }
 147
 148 /*
 149  * We reference the rpc_cred of the first WRITE that triggers the need for
 150  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 151  * rfc5661 is not clear about which credential should be used.
 152  */
 153 static void
 154 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
 155 {
 156         if (FILELAYOUT_LSEG(wdata->lseg)->commit_through_mds ||
 157             wdata->res.verf->committed == NFS_FILE_SYNC)
 158                 return;
 159
 160         pnfs_set_layoutcommit(wdata);
 161         dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, wdata->inode->i_ino,
 162                 (unsigned long) NFS_I(wdata->inode)->layout->plh_lwb);
 163 }
 164
 165 /*
 166  * Call ops for the async read/write cases
 167  * In the case of dense layouts, the offset needs to be reset to its
 168  * original value.
 169  */
 170 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 171 {
 172         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 173
 174         rdata->read_done_cb = filelayout_read_done_cb;
 175
 176         if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
 177                                 &rdata->args.seq_args, &rdata->res.seq_res,
 178                                 task))
 179                 return;
 180
 181         rpc_call_start(task);
 182 }
 183
 184 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 185 {
 186         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 187
 188         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 189
 190         /* Note this may cause RPC to be resent */
 191         rdata->mds_ops->rpc_call_done(task, data);
 192 }
 193
 194 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
 195 {
 196         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 197
 198         rpc_count_iostats(task, NFS_SERVER(rdata->inode)->client->cl_metrics);
 199 }
 200
 201 static void filelayout_read_release(void *data)
 202 {
 203         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 204
 205         rdata->mds_ops->rpc_release(data);
 206 }
 207
 208 static int filelayout_write_done_cb(struct rpc_task *task,
 209                                 struct nfs_write_data *data)
 210 {
 211         int reset = 0;
 212
 213         if (filelayout_async_handle_error(task, data->args.context->state,
 214                                           data->ds_clp, &reset) == -EAGAIN) {
 215                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 216                         __func__, data->ds_clp, data->ds_clp->cl_session);
 217                 if (reset) {
 218                         pnfs_set_lo_fail(data->lseg);
 219                         nfs4_reset_write(task, data);
 220                 }
 221                 rpc_restart_call_prepare(task);
 222                 return -EAGAIN;
 223         }
 224
 225         filelayout_set_layoutcommit(data);
 226         return 0;
 227 }
 228
 229 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 230 static void prepare_to_resend_writes(struct nfs_write_data *data)
 231 {
 232         struct nfs_page *first = nfs_list_entry(data->pages.next);
 233
 234         data->task.tk_status = 0;
 235         memcpy(data->verf.verifier, first->wb_verf.verifier,
 236                sizeof(first->wb_verf.verifier));
 237         data->verf.verifier[0]++; /* ensure verifier mismatch */
 238 }
 239
 240 static int filelayout_commit_done_cb(struct rpc_task *task,
 241                                      struct nfs_write_data *data)
 242 {
 243         int reset = 0;
 244
 245         if (filelayout_async_handle_error(task, data->args.context->state,
 246                                           data->ds_clp, &reset) == -EAGAIN) {
 247                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 248                         __func__, data->ds_clp, data->ds_clp->cl_session);
 249                 if (reset) {
 250                         prepare_to_resend_writes(data);
 251                         pnfs_set_lo_fail(data->lseg);
 252                 } else
 253                         rpc_restart_call_prepare(task);
 254                 return -EAGAIN;
 255         }
 256
 257         return 0;
 258 }
 259
 260 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 261 {
 262         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 263
 264         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 265                                 &wdata->args.seq_args, &wdata->res.seq_res,
 266                                 task))
 267                 return;
 268
 269         rpc_call_start(task);
 270 }
 271
 272 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 273 {
 274         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 275
 276         /* Note this may cause RPC to be resent */
 277         wdata->mds_ops->rpc_call_done(task, data);
 278 }
 279
 280 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
 281 {
 282         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 283
 284         rpc_count_iostats(task, NFS_SERVER(wdata->inode)->client->cl_metrics);
 285 }
 286
 287 static void filelayout_write_release(void *data)
 288 {
 289         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 290
 291         wdata->mds_ops->rpc_release(data);
 292 }
 293
 294 static void filelayout_commit_release(void *data)
 295 {
 296         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 297
 298         nfs_commit_release_pages(wdata);
 299         if (atomic_dec_and_test(&NFS_I(wdata->inode)->commits_outstanding))
 300                 nfs_commit_clear_lock(NFS_I(wdata->inode));
 301         nfs_commitdata_release(wdata);
 302 }
 303
 304 struct rpc_call_ops filelayout_read_call_ops = {
 305         .rpc_call_prepare = filelayout_read_prepare,
 306         .rpc_call_done = filelayout_read_call_done,
 307         .rpc_count_stats = filelayout_read_count_stats,
 308         .rpc_release = filelayout_read_release,
 309 };
 310
 311 struct rpc_call_ops filelayout_write_call_ops = {
 312         .rpc_call_prepare = filelayout_write_prepare,
 313         .rpc_call_done = filelayout_write_call_done,
 314         .rpc_count_stats = filelayout_write_count_stats,
 315         .rpc_release = filelayout_write_release,
 316 };
 317
 318 struct rpc_call_ops filelayout_commit_call_ops = {
 319         .rpc_call_prepare = filelayout_write_prepare,
 320         .rpc_call_done = filelayout_write_call_done,
 321         .rpc_count_stats = filelayout_write_count_stats,
 322         .rpc_release = filelayout_commit_release,
 323 };
 324
 325 static enum pnfs_try_status
 326 filelayout_read_pagelist(struct nfs_read_data *data)
 327 {
 328         struct pnfs_layout_segment *lseg = data->lseg;
 329         struct nfs4_pnfs_ds *ds;
 330         loff_t offset = data->args.offset;
 331         u32 j, idx;
 332         struct nfs_fh *fh;
 333         int status;
 334
 335         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 336                 __func__, data->inode->i_ino,
 337                 data->args.pgbase, (size_t)data->args.count, offset);
 338
 339         if (test_bit(NFS_DEVICEID_INVALID, &FILELAYOUT_DEVID_NODE(lseg)->flags))
 340                 return PNFS_NOT_ATTEMPTED;
 341
 342         /* Retrieve the correct rpc_client for the byte range */
 343         j = nfs4_fl_calc_j_index(lseg, offset);
 344         idx = nfs4_fl_calc_ds_index(lseg, j);
 345         ds = nfs4_fl_prepare_ds(lseg, idx);
 346         if (!ds) {
 347                 /* Either layout fh index faulty, or ds connect failed */
 348                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 349                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 350                 return PNFS_NOT_ATTEMPTED;
 351         }
 352         dprintk("%s USE DS: %s\n", __func__, ds->ds_remotestr);
 353
 354         /* No multipath support. Use first DS */
 355         data->ds_clp = ds->ds_clp;
 356         fh = nfs4_fl_select_ds_fh(lseg, j);
 357         if (fh)
 358                 data->args.fh = fh;
 359
 360         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 361         data->mds_offset = offset;
 362
 363         /* Perform an asynchronous read to ds */
 364         status = nfs_initiate_read(data, ds->ds_clp->cl_rpcclient,
 365                                    &filelayout_read_call_ops);
 366         BUG_ON(status != 0);
 367         return PNFS_ATTEMPTED;
 368 }
 369
 370 /* Perform async writes. */
 371 static enum pnfs_try_status
 372 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
 373 {
 374         struct pnfs_layout_segment *lseg = data->lseg;
 375         struct nfs4_pnfs_ds *ds;
 376         loff_t offset = data->args.offset;
 377         u32 j, idx;
 378         struct nfs_fh *fh;
 379         int status;
 380
 381         if (test_bit(NFS_DEVICEID_INVALID, &FILELAYOUT_DEVID_NODE(lseg)->flags))
 382                 return PNFS_NOT_ATTEMPTED;
 383
 384         /* Retrieve the correct rpc_client for the byte range */
 385         j = nfs4_fl_calc_j_index(lseg, offset);
 386         idx = nfs4_fl_calc_ds_index(lseg, j);
 387         ds = nfs4_fl_prepare_ds(lseg, idx);
 388         if (!ds) {
 389                 printk(KERN_ERR "NFS: %s: prepare_ds failed, use MDS\n",
 390                         __func__);
 391                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 392                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 393                 return PNFS_NOT_ATTEMPTED;
 394         }
 395         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s\n", __func__,
 396                 data->inode->i_ino, sync, (size_t) data->args.count, offset,
 397                 ds->ds_remotestr);
 398
 399         data->write_done_cb = filelayout_write_done_cb;
 400         data->ds_clp = ds->ds_clp;
 401         fh = nfs4_fl_select_ds_fh(lseg, j);
 402         if (fh)
 403                 data->args.fh = fh;
 404         /*
 405          * Get the file offset on the dserver. Set the write offset to
 406          * this offset and save the original offset.
 407          */
 408         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 409
 410         /* Perform an asynchronous write */
 411         status = nfs_initiate_write(data, ds->ds_clp->cl_rpcclient,
 412                                     &filelayout_write_call_ops, sync);
 413         BUG_ON(status != 0);
 414         return PNFS_ATTEMPTED;
 415 }
 416
 417 /*
 418  * filelayout_check_layout()
 419  *
 420  * Make sure layout segment parameters are sane WRT the device.
 421  * At this point no generic layer initialization of the lseg has occurred,
 422  * and nothing has been added to the layout_hdr cache.
 423  *
 424  */
 425 static int
 426 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 427                         struct nfs4_filelayout_segment *fl,
 428                         struct nfs4_layoutget_res *lgr,
 429                         struct nfs4_deviceid *id,
 430                         gfp_t gfp_flags)
 431 {
 432         struct nfs4_deviceid_node *d;
 433         struct nfs4_file_layout_dsaddr *dsaddr;
 434         int status = -EINVAL;
 435         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 436
 437         dprintk("--> %s\n", __func__);
 438
 439         /* FIXME: remove this check when layout segment support is added */
 440         if (lgr->range.offset != 0 ||
 441             lgr->range.length != NFS4_MAX_UINT64) {
 442                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 443                         __func__);
 444                 goto out;
 445         }
 446
 447         if (fl->pattern_offset > lgr->range.offset) {
 448                 dprintk("%s pattern_offset %lld too large\n",
 449                                 __func__, fl->pattern_offset);
 450                 goto out;
 451         }
 452
 453         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 454                 dprintk("%s Invalid stripe unit (%u)\n",
 455                         __func__, fl->stripe_unit);
 456                 goto out;
 457         }
 458
 459         /* find and reference the deviceid */
 460         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
 461                                    NFS_SERVER(lo->plh_inode)->nfs_client, id);
 462         if (d == NULL) {
 463                 dsaddr = get_device_info(lo->plh_inode, id, gfp_flags);
 464                 if (dsaddr == NULL)
 465                         goto out;
 466         } else
 467                 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 468         /* Found deviceid is being reaped */
 469         if (test_bit(NFS_DEVICEID_INVALID, &dsaddr->id_node.flags))
 470                         goto out_put;
 471
 472         fl->dsaddr = dsaddr;
 473
 474         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 475                 dprintk("%s Bad first_stripe_index %u\n",
 476                                 __func__, fl->first_stripe_index);
 477                 goto out_put;
 478         }
 479
 480         if ((fl->stripe_type == STRIPE_SPARSE &&
 481             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 482             (fl->stripe_type == STRIPE_DENSE &&
 483             fl->num_fh != dsaddr->stripe_count)) {
 484                 dprintk("%s num_fh %u not valid for given packing\n",
 485                         __func__, fl->num_fh);
 486                 goto out_put;
 487         }
 488
 489         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
 490                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
 491                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
 492                         nfss->wsize);
 493         }
 494
 495         status = 0;
 496 out:
 497         dprintk("--> %s returns %d\n", __func__, status);
 498         return status;
 499 out_put:
 500         nfs4_fl_put_deviceid(dsaddr);
 501         goto out;
 502 }
 503
 504 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 505 {
 506         int i;
 507
 508         for (i = 0; i < fl->num_fh; i++) {
 509                 if (!fl->fh_array[i])
 510                         break;
 511                 kfree(fl->fh_array[i]);
 512         }
 513         kfree(fl->fh_array);
 514         fl->fh_array = NULL;
 515 }
 516
 517 static void
 518 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 519 {
 520         filelayout_free_fh_array(fl);
 521         kfree(fl);
 522 }
 523
 524 static int
 525 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 526                          struct nfs4_filelayout_segment *fl,
 527                          struct nfs4_layoutget_res *lgr,
 528                          struct nfs4_deviceid *id,
 529                          gfp_t gfp_flags)
 530 {
 531         struct xdr_stream stream;
 532         struct xdr_buf buf;
 533         struct page *scratch;
 534         __be32 *p;
 535         uint32_t nfl_util;
 536         int i;
 537
 538         dprintk("%s: set_layout_map Begin\n", __func__);
 539
 540         scratch = alloc_page(gfp_flags);
 541         if (!scratch)
 542                 return -ENOMEM;
 543
 544         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 545         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 546
 547         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 548          * num_fh (4) */
 549         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 550         if (unlikely(!p))
 551                 goto out_err;
 552
 553         memcpy(id, p, sizeof(*id));
 554         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 555         nfs4_print_deviceid(id);
 556
 557         nfl_util = be32_to_cpup(p++);
 558         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 559                 fl->commit_through_mds = 1;
 560         if (nfl_util & NFL4_UFLG_DENSE)
 561                 fl->stripe_type = STRIPE_DENSE;
 562         else
 563                 fl->stripe_type = STRIPE_SPARSE;
 564         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 565
 566         fl->first_stripe_index = be32_to_cpup(p++);
 567         p = xdr_decode_hyper(p, &fl->pattern_offset);
 568         fl->num_fh = be32_to_cpup(p++);
 569
 570         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 571                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 572                 fl->pattern_offset);
 573
 574         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 575          * Futher checking is done in filelayout_check_layout */
 576         if (fl->num_fh >
 577             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 578                 goto out_err;
 579
 580         if (fl->num_fh > 0) {
 581                 fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *),
 582                                        gfp_flags);
 583                 if (!fl->fh_array)
 584                         goto out_err;
 585         }
 586
 587         for (i = 0; i < fl->num_fh; i++) {
 588                 /* Do we want to use a mempool here? */
 589                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 590                 if (!fl->fh_array[i])
 591                         goto out_err_free;
 592
 593                 p = xdr_inline_decode(&stream, 4);
 594                 if (unlikely(!p))
 595                         goto out_err_free;
 596                 fl->fh_array[i]->size = be32_to_cpup(p++);
 597                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 598                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
 599                                i, fl->fh_array[i]->size);
 600                         goto out_err_free;
 601                 }
 602
 603                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 604                 if (unlikely(!p))
 605                         goto out_err_free;
 606                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 607                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 608                         fl->fh_array[i]->size);
 609         }
 610
 611         __free_page(scratch);
 612         return 0;
 613
 614 out_err_free:
 615         filelayout_free_fh_array(fl);
 616 out_err:
 617         __free_page(scratch);
 618         return -EIO;
 619 }
 620
 621 static void
 622 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 623 {
 624         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 625
 626         dprintk("--> %s\n", __func__);
 627         nfs4_fl_put_deviceid(fl->dsaddr);
 628         kfree(fl->commit_buckets);
 629         _filelayout_free_lseg(fl);
 630 }
 631
 632 static struct pnfs_layout_segment *
 633 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 634                       struct nfs4_layoutget_res *lgr,
 635                       gfp_t gfp_flags)
 636 {
 637         struct nfs4_filelayout_segment *fl;
 638         int rc;
 639         struct nfs4_deviceid id;
 640
 641         dprintk("--> %s\n", __func__);
 642         fl = kzalloc(sizeof(*fl), gfp_flags);
 643         if (!fl)
 644                 return NULL;
 645
 646         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 647         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 648                 _filelayout_free_lseg(fl);
 649                 return NULL;
 650         }
 651
 652         /* This assumes there is only one IOMODE_RW lseg.  What
 653          * we really want to do is have a layout_hdr level
 654          * dictionary of <multipath_list4, fh> keys, each
 655          * associated with a struct list_head, populated by calls
 656          * to filelayout_write_pagelist().
 657          * */
 658         if ((!fl->commit_through_mds) && (lgr->range.iomode == IOMODE_RW)) {
 659                 int i;
 660                 int size = (fl->stripe_type == STRIPE_SPARSE) ?
 661                         fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 662
 663                 fl->commit_buckets = kcalloc(size, sizeof(struct list_head), gfp_flags);
 664                 if (!fl->commit_buckets) {
 665                         filelayout_free_lseg(&fl->generic_hdr);
 666                         return NULL;
 667                 }
 668                 fl->number_of_buckets = size;
 669                 for (i = 0; i < size; i++)
 670                         INIT_LIST_HEAD(&fl->commit_buckets[i]);
 671         }
 672         return &fl->generic_hdr;
 673 }
 674
 675 /*
 676  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 677  *
 678  * return true  : coalesce page
 679  * return false : don't coalesce page
 680  */
 681 static bool
 682 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 683                    struct nfs_page *req)
 684 {
 685         u64 p_stripe, r_stripe;
 686         u32 stripe_unit;
 687
 688         if (!pnfs_generic_pg_test(pgio, prev, req) ||
 689             !nfs_generic_pg_test(pgio, prev, req))
 690                 return false;
 691
 692         p_stripe = (u64)prev->wb_index << PAGE_CACHE_SHIFT;
 693         r_stripe = (u64)req->wb_index << PAGE_CACHE_SHIFT;
 694         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 695
 696         do_div(p_stripe, stripe_unit);
 697         do_div(r_stripe, stripe_unit);
 698
 699         return (p_stripe == r_stripe);
 700 }
 701
 702 void
 703 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 704                         struct nfs_page *req)
 705 {
 706         BUG_ON(pgio->pg_lseg != NULL);
 707
 708         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 709                                            req->wb_context,
 710                                            0,
 711                                            NFS4_MAX_UINT64,
 712                                            IOMODE_READ,
 713                                            GFP_KERNEL);
 714         /* If no lseg, fall back to read through mds */
 715         if (pgio->pg_lseg == NULL)
 716                 nfs_pageio_reset_read_mds(pgio);
 717 }
 718
 719 void
 720 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 721                          struct nfs_page *req)
 722 {
 723         BUG_ON(pgio->pg_lseg != NULL);
 724
 725         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 726                                            req->wb_context,
 727                                            0,
 728                                            NFS4_MAX_UINT64,
 729                                            IOMODE_RW,
 730                                            GFP_NOFS);
 731         /* If no lseg, fall back to write through mds */
 732         if (pgio->pg_lseg == NULL)
 733                 nfs_pageio_reset_write_mds(pgio);
 734 }
 735
 736 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 737         .pg_init = filelayout_pg_init_read,
 738         .pg_test = filelayout_pg_test,
 739         .pg_doio = pnfs_generic_pg_readpages,
 740 };
 741
 742 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 743         .pg_init = filelayout_pg_init_write,
 744         .pg_test = filelayout_pg_test,
 745         .pg_doio = pnfs_generic_pg_writepages,
 746 };
 747
 748 static bool filelayout_mark_pnfs_commit(struct pnfs_layout_segment *lseg)
 749 {
 750         return !FILELAYOUT_LSEG(lseg)->commit_through_mds;
 751 }
 752
 753 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 754 {
 755         if (fl->stripe_type == STRIPE_SPARSE)
 756                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
 757         else
 758                 return j;
 759 }
 760
 761 struct list_head *filelayout_choose_commit_list(struct nfs_page *req)
 762 {
 763         struct pnfs_layout_segment *lseg = req->wb_commit_lseg;
 764         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 765         u32 i, j;
 766         struct list_head *list;
 767
 768         /* Note that we are calling nfs4_fl_calc_j_index on each page
 769          * that ends up being committed to a data server.  An attractive
 770          * alternative is to add a field to nfs_write_data and nfs_page
 771          * to store the value calculated in filelayout_write_pagelist
 772          * and just use that here.
 773          */
 774         j = nfs4_fl_calc_j_index(lseg,
 775                                  (loff_t)req->wb_index << PAGE_CACHE_SHIFT);
 776         i = select_bucket_index(fl, j);
 777         list = &fl->commit_buckets[i];
 778         if (list_empty(list)) {
 779                 /* Non-empty buckets hold a reference on the lseg */
 780                 get_lseg(lseg);
 781         }
 782         return list;
 783 }
 784
 785 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 786 {
 787         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 788
 789         if (flseg->stripe_type == STRIPE_SPARSE)
 790                 return i;
 791         else
 792                 return nfs4_fl_calc_ds_index(lseg, i);
 793 }
 794
 795 static struct nfs_fh *
 796 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 797 {
 798         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 799
 800         if (flseg->stripe_type == STRIPE_SPARSE) {
 801                 if (flseg->num_fh == 1)
 802                         i = 0;
 803                 else if (flseg->num_fh == 0)
 804                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
 805                         return NULL;
 806         }
 807         return flseg->fh_array[i];
 808 }
 809
 810 static int filelayout_initiate_commit(struct nfs_write_data *data, int how)
 811 {
 812         struct pnfs_layout_segment *lseg = data->lseg;
 813         struct nfs4_pnfs_ds *ds;
 814         u32 idx;
 815         struct nfs_fh *fh;
 816
 817         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
 818         ds = nfs4_fl_prepare_ds(lseg, idx);
 819         if (!ds) {
 820                 printk(KERN_ERR "NFS: %s: prepare_ds failed, use MDS\n",
 821                         __func__);
 822                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 823                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 824                 prepare_to_resend_writes(data);
 825                 data->mds_ops->rpc_release(data);
 826                 return -EAGAIN;
 827         }
 828         dprintk("%s ino %lu, how %d\n", __func__, data->inode->i_ino, how);
 829         data->write_done_cb = filelayout_commit_done_cb;
 830         data->ds_clp = ds->ds_clp;
 831         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
 832         if (fh)
 833                 data->args.fh = fh;
 834         return nfs_initiate_commit(data, ds->ds_clp->cl_rpcclient,
 835                                    &filelayout_commit_call_ops, how);
 836 }
 837
 838 /*
 839  * This is only useful while we are using whole file layouts.
 840  */
 841 static struct pnfs_layout_segment *find_only_write_lseg(struct inode *inode)
 842 {
 843         struct pnfs_layout_segment *lseg, *rv = NULL;
 844
 845         spin_lock(&inode->i_lock);
 846         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
 847                 if (lseg->pls_range.iomode == IOMODE_RW)
 848                         rv = get_lseg(lseg);
 849         spin_unlock(&inode->i_lock);
 850         return rv;
 851 }
 852
 853 static int alloc_ds_commits(struct inode *inode, struct list_head *list)
 854 {
 855         struct pnfs_layout_segment *lseg;
 856         struct nfs4_filelayout_segment *fl;
 857         struct nfs_write_data *data;
 858         int i, j;
 859
 860         /* Won't need this when non-whole file layout segments are supported
 861          * instead we will use a pnfs_layout_hdr structure */
 862         lseg = find_only_write_lseg(inode);
 863         if (!lseg)
 864                 return 0;
 865         fl = FILELAYOUT_LSEG(lseg);
 866         for (i = 0; i < fl->number_of_buckets; i++) {
 867                 if (list_empty(&fl->commit_buckets[i]))
 868                         continue;
 869                 data = nfs_commitdata_alloc();
 870                 if (!data)
 871                         goto out_bad;
 872                 data->ds_commit_index = i;
 873                 data->lseg = lseg;
 874                 list_add(&data->pages, list);
 875         }
 876         put_lseg(lseg);
 877         return 0;
 878
 879 out_bad:
 880         for (j = i; j < fl->number_of_buckets; j++) {
 881                 if (list_empty(&fl->commit_buckets[i]))
 882                         continue;
 883                 nfs_retry_commit(&fl->commit_buckets[i], lseg);
 884                 put_lseg(lseg);  /* associated with emptying bucket */
 885         }
 886         put_lseg(lseg);
 887         /* Caller will clean up entries put on list */
 888         return -ENOMEM;
 889 }
 890
 891 /* This follows nfs_commit_list pretty closely */
 892 static int
 893 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
 894                            int how)
 895 {
 896         struct nfs_write_data   *data, *tmp;
 897         LIST_HEAD(list);
 898
 899         if (!list_empty(mds_pages)) {
 900                 data = nfs_commitdata_alloc();
 901                 if (!data)
 902                         goto out_bad;
 903                 data->lseg = NULL;
 904                 list_add(&data->pages, &list);
 905         }
 906
 907         if (alloc_ds_commits(inode, &list))
 908                 goto out_bad;
 909
 910         list_for_each_entry_safe(data, tmp, &list, pages) {
 911                 list_del_init(&data->pages);
 912                 atomic_inc(&NFS_I(inode)->commits_outstanding);
 913                 if (!data->lseg) {
 914                         nfs_init_commit(data, mds_pages, NULL);
 915                         nfs_initiate_commit(data, NFS_CLIENT(inode),
 916                                             data->mds_ops, how);
 917                 } else {
 918                         nfs_init_commit(data, &FILELAYOUT_LSEG(data->lseg)->commit_buckets[data->ds_commit_index], data->lseg);
 919                         filelayout_initiate_commit(data, how);
 920                 }
 921         }
 922         return 0;
 923  out_bad:
 924         list_for_each_entry_safe(data, tmp, &list, pages) {
 925                 nfs_retry_commit(&data->pages, data->lseg);
 926                 list_del_init(&data->pages);
 927                 nfs_commit_free(data);
 928         }
 929         nfs_retry_commit(mds_pages, NULL);
 930         nfs_commit_clear_lock(NFS_I(inode));
 931         return -ENOMEM;
 932 }
 933
 934 static void
 935 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
 936 {
 937         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
 938 }
 939
 940 static struct pnfs_layoutdriver_type filelayout_type = {
 941         .id                     = LAYOUT_NFSV4_1_FILES,
 942         .name                   = "LAYOUT_NFSV4_1_FILES",
 943         .owner                  = THIS_MODULE,
 944         .alloc_lseg             = filelayout_alloc_lseg,
 945         .free_lseg              = filelayout_free_lseg,
 946         .pg_read_ops            = &filelayout_pg_read_ops,
 947         .pg_write_ops           = &filelayout_pg_write_ops,
 948         .mark_pnfs_commit       = filelayout_mark_pnfs_commit,
 949         .choose_commit_list     = filelayout_choose_commit_list,
 950         .commit_pagelist        = filelayout_commit_pagelist,
 951         .read_pagelist          = filelayout_read_pagelist,
 952         .write_pagelist         = filelayout_write_pagelist,
 953         .free_deviceid_node     = filelayout_free_deveiceid_node,
 954 };
 955
 956 static int __init nfs4filelayout_init(void)
 957 {
 958         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
 959                __func__);
 960         return pnfs_register_layoutdriver(&filelayout_type);
 961 }
 962
 963 static void __exit nfs4filelayout_exit(void)
 964 {
 965         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
 966                __func__);
 967         pnfs_unregister_layoutdriver(&filelayout_type);
 968 }
 969
 970 MODULE_ALIAS("nfs-layouttype4-1");
 971
 972 module_init(nfs4filelayout_init);
 973 module_exit(nfs4filelayout_exit);