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 "nfs4session.h"
  39 #include "internal.h"
  40 #include "delegation.h"
  41 #include "nfs4filelayout.h"
  42
  43 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  44
  45 MODULE_LICENSE("GPL");
  46 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
  47 MODULE_DESCRIPTION("The NFSv4 file layout driver");
  48
  49 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
  50
  51 static loff_t
  52 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
  53                             loff_t offset)
  54 {
  55         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
  56         u64 stripe_no;
  57         u32 rem;
  58
  59         offset -= flseg->pattern_offset;
  60         stripe_no = div_u64(offset, stripe_width);
  61         div_u64_rem(offset, flseg->stripe_unit, &rem);
  62
  63         return stripe_no * flseg->stripe_unit + rem;
  64 }
  65
  66 /* This function is used by the layout driver to calculate the
  67  * offset of the file on the dserver based on whether the
  68  * layout type is STRIPE_DENSE or STRIPE_SPARSE
  69  */
  70 static loff_t
  71 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
  72 {
  73         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
  74
  75         switch (flseg->stripe_type) {
  76         case STRIPE_SPARSE:
  77                 return offset;
  78
  79         case STRIPE_DENSE:
  80                 return filelayout_get_dense_offset(flseg, offset);
  81         }
  82
  83         BUG();
  84 }
  85
  86 static void filelayout_reset_write(struct nfs_write_data *data)
  87 {
  88         struct nfs_pgio_header *hdr = data->header;
  89         struct rpc_task *task = &data->task;
  90
  91         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
  92                 dprintk("%s Reset task %5u for i/o through MDS "
  93                         "(req %s/%lld, %u bytes @ offset %llu)\n", __func__,
  94                         data->task.tk_pid,
  95                         hdr->inode->i_sb->s_id,
  96                         (long long)NFS_FILEID(hdr->inode),
  97                         data->args.count,
  98                         (unsigned long long)data->args.offset);
  99
 100                 task->tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
 101                                                         &hdr->pages,
 102                                                         hdr->completion_ops);
 103         }
 104 }
 105
 106 static void filelayout_reset_read(struct nfs_read_data *data)
 107 {
 108         struct nfs_pgio_header *hdr = data->header;
 109         struct rpc_task *task = &data->task;
 110
 111         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 112                 dprintk("%s Reset task %5u for i/o through MDS "
 113                         "(req %s/%lld, %u bytes @ offset %llu)\n", __func__,
 114                         data->task.tk_pid,
 115                         hdr->inode->i_sb->s_id,
 116                         (long long)NFS_FILEID(hdr->inode),
 117                         data->args.count,
 118                         (unsigned long long)data->args.offset);
 119
 120                 task->tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
 121                                                         &hdr->pages,
 122                                                         hdr->completion_ops);
 123         }
 124 }
 125
 126 static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo)
 127 {
 128         if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 129                 return;
 130         clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags);
 131         pnfs_return_layout(inode);
 132 }
 133
 134 static int filelayout_async_handle_error(struct rpc_task *task,
 135                                          struct nfs4_state *state,
 136                                          struct nfs_client *clp,
 137                                          struct pnfs_layout_segment *lseg)
 138 {
 139         struct pnfs_layout_hdr *lo = lseg->pls_layout;
 140         struct inode *inode = lo->plh_inode;
 141         struct nfs_server *mds_server = NFS_SERVER(inode);
 142         struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
 143         struct nfs_client *mds_client = mds_server->nfs_client;
 144         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
 145
 146         if (task->tk_status >= 0)
 147                 return 0;
 148
 149         switch (task->tk_status) {
 150         /* MDS state errors */
 151         case -NFS4ERR_DELEG_REVOKED:
 152         case -NFS4ERR_ADMIN_REVOKED:
 153         case -NFS4ERR_BAD_STATEID:
 154                 if (state == NULL)
 155                         break;
 156                 nfs_remove_bad_delegation(state->inode);
 157         case -NFS4ERR_OPENMODE:
 158                 if (state == NULL)
 159                         break;
 160                 nfs4_schedule_stateid_recovery(mds_server, state);
 161                 goto wait_on_recovery;
 162         case -NFS4ERR_EXPIRED:
 163                 if (state != NULL)
 164                         nfs4_schedule_stateid_recovery(mds_server, state);
 165                 nfs4_schedule_lease_recovery(mds_client);
 166                 goto wait_on_recovery;
 167         /* DS session errors */
 168         case -NFS4ERR_BADSESSION:
 169         case -NFS4ERR_BADSLOT:
 170         case -NFS4ERR_BAD_HIGH_SLOT:
 171         case -NFS4ERR_DEADSESSION:
 172         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
 173         case -NFS4ERR_SEQ_FALSE_RETRY:
 174         case -NFS4ERR_SEQ_MISORDERED:
 175                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 176                         "flags 0x%x\n", __func__, task->tk_status,
 177                         clp->cl_exchange_flags);
 178                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
 179                 break;
 180         case -NFS4ERR_DELAY:
 181         case -NFS4ERR_GRACE:
 182         case -EKEYEXPIRED:
 183                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 184                 break;
 185         case -NFS4ERR_RETRY_UNCACHED_REP:
 186                 break;
 187         /* Invalidate Layout errors */
 188         case -NFS4ERR_PNFS_NO_LAYOUT:
 189         case -ESTALE:           /* mapped NFS4ERR_STALE */
 190         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
 191         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
 192         case -NFS4ERR_FHEXPIRED:
 193         case -NFS4ERR_WRONG_TYPE:
 194                 dprintk("%s Invalid layout error %d\n", __func__,
 195                         task->tk_status);
 196                 /*
 197                  * Destroy layout so new i/o will get a new layout.
 198                  * Layout will not be destroyed until all current lseg
 199                  * references are put. Mark layout as invalid to resend failed
 200                  * i/o and all i/o waiting on the slot table to the MDS until
 201                  * layout is destroyed and a new valid layout is obtained.
 202                  */
 203                 pnfs_destroy_layout(NFS_I(inode));
 204                 rpc_wake_up(&tbl->slot_tbl_waitq);
 205                 goto reset;
 206         /* RPC connection errors */
 207         case -ECONNREFUSED:
 208         case -EHOSTDOWN:
 209         case -EHOSTUNREACH:
 210         case -ENETUNREACH:
 211         case -EIO:
 212         case -ETIMEDOUT:
 213         case -EPIPE:
 214                 dprintk("%s DS connection error %d\n", __func__,
 215                         task->tk_status);
 216                 nfs4_mark_deviceid_unavailable(devid);
 217                 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
 218                 rpc_wake_up(&tbl->slot_tbl_waitq);
 219                 /* fall through */
 220         default:
 221 reset:
 222                 dprintk("%s Retry through MDS. Error %d\n", __func__,
 223                         task->tk_status);
 224                 return -NFS4ERR_RESET_TO_MDS;
 225         }
 226 out:
 227         task->tk_status = 0;
 228         return -EAGAIN;
 229 wait_on_recovery:
 230         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
 231         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
 232                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
 233         goto out;
 234 }
 235
 236 /* NFS_PROTO call done callback routines */
 237
 238 static int filelayout_read_done_cb(struct rpc_task *task,
 239                                 struct nfs_read_data *data)
 240 {
 241         struct nfs_pgio_header *hdr = data->header;
 242         int err;
 243
 244         err = filelayout_async_handle_error(task, data->args.context->state,
 245                                             data->ds_clp, hdr->lseg);
 246
 247         switch (err) {
 248         case -NFS4ERR_RESET_TO_MDS:
 249                 filelayout_reset_read(data);
 250                 return task->tk_status;
 251         case -EAGAIN:
 252                 rpc_restart_call_prepare(task);
 253                 return -EAGAIN;
 254         }
 255
 256         return 0;
 257 }
 258
 259 /*
 260  * We reference the rpc_cred of the first WRITE that triggers the need for
 261  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 262  * rfc5661 is not clear about which credential should be used.
 263  */
 264 static void
 265 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
 266 {
 267         struct nfs_pgio_header *hdr = wdata->header;
 268
 269         if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
 270             wdata->res.verf->committed == NFS_FILE_SYNC)
 271                 return;
 272
 273         pnfs_set_layoutcommit(wdata);
 274         dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
 275                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
 276 }
 277
 278 bool
 279 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
 280 {
 281         return filelayout_test_devid_invalid(node) ||
 282                 nfs4_test_deviceid_unavailable(node);
 283 }
 284
 285 static bool
 286 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
 287 {
 288         struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
 289
 290         return filelayout_test_devid_unavailable(node);
 291 }
 292
 293 /*
 294  * Call ops for the async read/write cases
 295  * In the case of dense layouts, the offset needs to be reset to its
 296  * original value.
 297  */
 298 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 299 {
 300         struct nfs_read_data *rdata = data;
 301
 302         if (filelayout_reset_to_mds(rdata->header->lseg)) {
 303                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 304                 filelayout_reset_read(rdata);
 305                 rpc_exit(task, 0);
 306                 return;
 307         }
 308         rdata->read_done_cb = filelayout_read_done_cb;
 309
 310         if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
 311                                 &rdata->args.seq_args, &rdata->res.seq_res,
 312                                 task))
 313                 return;
 314
 315         rpc_call_start(task);
 316 }
 317
 318 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 319 {
 320         struct nfs_read_data *rdata = data;
 321
 322         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 323
 324         if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) &&
 325             task->tk_status == 0)
 326                 return;
 327
 328         /* Note this may cause RPC to be resent */
 329         rdata->header->mds_ops->rpc_call_done(task, data);
 330 }
 331
 332 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
 333 {
 334         struct nfs_read_data *rdata = data;
 335
 336         rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
 337 }
 338
 339 static void filelayout_read_release(void *data)
 340 {
 341         struct nfs_read_data *rdata = data;
 342         struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout;
 343
 344         filelayout_fenceme(lo->plh_inode, lo);
 345         nfs_put_client(rdata->ds_clp);
 346         rdata->header->mds_ops->rpc_release(data);
 347 }
 348
 349 static int filelayout_write_done_cb(struct rpc_task *task,
 350                                 struct nfs_write_data *data)
 351 {
 352         struct nfs_pgio_header *hdr = data->header;
 353         int err;
 354
 355         err = filelayout_async_handle_error(task, data->args.context->state,
 356                                             data->ds_clp, hdr->lseg);
 357
 358         switch (err) {
 359         case -NFS4ERR_RESET_TO_MDS:
 360                 filelayout_reset_write(data);
 361                 return task->tk_status;
 362         case -EAGAIN:
 363                 rpc_restart_call_prepare(task);
 364                 return -EAGAIN;
 365         }
 366
 367         filelayout_set_layoutcommit(data);
 368         return 0;
 369 }
 370
 371 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 372 static void prepare_to_resend_writes(struct nfs_commit_data *data)
 373 {
 374         struct nfs_page *first = nfs_list_entry(data->pages.next);
 375
 376         data->task.tk_status = 0;
 377         memcpy(&data->verf.verifier, &first->wb_verf,
 378                sizeof(data->verf.verifier));
 379         data->verf.verifier.data[0]++; /* ensure verifier mismatch */
 380 }
 381
 382 static int filelayout_commit_done_cb(struct rpc_task *task,
 383                                      struct nfs_commit_data *data)
 384 {
 385         int err;
 386
 387         err = filelayout_async_handle_error(task, NULL, data->ds_clp,
 388                                             data->lseg);
 389
 390         switch (err) {
 391         case -NFS4ERR_RESET_TO_MDS:
 392                 prepare_to_resend_writes(data);
 393                 return -EAGAIN;
 394         case -EAGAIN:
 395                 rpc_restart_call_prepare(task);
 396                 return -EAGAIN;
 397         }
 398
 399         return 0;
 400 }
 401
 402 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 403 {
 404         struct nfs_write_data *wdata = data;
 405
 406         if (filelayout_reset_to_mds(wdata->header->lseg)) {
 407                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 408                 filelayout_reset_write(wdata);
 409                 rpc_exit(task, 0);
 410                 return;
 411         }
 412         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 413                                 &wdata->args.seq_args, &wdata->res.seq_res,
 414                                 task))
 415                 return;
 416
 417         rpc_call_start(task);
 418 }
 419
 420 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 421 {
 422         struct nfs_write_data *wdata = data;
 423
 424         if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) &&
 425             task->tk_status == 0)
 426                 return;
 427
 428         /* Note this may cause RPC to be resent */
 429         wdata->header->mds_ops->rpc_call_done(task, data);
 430 }
 431
 432 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
 433 {
 434         struct nfs_write_data *wdata = data;
 435
 436         rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
 437 }
 438
 439 static void filelayout_write_release(void *data)
 440 {
 441         struct nfs_write_data *wdata = data;
 442         struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout;
 443
 444         filelayout_fenceme(lo->plh_inode, lo);
 445         nfs_put_client(wdata->ds_clp);
 446         wdata->header->mds_ops->rpc_release(data);
 447 }
 448
 449 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
 450 {
 451         struct nfs_commit_data *wdata = data;
 452
 453         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 454                                 &wdata->args.seq_args, &wdata->res.seq_res,
 455                                 task))
 456                 return;
 457
 458         rpc_call_start(task);
 459 }
 460
 461 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
 462 {
 463         struct nfs_commit_data *wdata = data;
 464
 465         /* Note this may cause RPC to be resent */
 466         wdata->mds_ops->rpc_call_done(task, data);
 467 }
 468
 469 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
 470 {
 471         struct nfs_commit_data *cdata = data;
 472
 473         rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
 474 }
 475
 476 static void filelayout_commit_release(void *calldata)
 477 {
 478         struct nfs_commit_data *data = calldata;
 479
 480         data->completion_ops->completion(data);
 481         pnfs_put_lseg(data->lseg);
 482         nfs_put_client(data->ds_clp);
 483         nfs_commitdata_release(data);
 484 }
 485
 486 static const struct rpc_call_ops filelayout_read_call_ops = {
 487         .rpc_call_prepare = filelayout_read_prepare,
 488         .rpc_call_done = filelayout_read_call_done,
 489         .rpc_count_stats = filelayout_read_count_stats,
 490         .rpc_release = filelayout_read_release,
 491 };
 492
 493 static const struct rpc_call_ops filelayout_write_call_ops = {
 494         .rpc_call_prepare = filelayout_write_prepare,
 495         .rpc_call_done = filelayout_write_call_done,
 496         .rpc_count_stats = filelayout_write_count_stats,
 497         .rpc_release = filelayout_write_release,
 498 };
 499
 500 static const struct rpc_call_ops filelayout_commit_call_ops = {
 501         .rpc_call_prepare = filelayout_commit_prepare,
 502         .rpc_call_done = filelayout_write_commit_done,
 503         .rpc_count_stats = filelayout_commit_count_stats,
 504         .rpc_release = filelayout_commit_release,
 505 };
 506
 507 static enum pnfs_try_status
 508 filelayout_read_pagelist(struct nfs_read_data *data)
 509 {
 510         struct nfs_pgio_header *hdr = data->header;
 511         struct pnfs_layout_segment *lseg = hdr->lseg;
 512         struct nfs4_pnfs_ds *ds;
 513         loff_t offset = data->args.offset;
 514         u32 j, idx;
 515         struct nfs_fh *fh;
 516
 517         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 518                 __func__, hdr->inode->i_ino,
 519                 data->args.pgbase, (size_t)data->args.count, offset);
 520
 521         /* Retrieve the correct rpc_client for the byte range */
 522         j = nfs4_fl_calc_j_index(lseg, offset);
 523         idx = nfs4_fl_calc_ds_index(lseg, j);
 524         ds = nfs4_fl_prepare_ds(lseg, idx);
 525         if (!ds)
 526                 return PNFS_NOT_ATTEMPTED;
 527         dprintk("%s USE DS: %s cl_count %d\n", __func__,
 528                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 529
 530         /* No multipath support. Use first DS */
 531         atomic_inc(&ds->ds_clp->cl_count);
 532         data->ds_clp = ds->ds_clp;
 533         fh = nfs4_fl_select_ds_fh(lseg, j);
 534         if (fh)
 535                 data->args.fh = fh;
 536
 537         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 538         data->mds_offset = offset;
 539
 540         /* Perform an asynchronous read to ds */
 541         nfs_initiate_read(ds->ds_clp->cl_rpcclient, data,
 542                                   &filelayout_read_call_ops, RPC_TASK_SOFTCONN);
 543         return PNFS_ATTEMPTED;
 544 }
 545
 546 /* Perform async writes. */
 547 static enum pnfs_try_status
 548 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
 549 {
 550         struct nfs_pgio_header *hdr = data->header;
 551         struct pnfs_layout_segment *lseg = hdr->lseg;
 552         struct nfs4_pnfs_ds *ds;
 553         loff_t offset = data->args.offset;
 554         u32 j, idx;
 555         struct nfs_fh *fh;
 556
 557         /* Retrieve the correct rpc_client for the byte range */
 558         j = nfs4_fl_calc_j_index(lseg, offset);
 559         idx = nfs4_fl_calc_ds_index(lseg, j);
 560         ds = nfs4_fl_prepare_ds(lseg, idx);
 561         if (!ds)
 562                 return PNFS_NOT_ATTEMPTED;
 563         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
 564                 __func__, hdr->inode->i_ino, sync, (size_t) data->args.count,
 565                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 566
 567         data->write_done_cb = filelayout_write_done_cb;
 568         atomic_inc(&ds->ds_clp->cl_count);
 569         data->ds_clp = ds->ds_clp;
 570         fh = nfs4_fl_select_ds_fh(lseg, j);
 571         if (fh)
 572                 data->args.fh = fh;
 573         /*
 574          * Get the file offset on the dserver. Set the write offset to
 575          * this offset and save the original offset.
 576          */
 577         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 578
 579         /* Perform an asynchronous write */
 580         nfs_initiate_write(ds->ds_clp->cl_rpcclient, data,
 581                                     &filelayout_write_call_ops, sync,
 582                                     RPC_TASK_SOFTCONN);
 583         return PNFS_ATTEMPTED;
 584 }
 585
 586 /*
 587  * filelayout_check_layout()
 588  *
 589  * Make sure layout segment parameters are sane WRT the device.
 590  * At this point no generic layer initialization of the lseg has occurred,
 591  * and nothing has been added to the layout_hdr cache.
 592  *
 593  */
 594 static int
 595 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 596                         struct nfs4_filelayout_segment *fl,
 597                         struct nfs4_layoutget_res *lgr,
 598                         struct nfs4_deviceid *id,
 599                         gfp_t gfp_flags)
 600 {
 601         struct nfs4_deviceid_node *d;
 602         struct nfs4_file_layout_dsaddr *dsaddr;
 603         int status = -EINVAL;
 604         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 605
 606         dprintk("--> %s\n", __func__);
 607
 608         /* FIXME: remove this check when layout segment support is added */
 609         if (lgr->range.offset != 0 ||
 610             lgr->range.length != NFS4_MAX_UINT64) {
 611                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 612                         __func__);
 613                 goto out;
 614         }
 615
 616         if (fl->pattern_offset > lgr->range.offset) {
 617                 dprintk("%s pattern_offset %lld too large\n",
 618                                 __func__, fl->pattern_offset);
 619                 goto out;
 620         }
 621
 622         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 623                 dprintk("%s Invalid stripe unit (%u)\n",
 624                         __func__, fl->stripe_unit);
 625                 goto out;
 626         }
 627
 628         /* find and reference the deviceid */
 629         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
 630                                    NFS_SERVER(lo->plh_inode)->nfs_client, id);
 631         if (d == NULL) {
 632                 dsaddr = filelayout_get_device_info(lo->plh_inode, id, gfp_flags);
 633                 if (dsaddr == NULL)
 634                         goto out;
 635         } else
 636                 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 637         /* Found deviceid is unavailable */
 638         if (filelayout_test_devid_unavailable(&dsaddr->id_node))
 639                         goto out_put;
 640
 641         fl->dsaddr = dsaddr;
 642
 643         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 644                 dprintk("%s Bad first_stripe_index %u\n",
 645                                 __func__, fl->first_stripe_index);
 646                 goto out_put;
 647         }
 648
 649         if ((fl->stripe_type == STRIPE_SPARSE &&
 650             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 651             (fl->stripe_type == STRIPE_DENSE &&
 652             fl->num_fh != dsaddr->stripe_count)) {
 653                 dprintk("%s num_fh %u not valid for given packing\n",
 654                         __func__, fl->num_fh);
 655                 goto out_put;
 656         }
 657
 658         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
 659                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
 660                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
 661                         nfss->wsize);
 662         }
 663
 664         status = 0;
 665 out:
 666         dprintk("--> %s returns %d\n", __func__, status);
 667         return status;
 668 out_put:
 669         nfs4_fl_put_deviceid(dsaddr);
 670         goto out;
 671 }
 672
 673 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 674 {
 675         int i;
 676
 677         for (i = 0; i < fl->num_fh; i++) {
 678                 if (!fl->fh_array[i])
 679                         break;
 680                 kfree(fl->fh_array[i]);
 681         }
 682         kfree(fl->fh_array);
 683         fl->fh_array = NULL;
 684 }
 685
 686 static void
 687 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 688 {
 689         filelayout_free_fh_array(fl);
 690         kfree(fl);
 691 }
 692
 693 static int
 694 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 695                          struct nfs4_filelayout_segment *fl,
 696                          struct nfs4_layoutget_res *lgr,
 697                          struct nfs4_deviceid *id,
 698                          gfp_t gfp_flags)
 699 {
 700         struct xdr_stream stream;
 701         struct xdr_buf buf;
 702         struct page *scratch;
 703         __be32 *p;
 704         uint32_t nfl_util;
 705         int i;
 706
 707         dprintk("%s: set_layout_map Begin\n", __func__);
 708
 709         scratch = alloc_page(gfp_flags);
 710         if (!scratch)
 711                 return -ENOMEM;
 712
 713         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 714         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 715
 716         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 717          * num_fh (4) */
 718         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 719         if (unlikely(!p))
 720                 goto out_err;
 721
 722         memcpy(id, p, sizeof(*id));
 723         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 724         nfs4_print_deviceid(id);
 725
 726         nfl_util = be32_to_cpup(p++);
 727         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 728                 fl->commit_through_mds = 1;
 729         if (nfl_util & NFL4_UFLG_DENSE)
 730                 fl->stripe_type = STRIPE_DENSE;
 731         else
 732                 fl->stripe_type = STRIPE_SPARSE;
 733         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 734
 735         fl->first_stripe_index = be32_to_cpup(p++);
 736         p = xdr_decode_hyper(p, &fl->pattern_offset);
 737         fl->num_fh = be32_to_cpup(p++);
 738
 739         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 740                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 741                 fl->pattern_offset);
 742
 743         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 744          * Futher checking is done in filelayout_check_layout */
 745         if (fl->num_fh >
 746             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 747                 goto out_err;
 748
 749         if (fl->num_fh > 0) {
 750                 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
 751                                        gfp_flags);
 752                 if (!fl->fh_array)
 753                         goto out_err;
 754         }
 755
 756         for (i = 0; i < fl->num_fh; i++) {
 757                 /* Do we want to use a mempool here? */
 758                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 759                 if (!fl->fh_array[i])
 760                         goto out_err_free;
 761
 762                 p = xdr_inline_decode(&stream, 4);
 763                 if (unlikely(!p))
 764                         goto out_err_free;
 765                 fl->fh_array[i]->size = be32_to_cpup(p++);
 766                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 767                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
 768                                i, fl->fh_array[i]->size);
 769                         goto out_err_free;
 770                 }
 771
 772                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 773                 if (unlikely(!p))
 774                         goto out_err_free;
 775                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 776                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 777                         fl->fh_array[i]->size);
 778         }
 779
 780         __free_page(scratch);
 781         return 0;
 782
 783 out_err_free:
 784         filelayout_free_fh_array(fl);
 785 out_err:
 786         __free_page(scratch);
 787         return -EIO;
 788 }
 789
 790 static void
 791 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 792 {
 793         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 794
 795         dprintk("--> %s\n", __func__);
 796         nfs4_fl_put_deviceid(fl->dsaddr);
 797         /* This assumes a single RW lseg */
 798         if (lseg->pls_range.iomode == IOMODE_RW) {
 799                 struct nfs4_filelayout *flo;
 800
 801                 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
 802                 flo->commit_info.nbuckets = 0;
 803                 kfree(flo->commit_info.buckets);
 804                 flo->commit_info.buckets = NULL;
 805         }
 806         _filelayout_free_lseg(fl);
 807 }
 808
 809 static int
 810 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 811                              struct nfs_commit_info *cinfo,
 812                              gfp_t gfp_flags)
 813 {
 814         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 815         struct pnfs_commit_bucket *buckets;
 816         int size;
 817
 818         if (fl->commit_through_mds)
 819                 return 0;
 820         if (cinfo->ds->nbuckets != 0) {
 821                 /* This assumes there is only one IOMODE_RW lseg.  What
 822                  * we really want to do is have a layout_hdr level
 823                  * dictionary of <multipath_list4, fh> keys, each
 824                  * associated with a struct list_head, populated by calls
 825                  * to filelayout_write_pagelist().
 826                  * */
 827                 return 0;
 828         }
 829
 830         size = (fl->stripe_type == STRIPE_SPARSE) ?
 831                 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 832
 833         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 834                           gfp_flags);
 835         if (!buckets)
 836                 return -ENOMEM;
 837         else {
 838                 int i;
 839
 840                 spin_lock(cinfo->lock);
 841                 if (cinfo->ds->nbuckets != 0)
 842                         kfree(buckets);
 843                 else {
 844                         cinfo->ds->buckets = buckets;
 845                         cinfo->ds->nbuckets = size;
 846                         for (i = 0; i < size; i++) {
 847                                 INIT_LIST_HEAD(&buckets[i].written);
 848                                 INIT_LIST_HEAD(&buckets[i].committing);
 849                         }
 850                 }
 851                 spin_unlock(cinfo->lock);
 852                 return 0;
 853         }
 854 }
 855
 856 static struct pnfs_layout_segment *
 857 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 858                       struct nfs4_layoutget_res *lgr,
 859                       gfp_t gfp_flags)
 860 {
 861         struct nfs4_filelayout_segment *fl;
 862         int rc;
 863         struct nfs4_deviceid id;
 864
 865         dprintk("--> %s\n", __func__);
 866         fl = kzalloc(sizeof(*fl), gfp_flags);
 867         if (!fl)
 868                 return NULL;
 869
 870         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 871         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 872                 _filelayout_free_lseg(fl);
 873                 return NULL;
 874         }
 875         return &fl->generic_hdr;
 876 }
 877
 878 /*
 879  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 880  *
 881  * return true  : coalesce page
 882  * return false : don't coalesce page
 883  */
 884 static bool
 885 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 886                    struct nfs_page *req)
 887 {
 888         u64 p_stripe, r_stripe;
 889         u32 stripe_unit;
 890
 891         if (!pnfs_generic_pg_test(pgio, prev, req) ||
 892             !nfs_generic_pg_test(pgio, prev, req))
 893                 return false;
 894
 895         p_stripe = (u64)req_offset(prev);
 896         r_stripe = (u64)req_offset(req);
 897         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 898
 899         do_div(p_stripe, stripe_unit);
 900         do_div(r_stripe, stripe_unit);
 901
 902         return (p_stripe == r_stripe);
 903 }
 904
 905 static void
 906 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 907                         struct nfs_page *req)
 908 {
 909         WARN_ON_ONCE(pgio->pg_lseg != NULL);
 910
 911         if (req->wb_offset != req->wb_pgbase) {
 912                 /*
 913                  * Handling unaligned pages is difficult, because have to
 914                  * somehow split a req in two in certain cases in the
 915                  * pg.test code.  Avoid this by just not using pnfs
 916                  * in this case.
 917                  */
 918                 nfs_pageio_reset_read_mds(pgio);
 919                 return;
 920         }
 921         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 922                                            req->wb_context,
 923                                            0,
 924                                            NFS4_MAX_UINT64,
 925                                            IOMODE_READ,
 926                                            GFP_KERNEL);
 927         /* If no lseg, fall back to read through mds */
 928         if (pgio->pg_lseg == NULL)
 929                 nfs_pageio_reset_read_mds(pgio);
 930 }
 931
 932 static void
 933 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 934                          struct nfs_page *req)
 935 {
 936         struct nfs_commit_info cinfo;
 937         int status;
 938
 939         WARN_ON_ONCE(pgio->pg_lseg != NULL);
 940
 941         if (req->wb_offset != req->wb_pgbase)
 942                 goto out_mds;
 943         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 944                                            req->wb_context,
 945                                            0,
 946                                            NFS4_MAX_UINT64,
 947                                            IOMODE_RW,
 948                                            GFP_NOFS);
 949         /* If no lseg, fall back to write through mds */
 950         if (pgio->pg_lseg == NULL)
 951                 goto out_mds;
 952         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 953         status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 954         if (status < 0) {
 955                 pnfs_put_lseg(pgio->pg_lseg);
 956                 pgio->pg_lseg = NULL;
 957                 goto out_mds;
 958         }
 959         return;
 960 out_mds:
 961         nfs_pageio_reset_write_mds(pgio);
 962 }
 963
 964 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 965         .pg_init = filelayout_pg_init_read,
 966         .pg_test = filelayout_pg_test,
 967         .pg_doio = pnfs_generic_pg_readpages,
 968 };
 969
 970 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 971         .pg_init = filelayout_pg_init_write,
 972         .pg_test = filelayout_pg_test,
 973         .pg_doio = pnfs_generic_pg_writepages,
 974 };
 975
 976 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 977 {
 978         if (fl->stripe_type == STRIPE_SPARSE)
 979                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
 980         else
 981                 return j;
 982 }
 983
 984 /* The generic layer is about to remove the req from the commit list.
 985  * If this will make the bucket empty, it will need to put the lseg reference.
 986  */
 987 static void
 988 filelayout_clear_request_commit(struct nfs_page *req,
 989                                 struct nfs_commit_info *cinfo)
 990 {
 991         struct pnfs_layout_segment *freeme = NULL;
 992
 993         spin_lock(cinfo->lock);
 994         if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
 995                 goto out;
 996         cinfo->ds->nwritten--;
 997         if (list_is_singular(&req->wb_list)) {
 998                 struct pnfs_commit_bucket *bucket;
 999
1000                 bucket = list_first_entry(&req->wb_list,
1001                                           struct pnfs_commit_bucket,
1002                                           written);
1003                 freeme = bucket->wlseg;
1004                 bucket->wlseg = NULL;
1005         }
1006 out:
1007         nfs_request_remove_commit_list(req, cinfo);
1008         spin_unlock(cinfo->lock);
1009         pnfs_put_lseg(freeme);
1010 }
1011
1012 static struct list_head *
1013 filelayout_choose_commit_list(struct nfs_page *req,
1014                               struct pnfs_layout_segment *lseg,
1015                               struct nfs_commit_info *cinfo)
1016 {
1017         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1018         u32 i, j;
1019         struct list_head *list;
1020         struct pnfs_commit_bucket *buckets;
1021
1022         if (fl->commit_through_mds)
1023                 return &cinfo->mds->list;
1024
1025         /* Note that we are calling nfs4_fl_calc_j_index on each page
1026          * that ends up being committed to a data server.  An attractive
1027          * alternative is to add a field to nfs_write_data and nfs_page
1028          * to store the value calculated in filelayout_write_pagelist
1029          * and just use that here.
1030          */
1031         j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1032         i = select_bucket_index(fl, j);
1033         buckets = cinfo->ds->buckets;
1034         list = &buckets[i].written;
1035         if (list_empty(list)) {
1036                 /* Non-empty buckets hold a reference on the lseg.  That ref
1037                  * is normally transferred to the COMMIT call and released
1038                  * there.  It could also be released if the last req is pulled
1039                  * off due to a rewrite, in which case it will be done in
1040                  * filelayout_clear_request_commit
1041                  */
1042                 buckets[i].wlseg = pnfs_get_lseg(lseg);
1043         }
1044         set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1045         cinfo->ds->nwritten++;
1046         return list;
1047 }
1048
1049 static void
1050 filelayout_mark_request_commit(struct nfs_page *req,
1051                                struct pnfs_layout_segment *lseg,
1052                                struct nfs_commit_info *cinfo)
1053 {
1054         struct list_head *list;
1055
1056         list = filelayout_choose_commit_list(req, lseg, cinfo);
1057         nfs_request_add_commit_list(req, list, cinfo);
1058 }
1059
1060 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1061 {
1062         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1063
1064         if (flseg->stripe_type == STRIPE_SPARSE)
1065                 return i;
1066         else
1067                 return nfs4_fl_calc_ds_index(lseg, i);
1068 }
1069
1070 static struct nfs_fh *
1071 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1072 {
1073         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1074
1075         if (flseg->stripe_type == STRIPE_SPARSE) {
1076                 if (flseg->num_fh == 1)
1077                         i = 0;
1078                 else if (flseg->num_fh == 0)
1079                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1080                         return NULL;
1081         }
1082         return flseg->fh_array[i];
1083 }
1084
1085 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1086 {
1087         struct pnfs_layout_segment *lseg = data->lseg;
1088         struct nfs4_pnfs_ds *ds;
1089         u32 idx;
1090         struct nfs_fh *fh;
1091
1092         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1093         ds = nfs4_fl_prepare_ds(lseg, idx);
1094         if (!ds) {
1095                 prepare_to_resend_writes(data);
1096                 filelayout_commit_release(data);
1097                 return -EAGAIN;
1098         }
1099         dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1100                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1101         data->commit_done_cb = filelayout_commit_done_cb;
1102         atomic_inc(&ds->ds_clp->cl_count);
1103         data->ds_clp = ds->ds_clp;
1104         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1105         if (fh)
1106                 data->args.fh = fh;
1107         return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data,
1108                                    &filelayout_commit_call_ops, how,
1109                                    RPC_TASK_SOFTCONN);
1110 }
1111
1112 static int
1113 transfer_commit_list(struct list_head *src, struct list_head *dst,
1114                      struct nfs_commit_info *cinfo, int max)
1115 {
1116         struct nfs_page *req, *tmp;
1117         int ret = 0;
1118
1119         list_for_each_entry_safe(req, tmp, src, wb_list) {
1120                 if (!nfs_lock_request(req))
1121                         continue;
1122                 kref_get(&req->wb_kref);
1123                 if (cond_resched_lock(cinfo->lock))
1124                         list_safe_reset_next(req, tmp, wb_list);
1125                 nfs_request_remove_commit_list(req, cinfo);
1126                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1127                 nfs_list_add_request(req, dst);
1128                 ret++;
1129                 if ((ret == max) && !cinfo->dreq)
1130                         break;
1131         }
1132         return ret;
1133 }
1134
1135 static int
1136 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1137                                struct nfs_commit_info *cinfo,
1138                                int max)
1139 {
1140         struct list_head *src = &bucket->written;
1141         struct list_head *dst = &bucket->committing;
1142         int ret;
1143
1144         ret = transfer_commit_list(src, dst, cinfo, max);
1145         if (ret) {
1146                 cinfo->ds->nwritten -= ret;
1147                 cinfo->ds->ncommitting += ret;
1148                 bucket->clseg = bucket->wlseg;
1149                 if (list_empty(src))
1150                         bucket->wlseg = NULL;
1151                 else
1152                         pnfs_get_lseg(bucket->clseg);
1153         }
1154         return ret;
1155 }
1156
1157 /* Move reqs from written to committing lists, returning count of number moved.
1158  * Note called with cinfo->lock held.
1159  */
1160 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1161                                         int max)
1162 {
1163         int i, rv = 0, cnt;
1164
1165         for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1166                 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1167                                                      cinfo, max);
1168                 max -= cnt;
1169                 rv += cnt;
1170         }
1171         return rv;
1172 }
1173
1174 /* Pull everything off the committing lists and dump into @dst */
1175 static void filelayout_recover_commit_reqs(struct list_head *dst,
1176                                            struct nfs_commit_info *cinfo)
1177 {
1178         struct pnfs_commit_bucket *b;
1179         int i;
1180
1181         /* NOTE cinfo->lock is NOT held, relying on fact that this is
1182          * only called on single thread per dreq.
1183          * Can't take the lock because need to do pnfs_put_lseg
1184          */
1185         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1186                 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1187                         pnfs_put_lseg(b->wlseg);
1188                         b->wlseg = NULL;
1189                 }
1190         }
1191         cinfo->ds->nwritten = 0;
1192 }
1193
1194 static unsigned int
1195 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1196 {
1197         struct pnfs_ds_commit_info *fl_cinfo;
1198         struct pnfs_commit_bucket *bucket;
1199         struct nfs_commit_data *data;
1200         int i, j;
1201         unsigned int nreq = 0;
1202
1203         fl_cinfo = cinfo->ds;
1204         bucket = fl_cinfo->buckets;
1205         for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1206                 if (list_empty(&bucket->committing))
1207                         continue;
1208                 data = nfs_commitdata_alloc();
1209                 if (!data)
1210                         break;
1211                 data->ds_commit_index = i;
1212                 data->lseg = bucket->clseg;
1213                 bucket->clseg = NULL;
1214                 list_add(&data->pages, list);
1215                 nreq++;
1216         }
1217
1218         /* Clean up on error */
1219         for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1220                 if (list_empty(&bucket->committing))
1221                         continue;
1222                 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1223                 pnfs_put_lseg(bucket->clseg);
1224                 bucket->clseg = NULL;
1225         }
1226         /* Caller will clean up entries put on list */
1227         return nreq;
1228 }
1229
1230 /* This follows nfs_commit_list pretty closely */
1231 static int
1232 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1233                            int how, struct nfs_commit_info *cinfo)
1234 {
1235         struct nfs_commit_data *data, *tmp;
1236         LIST_HEAD(list);
1237         unsigned int nreq = 0;
1238
1239         if (!list_empty(mds_pages)) {
1240                 data = nfs_commitdata_alloc();
1241                 if (data != NULL) {
1242                         data->lseg = NULL;
1243                         list_add(&data->pages, &list);
1244                         nreq++;
1245                 } else
1246                         nfs_retry_commit(mds_pages, NULL, cinfo);
1247         }
1248
1249         nreq += alloc_ds_commits(cinfo, &list);
1250
1251         if (nreq == 0) {
1252                 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1253                 goto out;
1254         }
1255
1256         atomic_add(nreq, &cinfo->mds->rpcs_out);
1257
1258         list_for_each_entry_safe(data, tmp, &list, pages) {
1259                 list_del_init(&data->pages);
1260                 if (!data->lseg) {
1261                         nfs_init_commit(data, mds_pages, NULL, cinfo);
1262                         nfs_initiate_commit(NFS_CLIENT(inode), data,
1263                                             data->mds_ops, how, 0);
1264                 } else {
1265                         struct pnfs_commit_bucket *buckets;
1266
1267                         buckets = cinfo->ds->buckets;
1268                         nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1269                         filelayout_initiate_commit(data, how);
1270                 }
1271         }
1272 out:
1273         cinfo->ds->ncommitting = 0;
1274         return PNFS_ATTEMPTED;
1275 }
1276
1277 static void
1278 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1279 {
1280         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1281 }
1282
1283 static struct pnfs_layout_hdr *
1284 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1285 {
1286         struct nfs4_filelayout *flo;
1287
1288         flo = kzalloc(sizeof(*flo), gfp_flags);
1289         return &flo->generic_hdr;
1290 }
1291
1292 static void
1293 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1294 {
1295         kfree(FILELAYOUT_FROM_HDR(lo));
1296 }
1297
1298 static struct pnfs_ds_commit_info *
1299 filelayout_get_ds_info(struct inode *inode)
1300 {
1301         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1302
1303         if (layout == NULL)
1304                 return NULL;
1305         else
1306                 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1307 }
1308
1309 static struct pnfs_layoutdriver_type filelayout_type = {
1310         .id                     = LAYOUT_NFSV4_1_FILES,
1311         .name                   = "LAYOUT_NFSV4_1_FILES",
1312         .owner                  = THIS_MODULE,
1313         .alloc_layout_hdr       = filelayout_alloc_layout_hdr,
1314         .free_layout_hdr        = filelayout_free_layout_hdr,
1315         .alloc_lseg             = filelayout_alloc_lseg,
1316         .free_lseg              = filelayout_free_lseg,
1317         .pg_read_ops            = &filelayout_pg_read_ops,
1318         .pg_write_ops           = &filelayout_pg_write_ops,
1319         .get_ds_info            = &filelayout_get_ds_info,
1320         .mark_request_commit    = filelayout_mark_request_commit,
1321         .clear_request_commit   = filelayout_clear_request_commit,
1322         .scan_commit_lists      = filelayout_scan_commit_lists,
1323         .recover_commit_reqs    = filelayout_recover_commit_reqs,
1324         .commit_pagelist        = filelayout_commit_pagelist,
1325         .read_pagelist          = filelayout_read_pagelist,
1326         .write_pagelist         = filelayout_write_pagelist,
1327         .free_deviceid_node     = filelayout_free_deveiceid_node,
1328 };
1329
1330 static int __init nfs4filelayout_init(void)
1331 {
1332         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1333                __func__);
1334         return pnfs_register_layoutdriver(&filelayout_type);
1335 }
1336
1337 static void __exit nfs4filelayout_exit(void)
1338 {
1339         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1340                __func__);
1341         pnfs_unregister_layoutdriver(&filelayout_type);
1342 }
1343
1344 MODULE_ALIAS("nfs-layouttype4-1");
1345
1346 module_init(nfs4filelayout_init);
1347 module_exit(nfs4filelayout_exit);