fs/nfs/nfs4filelayout.c

   1 /*
   2  *  Module for the pnfs nfs4 file layout driver.
   3  *  Defines all I/O and Policy interface operations, plus code
   4  *  to register itself with the pNFS client.
   5  *
   6  *  Copyright (c) 2002
   7  *  The Regents of the University of Michigan
   8  *  All Rights Reserved
   9  *
  10  *  Dean Hildebrand <dhildebz@umich.edu>
  11  *
  12  *  Permission is granted to use, copy, create derivative works, and
  13  *  redistribute this software and such derivative works for any purpose,
  14  *  so long as the name of the University of Michigan is not used in
  15  *  any advertising or publicity pertaining to the use or distribution
  16  *  of this software without specific, written prior authorization. If
  17  *  the above copyright notice or any other identification of the
  18  *  University of Michigan is included in any copy of any portion of
  19  *  this software, then the disclaimer below must also be included.
  20  *
  21  *  This software is provided as is, without representation or warranty
  22  *  of any kind either express or implied, including without limitation
  23  *  the implied warranties of merchantability, fitness for a particular
  24  *  purpose, or noninfringement.  The Regents of the University of
  25  *  Michigan shall not be liable for any damages, including special,
  26  *  indirect, incidental, or consequential damages, with respect to any
  27  *  claim arising out of or in connection with the use of the software,
  28  *  even if it has been or is hereafter advised of the possibility of
  29  *  such damages.
  30  */
  31
  32 #include <linux/nfs_fs.h>
  33 #include <linux/nfs_page.h>
  34 #include <linux/module.h>
  35
  36 #include <linux/sunrpc/metrics.h>
  37
  38 #include "nfs4session.h"
  39 #include "internal.h"
  40 #include "delegation.h"
  41 #include "nfs4filelayout.h"
  42
  43 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  44
  45 MODULE_LICENSE("GPL");
  46 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
  47 MODULE_DESCRIPTION("The NFSv4 file layout driver");
  48
  49 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
  50
  51 static loff_t
  52 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
  53                             loff_t offset)
  54 {
  55         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
  56         u64 stripe_no;
  57         u32 rem;
  58
  59         offset -= flseg->pattern_offset;
  60         stripe_no = div_u64(offset, stripe_width);
  61         div_u64_rem(offset, flseg->stripe_unit, &rem);
  62
  63         return stripe_no * flseg->stripe_unit + rem;
  64 }
  65
  66 /* This function is used by the layout driver to calculate the
  67  * offset of the file on the dserver based on whether the
  68  * layout type is STRIPE_DENSE or STRIPE_SPARSE
  69  */
  70 static loff_t
  71 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
  72 {
  73         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
  74
  75         switch (flseg->stripe_type) {
  76         case STRIPE_SPARSE:
  77                 return offset;
  78
  79         case STRIPE_DENSE:
  80                 return filelayout_get_dense_offset(flseg, offset);
  81         }
  82
  83         BUG();
  84 }
  85
  86 static void filelayout_reset_write(struct nfs_write_data *data)
  87 {
  88         struct nfs_pgio_header *hdr = data->header;
  89         struct rpc_task *task = &data->task;
  90
  91         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
  92                 dprintk("%s Reset task %5u for i/o through MDS "
  93                         "(req %s/%lld, %u bytes @ offset %llu)\n", __func__,
  94                         data->task.tk_pid,
  95                         hdr->inode->i_sb->s_id,
  96                         (long long)NFS_FILEID(hdr->inode),
  97                         data->args.count,
  98                         (unsigned long long)data->args.offset);
  99
 100                 task->tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
 101                                                         &hdr->pages,
 102                                                         hdr->completion_ops);
 103         }
 104 }
 105
 106 static void filelayout_reset_read(struct nfs_read_data *data)
 107 {
 108         struct nfs_pgio_header *hdr = data->header;
 109         struct rpc_task *task = &data->task;
 110
 111         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 112                 dprintk("%s Reset task %5u for i/o through MDS "
 113                         "(req %s/%lld, %u bytes @ offset %llu)\n", __func__,
 114                         data->task.tk_pid,
 115                         hdr->inode->i_sb->s_id,
 116                         (long long)NFS_FILEID(hdr->inode),
 117                         data->args.count,
 118                         (unsigned long long)data->args.offset);
 119
 120                 task->tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
 121                                                         &hdr->pages,
 122                                                         hdr->completion_ops);
 123         }
 124 }
 125
 126 static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo)
 127 {
 128         if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 129                 return;
 130         clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags);
 131         pnfs_return_layout(inode);
 132 }
 133
 134 static int filelayout_async_handle_error(struct rpc_task *task,
 135                                          struct nfs4_state *state,
 136                                          struct nfs_client *clp,
 137                                          struct pnfs_layout_segment *lseg)
 138 {
 139         struct pnfs_layout_hdr *lo = lseg->pls_layout;
 140         struct inode *inode = lo->plh_inode;
 141         struct nfs_server *mds_server = NFS_SERVER(inode);
 142         struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
 143         struct nfs_client *mds_client = mds_server->nfs_client;
 144         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
 145
 146         if (task->tk_status >= 0)
 147                 return 0;
 148
 149         switch (task->tk_status) {
 150         /* MDS state errors */
 151         case -NFS4ERR_DELEG_REVOKED:
 152         case -NFS4ERR_ADMIN_REVOKED:
 153         case -NFS4ERR_BAD_STATEID:
 154                 if (state == NULL)
 155                         break;
 156                 nfs_remove_bad_delegation(state->inode);
 157         case -NFS4ERR_OPENMODE:
 158                 if (state == NULL)
 159                         break;
 160                 nfs4_schedule_stateid_recovery(mds_server, state);
 161                 goto wait_on_recovery;
 162         case -NFS4ERR_EXPIRED:
 163                 if (state != NULL)
 164                         nfs4_schedule_stateid_recovery(mds_server, state);
 165                 nfs4_schedule_lease_recovery(mds_client);
 166                 goto wait_on_recovery;
 167         /* DS session errors */
 168         case -NFS4ERR_BADSESSION:
 169         case -NFS4ERR_BADSLOT:
 170         case -NFS4ERR_BAD_HIGH_SLOT:
 171         case -NFS4ERR_DEADSESSION:
 172         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
 173         case -NFS4ERR_SEQ_FALSE_RETRY:
 174         case -NFS4ERR_SEQ_MISORDERED:
 175                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 176                         "flags 0x%x\n", __func__, task->tk_status,
 177                         clp->cl_exchange_flags);
 178                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
 179                 break;
 180         case -NFS4ERR_DELAY:
 181         case -NFS4ERR_GRACE:
 182                 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         nfs41_setup_sequence(rdata->ds_clp->cl_session,
 310                         &rdata->args.seq_args,
 311                         &rdata->res.seq_res,
 312                         task);
 313 }
 314
 315 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 316 {
 317         struct nfs_read_data *rdata = data;
 318
 319         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 320
 321         if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) &&
 322             task->tk_status == 0)
 323                 return;
 324
 325         /* Note this may cause RPC to be resent */
 326         rdata->header->mds_ops->rpc_call_done(task, data);
 327 }
 328
 329 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
 330 {
 331         struct nfs_read_data *rdata = data;
 332
 333         rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
 334 }
 335
 336 static void filelayout_read_release(void *data)
 337 {
 338         struct nfs_read_data *rdata = data;
 339         struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout;
 340
 341         filelayout_fenceme(lo->plh_inode, lo);
 342         nfs_put_client(rdata->ds_clp);
 343         rdata->header->mds_ops->rpc_release(data);
 344 }
 345
 346 static int filelayout_write_done_cb(struct rpc_task *task,
 347                                 struct nfs_write_data *data)
 348 {
 349         struct nfs_pgio_header *hdr = data->header;
 350         int err;
 351
 352         err = filelayout_async_handle_error(task, data->args.context->state,
 353                                             data->ds_clp, hdr->lseg);
 354
 355         switch (err) {
 356         case -NFS4ERR_RESET_TO_MDS:
 357                 filelayout_reset_write(data);
 358                 return task->tk_status;
 359         case -EAGAIN:
 360                 rpc_restart_call_prepare(task);
 361                 return -EAGAIN;
 362         }
 363
 364         filelayout_set_layoutcommit(data);
 365         return 0;
 366 }
 367
 368 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 369 static void prepare_to_resend_writes(struct nfs_commit_data *data)
 370 {
 371         struct nfs_page *first = nfs_list_entry(data->pages.next);
 372
 373         data->task.tk_status = 0;
 374         memcpy(&data->verf.verifier, &first->wb_verf,
 375                sizeof(data->verf.verifier));
 376         data->verf.verifier.data[0]++; /* ensure verifier mismatch */
 377 }
 378
 379 static int filelayout_commit_done_cb(struct rpc_task *task,
 380                                      struct nfs_commit_data *data)
 381 {
 382         int err;
 383
 384         err = filelayout_async_handle_error(task, NULL, data->ds_clp,
 385                                             data->lseg);
 386
 387         switch (err) {
 388         case -NFS4ERR_RESET_TO_MDS:
 389                 prepare_to_resend_writes(data);
 390                 return -EAGAIN;
 391         case -EAGAIN:
 392                 rpc_restart_call_prepare(task);
 393                 return -EAGAIN;
 394         }
 395
 396         return 0;
 397 }
 398
 399 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 400 {
 401         struct nfs_write_data *wdata = data;
 402
 403         if (filelayout_reset_to_mds(wdata->header->lseg)) {
 404                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 405                 filelayout_reset_write(wdata);
 406                 rpc_exit(task, 0);
 407                 return;
 408         }
 409         nfs41_setup_sequence(wdata->ds_clp->cl_session,
 410                         &wdata->args.seq_args,
 411                         &wdata->res.seq_res,
 412                         task);
 413 }
 414
 415 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 416 {
 417         struct nfs_write_data *wdata = data;
 418
 419         if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) &&
 420             task->tk_status == 0)
 421                 return;
 422
 423         /* Note this may cause RPC to be resent */
 424         wdata->header->mds_ops->rpc_call_done(task, data);
 425 }
 426
 427 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
 428 {
 429         struct nfs_write_data *wdata = data;
 430
 431         rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
 432 }
 433
 434 static void filelayout_write_release(void *data)
 435 {
 436         struct nfs_write_data *wdata = data;
 437         struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout;
 438
 439         filelayout_fenceme(lo->plh_inode, lo);
 440         nfs_put_client(wdata->ds_clp);
 441         wdata->header->mds_ops->rpc_release(data);
 442 }
 443
 444 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
 445 {
 446         struct nfs_commit_data *wdata = data;
 447
 448         nfs41_setup_sequence(wdata->ds_clp->cl_session,
 449                         &wdata->args.seq_args,
 450                         &wdata->res.seq_res,
 451                         task);
 452 }
 453
 454 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
 455 {
 456         struct nfs_commit_data *wdata = data;
 457
 458         /* Note this may cause RPC to be resent */
 459         wdata->mds_ops->rpc_call_done(task, data);
 460 }
 461
 462 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
 463 {
 464         struct nfs_commit_data *cdata = data;
 465
 466         rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
 467 }
 468
 469 static void filelayout_commit_release(void *calldata)
 470 {
 471         struct nfs_commit_data *data = calldata;
 472
 473         data->completion_ops->completion(data);
 474         pnfs_put_lseg(data->lseg);
 475         nfs_put_client(data->ds_clp);
 476         nfs_commitdata_release(data);
 477 }
 478
 479 static const struct rpc_call_ops filelayout_read_call_ops = {
 480         .rpc_call_prepare = filelayout_read_prepare,
 481         .rpc_call_done = filelayout_read_call_done,
 482         .rpc_count_stats = filelayout_read_count_stats,
 483         .rpc_release = filelayout_read_release,
 484 };
 485
 486 static const struct rpc_call_ops filelayout_write_call_ops = {
 487         .rpc_call_prepare = filelayout_write_prepare,
 488         .rpc_call_done = filelayout_write_call_done,
 489         .rpc_count_stats = filelayout_write_count_stats,
 490         .rpc_release = filelayout_write_release,
 491 };
 492
 493 static const struct rpc_call_ops filelayout_commit_call_ops = {
 494         .rpc_call_prepare = filelayout_commit_prepare,
 495         .rpc_call_done = filelayout_write_commit_done,
 496         .rpc_count_stats = filelayout_commit_count_stats,
 497         .rpc_release = filelayout_commit_release,
 498 };
 499
 500 static enum pnfs_try_status
 501 filelayout_read_pagelist(struct nfs_read_data *data)
 502 {
 503         struct nfs_pgio_header *hdr = data->header;
 504         struct pnfs_layout_segment *lseg = hdr->lseg;
 505         struct nfs4_pnfs_ds *ds;
 506         loff_t offset = data->args.offset;
 507         u32 j, idx;
 508         struct nfs_fh *fh;
 509
 510         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 511                 __func__, hdr->inode->i_ino,
 512                 data->args.pgbase, (size_t)data->args.count, offset);
 513
 514         /* Retrieve the correct rpc_client for the byte range */
 515         j = nfs4_fl_calc_j_index(lseg, offset);
 516         idx = nfs4_fl_calc_ds_index(lseg, j);
 517         ds = nfs4_fl_prepare_ds(lseg, idx);
 518         if (!ds)
 519                 return PNFS_NOT_ATTEMPTED;
 520         dprintk("%s USE DS: %s cl_count %d\n", __func__,
 521                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 522
 523         /* No multipath support. Use first DS */
 524         atomic_inc(&ds->ds_clp->cl_count);
 525         data->ds_clp = ds->ds_clp;
 526         fh = nfs4_fl_select_ds_fh(lseg, j);
 527         if (fh)
 528                 data->args.fh = fh;
 529
 530         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 531         data->mds_offset = offset;
 532
 533         /* Perform an asynchronous read to ds */
 534         nfs_initiate_read(ds->ds_clp->cl_rpcclient, data,
 535                                   &filelayout_read_call_ops, RPC_TASK_SOFTCONN);
 536         return PNFS_ATTEMPTED;
 537 }
 538
 539 /* Perform async writes. */
 540 static enum pnfs_try_status
 541 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
 542 {
 543         struct nfs_pgio_header *hdr = data->header;
 544         struct pnfs_layout_segment *lseg = hdr->lseg;
 545         struct nfs4_pnfs_ds *ds;
 546         loff_t offset = data->args.offset;
 547         u32 j, idx;
 548         struct nfs_fh *fh;
 549
 550         /* Retrieve the correct rpc_client for the byte range */
 551         j = nfs4_fl_calc_j_index(lseg, offset);
 552         idx = nfs4_fl_calc_ds_index(lseg, j);
 553         ds = nfs4_fl_prepare_ds(lseg, idx);
 554         if (!ds)
 555                 return PNFS_NOT_ATTEMPTED;
 556         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
 557                 __func__, hdr->inode->i_ino, sync, (size_t) data->args.count,
 558                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 559
 560         data->write_done_cb = filelayout_write_done_cb;
 561         atomic_inc(&ds->ds_clp->cl_count);
 562         data->ds_clp = ds->ds_clp;
 563         fh = nfs4_fl_select_ds_fh(lseg, j);
 564         if (fh)
 565                 data->args.fh = fh;
 566         /*
 567          * Get the file offset on the dserver. Set the write offset to
 568          * this offset and save the original offset.
 569          */
 570         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 571
 572         /* Perform an asynchronous write */
 573         nfs_initiate_write(ds->ds_clp->cl_rpcclient, data,
 574                                     &filelayout_write_call_ops, sync,
 575                                     RPC_TASK_SOFTCONN);
 576         return PNFS_ATTEMPTED;
 577 }
 578
 579 /*
 580  * filelayout_check_layout()
 581  *
 582  * Make sure layout segment parameters are sane WRT the device.
 583  * At this point no generic layer initialization of the lseg has occurred,
 584  * and nothing has been added to the layout_hdr cache.
 585  *
 586  */
 587 static int
 588 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 589                         struct nfs4_filelayout_segment *fl,
 590                         struct nfs4_layoutget_res *lgr,
 591                         struct nfs4_deviceid *id,
 592                         gfp_t gfp_flags)
 593 {
 594         struct nfs4_deviceid_node *d;
 595         struct nfs4_file_layout_dsaddr *dsaddr;
 596         int status = -EINVAL;
 597         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 598
 599         dprintk("--> %s\n", __func__);
 600
 601         /* FIXME: remove this check when layout segment support is added */
 602         if (lgr->range.offset != 0 ||
 603             lgr->range.length != NFS4_MAX_UINT64) {
 604                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 605                         __func__);
 606                 goto out;
 607         }
 608
 609         if (fl->pattern_offset > lgr->range.offset) {
 610                 dprintk("%s pattern_offset %lld too large\n",
 611                                 __func__, fl->pattern_offset);
 612                 goto out;
 613         }
 614
 615         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 616                 dprintk("%s Invalid stripe unit (%u)\n",
 617                         __func__, fl->stripe_unit);
 618                 goto out;
 619         }
 620
 621         /* find and reference the deviceid */
 622         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
 623                                    NFS_SERVER(lo->plh_inode)->nfs_client, id);
 624         if (d == NULL) {
 625                 dsaddr = filelayout_get_device_info(lo->plh_inode, id, gfp_flags);
 626                 if (dsaddr == NULL)
 627                         goto out;
 628         } else
 629                 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 630         /* Found deviceid is unavailable */
 631         if (filelayout_test_devid_unavailable(&dsaddr->id_node))
 632                         goto out_put;
 633
 634         fl->dsaddr = dsaddr;
 635
 636         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 637                 dprintk("%s Bad first_stripe_index %u\n",
 638                                 __func__, fl->first_stripe_index);
 639                 goto out_put;
 640         }
 641
 642         if ((fl->stripe_type == STRIPE_SPARSE &&
 643             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 644             (fl->stripe_type == STRIPE_DENSE &&
 645             fl->num_fh != dsaddr->stripe_count)) {
 646                 dprintk("%s num_fh %u not valid for given packing\n",
 647                         __func__, fl->num_fh);
 648                 goto out_put;
 649         }
 650
 651         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
 652                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
 653                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
 654                         nfss->wsize);
 655         }
 656
 657         status = 0;
 658 out:
 659         dprintk("--> %s returns %d\n", __func__, status);
 660         return status;
 661 out_put:
 662         nfs4_fl_put_deviceid(dsaddr);
 663         goto out;
 664 }
 665
 666 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 667 {
 668         int i;
 669
 670         for (i = 0; i < fl->num_fh; i++) {
 671                 if (!fl->fh_array[i])
 672                         break;
 673                 kfree(fl->fh_array[i]);
 674         }
 675         kfree(fl->fh_array);
 676         fl->fh_array = NULL;
 677 }
 678
 679 static void
 680 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 681 {
 682         filelayout_free_fh_array(fl);
 683         kfree(fl);
 684 }
 685
 686 static int
 687 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 688                          struct nfs4_filelayout_segment *fl,
 689                          struct nfs4_layoutget_res *lgr,
 690                          struct nfs4_deviceid *id,
 691                          gfp_t gfp_flags)
 692 {
 693         struct xdr_stream stream;
 694         struct xdr_buf buf;
 695         struct page *scratch;
 696         __be32 *p;
 697         uint32_t nfl_util;
 698         int i;
 699
 700         dprintk("%s: set_layout_map Begin\n", __func__);
 701
 702         scratch = alloc_page(gfp_flags);
 703         if (!scratch)
 704                 return -ENOMEM;
 705
 706         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 707         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 708
 709         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 710          * num_fh (4) */
 711         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 712         if (unlikely(!p))
 713                 goto out_err;
 714
 715         memcpy(id, p, sizeof(*id));
 716         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 717         nfs4_print_deviceid(id);
 718
 719         nfl_util = be32_to_cpup(p++);
 720         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 721                 fl->commit_through_mds = 1;
 722         if (nfl_util & NFL4_UFLG_DENSE)
 723                 fl->stripe_type = STRIPE_DENSE;
 724         else
 725                 fl->stripe_type = STRIPE_SPARSE;
 726         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 727
 728         fl->first_stripe_index = be32_to_cpup(p++);
 729         p = xdr_decode_hyper(p, &fl->pattern_offset);
 730         fl->num_fh = be32_to_cpup(p++);
 731
 732         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 733                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 734                 fl->pattern_offset);
 735
 736         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 737          * Futher checking is done in filelayout_check_layout */
 738         if (fl->num_fh >
 739             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 740                 goto out_err;
 741
 742         if (fl->num_fh > 0) {
 743                 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
 744                                        gfp_flags);
 745                 if (!fl->fh_array)
 746                         goto out_err;
 747         }
 748
 749         for (i = 0; i < fl->num_fh; i++) {
 750                 /* Do we want to use a mempool here? */
 751                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 752                 if (!fl->fh_array[i])
 753                         goto out_err_free;
 754
 755                 p = xdr_inline_decode(&stream, 4);
 756                 if (unlikely(!p))
 757                         goto out_err_free;
 758                 fl->fh_array[i]->size = be32_to_cpup(p++);
 759                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 760                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
 761                                i, fl->fh_array[i]->size);
 762                         goto out_err_free;
 763                 }
 764
 765                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 766                 if (unlikely(!p))
 767                         goto out_err_free;
 768                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 769                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 770                         fl->fh_array[i]->size);
 771         }
 772
 773         __free_page(scratch);
 774         return 0;
 775
 776 out_err_free:
 777         filelayout_free_fh_array(fl);
 778 out_err:
 779         __free_page(scratch);
 780         return -EIO;
 781 }
 782
 783 static void
 784 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 785 {
 786         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 787
 788         dprintk("--> %s\n", __func__);
 789         nfs4_fl_put_deviceid(fl->dsaddr);
 790         /* This assumes a single RW lseg */
 791         if (lseg->pls_range.iomode == IOMODE_RW) {
 792                 struct nfs4_filelayout *flo;
 793
 794                 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
 795                 flo->commit_info.nbuckets = 0;
 796                 kfree(flo->commit_info.buckets);
 797                 flo->commit_info.buckets = NULL;
 798         }
 799         _filelayout_free_lseg(fl);
 800 }
 801
 802 static int
 803 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 804                              struct nfs_commit_info *cinfo,
 805                              gfp_t gfp_flags)
 806 {
 807         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 808         struct pnfs_commit_bucket *buckets;
 809         int size;
 810
 811         if (fl->commit_through_mds)
 812                 return 0;
 813         if (cinfo->ds->nbuckets != 0) {
 814                 /* This assumes there is only one IOMODE_RW lseg.  What
 815                  * we really want to do is have a layout_hdr level
 816                  * dictionary of <multipath_list4, fh> keys, each
 817                  * associated with a struct list_head, populated by calls
 818                  * to filelayout_write_pagelist().
 819                  * */
 820                 return 0;
 821         }
 822
 823         size = (fl->stripe_type == STRIPE_SPARSE) ?
 824                 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 825
 826         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 827                           gfp_flags);
 828         if (!buckets)
 829                 return -ENOMEM;
 830         else {
 831                 int i;
 832
 833                 spin_lock(cinfo->lock);
 834                 if (cinfo->ds->nbuckets != 0)
 835                         kfree(buckets);
 836                 else {
 837                         cinfo->ds->buckets = buckets;
 838                         cinfo->ds->nbuckets = size;
 839                         for (i = 0; i < size; i++) {
 840                                 INIT_LIST_HEAD(&buckets[i].written);
 841                                 INIT_LIST_HEAD(&buckets[i].committing);
 842                         }
 843                 }
 844                 spin_unlock(cinfo->lock);
 845                 return 0;
 846         }
 847 }
 848
 849 static struct pnfs_layout_segment *
 850 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 851                       struct nfs4_layoutget_res *lgr,
 852                       gfp_t gfp_flags)
 853 {
 854         struct nfs4_filelayout_segment *fl;
 855         int rc;
 856         struct nfs4_deviceid id;
 857
 858         dprintk("--> %s\n", __func__);
 859         fl = kzalloc(sizeof(*fl), gfp_flags);
 860         if (!fl)
 861                 return NULL;
 862
 863         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 864         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 865                 _filelayout_free_lseg(fl);
 866                 return NULL;
 867         }
 868         return &fl->generic_hdr;
 869 }
 870
 871 /*
 872  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 873  *
 874  * return true  : coalesce page
 875  * return false : don't coalesce page
 876  */
 877 static bool
 878 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 879                    struct nfs_page *req)
 880 {
 881         u64 p_stripe, r_stripe;
 882         u32 stripe_unit;
 883
 884         if (!pnfs_generic_pg_test(pgio, prev, req) ||
 885             !nfs_generic_pg_test(pgio, prev, req))
 886                 return false;
 887
 888         p_stripe = (u64)req_offset(prev);
 889         r_stripe = (u64)req_offset(req);
 890         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 891
 892         do_div(p_stripe, stripe_unit);
 893         do_div(r_stripe, stripe_unit);
 894
 895         return (p_stripe == r_stripe);
 896 }
 897
 898 static void
 899 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 900                         struct nfs_page *req)
 901 {
 902         WARN_ON_ONCE(pgio->pg_lseg != NULL);
 903
 904         if (req->wb_offset != req->wb_pgbase) {
 905                 /*
 906                  * Handling unaligned pages is difficult, because have to
 907                  * somehow split a req in two in certain cases in the
 908                  * pg.test code.  Avoid this by just not using pnfs
 909                  * in this case.
 910                  */
 911                 nfs_pageio_reset_read_mds(pgio);
 912                 return;
 913         }
 914         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 915                                            req->wb_context,
 916                                            0,
 917                                            NFS4_MAX_UINT64,
 918                                            IOMODE_READ,
 919                                            GFP_KERNEL);
 920         /* If no lseg, fall back to read through mds */
 921         if (pgio->pg_lseg == NULL)
 922                 nfs_pageio_reset_read_mds(pgio);
 923 }
 924
 925 static void
 926 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 927                          struct nfs_page *req)
 928 {
 929         struct nfs_commit_info cinfo;
 930         int status;
 931
 932         WARN_ON_ONCE(pgio->pg_lseg != NULL);
 933
 934         if (req->wb_offset != req->wb_pgbase)
 935                 goto out_mds;
 936         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 937                                            req->wb_context,
 938                                            0,
 939                                            NFS4_MAX_UINT64,
 940                                            IOMODE_RW,
 941                                            GFP_NOFS);
 942         /* If no lseg, fall back to write through mds */
 943         if (pgio->pg_lseg == NULL)
 944                 goto out_mds;
 945         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 946         status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 947         if (status < 0) {
 948                 pnfs_put_lseg(pgio->pg_lseg);
 949                 pgio->pg_lseg = NULL;
 950                 goto out_mds;
 951         }
 952         return;
 953 out_mds:
 954         nfs_pageio_reset_write_mds(pgio);
 955 }
 956
 957 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 958         .pg_init = filelayout_pg_init_read,
 959         .pg_test = filelayout_pg_test,
 960         .pg_doio = pnfs_generic_pg_readpages,
 961 };
 962
 963 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 964         .pg_init = filelayout_pg_init_write,
 965         .pg_test = filelayout_pg_test,
 966         .pg_doio = pnfs_generic_pg_writepages,
 967 };
 968
 969 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 970 {
 971         if (fl->stripe_type == STRIPE_SPARSE)
 972                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
 973         else
 974                 return j;
 975 }
 976
 977 /* The generic layer is about to remove the req from the commit list.
 978  * If this will make the bucket empty, it will need to put the lseg reference.
 979  */
 980 static void
 981 filelayout_clear_request_commit(struct nfs_page *req,
 982                                 struct nfs_commit_info *cinfo)
 983 {
 984         struct pnfs_layout_segment *freeme = NULL;
 985
 986         spin_lock(cinfo->lock);
 987         if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
 988                 goto out;
 989         cinfo->ds->nwritten--;
 990         if (list_is_singular(&req->wb_list)) {
 991                 struct pnfs_commit_bucket *bucket;
 992
 993                 bucket = list_first_entry(&req->wb_list,
 994                                           struct pnfs_commit_bucket,
 995                                           written);
 996                 freeme = bucket->wlseg;
 997                 bucket->wlseg = NULL;
 998         }
 999 out:
1000         nfs_request_remove_commit_list(req, cinfo);
1001         spin_unlock(cinfo->lock);
1002         pnfs_put_lseg(freeme);
1003 }
1004
1005 static struct list_head *
1006 filelayout_choose_commit_list(struct nfs_page *req,
1007                               struct pnfs_layout_segment *lseg,
1008                               struct nfs_commit_info *cinfo)
1009 {
1010         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1011         u32 i, j;
1012         struct list_head *list;
1013         struct pnfs_commit_bucket *buckets;
1014
1015         if (fl->commit_through_mds)
1016                 return &cinfo->mds->list;
1017
1018         /* Note that we are calling nfs4_fl_calc_j_index on each page
1019          * that ends up being committed to a data server.  An attractive
1020          * alternative is to add a field to nfs_write_data and nfs_page
1021          * to store the value calculated in filelayout_write_pagelist
1022          * and just use that here.
1023          */
1024         j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1025         i = select_bucket_index(fl, j);
1026         buckets = cinfo->ds->buckets;
1027         list = &buckets[i].written;
1028         if (list_empty(list)) {
1029                 /* Non-empty buckets hold a reference on the lseg.  That ref
1030                  * is normally transferred to the COMMIT call and released
1031                  * there.  It could also be released if the last req is pulled
1032                  * off due to a rewrite, in which case it will be done in
1033                  * filelayout_clear_request_commit
1034                  */
1035                 buckets[i].wlseg = pnfs_get_lseg(lseg);
1036         }
1037         set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1038         cinfo->ds->nwritten++;
1039         return list;
1040 }
1041
1042 static void
1043 filelayout_mark_request_commit(struct nfs_page *req,
1044                                struct pnfs_layout_segment *lseg,
1045                                struct nfs_commit_info *cinfo)
1046 {
1047         struct list_head *list;
1048
1049         list = filelayout_choose_commit_list(req, lseg, cinfo);
1050         nfs_request_add_commit_list(req, list, cinfo);
1051 }
1052
1053 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1054 {
1055         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1056
1057         if (flseg->stripe_type == STRIPE_SPARSE)
1058                 return i;
1059         else
1060                 return nfs4_fl_calc_ds_index(lseg, i);
1061 }
1062
1063 static struct nfs_fh *
1064 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1065 {
1066         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1067
1068         if (flseg->stripe_type == STRIPE_SPARSE) {
1069                 if (flseg->num_fh == 1)
1070                         i = 0;
1071                 else if (flseg->num_fh == 0)
1072                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1073                         return NULL;
1074         }
1075         return flseg->fh_array[i];
1076 }
1077
1078 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1079 {
1080         struct pnfs_layout_segment *lseg = data->lseg;
1081         struct nfs4_pnfs_ds *ds;
1082         u32 idx;
1083         struct nfs_fh *fh;
1084
1085         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1086         ds = nfs4_fl_prepare_ds(lseg, idx);
1087         if (!ds) {
1088                 prepare_to_resend_writes(data);
1089                 filelayout_commit_release(data);
1090                 return -EAGAIN;
1091         }
1092         dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1093                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1094         data->commit_done_cb = filelayout_commit_done_cb;
1095         atomic_inc(&ds->ds_clp->cl_count);
1096         data->ds_clp = ds->ds_clp;
1097         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1098         if (fh)
1099                 data->args.fh = fh;
1100         return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data,
1101                                    &filelayout_commit_call_ops, how,
1102                                    RPC_TASK_SOFTCONN);
1103 }
1104
1105 static int
1106 transfer_commit_list(struct list_head *src, struct list_head *dst,
1107                      struct nfs_commit_info *cinfo, int max)
1108 {
1109         struct nfs_page *req, *tmp;
1110         int ret = 0;
1111
1112         list_for_each_entry_safe(req, tmp, src, wb_list) {
1113                 if (!nfs_lock_request(req))
1114                         continue;
1115                 kref_get(&req->wb_kref);
1116                 if (cond_resched_lock(cinfo->lock))
1117                         list_safe_reset_next(req, tmp, wb_list);
1118                 nfs_request_remove_commit_list(req, cinfo);
1119                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1120                 nfs_list_add_request(req, dst);
1121                 ret++;
1122                 if ((ret == max) && !cinfo->dreq)
1123                         break;
1124         }
1125         return ret;
1126 }
1127
1128 static int
1129 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1130                                struct nfs_commit_info *cinfo,
1131                                int max)
1132 {
1133         struct list_head *src = &bucket->written;
1134         struct list_head *dst = &bucket->committing;
1135         int ret;
1136
1137         ret = transfer_commit_list(src, dst, cinfo, max);
1138         if (ret) {
1139                 cinfo->ds->nwritten -= ret;
1140                 cinfo->ds->ncommitting += ret;
1141                 bucket->clseg = bucket->wlseg;
1142                 if (list_empty(src))
1143                         bucket->wlseg = NULL;
1144                 else
1145                         pnfs_get_lseg(bucket->clseg);
1146         }
1147         return ret;
1148 }
1149
1150 /* Move reqs from written to committing lists, returning count of number moved.
1151  * Note called with cinfo->lock held.
1152  */
1153 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1154                                         int max)
1155 {
1156         int i, rv = 0, cnt;
1157
1158         for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1159                 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1160                                                      cinfo, max);
1161                 max -= cnt;
1162                 rv += cnt;
1163         }
1164         return rv;
1165 }
1166
1167 /* Pull everything off the committing lists and dump into @dst */
1168 static void filelayout_recover_commit_reqs(struct list_head *dst,
1169                                            struct nfs_commit_info *cinfo)
1170 {
1171         struct pnfs_commit_bucket *b;
1172         int i;
1173
1174         /* NOTE cinfo->lock is NOT held, relying on fact that this is
1175          * only called on single thread per dreq.
1176          * Can't take the lock because need to do pnfs_put_lseg
1177          */
1178         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1179                 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1180                         pnfs_put_lseg(b->wlseg);
1181                         b->wlseg = NULL;
1182                 }
1183         }
1184         cinfo->ds->nwritten = 0;
1185 }
1186
1187 static unsigned int
1188 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1189 {
1190         struct pnfs_ds_commit_info *fl_cinfo;
1191         struct pnfs_commit_bucket *bucket;
1192         struct nfs_commit_data *data;
1193         int i, j;
1194         unsigned int nreq = 0;
1195
1196         fl_cinfo = cinfo->ds;
1197         bucket = fl_cinfo->buckets;
1198         for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1199                 if (list_empty(&bucket->committing))
1200                         continue;
1201                 data = nfs_commitdata_alloc();
1202                 if (!data)
1203                         break;
1204                 data->ds_commit_index = i;
1205                 data->lseg = bucket->clseg;
1206                 bucket->clseg = NULL;
1207                 list_add(&data->pages, list);
1208                 nreq++;
1209         }
1210
1211         /* Clean up on error */
1212         for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1213                 if (list_empty(&bucket->committing))
1214                         continue;
1215                 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1216                 pnfs_put_lseg(bucket->clseg);
1217                 bucket->clseg = NULL;
1218         }
1219         /* Caller will clean up entries put on list */
1220         return nreq;
1221 }
1222
1223 /* This follows nfs_commit_list pretty closely */
1224 static int
1225 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1226                            int how, struct nfs_commit_info *cinfo)
1227 {
1228         struct nfs_commit_data *data, *tmp;
1229         LIST_HEAD(list);
1230         unsigned int nreq = 0;
1231
1232         if (!list_empty(mds_pages)) {
1233                 data = nfs_commitdata_alloc();
1234                 if (data != NULL) {
1235                         data->lseg = NULL;
1236                         list_add(&data->pages, &list);
1237                         nreq++;
1238                 } else
1239                         nfs_retry_commit(mds_pages, NULL, cinfo);
1240         }
1241
1242         nreq += alloc_ds_commits(cinfo, &list);
1243
1244         if (nreq == 0) {
1245                 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1246                 goto out;
1247         }
1248
1249         atomic_add(nreq, &cinfo->mds->rpcs_out);
1250
1251         list_for_each_entry_safe(data, tmp, &list, pages) {
1252                 list_del_init(&data->pages);
1253                 if (!data->lseg) {
1254                         nfs_init_commit(data, mds_pages, NULL, cinfo);
1255                         nfs_initiate_commit(NFS_CLIENT(inode), data,
1256                                             data->mds_ops, how, 0);
1257                 } else {
1258                         struct pnfs_commit_bucket *buckets;
1259
1260                         buckets = cinfo->ds->buckets;
1261                         nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1262                         filelayout_initiate_commit(data, how);
1263                 }
1264         }
1265 out:
1266         cinfo->ds->ncommitting = 0;
1267         return PNFS_ATTEMPTED;
1268 }
1269
1270 static void
1271 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1272 {
1273         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1274 }
1275
1276 static struct pnfs_layout_hdr *
1277 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1278 {
1279         struct nfs4_filelayout *flo;
1280
1281         flo = kzalloc(sizeof(*flo), gfp_flags);
1282         return &flo->generic_hdr;
1283 }
1284
1285 static void
1286 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1287 {
1288         kfree(FILELAYOUT_FROM_HDR(lo));
1289 }
1290
1291 static struct pnfs_ds_commit_info *
1292 filelayout_get_ds_info(struct inode *inode)
1293 {
1294         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1295
1296         if (layout == NULL)
1297                 return NULL;
1298         else
1299                 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1300 }
1301
1302 static struct pnfs_layoutdriver_type filelayout_type = {
1303         .id                     = LAYOUT_NFSV4_1_FILES,
1304         .name                   = "LAYOUT_NFSV4_1_FILES",
1305         .owner                  = THIS_MODULE,
1306         .alloc_layout_hdr       = filelayout_alloc_layout_hdr,
1307         .free_layout_hdr        = filelayout_free_layout_hdr,
1308         .alloc_lseg             = filelayout_alloc_lseg,
1309         .free_lseg              = filelayout_free_lseg,
1310         .pg_read_ops            = &filelayout_pg_read_ops,
1311         .pg_write_ops           = &filelayout_pg_write_ops,
1312         .get_ds_info            = &filelayout_get_ds_info,
1313         .mark_request_commit    = filelayout_mark_request_commit,
1314         .clear_request_commit   = filelayout_clear_request_commit,
1315         .scan_commit_lists      = filelayout_scan_commit_lists,
1316         .recover_commit_reqs    = filelayout_recover_commit_reqs,
1317         .commit_pagelist        = filelayout_commit_pagelist,
1318         .read_pagelist          = filelayout_read_pagelist,
1319         .write_pagelist         = filelayout_write_pagelist,
1320         .free_deviceid_node     = filelayout_free_deveiceid_node,
1321 };
1322
1323 static int __init nfs4filelayout_init(void)
1324 {
1325         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1326                __func__);
1327         return pnfs_register_layoutdriver(&filelayout_type);
1328 }
1329
1330 static void __exit nfs4filelayout_exit(void)
1331 {
1332         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1333                __func__);
1334         pnfs_unregister_layoutdriver(&filelayout_type);
1335 }
1336
1337 MODULE_ALIAS("nfs-layouttype4-1");
1338
1339 module_init(nfs4filelayout_init);
1340 module_exit(nfs4filelayout_exit);