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 "delegation.h"
  40 #include "nfs4filelayout.h"
  41
  42 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  43
  44 MODULE_LICENSE("GPL");
  45 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
  46 MODULE_DESCRIPTION("The NFSv4 file layout driver");
  47
  48 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
  49
  50 static loff_t
  51 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
  52                             loff_t offset)
  53 {
  54         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
  55         u64 stripe_no;
  56         u32 rem;
  57
  58         offset -= flseg->pattern_offset;
  59         stripe_no = div_u64(offset, stripe_width);
  60         div_u64_rem(offset, flseg->stripe_unit, &rem);
  61
  62         return stripe_no * flseg->stripe_unit + rem;
  63 }
  64
  65 /* This function is used by the layout driver to calculate the
  66  * offset of the file on the dserver based on whether the
  67  * layout type is STRIPE_DENSE or STRIPE_SPARSE
  68  */
  69 static loff_t
  70 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
  71 {
  72         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
  73
  74         switch (flseg->stripe_type) {
  75         case STRIPE_SPARSE:
  76                 return offset;
  77
  78         case STRIPE_DENSE:
  79                 return filelayout_get_dense_offset(flseg, offset);
  80         }
  81
  82         BUG();
  83 }
  84
  85 static void filelayout_reset_write(struct nfs_write_data *data)
  86 {
  87         struct nfs_pgio_header *hdr = data->header;
  88         struct rpc_task *task = &data->task;
  89
  90         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
  91                 dprintk("%s Reset task %5u for i/o through MDS "
  92                         "(req %s/%lld, %u bytes @ offset %llu)\n", __func__,
  93                         data->task.tk_pid,
  94                         hdr->inode->i_sb->s_id,
  95                         (long long)NFS_FILEID(hdr->inode),
  96                         data->args.count,
  97                         (unsigned long long)data->args.offset);
  98
  99                 task->tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
 100                                                         &hdr->pages,
 101                                                         hdr->completion_ops);
 102         }
 103 }
 104
 105 static void filelayout_reset_read(struct nfs_read_data *data)
 106 {
 107         struct nfs_pgio_header *hdr = data->header;
 108         struct rpc_task *task = &data->task;
 109
 110         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 111                 dprintk("%s Reset task %5u for i/o through MDS "
 112                         "(req %s/%lld, %u bytes @ offset %llu)\n", __func__,
 113                         data->task.tk_pid,
 114                         hdr->inode->i_sb->s_id,
 115                         (long long)NFS_FILEID(hdr->inode),
 116                         data->args.count,
 117                         (unsigned long long)data->args.offset);
 118
 119                 task->tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
 120                                                         &hdr->pages,
 121                                                         hdr->completion_ops);
 122         }
 123 }
 124
 125 static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo)
 126 {
 127         if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 128                 return;
 129         clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags);
 130         pnfs_return_layout(inode);
 131 }
 132
 133 static int filelayout_async_handle_error(struct rpc_task *task,
 134                                          struct nfs4_state *state,
 135                                          struct nfs_client *clp,
 136                                          struct pnfs_layout_segment *lseg)
 137 {
 138         struct pnfs_layout_hdr *lo = lseg->pls_layout;
 139         struct inode *inode = lo->plh_inode;
 140         struct nfs_server *mds_server = NFS_SERVER(inode);
 141         struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
 142         struct nfs_client *mds_client = mds_server->nfs_client;
 143         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
 144
 145         if (task->tk_status >= 0)
 146                 return 0;
 147
 148         switch (task->tk_status) {
 149         /* MDS state errors */
 150         case -NFS4ERR_DELEG_REVOKED:
 151         case -NFS4ERR_ADMIN_REVOKED:
 152         case -NFS4ERR_BAD_STATEID:
 153                 if (state == NULL)
 154                         break;
 155                 nfs_remove_bad_delegation(state->inode);
 156         case -NFS4ERR_OPENMODE:
 157                 if (state == NULL)
 158                         break;
 159                 nfs4_schedule_stateid_recovery(mds_server, state);
 160                 goto wait_on_recovery;
 161         case -NFS4ERR_EXPIRED:
 162                 if (state != NULL)
 163                         nfs4_schedule_stateid_recovery(mds_server, state);
 164                 nfs4_schedule_lease_recovery(mds_client);
 165                 goto wait_on_recovery;
 166         /* DS session errors */
 167         case -NFS4ERR_BADSESSION:
 168         case -NFS4ERR_BADSLOT:
 169         case -NFS4ERR_BAD_HIGH_SLOT:
 170         case -NFS4ERR_DEADSESSION:
 171         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
 172         case -NFS4ERR_SEQ_FALSE_RETRY:
 173         case -NFS4ERR_SEQ_MISORDERED:
 174                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 175                         "flags 0x%x\n", __func__, task->tk_status,
 176                         clp->cl_exchange_flags);
 177                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
 178                 break;
 179         case -NFS4ERR_DELAY:
 180         case -NFS4ERR_GRACE:
 181         case -EKEYEXPIRED:
 182                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 183                 break;
 184         case -NFS4ERR_RETRY_UNCACHED_REP:
 185                 break;
 186         /* Invalidate Layout errors */
 187         case -NFS4ERR_PNFS_NO_LAYOUT:
 188         case -ESTALE:           /* mapped NFS4ERR_STALE */
 189         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
 190         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
 191         case -NFS4ERR_FHEXPIRED:
 192         case -NFS4ERR_WRONG_TYPE:
 193                 dprintk("%s Invalid layout error %d\n", __func__,
 194                         task->tk_status);
 195                 /*
 196                  * Destroy layout so new i/o will get a new layout.
 197                  * Layout will not be destroyed until all current lseg
 198                  * references are put. Mark layout as invalid to resend failed
 199                  * i/o and all i/o waiting on the slot table to the MDS until
 200                  * layout is destroyed and a new valid layout is obtained.
 201                  */
 202                 pnfs_destroy_layout(NFS_I(inode));
 203                 rpc_wake_up(&tbl->slot_tbl_waitq);
 204                 goto reset;
 205         /* RPC connection errors */
 206         case -ECONNREFUSED:
 207         case -EHOSTDOWN:
 208         case -EHOSTUNREACH:
 209         case -ENETUNREACH:
 210         case -EIO:
 211         case -ETIMEDOUT:
 212         case -EPIPE:
 213                 dprintk("%s DS connection error %d\n", __func__,
 214                         task->tk_status);
 215                 nfs4_mark_deviceid_unavailable(devid);
 216                 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
 217                 rpc_wake_up(&tbl->slot_tbl_waitq);
 218                 /* fall through */
 219         default:
 220 reset:
 221                 dprintk("%s Retry through MDS. Error %d\n", __func__,
 222                         task->tk_status);
 223                 return -NFS4ERR_RESET_TO_MDS;
 224         }
 225 out:
 226         task->tk_status = 0;
 227         return -EAGAIN;
 228 wait_on_recovery:
 229         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
 230         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
 231                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
 232         goto out;
 233 }
 234
 235 /* NFS_PROTO call done callback routines */
 236
 237 static int filelayout_read_done_cb(struct rpc_task *task,
 238                                 struct nfs_read_data *data)
 239 {
 240         struct nfs_pgio_header *hdr = data->header;
 241         int err;
 242
 243         err = filelayout_async_handle_error(task, data->args.context->state,
 244                                             data->ds_clp, hdr->lseg);
 245
 246         switch (err) {
 247         case -NFS4ERR_RESET_TO_MDS:
 248                 filelayout_reset_read(data);
 249                 return task->tk_status;
 250         case -EAGAIN:
 251                 rpc_restart_call_prepare(task);
 252                 return -EAGAIN;
 253         }
 254
 255         return 0;
 256 }
 257
 258 /*
 259  * We reference the rpc_cred of the first WRITE that triggers the need for
 260  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 261  * rfc5661 is not clear about which credential should be used.
 262  */
 263 static void
 264 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
 265 {
 266         struct nfs_pgio_header *hdr = wdata->header;
 267
 268         if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
 269             wdata->res.verf->committed == NFS_FILE_SYNC)
 270                 return;
 271
 272         pnfs_set_layoutcommit(wdata);
 273         dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
 274                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
 275 }
 276
 277 bool
 278 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
 279 {
 280         return filelayout_test_devid_invalid(node) ||
 281                 nfs4_test_deviceid_unavailable(node);
 282 }
 283
 284 static bool
 285 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
 286 {
 287         struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
 288
 289         return filelayout_test_devid_unavailable(node);
 290 }
 291
 292 /*
 293  * Call ops for the async read/write cases
 294  * In the case of dense layouts, the offset needs to be reset to its
 295  * original value.
 296  */
 297 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 298 {
 299         struct nfs_read_data *rdata = data;
 300
 301         if (filelayout_reset_to_mds(rdata->header->lseg)) {
 302                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 303                 filelayout_reset_read(rdata);
 304                 rpc_exit(task, 0);
 305                 return;
 306         }
 307         rdata->read_done_cb = filelayout_read_done_cb;
 308
 309         if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
 310                                 &rdata->args.seq_args, &rdata->res.seq_res,
 311                                 task))
 312                 return;
 313
 314         rpc_call_start(task);
 315 }
 316
 317 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 318 {
 319         struct nfs_read_data *rdata = data;
 320
 321         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 322
 323         if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) &&
 324             task->tk_status == 0)
 325                 return;
 326
 327         /* Note this may cause RPC to be resent */
 328         rdata->header->mds_ops->rpc_call_done(task, data);
 329 }
 330
 331 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
 332 {
 333         struct nfs_read_data *rdata = data;
 334
 335         rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
 336 }
 337
 338 static void filelayout_read_release(void *data)
 339 {
 340         struct nfs_read_data *rdata = data;
 341         struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout;
 342
 343         filelayout_fenceme(lo->plh_inode, lo);
 344         nfs_put_client(rdata->ds_clp);
 345         rdata->header->mds_ops->rpc_release(data);
 346 }
 347
 348 static int filelayout_write_done_cb(struct rpc_task *task,
 349                                 struct nfs_write_data *data)
 350 {
 351         struct nfs_pgio_header *hdr = data->header;
 352         int err;
 353
 354         err = filelayout_async_handle_error(task, data->args.context->state,
 355                                             data->ds_clp, hdr->lseg);
 356
 357         switch (err) {
 358         case -NFS4ERR_RESET_TO_MDS:
 359                 filelayout_reset_write(data);
 360                 return task->tk_status;
 361         case -EAGAIN:
 362                 rpc_restart_call_prepare(task);
 363                 return -EAGAIN;
 364         }
 365
 366         filelayout_set_layoutcommit(data);
 367         return 0;
 368 }
 369
 370 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 371 static void prepare_to_resend_writes(struct nfs_commit_data *data)
 372 {
 373         struct nfs_page *first = nfs_list_entry(data->pages.next);
 374
 375         data->task.tk_status = 0;
 376         memcpy(&data->verf.verifier, &first->wb_verf,
 377                sizeof(data->verf.verifier));
 378         data->verf.verifier.data[0]++; /* ensure verifier mismatch */
 379 }
 380
 381 static int filelayout_commit_done_cb(struct rpc_task *task,
 382                                      struct nfs_commit_data *data)
 383 {
 384         int err;
 385
 386         err = filelayout_async_handle_error(task, NULL, data->ds_clp,
 387                                             data->lseg);
 388
 389         switch (err) {
 390         case -NFS4ERR_RESET_TO_MDS:
 391                 prepare_to_resend_writes(data);
 392                 return -EAGAIN;
 393         case -EAGAIN:
 394                 rpc_restart_call_prepare(task);
 395                 return -EAGAIN;
 396         }
 397
 398         return 0;
 399 }
 400
 401 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 402 {
 403         struct nfs_write_data *wdata = data;
 404
 405         if (filelayout_reset_to_mds(wdata->header->lseg)) {
 406                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 407                 filelayout_reset_write(wdata);
 408                 rpc_exit(task, 0);
 409                 return;
 410         }
 411         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 412                                 &wdata->args.seq_args, &wdata->res.seq_res,
 413                                 task))
 414                 return;
 415
 416         rpc_call_start(task);
 417 }
 418
 419 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 420 {
 421         struct nfs_write_data *wdata = data;
 422
 423         if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) &&
 424             task->tk_status == 0)
 425                 return;
 426
 427         /* Note this may cause RPC to be resent */
 428         wdata->header->mds_ops->rpc_call_done(task, data);
 429 }
 430
 431 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
 432 {
 433         struct nfs_write_data *wdata = data;
 434
 435         rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
 436 }
 437
 438 static void filelayout_write_release(void *data)
 439 {
 440         struct nfs_write_data *wdata = data;
 441         struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout;
 442
 443         filelayout_fenceme(lo->plh_inode, lo);
 444         nfs_put_client(wdata->ds_clp);
 445         wdata->header->mds_ops->rpc_release(data);
 446 }
 447
 448 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
 449 {
 450         struct nfs_commit_data *wdata = data;
 451
 452         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 453                                 &wdata->args.seq_args, &wdata->res.seq_res,
 454                                 task))
 455                 return;
 456
 457         rpc_call_start(task);
 458 }
 459
 460 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
 461 {
 462         struct nfs_commit_data *wdata = data;
 463
 464         /* Note this may cause RPC to be resent */
 465         wdata->mds_ops->rpc_call_done(task, data);
 466 }
 467
 468 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
 469 {
 470         struct nfs_commit_data *cdata = data;
 471
 472         rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
 473 }
 474
 475 static void filelayout_commit_release(void *calldata)
 476 {
 477         struct nfs_commit_data *data = calldata;
 478
 479         data->completion_ops->completion(data);
 480         pnfs_put_lseg(data->lseg);
 481         nfs_put_client(data->ds_clp);
 482         nfs_commitdata_release(data);
 483 }
 484
 485 static const struct rpc_call_ops filelayout_read_call_ops = {
 486         .rpc_call_prepare = filelayout_read_prepare,
 487         .rpc_call_done = filelayout_read_call_done,
 488         .rpc_count_stats = filelayout_read_count_stats,
 489         .rpc_release = filelayout_read_release,
 490 };
 491
 492 static const struct rpc_call_ops filelayout_write_call_ops = {
 493         .rpc_call_prepare = filelayout_write_prepare,
 494         .rpc_call_done = filelayout_write_call_done,
 495         .rpc_count_stats = filelayout_write_count_stats,
 496         .rpc_release = filelayout_write_release,
 497 };
 498
 499 static const struct rpc_call_ops filelayout_commit_call_ops = {
 500         .rpc_call_prepare = filelayout_commit_prepare,
 501         .rpc_call_done = filelayout_write_commit_done,
 502         .rpc_count_stats = filelayout_commit_count_stats,
 503         .rpc_release = filelayout_commit_release,
 504 };
 505
 506 static enum pnfs_try_status
 507 filelayout_read_pagelist(struct nfs_read_data *data)
 508 {
 509         struct nfs_pgio_header *hdr = data->header;
 510         struct pnfs_layout_segment *lseg = hdr->lseg;
 511         struct nfs4_pnfs_ds *ds;
 512         loff_t offset = data->args.offset;
 513         u32 j, idx;
 514         struct nfs_fh *fh;
 515         int status;
 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         status = nfs_initiate_read(ds->ds_clp->cl_rpcclient, data,
 542                                   &filelayout_read_call_ops, RPC_TASK_SOFTCONN);
 543         BUG_ON(status != 0);
 544         return PNFS_ATTEMPTED;
 545 }
 546
 547 /* Perform async writes. */
 548 static enum pnfs_try_status
 549 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
 550 {
 551         struct nfs_pgio_header *hdr = data->header;
 552         struct pnfs_layout_segment *lseg = hdr->lseg;
 553         struct nfs4_pnfs_ds *ds;
 554         loff_t offset = data->args.offset;
 555         u32 j, idx;
 556         struct nfs_fh *fh;
 557         int status;
 558
 559         /* Retrieve the correct rpc_client for the byte range */
 560         j = nfs4_fl_calc_j_index(lseg, offset);
 561         idx = nfs4_fl_calc_ds_index(lseg, j);
 562         ds = nfs4_fl_prepare_ds(lseg, idx);
 563         if (!ds)
 564                 return PNFS_NOT_ATTEMPTED;
 565         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
 566                 __func__, hdr->inode->i_ino, sync, (size_t) data->args.count,
 567                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 568
 569         data->write_done_cb = filelayout_write_done_cb;
 570         atomic_inc(&ds->ds_clp->cl_count);
 571         data->ds_clp = ds->ds_clp;
 572         fh = nfs4_fl_select_ds_fh(lseg, j);
 573         if (fh)
 574                 data->args.fh = fh;
 575         /*
 576          * Get the file offset on the dserver. Set the write offset to
 577          * this offset and save the original offset.
 578          */
 579         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 580
 581         /* Perform an asynchronous write */
 582         status = nfs_initiate_write(ds->ds_clp->cl_rpcclient, data,
 583                                     &filelayout_write_call_ops, sync,
 584                                     RPC_TASK_SOFTCONN);
 585         BUG_ON(status != 0);
 586         return PNFS_ATTEMPTED;
 587 }
 588
 589 /*
 590  * filelayout_check_layout()
 591  *
 592  * Make sure layout segment parameters are sane WRT the device.
 593  * At this point no generic layer initialization of the lseg has occurred,
 594  * and nothing has been added to the layout_hdr cache.
 595  *
 596  */
 597 static int
 598 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 599                         struct nfs4_filelayout_segment *fl,
 600                         struct nfs4_layoutget_res *lgr,
 601                         struct nfs4_deviceid *id,
 602                         gfp_t gfp_flags)
 603 {
 604         struct nfs4_deviceid_node *d;
 605         struct nfs4_file_layout_dsaddr *dsaddr;
 606         int status = -EINVAL;
 607         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 608
 609         dprintk("--> %s\n", __func__);
 610
 611         /* FIXME: remove this check when layout segment support is added */
 612         if (lgr->range.offset != 0 ||
 613             lgr->range.length != NFS4_MAX_UINT64) {
 614                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 615                         __func__);
 616                 goto out;
 617         }
 618
 619         if (fl->pattern_offset > lgr->range.offset) {
 620                 dprintk("%s pattern_offset %lld too large\n",
 621                                 __func__, fl->pattern_offset);
 622                 goto out;
 623         }
 624
 625         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 626                 dprintk("%s Invalid stripe unit (%u)\n",
 627                         __func__, fl->stripe_unit);
 628                 goto out;
 629         }
 630
 631         /* find and reference the deviceid */
 632         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
 633                                    NFS_SERVER(lo->plh_inode)->nfs_client, id);
 634         if (d == NULL) {
 635                 dsaddr = filelayout_get_device_info(lo->plh_inode, id, gfp_flags);
 636                 if (dsaddr == NULL)
 637                         goto out;
 638         } else
 639                 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 640         /* Found deviceid is unavailable */
 641         if (filelayout_test_devid_unavailable(&dsaddr->id_node))
 642                         goto out_put;
 643
 644         fl->dsaddr = dsaddr;
 645
 646         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 647                 dprintk("%s Bad first_stripe_index %u\n",
 648                                 __func__, fl->first_stripe_index);
 649                 goto out_put;
 650         }
 651
 652         if ((fl->stripe_type == STRIPE_SPARSE &&
 653             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 654             (fl->stripe_type == STRIPE_DENSE &&
 655             fl->num_fh != dsaddr->stripe_count)) {
 656                 dprintk("%s num_fh %u not valid for given packing\n",
 657                         __func__, fl->num_fh);
 658                 goto out_put;
 659         }
 660
 661         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
 662                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
 663                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
 664                         nfss->wsize);
 665         }
 666
 667         status = 0;
 668 out:
 669         dprintk("--> %s returns %d\n", __func__, status);
 670         return status;
 671 out_put:
 672         nfs4_fl_put_deviceid(dsaddr);
 673         goto out;
 674 }
 675
 676 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 677 {
 678         int i;
 679
 680         for (i = 0; i < fl->num_fh; i++) {
 681                 if (!fl->fh_array[i])
 682                         break;
 683                 kfree(fl->fh_array[i]);
 684         }
 685         kfree(fl->fh_array);
 686         fl->fh_array = NULL;
 687 }
 688
 689 static void
 690 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 691 {
 692         filelayout_free_fh_array(fl);
 693         kfree(fl);
 694 }
 695
 696 static int
 697 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 698                          struct nfs4_filelayout_segment *fl,
 699                          struct nfs4_layoutget_res *lgr,
 700                          struct nfs4_deviceid *id,
 701                          gfp_t gfp_flags)
 702 {
 703         struct xdr_stream stream;
 704         struct xdr_buf buf;
 705         struct page *scratch;
 706         __be32 *p;
 707         uint32_t nfl_util;
 708         int i;
 709
 710         dprintk("%s: set_layout_map Begin\n", __func__);
 711
 712         scratch = alloc_page(gfp_flags);
 713         if (!scratch)
 714                 return -ENOMEM;
 715
 716         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 717         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 718
 719         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 720          * num_fh (4) */
 721         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 722         if (unlikely(!p))
 723                 goto out_err;
 724
 725         memcpy(id, p, sizeof(*id));
 726         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 727         nfs4_print_deviceid(id);
 728
 729         nfl_util = be32_to_cpup(p++);
 730         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 731                 fl->commit_through_mds = 1;
 732         if (nfl_util & NFL4_UFLG_DENSE)
 733                 fl->stripe_type = STRIPE_DENSE;
 734         else
 735                 fl->stripe_type = STRIPE_SPARSE;
 736         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 737
 738         fl->first_stripe_index = be32_to_cpup(p++);
 739         p = xdr_decode_hyper(p, &fl->pattern_offset);
 740         fl->num_fh = be32_to_cpup(p++);
 741
 742         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 743                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 744                 fl->pattern_offset);
 745
 746         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 747          * Futher checking is done in filelayout_check_layout */
 748         if (fl->num_fh >
 749             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 750                 goto out_err;
 751
 752         if (fl->num_fh > 0) {
 753                 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
 754                                        gfp_flags);
 755                 if (!fl->fh_array)
 756                         goto out_err;
 757         }
 758
 759         for (i = 0; i < fl->num_fh; i++) {
 760                 /* Do we want to use a mempool here? */
 761                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 762                 if (!fl->fh_array[i])
 763                         goto out_err_free;
 764
 765                 p = xdr_inline_decode(&stream, 4);
 766                 if (unlikely(!p))
 767                         goto out_err_free;
 768                 fl->fh_array[i]->size = be32_to_cpup(p++);
 769                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 770                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
 771                                i, fl->fh_array[i]->size);
 772                         goto out_err_free;
 773                 }
 774
 775                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 776                 if (unlikely(!p))
 777                         goto out_err_free;
 778                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 779                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 780                         fl->fh_array[i]->size);
 781         }
 782
 783         __free_page(scratch);
 784         return 0;
 785
 786 out_err_free:
 787         filelayout_free_fh_array(fl);
 788 out_err:
 789         __free_page(scratch);
 790         return -EIO;
 791 }
 792
 793 static void
 794 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 795 {
 796         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 797
 798         dprintk("--> %s\n", __func__);
 799         nfs4_fl_put_deviceid(fl->dsaddr);
 800         /* This assumes a single RW lseg */
 801         if (lseg->pls_range.iomode == IOMODE_RW) {
 802                 struct nfs4_filelayout *flo;
 803
 804                 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
 805                 flo->commit_info.nbuckets = 0;
 806                 kfree(flo->commit_info.buckets);
 807                 flo->commit_info.buckets = NULL;
 808         }
 809         _filelayout_free_lseg(fl);
 810 }
 811
 812 static int
 813 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 814                              struct nfs_commit_info *cinfo,
 815                              gfp_t gfp_flags)
 816 {
 817         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 818         struct pnfs_commit_bucket *buckets;
 819         int size;
 820
 821         if (fl->commit_through_mds)
 822                 return 0;
 823         if (cinfo->ds->nbuckets != 0) {
 824                 /* This assumes there is only one IOMODE_RW lseg.  What
 825                  * we really want to do is have a layout_hdr level
 826                  * dictionary of <multipath_list4, fh> keys, each
 827                  * associated with a struct list_head, populated by calls
 828                  * to filelayout_write_pagelist().
 829                  * */
 830                 return 0;
 831         }
 832
 833         size = (fl->stripe_type == STRIPE_SPARSE) ?
 834                 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 835
 836         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 837                           gfp_flags);
 838         if (!buckets)
 839                 return -ENOMEM;
 840         else {
 841                 int i;
 842
 843                 spin_lock(cinfo->lock);
 844                 if (cinfo->ds->nbuckets != 0)
 845                         kfree(buckets);
 846                 else {
 847                         cinfo->ds->buckets = buckets;
 848                         cinfo->ds->nbuckets = size;
 849                         for (i = 0; i < size; i++) {
 850                                 INIT_LIST_HEAD(&buckets[i].written);
 851                                 INIT_LIST_HEAD(&buckets[i].committing);
 852                         }
 853                 }
 854                 spin_unlock(cinfo->lock);
 855                 return 0;
 856         }
 857 }
 858
 859 static struct pnfs_layout_segment *
 860 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 861                       struct nfs4_layoutget_res *lgr,
 862                       gfp_t gfp_flags)
 863 {
 864         struct nfs4_filelayout_segment *fl;
 865         int rc;
 866         struct nfs4_deviceid id;
 867
 868         dprintk("--> %s\n", __func__);
 869         fl = kzalloc(sizeof(*fl), gfp_flags);
 870         if (!fl)
 871                 return NULL;
 872
 873         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 874         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 875                 _filelayout_free_lseg(fl);
 876                 return NULL;
 877         }
 878         return &fl->generic_hdr;
 879 }
 880
 881 /*
 882  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 883  *
 884  * return true  : coalesce page
 885  * return false : don't coalesce page
 886  */
 887 static bool
 888 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 889                    struct nfs_page *req)
 890 {
 891         u64 p_stripe, r_stripe;
 892         u32 stripe_unit;
 893
 894         if (!pnfs_generic_pg_test(pgio, prev, req) ||
 895             !nfs_generic_pg_test(pgio, prev, req))
 896                 return false;
 897
 898         p_stripe = (u64)req_offset(prev);
 899         r_stripe = (u64)req_offset(req);
 900         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 901
 902         do_div(p_stripe, stripe_unit);
 903         do_div(r_stripe, stripe_unit);
 904
 905         return (p_stripe == r_stripe);
 906 }
 907
 908 static void
 909 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 910                         struct nfs_page *req)
 911 {
 912         BUG_ON(pgio->pg_lseg != NULL);
 913
 914         if (req->wb_offset != req->wb_pgbase) {
 915                 /*
 916                  * Handling unaligned pages is difficult, because have to
 917                  * somehow split a req in two in certain cases in the
 918                  * pg.test code.  Avoid this by just not using pnfs
 919                  * in this case.
 920                  */
 921                 nfs_pageio_reset_read_mds(pgio);
 922                 return;
 923         }
 924         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 925                                            req->wb_context,
 926                                            0,
 927                                            NFS4_MAX_UINT64,
 928                                            IOMODE_READ,
 929                                            GFP_KERNEL);
 930         /* If no lseg, fall back to read through mds */
 931         if (pgio->pg_lseg == NULL)
 932                 nfs_pageio_reset_read_mds(pgio);
 933 }
 934
 935 static void
 936 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 937                          struct nfs_page *req)
 938 {
 939         struct nfs_commit_info cinfo;
 940         int status;
 941
 942         BUG_ON(pgio->pg_lseg != NULL);
 943
 944         if (req->wb_offset != req->wb_pgbase)
 945                 goto out_mds;
 946         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 947                                            req->wb_context,
 948                                            0,
 949                                            NFS4_MAX_UINT64,
 950                                            IOMODE_RW,
 951                                            GFP_NOFS);
 952         /* If no lseg, fall back to write through mds */
 953         if (pgio->pg_lseg == NULL)
 954                 goto out_mds;
 955         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 956         status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 957         if (status < 0) {
 958                 pnfs_put_lseg(pgio->pg_lseg);
 959                 pgio->pg_lseg = NULL;
 960                 goto out_mds;
 961         }
 962         return;
 963 out_mds:
 964         nfs_pageio_reset_write_mds(pgio);
 965 }
 966
 967 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 968         .pg_init = filelayout_pg_init_read,
 969         .pg_test = filelayout_pg_test,
 970         .pg_doio = pnfs_generic_pg_readpages,
 971 };
 972
 973 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 974         .pg_init = filelayout_pg_init_write,
 975         .pg_test = filelayout_pg_test,
 976         .pg_doio = pnfs_generic_pg_writepages,
 977 };
 978
 979 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 980 {
 981         if (fl->stripe_type == STRIPE_SPARSE)
 982                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
 983         else
 984                 return j;
 985 }
 986
 987 /* The generic layer is about to remove the req from the commit list.
 988  * If this will make the bucket empty, it will need to put the lseg reference.
 989  */
 990 static void
 991 filelayout_clear_request_commit(struct nfs_page *req,
 992                                 struct nfs_commit_info *cinfo)
 993 {
 994         struct pnfs_layout_segment *freeme = NULL;
 995
 996         spin_lock(cinfo->lock);
 997         if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
 998                 goto out;
 999         cinfo->ds->nwritten--;
1000         if (list_is_singular(&req->wb_list)) {
1001                 struct pnfs_commit_bucket *bucket;
1002
1003                 bucket = list_first_entry(&req->wb_list,
1004                                           struct pnfs_commit_bucket,
1005                                           written);
1006                 freeme = bucket->wlseg;
1007                 bucket->wlseg = NULL;
1008         }
1009 out:
1010         nfs_request_remove_commit_list(req, cinfo);
1011         spin_unlock(cinfo->lock);
1012         pnfs_put_lseg(freeme);
1013 }
1014
1015 static struct list_head *
1016 filelayout_choose_commit_list(struct nfs_page *req,
1017                               struct pnfs_layout_segment *lseg,
1018                               struct nfs_commit_info *cinfo)
1019 {
1020         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1021         u32 i, j;
1022         struct list_head *list;
1023         struct pnfs_commit_bucket *buckets;
1024
1025         if (fl->commit_through_mds)
1026                 return &cinfo->mds->list;
1027
1028         /* Note that we are calling nfs4_fl_calc_j_index on each page
1029          * that ends up being committed to a data server.  An attractive
1030          * alternative is to add a field to nfs_write_data and nfs_page
1031          * to store the value calculated in filelayout_write_pagelist
1032          * and just use that here.
1033          */
1034         j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1035         i = select_bucket_index(fl, j);
1036         buckets = cinfo->ds->buckets;
1037         list = &buckets[i].written;
1038         if (list_empty(list)) {
1039                 /* Non-empty buckets hold a reference on the lseg.  That ref
1040                  * is normally transferred to the COMMIT call and released
1041                  * there.  It could also be released if the last req is pulled
1042                  * off due to a rewrite, in which case it will be done in
1043                  * filelayout_clear_request_commit
1044                  */
1045                 buckets[i].wlseg = pnfs_get_lseg(lseg);
1046         }
1047         set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1048         cinfo->ds->nwritten++;
1049         return list;
1050 }
1051
1052 static void
1053 filelayout_mark_request_commit(struct nfs_page *req,
1054                                struct pnfs_layout_segment *lseg,
1055                                struct nfs_commit_info *cinfo)
1056 {
1057         struct list_head *list;
1058
1059         list = filelayout_choose_commit_list(req, lseg, cinfo);
1060         nfs_request_add_commit_list(req, list, cinfo);
1061 }
1062
1063 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1064 {
1065         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1066
1067         if (flseg->stripe_type == STRIPE_SPARSE)
1068                 return i;
1069         else
1070                 return nfs4_fl_calc_ds_index(lseg, i);
1071 }
1072
1073 static struct nfs_fh *
1074 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1075 {
1076         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1077
1078         if (flseg->stripe_type == STRIPE_SPARSE) {
1079                 if (flseg->num_fh == 1)
1080                         i = 0;
1081                 else if (flseg->num_fh == 0)
1082                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1083                         return NULL;
1084         }
1085         return flseg->fh_array[i];
1086 }
1087
1088 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1089 {
1090         struct pnfs_layout_segment *lseg = data->lseg;
1091         struct nfs4_pnfs_ds *ds;
1092         u32 idx;
1093         struct nfs_fh *fh;
1094
1095         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1096         ds = nfs4_fl_prepare_ds(lseg, idx);
1097         if (!ds) {
1098                 prepare_to_resend_writes(data);
1099                 filelayout_commit_release(data);
1100                 return -EAGAIN;
1101         }
1102         dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1103                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1104         data->commit_done_cb = filelayout_commit_done_cb;
1105         atomic_inc(&ds->ds_clp->cl_count);
1106         data->ds_clp = ds->ds_clp;
1107         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1108         if (fh)
1109                 data->args.fh = fh;
1110         return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data,
1111                                    &filelayout_commit_call_ops, how,
1112                                    RPC_TASK_SOFTCONN);
1113 }
1114
1115 static int
1116 transfer_commit_list(struct list_head *src, struct list_head *dst,
1117                      struct nfs_commit_info *cinfo, int max)
1118 {
1119         struct nfs_page *req, *tmp;
1120         int ret = 0;
1121
1122         list_for_each_entry_safe(req, tmp, src, wb_list) {
1123                 if (!nfs_lock_request(req))
1124                         continue;
1125                 kref_get(&req->wb_kref);
1126                 if (cond_resched_lock(cinfo->lock))
1127                         list_safe_reset_next(req, tmp, wb_list);
1128                 nfs_request_remove_commit_list(req, cinfo);
1129                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1130                 nfs_list_add_request(req, dst);
1131                 ret++;
1132                 if ((ret == max) && !cinfo->dreq)
1133                         break;
1134         }
1135         return ret;
1136 }
1137
1138 static int
1139 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1140                                struct nfs_commit_info *cinfo,
1141                                int max)
1142 {
1143         struct list_head *src = &bucket->written;
1144         struct list_head *dst = &bucket->committing;
1145         int ret;
1146
1147         ret = transfer_commit_list(src, dst, cinfo, max);
1148         if (ret) {
1149                 cinfo->ds->nwritten -= ret;
1150                 cinfo->ds->ncommitting += ret;
1151                 bucket->clseg = bucket->wlseg;
1152                 if (list_empty(src))
1153                         bucket->wlseg = NULL;
1154                 else
1155                         pnfs_get_lseg(bucket->clseg);
1156         }
1157         return ret;
1158 }
1159
1160 /* Move reqs from written to committing lists, returning count of number moved.
1161  * Note called with cinfo->lock held.
1162  */
1163 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1164                                         int max)
1165 {
1166         int i, rv = 0, cnt;
1167
1168         for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1169                 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1170                                                      cinfo, max);
1171                 max -= cnt;
1172                 rv += cnt;
1173         }
1174         return rv;
1175 }
1176
1177 /* Pull everything off the committing lists and dump into @dst */
1178 static void filelayout_recover_commit_reqs(struct list_head *dst,
1179                                            struct nfs_commit_info *cinfo)
1180 {
1181         struct pnfs_commit_bucket *b;
1182         int i;
1183
1184         /* NOTE cinfo->lock is NOT held, relying on fact that this is
1185          * only called on single thread per dreq.
1186          * Can't take the lock because need to do pnfs_put_lseg
1187          */
1188         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1189                 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1190                         BUG_ON(!list_empty(&b->written));
1191                         pnfs_put_lseg(b->wlseg);
1192                         b->wlseg = NULL;
1193                 }
1194         }
1195         cinfo->ds->nwritten = 0;
1196 }
1197
1198 static unsigned int
1199 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1200 {
1201         struct pnfs_ds_commit_info *fl_cinfo;
1202         struct pnfs_commit_bucket *bucket;
1203         struct nfs_commit_data *data;
1204         int i, j;
1205         unsigned int nreq = 0;
1206
1207         fl_cinfo = cinfo->ds;
1208         bucket = fl_cinfo->buckets;
1209         for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1210                 if (list_empty(&bucket->committing))
1211                         continue;
1212                 data = nfs_commitdata_alloc();
1213                 if (!data)
1214                         break;
1215                 data->ds_commit_index = i;
1216                 data->lseg = bucket->clseg;
1217                 bucket->clseg = NULL;
1218                 list_add(&data->pages, list);
1219                 nreq++;
1220         }
1221
1222         /* Clean up on error */
1223         for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1224                 if (list_empty(&bucket->committing))
1225                         continue;
1226                 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1227                 pnfs_put_lseg(bucket->clseg);
1228                 bucket->clseg = NULL;
1229         }
1230         /* Caller will clean up entries put on list */
1231         return nreq;
1232 }
1233
1234 /* This follows nfs_commit_list pretty closely */
1235 static int
1236 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1237                            int how, struct nfs_commit_info *cinfo)
1238 {
1239         struct nfs_commit_data *data, *tmp;
1240         LIST_HEAD(list);
1241         unsigned int nreq = 0;
1242
1243         if (!list_empty(mds_pages)) {
1244                 data = nfs_commitdata_alloc();
1245                 if (data != NULL) {
1246                         data->lseg = NULL;
1247                         list_add(&data->pages, &list);
1248                         nreq++;
1249                 } else
1250                         nfs_retry_commit(mds_pages, NULL, cinfo);
1251         }
1252
1253         nreq += alloc_ds_commits(cinfo, &list);
1254
1255         if (nreq == 0) {
1256                 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1257                 goto out;
1258         }
1259
1260         atomic_add(nreq, &cinfo->mds->rpcs_out);
1261
1262         list_for_each_entry_safe(data, tmp, &list, pages) {
1263                 list_del_init(&data->pages);
1264                 if (!data->lseg) {
1265                         nfs_init_commit(data, mds_pages, NULL, cinfo);
1266                         nfs_initiate_commit(NFS_CLIENT(inode), data,
1267                                             data->mds_ops, how, 0);
1268                 } else {
1269                         struct pnfs_commit_bucket *buckets;
1270
1271                         buckets = cinfo->ds->buckets;
1272                         nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1273                         filelayout_initiate_commit(data, how);
1274                 }
1275         }
1276 out:
1277         cinfo->ds->ncommitting = 0;
1278         return PNFS_ATTEMPTED;
1279 }
1280
1281 static void
1282 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1283 {
1284         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1285 }
1286
1287 static struct pnfs_layout_hdr *
1288 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1289 {
1290         struct nfs4_filelayout *flo;
1291
1292         flo = kzalloc(sizeof(*flo), gfp_flags);
1293         return &flo->generic_hdr;
1294 }
1295
1296 static void
1297 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1298 {
1299         kfree(FILELAYOUT_FROM_HDR(lo));
1300 }
1301
1302 static struct pnfs_ds_commit_info *
1303 filelayout_get_ds_info(struct inode *inode)
1304 {
1305         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1306
1307         if (layout == NULL)
1308                 return NULL;
1309         else
1310                 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1311 }
1312
1313 static struct pnfs_layoutdriver_type filelayout_type = {
1314         .id                     = LAYOUT_NFSV4_1_FILES,
1315         .name                   = "LAYOUT_NFSV4_1_FILES",
1316         .owner                  = THIS_MODULE,
1317         .alloc_layout_hdr       = filelayout_alloc_layout_hdr,
1318         .free_layout_hdr        = filelayout_free_layout_hdr,
1319         .alloc_lseg             = filelayout_alloc_lseg,
1320         .free_lseg              = filelayout_free_lseg,
1321         .pg_read_ops            = &filelayout_pg_read_ops,
1322         .pg_write_ops           = &filelayout_pg_write_ops,
1323         .get_ds_info            = &filelayout_get_ds_info,
1324         .mark_request_commit    = filelayout_mark_request_commit,
1325         .clear_request_commit   = filelayout_clear_request_commit,
1326         .scan_commit_lists      = filelayout_scan_commit_lists,
1327         .recover_commit_reqs    = filelayout_recover_commit_reqs,
1328         .commit_pagelist        = filelayout_commit_pagelist,
1329         .read_pagelist          = filelayout_read_pagelist,
1330         .write_pagelist         = filelayout_write_pagelist,
1331         .free_deviceid_node     = filelayout_free_deveiceid_node,
1332 };
1333
1334 static int __init nfs4filelayout_init(void)
1335 {
1336         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1337                __func__);
1338         return pnfs_register_layoutdriver(&filelayout_type);
1339 }
1340
1341 static void __exit nfs4filelayout_exit(void)
1342 {
1343         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1344                __func__);
1345         pnfs_unregister_layoutdriver(&filelayout_type);
1346 }
1347
1348 MODULE_ALIAS("nfs-layouttype4-1");
1349
1350 module_init(nfs4filelayout_init);
1351 module_exit(nfs4filelayout_exit);