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