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 #include "nfs4trace.h"
  43
  44 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  45
  46 MODULE_LICENSE("GPL");
  47 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
  48 MODULE_DESCRIPTION("The NFSv4 file layout driver");
  49
  50 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
  51
  52 static loff_t
  53 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
  54                             loff_t offset)
  55 {
  56         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
  57         u64 stripe_no;
  58         u32 rem;
  59
  60         offset -= flseg->pattern_offset;
  61         stripe_no = div_u64(offset, stripe_width);
  62         div_u64_rem(offset, flseg->stripe_unit, &rem);
  63
  64         return stripe_no * flseg->stripe_unit + rem;
  65 }
  66
  67 /* This function is used by the layout driver to calculate the
  68  * offset of the file on the dserver based on whether the
  69  * layout type is STRIPE_DENSE or STRIPE_SPARSE
  70  */
  71 static loff_t
  72 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
  73 {
  74         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
  75
  76         switch (flseg->stripe_type) {
  77         case STRIPE_SPARSE:
  78                 return offset;
  79
  80         case STRIPE_DENSE:
  81                 return filelayout_get_dense_offset(flseg, offset);
  82         }
  83
  84         BUG();
  85 }
  86
  87 static void filelayout_reset_write(struct nfs_pgio_data *data)
  88 {
  89         struct nfs_pgio_header *hdr = data->header;
  90         struct rpc_task *task = &data->task;
  91
  92         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
  93                 dprintk("%s Reset task %5u for i/o through MDS "
  94                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
  95                         data->task.tk_pid,
  96                         hdr->inode->i_sb->s_id,
  97                         (unsigned long long)NFS_FILEID(hdr->inode),
  98                         data->args.count,
  99                         (unsigned long long)data->args.offset);
 100
 101                 task->tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
 102                                                         &hdr->pages,
 103                                                         hdr->completion_ops,
 104                                                         hdr->dreq);
 105         }
 106 }
 107
 108 static void filelayout_reset_read(struct nfs_pgio_data *data)
 109 {
 110         struct nfs_pgio_header *hdr = data->header;
 111         struct rpc_task *task = &data->task;
 112
 113         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 114                 dprintk("%s Reset task %5u for i/o through MDS "
 115                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
 116                         data->task.tk_pid,
 117                         hdr->inode->i_sb->s_id,
 118                         (unsigned long long)NFS_FILEID(hdr->inode),
 119                         data->args.count,
 120                         (unsigned long long)data->args.offset);
 121
 122                 task->tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
 123                                                         &hdr->pages,
 124                                                         hdr->completion_ops,
 125                                                         hdr->dreq);
 126         }
 127 }
 128
 129 static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo)
 130 {
 131         if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 132                 return;
 133         pnfs_return_layout(inode);
 134 }
 135
 136 static int filelayout_async_handle_error(struct rpc_task *task,
 137                                          struct nfs4_state *state,
 138                                          struct nfs_client *clp,
 139                                          struct pnfs_layout_segment *lseg)
 140 {
 141         struct pnfs_layout_hdr *lo = lseg->pls_layout;
 142         struct inode *inode = lo->plh_inode;
 143         struct nfs_server *mds_server = NFS_SERVER(inode);
 144         struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
 145         struct nfs_client *mds_client = mds_server->nfs_client;
 146         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
 147
 148         if (task->tk_status >= 0)
 149                 return 0;
 150
 151         switch (task->tk_status) {
 152         /* MDS state errors */
 153         case -NFS4ERR_DELEG_REVOKED:
 154         case -NFS4ERR_ADMIN_REVOKED:
 155         case -NFS4ERR_BAD_STATEID:
 156                 if (state == NULL)
 157                         break;
 158                 nfs_remove_bad_delegation(state->inode);
 159         case -NFS4ERR_OPENMODE:
 160                 if (state == NULL)
 161                         break;
 162                 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
 163                         goto out_bad_stateid;
 164                 goto wait_on_recovery;
 165         case -NFS4ERR_EXPIRED:
 166                 if (state != NULL) {
 167                         if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
 168                                 goto out_bad_stateid;
 169                 }
 170                 nfs4_schedule_lease_recovery(mds_client);
 171                 goto wait_on_recovery;
 172         /* DS session errors */
 173         case -NFS4ERR_BADSESSION:
 174         case -NFS4ERR_BADSLOT:
 175         case -NFS4ERR_BAD_HIGH_SLOT:
 176         case -NFS4ERR_DEADSESSION:
 177         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
 178         case -NFS4ERR_SEQ_FALSE_RETRY:
 179         case -NFS4ERR_SEQ_MISORDERED:
 180                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 181                         "flags 0x%x\n", __func__, task->tk_status,
 182                         clp->cl_exchange_flags);
 183                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
 184                 break;
 185         case -NFS4ERR_DELAY:
 186         case -NFS4ERR_GRACE:
 187                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 188                 break;
 189         case -NFS4ERR_RETRY_UNCACHED_REP:
 190                 break;
 191         /* Invalidate Layout errors */
 192         case -NFS4ERR_PNFS_NO_LAYOUT:
 193         case -ESTALE:           /* mapped NFS4ERR_STALE */
 194         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
 195         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
 196         case -NFS4ERR_FHEXPIRED:
 197         case -NFS4ERR_WRONG_TYPE:
 198                 dprintk("%s Invalid layout error %d\n", __func__,
 199                         task->tk_status);
 200                 /*
 201                  * Destroy layout so new i/o will get a new layout.
 202                  * Layout will not be destroyed until all current lseg
 203                  * references are put. Mark layout as invalid to resend failed
 204                  * i/o and all i/o waiting on the slot table to the MDS until
 205                  * layout is destroyed and a new valid layout is obtained.
 206                  */
 207                 pnfs_destroy_layout(NFS_I(inode));
 208                 rpc_wake_up(&tbl->slot_tbl_waitq);
 209                 goto reset;
 210         /* RPC connection errors */
 211         case -ECONNREFUSED:
 212         case -EHOSTDOWN:
 213         case -EHOSTUNREACH:
 214         case -ENETUNREACH:
 215         case -EIO:
 216         case -ETIMEDOUT:
 217         case -EPIPE:
 218                 dprintk("%s DS connection error %d\n", __func__,
 219                         task->tk_status);
 220                 nfs4_mark_deviceid_unavailable(devid);
 221                 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
 222                 rpc_wake_up(&tbl->slot_tbl_waitq);
 223                 /* fall through */
 224         default:
 225 reset:
 226                 dprintk("%s Retry through MDS. Error %d\n", __func__,
 227                         task->tk_status);
 228                 return -NFS4ERR_RESET_TO_MDS;
 229         }
 230 out:
 231         task->tk_status = 0;
 232         return -EAGAIN;
 233 out_bad_stateid:
 234         task->tk_status = -EIO;
 235         return 0;
 236 wait_on_recovery:
 237         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
 238         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
 239                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
 240         goto out;
 241 }
 242
 243 /* NFS_PROTO call done callback routines */
 244
 245 static int filelayout_read_done_cb(struct rpc_task *task,
 246                                 struct nfs_pgio_data *data)
 247 {
 248         struct nfs_pgio_header *hdr = data->header;
 249         int err;
 250
 251         trace_nfs4_pnfs_read(data, task->tk_status);
 252         err = filelayout_async_handle_error(task, data->args.context->state,
 253                                             data->ds_clp, hdr->lseg);
 254
 255         switch (err) {
 256         case -NFS4ERR_RESET_TO_MDS:
 257                 filelayout_reset_read(data);
 258                 return task->tk_status;
 259         case -EAGAIN:
 260                 rpc_restart_call_prepare(task);
 261                 return -EAGAIN;
 262         }
 263
 264         return 0;
 265 }
 266
 267 /*
 268  * We reference the rpc_cred of the first WRITE that triggers the need for
 269  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 270  * rfc5661 is not clear about which credential should be used.
 271  */
 272 static void
 273 filelayout_set_layoutcommit(struct nfs_pgio_data *wdata)
 274 {
 275         struct nfs_pgio_header *hdr = wdata->header;
 276
 277         if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
 278             wdata->res.verf->committed == NFS_FILE_SYNC)
 279                 return;
 280
 281         pnfs_set_layoutcommit(wdata);
 282         dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
 283                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
 284 }
 285
 286 bool
 287 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
 288 {
 289         return filelayout_test_devid_invalid(node) ||
 290                 nfs4_test_deviceid_unavailable(node);
 291 }
 292
 293 static bool
 294 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
 295 {
 296         struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
 297
 298         return filelayout_test_devid_unavailable(node);
 299 }
 300
 301 /*
 302  * Call ops for the async read/write cases
 303  * In the case of dense layouts, the offset needs to be reset to its
 304  * original value.
 305  */
 306 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 307 {
 308         struct nfs_pgio_data *rdata = data;
 309
 310         if (unlikely(test_bit(NFS_CONTEXT_BAD, &rdata->args.context->flags))) {
 311                 rpc_exit(task, -EIO);
 312                 return;
 313         }
 314         if (filelayout_reset_to_mds(rdata->header->lseg)) {
 315                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 316                 filelayout_reset_read(rdata);
 317                 rpc_exit(task, 0);
 318                 return;
 319         }
 320         rdata->pgio_done_cb = filelayout_read_done_cb;
 321
 322         if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
 323                         &rdata->args.seq_args,
 324                         &rdata->res.seq_res,
 325                         task))
 326                 return;
 327         if (nfs4_set_rw_stateid(&rdata->args.stateid, rdata->args.context,
 328                         rdata->args.lock_context, FMODE_READ) == -EIO)
 329                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
 330 }
 331
 332 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 333 {
 334         struct nfs_pgio_data *rdata = data;
 335
 336         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 337
 338         if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) &&
 339             task->tk_status == 0) {
 340                 nfs41_sequence_done(task, &rdata->res.seq_res);
 341                 return;
 342         }
 343
 344         /* Note this may cause RPC to be resent */
 345         rdata->header->mds_ops->rpc_call_done(task, data);
 346 }
 347
 348 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
 349 {
 350         struct nfs_pgio_data *rdata = data;
 351
 352         rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
 353 }
 354
 355 static void filelayout_read_release(void *data)
 356 {
 357         struct nfs_pgio_data *rdata = data;
 358         struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout;
 359
 360         filelayout_fenceme(lo->plh_inode, lo);
 361         nfs_put_client(rdata->ds_clp);
 362         rdata->header->mds_ops->rpc_release(data);
 363 }
 364
 365 static int filelayout_write_done_cb(struct rpc_task *task,
 366                                 struct nfs_pgio_data *data)
 367 {
 368         struct nfs_pgio_header *hdr = data->header;
 369         int err;
 370
 371         trace_nfs4_pnfs_write(data, task->tk_status);
 372         err = filelayout_async_handle_error(task, data->args.context->state,
 373                                             data->ds_clp, hdr->lseg);
 374
 375         switch (err) {
 376         case -NFS4ERR_RESET_TO_MDS:
 377                 filelayout_reset_write(data);
 378                 return task->tk_status;
 379         case -EAGAIN:
 380                 rpc_restart_call_prepare(task);
 381                 return -EAGAIN;
 382         }
 383
 384         filelayout_set_layoutcommit(data);
 385         return 0;
 386 }
 387
 388 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 389 static void prepare_to_resend_writes(struct nfs_commit_data *data)
 390 {
 391         struct nfs_page *first = nfs_list_entry(data->pages.next);
 392
 393         data->task.tk_status = 0;
 394         memcpy(&data->verf.verifier, &first->wb_verf,
 395                sizeof(data->verf.verifier));
 396         data->verf.verifier.data[0]++; /* ensure verifier mismatch */
 397 }
 398
 399 static int filelayout_commit_done_cb(struct rpc_task *task,
 400                                      struct nfs_commit_data *data)
 401 {
 402         int err;
 403
 404         trace_nfs4_pnfs_commit_ds(data, task->tk_status);
 405         err = filelayout_async_handle_error(task, NULL, data->ds_clp,
 406                                             data->lseg);
 407
 408         switch (err) {
 409         case -NFS4ERR_RESET_TO_MDS:
 410                 prepare_to_resend_writes(data);
 411                 return -EAGAIN;
 412         case -EAGAIN:
 413                 rpc_restart_call_prepare(task);
 414                 return -EAGAIN;
 415         }
 416
 417         return 0;
 418 }
 419
 420 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 421 {
 422         struct nfs_pgio_data *wdata = data;
 423
 424         if (unlikely(test_bit(NFS_CONTEXT_BAD, &wdata->args.context->flags))) {
 425                 rpc_exit(task, -EIO);
 426                 return;
 427         }
 428         if (filelayout_reset_to_mds(wdata->header->lseg)) {
 429                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 430                 filelayout_reset_write(wdata);
 431                 rpc_exit(task, 0);
 432                 return;
 433         }
 434         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 435                         &wdata->args.seq_args,
 436                         &wdata->res.seq_res,
 437                         task))
 438                 return;
 439         if (nfs4_set_rw_stateid(&wdata->args.stateid, wdata->args.context,
 440                         wdata->args.lock_context, FMODE_WRITE) == -EIO)
 441                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
 442 }
 443
 444 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 445 {
 446         struct nfs_pgio_data *wdata = data;
 447
 448         if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) &&
 449             task->tk_status == 0) {
 450                 nfs41_sequence_done(task, &wdata->res.seq_res);
 451                 return;
 452         }
 453
 454         /* Note this may cause RPC to be resent */
 455         wdata->header->mds_ops->rpc_call_done(task, data);
 456 }
 457
 458 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
 459 {
 460         struct nfs_pgio_data *wdata = data;
 461
 462         rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
 463 }
 464
 465 static void filelayout_write_release(void *data)
 466 {
 467         struct nfs_pgio_data *wdata = data;
 468         struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout;
 469
 470         filelayout_fenceme(lo->plh_inode, lo);
 471         nfs_put_client(wdata->ds_clp);
 472         wdata->header->mds_ops->rpc_release(data);
 473 }
 474
 475 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
 476 {
 477         struct nfs_commit_data *wdata = data;
 478
 479         nfs41_setup_sequence(wdata->ds_clp->cl_session,
 480                         &wdata->args.seq_args,
 481                         &wdata->res.seq_res,
 482                         task);
 483 }
 484
 485 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
 486 {
 487         struct nfs_commit_data *wdata = data;
 488
 489         /* Note this may cause RPC to be resent */
 490         wdata->mds_ops->rpc_call_done(task, data);
 491 }
 492
 493 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
 494 {
 495         struct nfs_commit_data *cdata = data;
 496
 497         rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
 498 }
 499
 500 static void filelayout_commit_release(void *calldata)
 501 {
 502         struct nfs_commit_data *data = calldata;
 503
 504         data->completion_ops->completion(data);
 505         pnfs_put_lseg(data->lseg);
 506         nfs_put_client(data->ds_clp);
 507         nfs_commitdata_release(data);
 508 }
 509
 510 static const struct rpc_call_ops filelayout_read_call_ops = {
 511         .rpc_call_prepare = filelayout_read_prepare,
 512         .rpc_call_done = filelayout_read_call_done,
 513         .rpc_count_stats = filelayout_read_count_stats,
 514         .rpc_release = filelayout_read_release,
 515 };
 516
 517 static const struct rpc_call_ops filelayout_write_call_ops = {
 518         .rpc_call_prepare = filelayout_write_prepare,
 519         .rpc_call_done = filelayout_write_call_done,
 520         .rpc_count_stats = filelayout_write_count_stats,
 521         .rpc_release = filelayout_write_release,
 522 };
 523
 524 static const struct rpc_call_ops filelayout_commit_call_ops = {
 525         .rpc_call_prepare = filelayout_commit_prepare,
 526         .rpc_call_done = filelayout_write_commit_done,
 527         .rpc_count_stats = filelayout_commit_count_stats,
 528         .rpc_release = filelayout_commit_release,
 529 };
 530
 531 static enum pnfs_try_status
 532 filelayout_read_pagelist(struct nfs_pgio_data *data)
 533 {
 534         struct nfs_pgio_header *hdr = data->header;
 535         struct pnfs_layout_segment *lseg = hdr->lseg;
 536         struct nfs4_pnfs_ds *ds;
 537         struct rpc_clnt *ds_clnt;
 538         loff_t offset = data->args.offset;
 539         u32 j, idx;
 540         struct nfs_fh *fh;
 541
 542         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 543                 __func__, hdr->inode->i_ino,
 544                 data->args.pgbase, (size_t)data->args.count, offset);
 545
 546         /* Retrieve the correct rpc_client for the byte range */
 547         j = nfs4_fl_calc_j_index(lseg, offset);
 548         idx = nfs4_fl_calc_ds_index(lseg, j);
 549         ds = nfs4_fl_prepare_ds(lseg, idx);
 550         if (!ds)
 551                 return PNFS_NOT_ATTEMPTED;
 552
 553         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
 554         if (IS_ERR(ds_clnt))
 555                 return PNFS_NOT_ATTEMPTED;
 556
 557         dprintk("%s USE DS: %s cl_count %d\n", __func__,
 558                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 559
 560         /* No multipath support. Use first DS */
 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         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 568         data->mds_offset = offset;
 569
 570         /* Perform an asynchronous read to ds */
 571         nfs_initiate_pgio(ds_clnt, data,
 572                             &filelayout_read_call_ops, 0, RPC_TASK_SOFTCONN);
 573         return PNFS_ATTEMPTED;
 574 }
 575
 576 /* Perform async writes. */
 577 static enum pnfs_try_status
 578 filelayout_write_pagelist(struct nfs_pgio_data *data, int sync)
 579 {
 580         struct nfs_pgio_header *hdr = data->header;
 581         struct pnfs_layout_segment *lseg = hdr->lseg;
 582         struct nfs4_pnfs_ds *ds;
 583         struct rpc_clnt *ds_clnt;
 584         loff_t offset = data->args.offset;
 585         u32 j, idx;
 586         struct nfs_fh *fh;
 587
 588         /* Retrieve the correct rpc_client for the byte range */
 589         j = nfs4_fl_calc_j_index(lseg, offset);
 590         idx = nfs4_fl_calc_ds_index(lseg, j);
 591         ds = nfs4_fl_prepare_ds(lseg, idx);
 592         if (!ds)
 593                 return PNFS_NOT_ATTEMPTED;
 594
 595         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
 596         if (IS_ERR(ds_clnt))
 597                 return PNFS_NOT_ATTEMPTED;
 598
 599         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
 600                 __func__, hdr->inode->i_ino, sync, (size_t) data->args.count,
 601                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 602
 603         data->pgio_done_cb = filelayout_write_done_cb;
 604         atomic_inc(&ds->ds_clp->cl_count);
 605         data->ds_clp = ds->ds_clp;
 606         fh = nfs4_fl_select_ds_fh(lseg, j);
 607         if (fh)
 608                 data->args.fh = fh;
 609         /*
 610          * Get the file offset on the dserver. Set the write offset to
 611          * this offset and save the original offset.
 612          */
 613         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 614
 615         /* Perform an asynchronous write */
 616         nfs_initiate_pgio(ds_clnt, data,
 617                                     &filelayout_write_call_ops, sync,
 618                                     RPC_TASK_SOFTCONN);
 619         return PNFS_ATTEMPTED;
 620 }
 621
 622 /*
 623  * filelayout_check_layout()
 624  *
 625  * Make sure layout segment parameters are sane WRT the device.
 626  * At this point no generic layer initialization of the lseg has occurred,
 627  * and nothing has been added to the layout_hdr cache.
 628  *
 629  */
 630 static int
 631 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 632                         struct nfs4_filelayout_segment *fl,
 633                         struct nfs4_layoutget_res *lgr,
 634                         struct nfs4_deviceid *id,
 635                         gfp_t gfp_flags)
 636 {
 637         struct nfs4_deviceid_node *d;
 638         struct nfs4_file_layout_dsaddr *dsaddr;
 639         int status = -EINVAL;
 640         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 641
 642         dprintk("--> %s\n", __func__);
 643
 644         /* FIXME: remove this check when layout segment support is added */
 645         if (lgr->range.offset != 0 ||
 646             lgr->range.length != NFS4_MAX_UINT64) {
 647                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 648                         __func__);
 649                 goto out;
 650         }
 651
 652         if (fl->pattern_offset > lgr->range.offset) {
 653                 dprintk("%s pattern_offset %lld too large\n",
 654                                 __func__, fl->pattern_offset);
 655                 goto out;
 656         }
 657
 658         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 659                 dprintk("%s Invalid stripe unit (%u)\n",
 660                         __func__, fl->stripe_unit);
 661                 goto out;
 662         }
 663
 664         /* find and reference the deviceid */
 665         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
 666                                    NFS_SERVER(lo->plh_inode)->nfs_client, id);
 667         if (d == NULL) {
 668                 dsaddr = filelayout_get_device_info(lo->plh_inode, id,
 669                                 lo->plh_lc_cred, gfp_flags);
 670                 if (dsaddr == NULL)
 671                         goto out;
 672         } else
 673                 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 674         /* Found deviceid is unavailable */
 675         if (filelayout_test_devid_unavailable(&dsaddr->id_node))
 676                         goto out_put;
 677
 678         fl->dsaddr = dsaddr;
 679
 680         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 681                 dprintk("%s Bad first_stripe_index %u\n",
 682                                 __func__, fl->first_stripe_index);
 683                 goto out_put;
 684         }
 685
 686         if ((fl->stripe_type == STRIPE_SPARSE &&
 687             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 688             (fl->stripe_type == STRIPE_DENSE &&
 689             fl->num_fh != dsaddr->stripe_count)) {
 690                 dprintk("%s num_fh %u not valid for given packing\n",
 691                         __func__, fl->num_fh);
 692                 goto out_put;
 693         }
 694
 695         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
 696                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
 697                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
 698                         nfss->wsize);
 699         }
 700
 701         status = 0;
 702 out:
 703         dprintk("--> %s returns %d\n", __func__, status);
 704         return status;
 705 out_put:
 706         nfs4_fl_put_deviceid(dsaddr);
 707         goto out;
 708 }
 709
 710 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 711 {
 712         int i;
 713
 714         for (i = 0; i < fl->num_fh; i++) {
 715                 if (!fl->fh_array[i])
 716                         break;
 717                 kfree(fl->fh_array[i]);
 718         }
 719         kfree(fl->fh_array);
 720         fl->fh_array = NULL;
 721 }
 722
 723 static void
 724 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 725 {
 726         filelayout_free_fh_array(fl);
 727         kfree(fl);
 728 }
 729
 730 static int
 731 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 732                          struct nfs4_filelayout_segment *fl,
 733                          struct nfs4_layoutget_res *lgr,
 734                          struct nfs4_deviceid *id,
 735                          gfp_t gfp_flags)
 736 {
 737         struct xdr_stream stream;
 738         struct xdr_buf buf;
 739         struct page *scratch;
 740         __be32 *p;
 741         uint32_t nfl_util;
 742         int i;
 743
 744         dprintk("%s: set_layout_map Begin\n", __func__);
 745
 746         scratch = alloc_page(gfp_flags);
 747         if (!scratch)
 748                 return -ENOMEM;
 749
 750         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 751         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 752
 753         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 754          * num_fh (4) */
 755         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 756         if (unlikely(!p))
 757                 goto out_err;
 758
 759         memcpy(id, p, sizeof(*id));
 760         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 761         nfs4_print_deviceid(id);
 762
 763         nfl_util = be32_to_cpup(p++);
 764         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 765                 fl->commit_through_mds = 1;
 766         if (nfl_util & NFL4_UFLG_DENSE)
 767                 fl->stripe_type = STRIPE_DENSE;
 768         else
 769                 fl->stripe_type = STRIPE_SPARSE;
 770         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 771
 772         fl->first_stripe_index = be32_to_cpup(p++);
 773         p = xdr_decode_hyper(p, &fl->pattern_offset);
 774         fl->num_fh = be32_to_cpup(p++);
 775
 776         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 777                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 778                 fl->pattern_offset);
 779
 780         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 781          * Futher checking is done in filelayout_check_layout */
 782         if (fl->num_fh >
 783             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 784                 goto out_err;
 785
 786         if (fl->num_fh > 0) {
 787                 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
 788                                        gfp_flags);
 789                 if (!fl->fh_array)
 790                         goto out_err;
 791         }
 792
 793         for (i = 0; i < fl->num_fh; i++) {
 794                 /* Do we want to use a mempool here? */
 795                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 796                 if (!fl->fh_array[i])
 797                         goto out_err_free;
 798
 799                 p = xdr_inline_decode(&stream, 4);
 800                 if (unlikely(!p))
 801                         goto out_err_free;
 802                 fl->fh_array[i]->size = be32_to_cpup(p++);
 803                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 804                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
 805                                i, fl->fh_array[i]->size);
 806                         goto out_err_free;
 807                 }
 808
 809                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 810                 if (unlikely(!p))
 811                         goto out_err_free;
 812                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 813                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 814                         fl->fh_array[i]->size);
 815         }
 816
 817         __free_page(scratch);
 818         return 0;
 819
 820 out_err_free:
 821         filelayout_free_fh_array(fl);
 822 out_err:
 823         __free_page(scratch);
 824         return -EIO;
 825 }
 826
 827 static void
 828 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 829 {
 830         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 831
 832         dprintk("--> %s\n", __func__);
 833         nfs4_fl_put_deviceid(fl->dsaddr);
 834         /* This assumes a single RW lseg */
 835         if (lseg->pls_range.iomode == IOMODE_RW) {
 836                 struct nfs4_filelayout *flo;
 837
 838                 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
 839                 flo->commit_info.nbuckets = 0;
 840                 kfree(flo->commit_info.buckets);
 841                 flo->commit_info.buckets = NULL;
 842         }
 843         _filelayout_free_lseg(fl);
 844 }
 845
 846 static int
 847 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 848                              struct nfs_commit_info *cinfo,
 849                              gfp_t gfp_flags)
 850 {
 851         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 852         struct pnfs_commit_bucket *buckets;
 853         int size;
 854
 855         if (fl->commit_through_mds)
 856                 return 0;
 857         if (cinfo->ds->nbuckets != 0) {
 858                 /* This assumes there is only one IOMODE_RW lseg.  What
 859                  * we really want to do is have a layout_hdr level
 860                  * dictionary of <multipath_list4, fh> keys, each
 861                  * associated with a struct list_head, populated by calls
 862                  * to filelayout_write_pagelist().
 863                  * */
 864                 return 0;
 865         }
 866
 867         size = (fl->stripe_type == STRIPE_SPARSE) ?
 868                 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 869
 870         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 871                           gfp_flags);
 872         if (!buckets)
 873                 return -ENOMEM;
 874         else {
 875                 int i;
 876
 877                 spin_lock(cinfo->lock);
 878                 if (cinfo->ds->nbuckets != 0)
 879                         kfree(buckets);
 880                 else {
 881                         cinfo->ds->buckets = buckets;
 882                         cinfo->ds->nbuckets = size;
 883                         for (i = 0; i < size; i++) {
 884                                 INIT_LIST_HEAD(&buckets[i].written);
 885                                 INIT_LIST_HEAD(&buckets[i].committing);
 886                         }
 887                 }
 888                 spin_unlock(cinfo->lock);
 889                 return 0;
 890         }
 891 }
 892
 893 static struct pnfs_layout_segment *
 894 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 895                       struct nfs4_layoutget_res *lgr,
 896                       gfp_t gfp_flags)
 897 {
 898         struct nfs4_filelayout_segment *fl;
 899         int rc;
 900         struct nfs4_deviceid id;
 901
 902         dprintk("--> %s\n", __func__);
 903         fl = kzalloc(sizeof(*fl), gfp_flags);
 904         if (!fl)
 905                 return NULL;
 906
 907         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 908         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 909                 _filelayout_free_lseg(fl);
 910                 return NULL;
 911         }
 912         return &fl->generic_hdr;
 913 }
 914
 915 /*
 916  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 917  *
 918  * return true  : coalesce page
 919  * return false : don't coalesce page
 920  */
 921 static bool
 922 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 923                    struct nfs_page *req)
 924 {
 925         u64 p_stripe, r_stripe;
 926         u32 stripe_unit;
 927
 928         if (!pnfs_generic_pg_test(pgio, prev, req) ||
 929             !nfs_generic_pg_test(pgio, prev, req))
 930                 return false;
 931
 932         p_stripe = (u64)req_offset(prev);
 933         r_stripe = (u64)req_offset(req);
 934         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 935
 936         do_div(p_stripe, stripe_unit);
 937         do_div(r_stripe, stripe_unit);
 938
 939         return (p_stripe == r_stripe);
 940 }
 941
 942 static void
 943 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 944                         struct nfs_page *req)
 945 {
 946         WARN_ON_ONCE(pgio->pg_lseg != NULL);
 947
 948         if (req->wb_offset != req->wb_pgbase) {
 949                 /*
 950                  * Handling unaligned pages is difficult, because have to
 951                  * somehow split a req in two in certain cases in the
 952                  * pg.test code.  Avoid this by just not using pnfs
 953                  * in this case.
 954                  */
 955                 nfs_pageio_reset_read_mds(pgio);
 956                 return;
 957         }
 958         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 959                                            req->wb_context,
 960                                            0,
 961                                            NFS4_MAX_UINT64,
 962                                            IOMODE_READ,
 963                                            GFP_KERNEL);
 964         /* If no lseg, fall back to read through mds */
 965         if (pgio->pg_lseg == NULL)
 966                 nfs_pageio_reset_read_mds(pgio);
 967 }
 968
 969 static void
 970 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 971                          struct nfs_page *req)
 972 {
 973         struct nfs_commit_info cinfo;
 974         int status;
 975
 976         WARN_ON_ONCE(pgio->pg_lseg != NULL);
 977
 978         if (req->wb_offset != req->wb_pgbase)
 979                 goto out_mds;
 980         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 981                                            req->wb_context,
 982                                            0,
 983                                            NFS4_MAX_UINT64,
 984                                            IOMODE_RW,
 985                                            GFP_NOFS);
 986         /* If no lseg, fall back to write through mds */
 987         if (pgio->pg_lseg == NULL)
 988                 goto out_mds;
 989         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 990         status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 991         if (status < 0) {
 992                 pnfs_put_lseg(pgio->pg_lseg);
 993                 pgio->pg_lseg = NULL;
 994                 goto out_mds;
 995         }
 996         return;
 997 out_mds:
 998         nfs_pageio_reset_write_mds(pgio);
 999 }
1000
1001 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
1002         .pg_init = filelayout_pg_init_read,
1003         .pg_test = filelayout_pg_test,
1004         .pg_doio = pnfs_generic_pg_readpages,
1005 };
1006
1007 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
1008         .pg_init = filelayout_pg_init_write,
1009         .pg_test = filelayout_pg_test,
1010         .pg_doio = pnfs_generic_pg_writepages,
1011 };
1012
1013 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
1014 {
1015         if (fl->stripe_type == STRIPE_SPARSE)
1016                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
1017         else
1018                 return j;
1019 }
1020
1021 /* The generic layer is about to remove the req from the commit list.
1022  * If this will make the bucket empty, it will need to put the lseg reference.
1023  */
1024 static void
1025 filelayout_clear_request_commit(struct nfs_page *req,
1026                                 struct nfs_commit_info *cinfo)
1027 {
1028         struct pnfs_layout_segment *freeme = NULL;
1029
1030         spin_lock(cinfo->lock);
1031         if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
1032                 goto out;
1033         cinfo->ds->nwritten--;
1034         if (list_is_singular(&req->wb_list)) {
1035                 struct pnfs_commit_bucket *bucket;
1036
1037                 bucket = list_first_entry(&req->wb_list,
1038                                           struct pnfs_commit_bucket,
1039                                           written);
1040                 freeme = bucket->wlseg;
1041                 bucket->wlseg = NULL;
1042         }
1043 out:
1044         nfs_request_remove_commit_list(req, cinfo);
1045         spin_unlock(cinfo->lock);
1046         pnfs_put_lseg(freeme);
1047 }
1048
1049 static struct list_head *
1050 filelayout_choose_commit_list(struct nfs_page *req,
1051                               struct pnfs_layout_segment *lseg,
1052                               struct nfs_commit_info *cinfo)
1053 {
1054         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1055         u32 i, j;
1056         struct list_head *list;
1057         struct pnfs_commit_bucket *buckets;
1058
1059         if (fl->commit_through_mds)
1060                 return &cinfo->mds->list;
1061
1062         /* Note that we are calling nfs4_fl_calc_j_index on each page
1063          * that ends up being committed to a data server.  An attractive
1064          * alternative is to add a field to nfs_write_data and nfs_page
1065          * to store the value calculated in filelayout_write_pagelist
1066          * and just use that here.
1067          */
1068         j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1069         i = select_bucket_index(fl, j);
1070         buckets = cinfo->ds->buckets;
1071         list = &buckets[i].written;
1072         if (list_empty(list)) {
1073                 /* Non-empty buckets hold a reference on the lseg.  That ref
1074                  * is normally transferred to the COMMIT call and released
1075                  * there.  It could also be released if the last req is pulled
1076                  * off due to a rewrite, in which case it will be done in
1077                  * filelayout_clear_request_commit
1078                  */
1079                 buckets[i].wlseg = pnfs_get_lseg(lseg);
1080         }
1081         set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1082         cinfo->ds->nwritten++;
1083         return list;
1084 }
1085
1086 static void
1087 filelayout_mark_request_commit(struct nfs_page *req,
1088                                struct pnfs_layout_segment *lseg,
1089                                struct nfs_commit_info *cinfo)
1090 {
1091         struct list_head *list;
1092
1093         list = filelayout_choose_commit_list(req, lseg, cinfo);
1094         nfs_request_add_commit_list(req, list, cinfo);
1095 }
1096
1097 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1098 {
1099         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1100
1101         if (flseg->stripe_type == STRIPE_SPARSE)
1102                 return i;
1103         else
1104                 return nfs4_fl_calc_ds_index(lseg, i);
1105 }
1106
1107 static struct nfs_fh *
1108 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1109 {
1110         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1111
1112         if (flseg->stripe_type == STRIPE_SPARSE) {
1113                 if (flseg->num_fh == 1)
1114                         i = 0;
1115                 else if (flseg->num_fh == 0)
1116                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1117                         return NULL;
1118         }
1119         return flseg->fh_array[i];
1120 }
1121
1122 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1123 {
1124         struct pnfs_layout_segment *lseg = data->lseg;
1125         struct nfs4_pnfs_ds *ds;
1126         struct rpc_clnt *ds_clnt;
1127         u32 idx;
1128         struct nfs_fh *fh;
1129
1130         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1131         ds = nfs4_fl_prepare_ds(lseg, idx);
1132         if (!ds)
1133                 goto out_err;
1134
1135         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1136         if (IS_ERR(ds_clnt))
1137                 goto out_err;
1138
1139         dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1140                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1141         data->commit_done_cb = filelayout_commit_done_cb;
1142         atomic_inc(&ds->ds_clp->cl_count);
1143         data->ds_clp = ds->ds_clp;
1144         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1145         if (fh)
1146                 data->args.fh = fh;
1147         return nfs_initiate_commit(ds_clnt, data,
1148                                    &filelayout_commit_call_ops, how,
1149                                    RPC_TASK_SOFTCONN);
1150 out_err:
1151         prepare_to_resend_writes(data);
1152         filelayout_commit_release(data);
1153         return -EAGAIN;
1154 }
1155
1156 static int
1157 transfer_commit_list(struct list_head *src, struct list_head *dst,
1158                      struct nfs_commit_info *cinfo, int max)
1159 {
1160         struct nfs_page *req, *tmp;
1161         int ret = 0;
1162
1163         list_for_each_entry_safe(req, tmp, src, wb_list) {
1164                 if (!nfs_lock_request(req))
1165                         continue;
1166                 kref_get(&req->wb_kref);
1167                 if (cond_resched_lock(cinfo->lock))
1168                         list_safe_reset_next(req, tmp, wb_list);
1169                 nfs_request_remove_commit_list(req, cinfo);
1170                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1171                 nfs_list_add_request(req, dst);
1172                 ret++;
1173                 if ((ret == max) && !cinfo->dreq)
1174                         break;
1175         }
1176         return ret;
1177 }
1178
1179 static int
1180 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1181                                struct nfs_commit_info *cinfo,
1182                                int max)
1183 {
1184         struct list_head *src = &bucket->written;
1185         struct list_head *dst = &bucket->committing;
1186         int ret;
1187
1188         ret = transfer_commit_list(src, dst, cinfo, max);
1189         if (ret) {
1190                 cinfo->ds->nwritten -= ret;
1191                 cinfo->ds->ncommitting += ret;
1192                 bucket->clseg = bucket->wlseg;
1193                 if (list_empty(src))
1194                         bucket->wlseg = NULL;
1195                 else
1196                         pnfs_get_lseg(bucket->clseg);
1197         }
1198         return ret;
1199 }
1200
1201 /* Move reqs from written to committing lists, returning count of number moved.
1202  * Note called with cinfo->lock held.
1203  */
1204 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1205                                         int max)
1206 {
1207         int i, rv = 0, cnt;
1208
1209         for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1210                 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1211                                                      cinfo, max);
1212                 max -= cnt;
1213                 rv += cnt;
1214         }
1215         return rv;
1216 }
1217
1218 /* Pull everything off the committing lists and dump into @dst */
1219 static void filelayout_recover_commit_reqs(struct list_head *dst,
1220                                            struct nfs_commit_info *cinfo)
1221 {
1222         struct pnfs_commit_bucket *b;
1223         int i;
1224
1225         spin_lock(cinfo->lock);
1226         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1227                 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1228                         spin_unlock(cinfo->lock);
1229                         pnfs_put_lseg(b->wlseg);
1230                         b->wlseg = NULL;
1231                         spin_lock(cinfo->lock);
1232                 }
1233         }
1234         cinfo->ds->nwritten = 0;
1235         spin_unlock(cinfo->lock);
1236 }
1237
1238 static unsigned int
1239 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1240 {
1241         struct pnfs_ds_commit_info *fl_cinfo;
1242         struct pnfs_commit_bucket *bucket;
1243         struct nfs_commit_data *data;
1244         int i, j;
1245         unsigned int nreq = 0;
1246
1247         fl_cinfo = cinfo->ds;
1248         bucket = fl_cinfo->buckets;
1249         for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1250                 if (list_empty(&bucket->committing))
1251                         continue;
1252                 data = nfs_commitdata_alloc();
1253                 if (!data)
1254                         break;
1255                 data->ds_commit_index = i;
1256                 data->lseg = bucket->clseg;
1257                 bucket->clseg = NULL;
1258                 list_add(&data->pages, list);
1259                 nreq++;
1260         }
1261
1262         /* Clean up on error */
1263         for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1264                 if (list_empty(&bucket->committing))
1265                         continue;
1266                 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1267                 pnfs_put_lseg(bucket->clseg);
1268                 bucket->clseg = NULL;
1269         }
1270         /* Caller will clean up entries put on list */
1271         return nreq;
1272 }
1273
1274 /* This follows nfs_commit_list pretty closely */
1275 static int
1276 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1277                            int how, struct nfs_commit_info *cinfo)
1278 {
1279         struct nfs_commit_data *data, *tmp;
1280         LIST_HEAD(list);
1281         unsigned int nreq = 0;
1282
1283         if (!list_empty(mds_pages)) {
1284                 data = nfs_commitdata_alloc();
1285                 if (data != NULL) {
1286                         data->lseg = NULL;
1287                         list_add(&data->pages, &list);
1288                         nreq++;
1289                 } else
1290                         nfs_retry_commit(mds_pages, NULL, cinfo);
1291         }
1292
1293         nreq += alloc_ds_commits(cinfo, &list);
1294
1295         if (nreq == 0) {
1296                 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1297                 goto out;
1298         }
1299
1300         atomic_add(nreq, &cinfo->mds->rpcs_out);
1301
1302         list_for_each_entry_safe(data, tmp, &list, pages) {
1303                 list_del_init(&data->pages);
1304                 if (!data->lseg) {
1305                         nfs_init_commit(data, mds_pages, NULL, cinfo);
1306                         nfs_initiate_commit(NFS_CLIENT(inode), data,
1307                                             data->mds_ops, how, 0);
1308                 } else {
1309                         struct pnfs_commit_bucket *buckets;
1310
1311                         buckets = cinfo->ds->buckets;
1312                         nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1313                         filelayout_initiate_commit(data, how);
1314                 }
1315         }
1316 out:
1317         cinfo->ds->ncommitting = 0;
1318         return PNFS_ATTEMPTED;
1319 }
1320
1321 static void
1322 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1323 {
1324         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1325 }
1326
1327 static struct pnfs_layout_hdr *
1328 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1329 {
1330         struct nfs4_filelayout *flo;
1331
1332         flo = kzalloc(sizeof(*flo), gfp_flags);
1333         return &flo->generic_hdr;
1334 }
1335
1336 static void
1337 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1338 {
1339         kfree(FILELAYOUT_FROM_HDR(lo));
1340 }
1341
1342 static struct pnfs_ds_commit_info *
1343 filelayout_get_ds_info(struct inode *inode)
1344 {
1345         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1346
1347         if (layout == NULL)
1348                 return NULL;
1349         else
1350                 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1351 }
1352
1353 static struct pnfs_layoutdriver_type filelayout_type = {
1354         .id                     = LAYOUT_NFSV4_1_FILES,
1355         .name                   = "LAYOUT_NFSV4_1_FILES",
1356         .owner                  = THIS_MODULE,
1357         .alloc_layout_hdr       = filelayout_alloc_layout_hdr,
1358         .free_layout_hdr        = filelayout_free_layout_hdr,
1359         .alloc_lseg             = filelayout_alloc_lseg,
1360         .free_lseg              = filelayout_free_lseg,
1361         .pg_read_ops            = &filelayout_pg_read_ops,
1362         .pg_write_ops           = &filelayout_pg_write_ops,
1363         .get_ds_info            = &filelayout_get_ds_info,
1364         .mark_request_commit    = filelayout_mark_request_commit,
1365         .clear_request_commit   = filelayout_clear_request_commit,
1366         .scan_commit_lists      = filelayout_scan_commit_lists,
1367         .recover_commit_reqs    = filelayout_recover_commit_reqs,
1368         .commit_pagelist        = filelayout_commit_pagelist,
1369         .read_pagelist          = filelayout_read_pagelist,
1370         .write_pagelist         = filelayout_write_pagelist,
1371         .free_deviceid_node     = filelayout_free_deveiceid_node,
1372 };
1373
1374 static int __init nfs4filelayout_init(void)
1375 {
1376         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1377                __func__);
1378         return pnfs_register_layoutdriver(&filelayout_type);
1379 }
1380
1381 static void __exit nfs4filelayout_exit(void)
1382 {
1383         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1384                __func__);
1385         pnfs_unregister_layoutdriver(&filelayout_type);
1386 }
1387
1388 MODULE_ALIAS("nfs-layouttype4-1");
1389
1390 module_init(nfs4filelayout_init);
1391 module_exit(nfs4filelayout_exit);