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         pnfs_return_layout(inode);
 133 }
 134
 135 static int filelayout_async_handle_error(struct rpc_task *task,
 136                                          struct nfs4_state *state,
 137                                          struct nfs_client *clp,
 138                                          struct pnfs_layout_segment *lseg)
 139 {
 140         struct pnfs_layout_hdr *lo = lseg->pls_layout;
 141         struct inode *inode = lo->plh_inode;
 142         struct nfs_server *mds_server = NFS_SERVER(inode);
 143         struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
 144         struct nfs_client *mds_client = mds_server->nfs_client;
 145         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
 146
 147         if (task->tk_status >= 0)
 148                 return 0;
 149
 150         switch (task->tk_status) {
 151         /* MDS state errors */
 152         case -NFS4ERR_DELEG_REVOKED:
 153         case -NFS4ERR_ADMIN_REVOKED:
 154         case -NFS4ERR_BAD_STATEID:
 155                 if (state == NULL)
 156                         break;
 157                 nfs_remove_bad_delegation(state->inode);
 158         case -NFS4ERR_OPENMODE:
 159                 if (state == NULL)
 160                         break;
 161                 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
 162                         goto out_bad_stateid;
 163                 goto wait_on_recovery;
 164         case -NFS4ERR_EXPIRED:
 165                 if (state != NULL) {
 166                         if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
 167                                 goto out_bad_stateid;
 168                 }
 169                 nfs4_schedule_lease_recovery(mds_client);
 170                 goto wait_on_recovery;
 171         /* DS session errors */
 172         case -NFS4ERR_BADSESSION:
 173         case -NFS4ERR_BADSLOT:
 174         case -NFS4ERR_BAD_HIGH_SLOT:
 175         case -NFS4ERR_DEADSESSION:
 176         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
 177         case -NFS4ERR_SEQ_FALSE_RETRY:
 178         case -NFS4ERR_SEQ_MISORDERED:
 179                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 180                         "flags 0x%x\n", __func__, task->tk_status,
 181                         clp->cl_exchange_flags);
 182                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
 183                 break;
 184         case -NFS4ERR_DELAY:
 185         case -NFS4ERR_GRACE:
 186                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 187                 break;
 188         case -NFS4ERR_RETRY_UNCACHED_REP:
 189                 break;
 190         /* Invalidate Layout errors */
 191         case -NFS4ERR_PNFS_NO_LAYOUT:
 192         case -ESTALE:           /* mapped NFS4ERR_STALE */
 193         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
 194         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
 195         case -NFS4ERR_FHEXPIRED:
 196         case -NFS4ERR_WRONG_TYPE:
 197                 dprintk("%s Invalid layout error %d\n", __func__,
 198                         task->tk_status);
 199                 /*
 200                  * Destroy layout so new i/o will get a new layout.
 201                  * Layout will not be destroyed until all current lseg
 202                  * references are put. Mark layout as invalid to resend failed
 203                  * i/o and all i/o waiting on the slot table to the MDS until
 204                  * layout is destroyed and a new valid layout is obtained.
 205                  */
 206                 pnfs_destroy_layout(NFS_I(inode));
 207                 rpc_wake_up(&tbl->slot_tbl_waitq);
 208                 goto reset;
 209         /* RPC connection errors */
 210         case -ECONNREFUSED:
 211         case -EHOSTDOWN:
 212         case -EHOSTUNREACH:
 213         case -ENETUNREACH:
 214         case -EIO:
 215         case -ETIMEDOUT:
 216         case -EPIPE:
 217                 dprintk("%s DS connection error %d\n", __func__,
 218                         task->tk_status);
 219                 nfs4_mark_deviceid_unavailable(devid);
 220                 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
 221                 rpc_wake_up(&tbl->slot_tbl_waitq);
 222                 /* fall through */
 223         default:
 224 reset:
 225                 dprintk("%s Retry through MDS. Error %d\n", __func__,
 226                         task->tk_status);
 227                 return -NFS4ERR_RESET_TO_MDS;
 228         }
 229 out:
 230         task->tk_status = 0;
 231         return -EAGAIN;
 232 out_bad_stateid:
 233         task->tk_status = -EIO;
 234         return 0;
 235 wait_on_recovery:
 236         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
 237         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
 238                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
 239         goto out;
 240 }
 241
 242 /* NFS_PROTO call done callback routines */
 243
 244 static int filelayout_read_done_cb(struct rpc_task *task,
 245                                 struct nfs_read_data *data)
 246 {
 247         struct nfs_pgio_header *hdr = data->header;
 248         int err;
 249
 250         err = filelayout_async_handle_error(task, data->args.context->state,
 251                                             data->ds_clp, hdr->lseg);
 252
 253         switch (err) {
 254         case -NFS4ERR_RESET_TO_MDS:
 255                 filelayout_reset_read(data);
 256                 return task->tk_status;
 257         case -EAGAIN:
 258                 rpc_restart_call_prepare(task);
 259                 return -EAGAIN;
 260         }
 261
 262         return 0;
 263 }
 264
 265 /*
 266  * We reference the rpc_cred of the first WRITE that triggers the need for
 267  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 268  * rfc5661 is not clear about which credential should be used.
 269  */
 270 static void
 271 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
 272 {
 273         struct nfs_pgio_header *hdr = wdata->header;
 274
 275         if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
 276             wdata->res.verf->committed == NFS_FILE_SYNC)
 277                 return;
 278
 279         pnfs_set_layoutcommit(wdata);
 280         dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
 281                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
 282 }
 283
 284 bool
 285 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
 286 {
 287         return filelayout_test_devid_invalid(node) ||
 288                 nfs4_test_deviceid_unavailable(node);
 289 }
 290
 291 static bool
 292 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
 293 {
 294         struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
 295
 296         return filelayout_test_devid_unavailable(node);
 297 }
 298
 299 /*
 300  * Call ops for the async read/write cases
 301  * In the case of dense layouts, the offset needs to be reset to its
 302  * original value.
 303  */
 304 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 305 {
 306         struct nfs_read_data *rdata = data;
 307
 308         if (unlikely(test_bit(NFS_CONTEXT_BAD, &rdata->args.context->flags))) {
 309                 rpc_exit(task, -EIO);
 310                 return;
 311         }
 312         if (filelayout_reset_to_mds(rdata->header->lseg)) {
 313                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 314                 filelayout_reset_read(rdata);
 315                 rpc_exit(task, 0);
 316                 return;
 317         }
 318         rdata->read_done_cb = filelayout_read_done_cb;
 319
 320         if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
 321                         &rdata->args.seq_args,
 322                         &rdata->res.seq_res,
 323                         task))
 324                 return;
 325         nfs4_set_rw_stateid(&rdata->args.stateid, rdata->args.context,
 326                         rdata->args.lock_context, FMODE_READ);
 327 }
 328
 329 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 330 {
 331         struct nfs_read_data *rdata = data;
 332
 333         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 334
 335         if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) &&
 336             task->tk_status == 0)
 337                 return;
 338
 339         /* Note this may cause RPC to be resent */
 340         rdata->header->mds_ops->rpc_call_done(task, data);
 341 }
 342
 343 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
 344 {
 345         struct nfs_read_data *rdata = data;
 346
 347         rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
 348 }
 349
 350 static void filelayout_read_release(void *data)
 351 {
 352         struct nfs_read_data *rdata = data;
 353         struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout;
 354
 355         filelayout_fenceme(lo->plh_inode, lo);
 356         nfs_put_client(rdata->ds_clp);
 357         rdata->header->mds_ops->rpc_release(data);
 358 }
 359
 360 static int filelayout_write_done_cb(struct rpc_task *task,
 361                                 struct nfs_write_data *data)
 362 {
 363         struct nfs_pgio_header *hdr = data->header;
 364         int err;
 365
 366         err = filelayout_async_handle_error(task, data->args.context->state,
 367                                             data->ds_clp, hdr->lseg);
 368
 369         switch (err) {
 370         case -NFS4ERR_RESET_TO_MDS:
 371                 filelayout_reset_write(data);
 372                 return task->tk_status;
 373         case -EAGAIN:
 374                 rpc_restart_call_prepare(task);
 375                 return -EAGAIN;
 376         }
 377
 378         filelayout_set_layoutcommit(data);
 379         return 0;
 380 }
 381
 382 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 383 static void prepare_to_resend_writes(struct nfs_commit_data *data)
 384 {
 385         struct nfs_page *first = nfs_list_entry(data->pages.next);
 386
 387         data->task.tk_status = 0;
 388         memcpy(&data->verf.verifier, &first->wb_verf,
 389                sizeof(data->verf.verifier));
 390         data->verf.verifier.data[0]++; /* ensure verifier mismatch */
 391 }
 392
 393 static int filelayout_commit_done_cb(struct rpc_task *task,
 394                                      struct nfs_commit_data *data)
 395 {
 396         int err;
 397
 398         err = filelayout_async_handle_error(task, NULL, data->ds_clp,
 399                                             data->lseg);
 400
 401         switch (err) {
 402         case -NFS4ERR_RESET_TO_MDS:
 403                 prepare_to_resend_writes(data);
 404                 return -EAGAIN;
 405         case -EAGAIN:
 406                 rpc_restart_call_prepare(task);
 407                 return -EAGAIN;
 408         }
 409
 410         return 0;
 411 }
 412
 413 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 414 {
 415         struct nfs_write_data *wdata = data;
 416
 417         if (unlikely(test_bit(NFS_CONTEXT_BAD, &wdata->args.context->flags))) {
 418                 rpc_exit(task, -EIO);
 419                 return;
 420         }
 421         if (filelayout_reset_to_mds(wdata->header->lseg)) {
 422                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 423                 filelayout_reset_write(wdata);
 424                 rpc_exit(task, 0);
 425                 return;
 426         }
 427         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 428                         &wdata->args.seq_args,
 429                         &wdata->res.seq_res,
 430                         task))
 431                 return;
 432         nfs4_set_rw_stateid(&wdata->args.stateid, wdata->args.context,
 433                         wdata->args.lock_context, FMODE_WRITE);
 434 }
 435
 436 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 437 {
 438         struct nfs_write_data *wdata = data;
 439
 440         if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) &&
 441             task->tk_status == 0)
 442                 return;
 443
 444         /* Note this may cause RPC to be resent */
 445         wdata->header->mds_ops->rpc_call_done(task, data);
 446 }
 447
 448 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
 449 {
 450         struct nfs_write_data *wdata = data;
 451
 452         rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
 453 }
 454
 455 static void filelayout_write_release(void *data)
 456 {
 457         struct nfs_write_data *wdata = data;
 458         struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout;
 459
 460         filelayout_fenceme(lo->plh_inode, lo);
 461         nfs_put_client(wdata->ds_clp);
 462         wdata->header->mds_ops->rpc_release(data);
 463 }
 464
 465 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
 466 {
 467         struct nfs_commit_data *wdata = data;
 468
 469         nfs41_setup_sequence(wdata->ds_clp->cl_session,
 470                         &wdata->args.seq_args,
 471                         &wdata->res.seq_res,
 472                         task);
 473 }
 474
 475 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
 476 {
 477         struct nfs_commit_data *wdata = data;
 478
 479         /* Note this may cause RPC to be resent */
 480         wdata->mds_ops->rpc_call_done(task, data);
 481 }
 482
 483 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
 484 {
 485         struct nfs_commit_data *cdata = data;
 486
 487         rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
 488 }
 489
 490 static void filelayout_commit_release(void *calldata)
 491 {
 492         struct nfs_commit_data *data = calldata;
 493
 494         data->completion_ops->completion(data);
 495         pnfs_put_lseg(data->lseg);
 496         nfs_put_client(data->ds_clp);
 497         nfs_commitdata_release(data);
 498 }
 499
 500 static const struct rpc_call_ops filelayout_read_call_ops = {
 501         .rpc_call_prepare = filelayout_read_prepare,
 502         .rpc_call_done = filelayout_read_call_done,
 503         .rpc_count_stats = filelayout_read_count_stats,
 504         .rpc_release = filelayout_read_release,
 505 };
 506
 507 static const struct rpc_call_ops filelayout_write_call_ops = {
 508         .rpc_call_prepare = filelayout_write_prepare,
 509         .rpc_call_done = filelayout_write_call_done,
 510         .rpc_count_stats = filelayout_write_count_stats,
 511         .rpc_release = filelayout_write_release,
 512 };
 513
 514 static const struct rpc_call_ops filelayout_commit_call_ops = {
 515         .rpc_call_prepare = filelayout_commit_prepare,
 516         .rpc_call_done = filelayout_write_commit_done,
 517         .rpc_count_stats = filelayout_commit_count_stats,
 518         .rpc_release = filelayout_commit_release,
 519 };
 520
 521 static enum pnfs_try_status
 522 filelayout_read_pagelist(struct nfs_read_data *data)
 523 {
 524         struct nfs_pgio_header *hdr = data->header;
 525         struct pnfs_layout_segment *lseg = hdr->lseg;
 526         struct nfs4_pnfs_ds *ds;
 527         loff_t offset = data->args.offset;
 528         u32 j, idx;
 529         struct nfs_fh *fh;
 530
 531         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 532                 __func__, hdr->inode->i_ino,
 533                 data->args.pgbase, (size_t)data->args.count, offset);
 534
 535         /* Retrieve the correct rpc_client for the byte range */
 536         j = nfs4_fl_calc_j_index(lseg, offset);
 537         idx = nfs4_fl_calc_ds_index(lseg, j);
 538         ds = nfs4_fl_prepare_ds(lseg, idx);
 539         if (!ds)
 540                 return PNFS_NOT_ATTEMPTED;
 541         dprintk("%s USE DS: %s cl_count %d\n", __func__,
 542                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 543
 544         /* No multipath support. Use first DS */
 545         atomic_inc(&ds->ds_clp->cl_count);
 546         data->ds_clp = ds->ds_clp;
 547         fh = nfs4_fl_select_ds_fh(lseg, j);
 548         if (fh)
 549                 data->args.fh = fh;
 550
 551         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 552         data->mds_offset = offset;
 553
 554         /* Perform an asynchronous read to ds */
 555         nfs_initiate_read(ds->ds_clp->cl_rpcclient, data,
 556                                   &filelayout_read_call_ops, RPC_TASK_SOFTCONN);
 557         return PNFS_ATTEMPTED;
 558 }
 559
 560 /* Perform async writes. */
 561 static enum pnfs_try_status
 562 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
 563 {
 564         struct nfs_pgio_header *hdr = data->header;
 565         struct pnfs_layout_segment *lseg = hdr->lseg;
 566         struct nfs4_pnfs_ds *ds;
 567         loff_t offset = data->args.offset;
 568         u32 j, idx;
 569         struct nfs_fh *fh;
 570
 571         /* Retrieve the correct rpc_client for the byte range */
 572         j = nfs4_fl_calc_j_index(lseg, offset);
 573         idx = nfs4_fl_calc_ds_index(lseg, j);
 574         ds = nfs4_fl_prepare_ds(lseg, idx);
 575         if (!ds)
 576                 return PNFS_NOT_ATTEMPTED;
 577         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
 578                 __func__, hdr->inode->i_ino, sync, (size_t) data->args.count,
 579                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 580
 581         data->write_done_cb = filelayout_write_done_cb;
 582         atomic_inc(&ds->ds_clp->cl_count);
 583         data->ds_clp = ds->ds_clp;
 584         fh = nfs4_fl_select_ds_fh(lseg, j);
 585         if (fh)
 586                 data->args.fh = fh;
 587         /*
 588          * Get the file offset on the dserver. Set the write offset to
 589          * this offset and save the original offset.
 590          */
 591         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 592
 593         /* Perform an asynchronous write */
 594         nfs_initiate_write(ds->ds_clp->cl_rpcclient, data,
 595                                     &filelayout_write_call_ops, sync,
 596                                     RPC_TASK_SOFTCONN);
 597         return PNFS_ATTEMPTED;
 598 }
 599
 600 /*
 601  * filelayout_check_layout()
 602  *
 603  * Make sure layout segment parameters are sane WRT the device.
 604  * At this point no generic layer initialization of the lseg has occurred,
 605  * and nothing has been added to the layout_hdr cache.
 606  *
 607  */
 608 static int
 609 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 610                         struct nfs4_filelayout_segment *fl,
 611                         struct nfs4_layoutget_res *lgr,
 612                         struct nfs4_deviceid *id,
 613                         gfp_t gfp_flags)
 614 {
 615         struct nfs4_deviceid_node *d;
 616         struct nfs4_file_layout_dsaddr *dsaddr;
 617         int status = -EINVAL;
 618         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 619
 620         dprintk("--> %s\n", __func__);
 621
 622         /* FIXME: remove this check when layout segment support is added */
 623         if (lgr->range.offset != 0 ||
 624             lgr->range.length != NFS4_MAX_UINT64) {
 625                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 626                         __func__);
 627                 goto out;
 628         }
 629
 630         if (fl->pattern_offset > lgr->range.offset) {
 631                 dprintk("%s pattern_offset %lld too large\n",
 632                                 __func__, fl->pattern_offset);
 633                 goto out;
 634         }
 635
 636         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 637                 dprintk("%s Invalid stripe unit (%u)\n",
 638                         __func__, fl->stripe_unit);
 639                 goto out;
 640         }
 641
 642         /* find and reference the deviceid */
 643         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
 644                                    NFS_SERVER(lo->plh_inode)->nfs_client, id);
 645         if (d == NULL) {
 646                 dsaddr = filelayout_get_device_info(lo->plh_inode, id, gfp_flags);
 647                 if (dsaddr == NULL)
 648                         goto out;
 649         } else
 650                 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 651         /* Found deviceid is unavailable */
 652         if (filelayout_test_devid_unavailable(&dsaddr->id_node))
 653                         goto out_put;
 654
 655         fl->dsaddr = dsaddr;
 656
 657         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 658                 dprintk("%s Bad first_stripe_index %u\n",
 659                                 __func__, fl->first_stripe_index);
 660                 goto out_put;
 661         }
 662
 663         if ((fl->stripe_type == STRIPE_SPARSE &&
 664             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 665             (fl->stripe_type == STRIPE_DENSE &&
 666             fl->num_fh != dsaddr->stripe_count)) {
 667                 dprintk("%s num_fh %u not valid for given packing\n",
 668                         __func__, fl->num_fh);
 669                 goto out_put;
 670         }
 671
 672         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
 673                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
 674                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
 675                         nfss->wsize);
 676         }
 677
 678         status = 0;
 679 out:
 680         dprintk("--> %s returns %d\n", __func__, status);
 681         return status;
 682 out_put:
 683         nfs4_fl_put_deviceid(dsaddr);
 684         goto out;
 685 }
 686
 687 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 688 {
 689         int i;
 690
 691         for (i = 0; i < fl->num_fh; i++) {
 692                 if (!fl->fh_array[i])
 693                         break;
 694                 kfree(fl->fh_array[i]);
 695         }
 696         kfree(fl->fh_array);
 697         fl->fh_array = NULL;
 698 }
 699
 700 static void
 701 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 702 {
 703         filelayout_free_fh_array(fl);
 704         kfree(fl);
 705 }
 706
 707 static int
 708 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 709                          struct nfs4_filelayout_segment *fl,
 710                          struct nfs4_layoutget_res *lgr,
 711                          struct nfs4_deviceid *id,
 712                          gfp_t gfp_flags)
 713 {
 714         struct xdr_stream stream;
 715         struct xdr_buf buf;
 716         struct page *scratch;
 717         __be32 *p;
 718         uint32_t nfl_util;
 719         int i;
 720
 721         dprintk("%s: set_layout_map Begin\n", __func__);
 722
 723         scratch = alloc_page(gfp_flags);
 724         if (!scratch)
 725                 return -ENOMEM;
 726
 727         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 728         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 729
 730         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 731          * num_fh (4) */
 732         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 733         if (unlikely(!p))
 734                 goto out_err;
 735
 736         memcpy(id, p, sizeof(*id));
 737         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 738         nfs4_print_deviceid(id);
 739
 740         nfl_util = be32_to_cpup(p++);
 741         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 742                 fl->commit_through_mds = 1;
 743         if (nfl_util & NFL4_UFLG_DENSE)
 744                 fl->stripe_type = STRIPE_DENSE;
 745         else
 746                 fl->stripe_type = STRIPE_SPARSE;
 747         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 748
 749         fl->first_stripe_index = be32_to_cpup(p++);
 750         p = xdr_decode_hyper(p, &fl->pattern_offset);
 751         fl->num_fh = be32_to_cpup(p++);
 752
 753         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 754                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 755                 fl->pattern_offset);
 756
 757         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 758          * Futher checking is done in filelayout_check_layout */
 759         if (fl->num_fh >
 760             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 761                 goto out_err;
 762
 763         if (fl->num_fh > 0) {
 764                 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
 765                                        gfp_flags);
 766                 if (!fl->fh_array)
 767                         goto out_err;
 768         }
 769
 770         for (i = 0; i < fl->num_fh; i++) {
 771                 /* Do we want to use a mempool here? */
 772                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 773                 if (!fl->fh_array[i])
 774                         goto out_err_free;
 775
 776                 p = xdr_inline_decode(&stream, 4);
 777                 if (unlikely(!p))
 778                         goto out_err_free;
 779                 fl->fh_array[i]->size = be32_to_cpup(p++);
 780                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 781                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
 782                                i, fl->fh_array[i]->size);
 783                         goto out_err_free;
 784                 }
 785
 786                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 787                 if (unlikely(!p))
 788                         goto out_err_free;
 789                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 790                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 791                         fl->fh_array[i]->size);
 792         }
 793
 794         __free_page(scratch);
 795         return 0;
 796
 797 out_err_free:
 798         filelayout_free_fh_array(fl);
 799 out_err:
 800         __free_page(scratch);
 801         return -EIO;
 802 }
 803
 804 static void
 805 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 806 {
 807         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 808
 809         dprintk("--> %s\n", __func__);
 810         nfs4_fl_put_deviceid(fl->dsaddr);
 811         /* This assumes a single RW lseg */
 812         if (lseg->pls_range.iomode == IOMODE_RW) {
 813                 struct nfs4_filelayout *flo;
 814
 815                 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
 816                 flo->commit_info.nbuckets = 0;
 817                 kfree(flo->commit_info.buckets);
 818                 flo->commit_info.buckets = NULL;
 819         }
 820         _filelayout_free_lseg(fl);
 821 }
 822
 823 static int
 824 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 825                              struct nfs_commit_info *cinfo,
 826                              gfp_t gfp_flags)
 827 {
 828         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 829         struct pnfs_commit_bucket *buckets;
 830         int size;
 831
 832         if (fl->commit_through_mds)
 833                 return 0;
 834         if (cinfo->ds->nbuckets != 0) {
 835                 /* This assumes there is only one IOMODE_RW lseg.  What
 836                  * we really want to do is have a layout_hdr level
 837                  * dictionary of <multipath_list4, fh> keys, each
 838                  * associated with a struct list_head, populated by calls
 839                  * to filelayout_write_pagelist().
 840                  * */
 841                 return 0;
 842         }
 843
 844         size = (fl->stripe_type == STRIPE_SPARSE) ?
 845                 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 846
 847         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 848                           gfp_flags);
 849         if (!buckets)
 850                 return -ENOMEM;
 851         else {
 852                 int i;
 853
 854                 spin_lock(cinfo->lock);
 855                 if (cinfo->ds->nbuckets != 0)
 856                         kfree(buckets);
 857                 else {
 858                         cinfo->ds->buckets = buckets;
 859                         cinfo->ds->nbuckets = size;
 860                         for (i = 0; i < size; i++) {
 861                                 INIT_LIST_HEAD(&buckets[i].written);
 862                                 INIT_LIST_HEAD(&buckets[i].committing);
 863                         }
 864                 }
 865                 spin_unlock(cinfo->lock);
 866                 return 0;
 867         }
 868 }
 869
 870 static struct pnfs_layout_segment *
 871 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 872                       struct nfs4_layoutget_res *lgr,
 873                       gfp_t gfp_flags)
 874 {
 875         struct nfs4_filelayout_segment *fl;
 876         int rc;
 877         struct nfs4_deviceid id;
 878
 879         dprintk("--> %s\n", __func__);
 880         fl = kzalloc(sizeof(*fl), gfp_flags);
 881         if (!fl)
 882                 return NULL;
 883
 884         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 885         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 886                 _filelayout_free_lseg(fl);
 887                 return NULL;
 888         }
 889         return &fl->generic_hdr;
 890 }
 891
 892 /*
 893  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 894  *
 895  * return true  : coalesce page
 896  * return false : don't coalesce page
 897  */
 898 static bool
 899 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 900                    struct nfs_page *req)
 901 {
 902         u64 p_stripe, r_stripe;
 903         u32 stripe_unit;
 904
 905         if (!pnfs_generic_pg_test(pgio, prev, req) ||
 906             !nfs_generic_pg_test(pgio, prev, req))
 907                 return false;
 908
 909         p_stripe = (u64)req_offset(prev);
 910         r_stripe = (u64)req_offset(req);
 911         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 912
 913         do_div(p_stripe, stripe_unit);
 914         do_div(r_stripe, stripe_unit);
 915
 916         return (p_stripe == r_stripe);
 917 }
 918
 919 static void
 920 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 921                         struct nfs_page *req)
 922 {
 923         WARN_ON_ONCE(pgio->pg_lseg != NULL);
 924
 925         if (req->wb_offset != req->wb_pgbase) {
 926                 /*
 927                  * Handling unaligned pages is difficult, because have to
 928                  * somehow split a req in two in certain cases in the
 929                  * pg.test code.  Avoid this by just not using pnfs
 930                  * in this case.
 931                  */
 932                 nfs_pageio_reset_read_mds(pgio);
 933                 return;
 934         }
 935         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 936                                            req->wb_context,
 937                                            0,
 938                                            NFS4_MAX_UINT64,
 939                                            IOMODE_READ,
 940                                            GFP_KERNEL);
 941         /* If no lseg, fall back to read through mds */
 942         if (pgio->pg_lseg == NULL)
 943                 nfs_pageio_reset_read_mds(pgio);
 944 }
 945
 946 static void
 947 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 948                          struct nfs_page *req)
 949 {
 950         struct nfs_commit_info cinfo;
 951         int status;
 952
 953         WARN_ON_ONCE(pgio->pg_lseg != NULL);
 954
 955         if (req->wb_offset != req->wb_pgbase)
 956                 goto out_mds;
 957         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 958                                            req->wb_context,
 959                                            0,
 960                                            NFS4_MAX_UINT64,
 961                                            IOMODE_RW,
 962                                            GFP_NOFS);
 963         /* If no lseg, fall back to write through mds */
 964         if (pgio->pg_lseg == NULL)
 965                 goto out_mds;
 966         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 967         status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 968         if (status < 0) {
 969                 pnfs_put_lseg(pgio->pg_lseg);
 970                 pgio->pg_lseg = NULL;
 971                 goto out_mds;
 972         }
 973         return;
 974 out_mds:
 975         nfs_pageio_reset_write_mds(pgio);
 976 }
 977
 978 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 979         .pg_init = filelayout_pg_init_read,
 980         .pg_test = filelayout_pg_test,
 981         .pg_doio = pnfs_generic_pg_readpages,
 982 };
 983
 984 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 985         .pg_init = filelayout_pg_init_write,
 986         .pg_test = filelayout_pg_test,
 987         .pg_doio = pnfs_generic_pg_writepages,
 988 };
 989
 990 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 991 {
 992         if (fl->stripe_type == STRIPE_SPARSE)
 993                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
 994         else
 995                 return j;
 996 }
 997
 998 /* The generic layer is about to remove the req from the commit list.
 999  * If this will make the bucket empty, it will need to put the lseg reference.
1000  */
1001 static void
1002 filelayout_clear_request_commit(struct nfs_page *req,
1003                                 struct nfs_commit_info *cinfo)
1004 {
1005         struct pnfs_layout_segment *freeme = NULL;
1006
1007         spin_lock(cinfo->lock);
1008         if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
1009                 goto out;
1010         cinfo->ds->nwritten--;
1011         if (list_is_singular(&req->wb_list)) {
1012                 struct pnfs_commit_bucket *bucket;
1013
1014                 bucket = list_first_entry(&req->wb_list,
1015                                           struct pnfs_commit_bucket,
1016                                           written);
1017                 freeme = bucket->wlseg;
1018                 bucket->wlseg = NULL;
1019         }
1020 out:
1021         nfs_request_remove_commit_list(req, cinfo);
1022         spin_unlock(cinfo->lock);
1023         pnfs_put_lseg(freeme);
1024 }
1025
1026 static struct list_head *
1027 filelayout_choose_commit_list(struct nfs_page *req,
1028                               struct pnfs_layout_segment *lseg,
1029                               struct nfs_commit_info *cinfo)
1030 {
1031         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1032         u32 i, j;
1033         struct list_head *list;
1034         struct pnfs_commit_bucket *buckets;
1035
1036         if (fl->commit_through_mds)
1037                 return &cinfo->mds->list;
1038
1039         /* Note that we are calling nfs4_fl_calc_j_index on each page
1040          * that ends up being committed to a data server.  An attractive
1041          * alternative is to add a field to nfs_write_data and nfs_page
1042          * to store the value calculated in filelayout_write_pagelist
1043          * and just use that here.
1044          */
1045         j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1046         i = select_bucket_index(fl, j);
1047         buckets = cinfo->ds->buckets;
1048         list = &buckets[i].written;
1049         if (list_empty(list)) {
1050                 /* Non-empty buckets hold a reference on the lseg.  That ref
1051                  * is normally transferred to the COMMIT call and released
1052                  * there.  It could also be released if the last req is pulled
1053                  * off due to a rewrite, in which case it will be done in
1054                  * filelayout_clear_request_commit
1055                  */
1056                 buckets[i].wlseg = pnfs_get_lseg(lseg);
1057         }
1058         set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1059         cinfo->ds->nwritten++;
1060         return list;
1061 }
1062
1063 static void
1064 filelayout_mark_request_commit(struct nfs_page *req,
1065                                struct pnfs_layout_segment *lseg,
1066                                struct nfs_commit_info *cinfo)
1067 {
1068         struct list_head *list;
1069
1070         list = filelayout_choose_commit_list(req, lseg, cinfo);
1071         nfs_request_add_commit_list(req, list, cinfo);
1072 }
1073
1074 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1075 {
1076         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1077
1078         if (flseg->stripe_type == STRIPE_SPARSE)
1079                 return i;
1080         else
1081                 return nfs4_fl_calc_ds_index(lseg, i);
1082 }
1083
1084 static struct nfs_fh *
1085 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1086 {
1087         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1088
1089         if (flseg->stripe_type == STRIPE_SPARSE) {
1090                 if (flseg->num_fh == 1)
1091                         i = 0;
1092                 else if (flseg->num_fh == 0)
1093                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1094                         return NULL;
1095         }
1096         return flseg->fh_array[i];
1097 }
1098
1099 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1100 {
1101         struct pnfs_layout_segment *lseg = data->lseg;
1102         struct nfs4_pnfs_ds *ds;
1103         u32 idx;
1104         struct nfs_fh *fh;
1105
1106         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1107         ds = nfs4_fl_prepare_ds(lseg, idx);
1108         if (!ds) {
1109                 prepare_to_resend_writes(data);
1110                 filelayout_commit_release(data);
1111                 return -EAGAIN;
1112         }
1113         dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1114                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1115         data->commit_done_cb = filelayout_commit_done_cb;
1116         atomic_inc(&ds->ds_clp->cl_count);
1117         data->ds_clp = ds->ds_clp;
1118         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1119         if (fh)
1120                 data->args.fh = fh;
1121         return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data,
1122                                    &filelayout_commit_call_ops, how,
1123                                    RPC_TASK_SOFTCONN);
1124 }
1125
1126 static int
1127 transfer_commit_list(struct list_head *src, struct list_head *dst,
1128                      struct nfs_commit_info *cinfo, int max)
1129 {
1130         struct nfs_page *req, *tmp;
1131         int ret = 0;
1132
1133         list_for_each_entry_safe(req, tmp, src, wb_list) {
1134                 if (!nfs_lock_request(req))
1135                         continue;
1136                 kref_get(&req->wb_kref);
1137                 if (cond_resched_lock(cinfo->lock))
1138                         list_safe_reset_next(req, tmp, wb_list);
1139                 nfs_request_remove_commit_list(req, cinfo);
1140                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1141                 nfs_list_add_request(req, dst);
1142                 ret++;
1143                 if ((ret == max) && !cinfo->dreq)
1144                         break;
1145         }
1146         return ret;
1147 }
1148
1149 static int
1150 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1151                                struct nfs_commit_info *cinfo,
1152                                int max)
1153 {
1154         struct list_head *src = &bucket->written;
1155         struct list_head *dst = &bucket->committing;
1156         int ret;
1157
1158         ret = transfer_commit_list(src, dst, cinfo, max);
1159         if (ret) {
1160                 cinfo->ds->nwritten -= ret;
1161                 cinfo->ds->ncommitting += ret;
1162                 bucket->clseg = bucket->wlseg;
1163                 if (list_empty(src))
1164                         bucket->wlseg = NULL;
1165                 else
1166                         pnfs_get_lseg(bucket->clseg);
1167         }
1168         return ret;
1169 }
1170
1171 /* Move reqs from written to committing lists, returning count of number moved.
1172  * Note called with cinfo->lock held.
1173  */
1174 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1175                                         int max)
1176 {
1177         int i, rv = 0, cnt;
1178
1179         for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1180                 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1181                                                      cinfo, max);
1182                 max -= cnt;
1183                 rv += cnt;
1184         }
1185         return rv;
1186 }
1187
1188 /* Pull everything off the committing lists and dump into @dst */
1189 static void filelayout_recover_commit_reqs(struct list_head *dst,
1190                                            struct nfs_commit_info *cinfo)
1191 {
1192         struct pnfs_commit_bucket *b;
1193         int i;
1194
1195         /* NOTE cinfo->lock is NOT held, relying on fact that this is
1196          * only called on single thread per dreq.
1197          * Can't take the lock because need to do pnfs_put_lseg
1198          */
1199         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1200                 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1201                         pnfs_put_lseg(b->wlseg);
1202                         b->wlseg = NULL;
1203                 }
1204         }
1205         cinfo->ds->nwritten = 0;
1206 }
1207
1208 static unsigned int
1209 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1210 {
1211         struct pnfs_ds_commit_info *fl_cinfo;
1212         struct pnfs_commit_bucket *bucket;
1213         struct nfs_commit_data *data;
1214         int i, j;
1215         unsigned int nreq = 0;
1216
1217         fl_cinfo = cinfo->ds;
1218         bucket = fl_cinfo->buckets;
1219         for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1220                 if (list_empty(&bucket->committing))
1221                         continue;
1222                 data = nfs_commitdata_alloc();
1223                 if (!data)
1224                         break;
1225                 data->ds_commit_index = i;
1226                 data->lseg = bucket->clseg;
1227                 bucket->clseg = NULL;
1228                 list_add(&data->pages, list);
1229                 nreq++;
1230         }
1231
1232         /* Clean up on error */
1233         for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1234                 if (list_empty(&bucket->committing))
1235                         continue;
1236                 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1237                 pnfs_put_lseg(bucket->clseg);
1238                 bucket->clseg = NULL;
1239         }
1240         /* Caller will clean up entries put on list */
1241         return nreq;
1242 }
1243
1244 /* This follows nfs_commit_list pretty closely */
1245 static int
1246 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1247                            int how, struct nfs_commit_info *cinfo)
1248 {
1249         struct nfs_commit_data *data, *tmp;
1250         LIST_HEAD(list);
1251         unsigned int nreq = 0;
1252
1253         if (!list_empty(mds_pages)) {
1254                 data = nfs_commitdata_alloc();
1255                 if (data != NULL) {
1256                         data->lseg = NULL;
1257                         list_add(&data->pages, &list);
1258                         nreq++;
1259                 } else
1260                         nfs_retry_commit(mds_pages, NULL, cinfo);
1261         }
1262
1263         nreq += alloc_ds_commits(cinfo, &list);
1264
1265         if (nreq == 0) {
1266                 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1267                 goto out;
1268         }
1269
1270         atomic_add(nreq, &cinfo->mds->rpcs_out);
1271
1272         list_for_each_entry_safe(data, tmp, &list, pages) {
1273                 list_del_init(&data->pages);
1274                 if (!data->lseg) {
1275                         nfs_init_commit(data, mds_pages, NULL, cinfo);
1276                         nfs_initiate_commit(NFS_CLIENT(inode), data,
1277                                             data->mds_ops, how, 0);
1278                 } else {
1279                         struct pnfs_commit_bucket *buckets;
1280
1281                         buckets = cinfo->ds->buckets;
1282                         nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1283                         filelayout_initiate_commit(data, how);
1284                 }
1285         }
1286 out:
1287         cinfo->ds->ncommitting = 0;
1288         return PNFS_ATTEMPTED;
1289 }
1290
1291 static void
1292 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1293 {
1294         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1295 }
1296
1297 static struct pnfs_layout_hdr *
1298 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1299 {
1300         struct nfs4_filelayout *flo;
1301
1302         flo = kzalloc(sizeof(*flo), gfp_flags);
1303         return &flo->generic_hdr;
1304 }
1305
1306 static void
1307 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1308 {
1309         kfree(FILELAYOUT_FROM_HDR(lo));
1310 }
1311
1312 static struct pnfs_ds_commit_info *
1313 filelayout_get_ds_info(struct inode *inode)
1314 {
1315         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1316
1317         if (layout == NULL)
1318                 return NULL;
1319         else
1320                 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1321 }
1322
1323 static struct pnfs_layoutdriver_type filelayout_type = {
1324         .id                     = LAYOUT_NFSV4_1_FILES,
1325         .name                   = "LAYOUT_NFSV4_1_FILES",
1326         .owner                  = THIS_MODULE,
1327         .alloc_layout_hdr       = filelayout_alloc_layout_hdr,
1328         .free_layout_hdr        = filelayout_free_layout_hdr,
1329         .alloc_lseg             = filelayout_alloc_lseg,
1330         .free_lseg              = filelayout_free_lseg,
1331         .pg_read_ops            = &filelayout_pg_read_ops,
1332         .pg_write_ops           = &filelayout_pg_write_ops,
1333         .get_ds_info            = &filelayout_get_ds_info,
1334         .mark_request_commit    = filelayout_mark_request_commit,
1335         .clear_request_commit   = filelayout_clear_request_commit,
1336         .scan_commit_lists      = filelayout_scan_commit_lists,
1337         .recover_commit_reqs    = filelayout_recover_commit_reqs,
1338         .commit_pagelist        = filelayout_commit_pagelist,
1339         .read_pagelist          = filelayout_read_pagelist,
1340         .write_pagelist         = filelayout_write_pagelist,
1341         .free_deviceid_node     = filelayout_free_deveiceid_node,
1342 };
1343
1344 static int __init nfs4filelayout_init(void)
1345 {
1346         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1347                __func__);
1348         return pnfs_register_layoutdriver(&filelayout_type);
1349 }
1350
1351 static void __exit nfs4filelayout_exit(void)
1352 {
1353         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1354                __func__);
1355         pnfs_unregister_layoutdriver(&filelayout_type);
1356 }
1357
1358 MODULE_ALIAS("nfs-layouttype4-1");
1359
1360 module_init(nfs4filelayout_init);
1361 module_exit(nfs4filelayout_exit);