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_write_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/%lld, %u bytes @ offset %llu)\n", __func__,
  95                         data->task.tk_pid,
  96                         hdr->inode->i_sb->s_id,
  97                         (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_read_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/%lld, %u bytes @ offset %llu)\n", __func__,
 116                         data->task.tk_pid,
 117                         hdr->inode->i_sb->s_id,
 118                         (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_read_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_write_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_read_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->read_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         nfs4_set_rw_stateid(&rdata->args.stateid, rdata->args.context,
 328                         rdata->args.lock_context, FMODE_READ);
 329 }
 330
 331 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 332 {
 333         struct nfs_read_data *rdata = data;
 334
 335         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 336
 337         if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) &&
 338             task->tk_status == 0)
 339                 return;
 340
 341         /* Note this may cause RPC to be resent */
 342         rdata->header->mds_ops->rpc_call_done(task, data);
 343 }
 344
 345 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
 346 {
 347         struct nfs_read_data *rdata = data;
 348
 349         rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
 350 }
 351
 352 static void filelayout_read_release(void *data)
 353 {
 354         struct nfs_read_data *rdata = data;
 355         struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout;
 356
 357         filelayout_fenceme(lo->plh_inode, lo);
 358         nfs_put_client(rdata->ds_clp);
 359         rdata->header->mds_ops->rpc_release(data);
 360 }
 361
 362 static int filelayout_write_done_cb(struct rpc_task *task,
 363                                 struct nfs_write_data *data)
 364 {
 365         struct nfs_pgio_header *hdr = data->header;
 366         int err;
 367
 368         trace_nfs4_pnfs_write(data, task->tk_status);
 369         err = filelayout_async_handle_error(task, data->args.context->state,
 370                                             data->ds_clp, hdr->lseg);
 371
 372         switch (err) {
 373         case -NFS4ERR_RESET_TO_MDS:
 374                 filelayout_reset_write(data);
 375                 return task->tk_status;
 376         case -EAGAIN:
 377                 rpc_restart_call_prepare(task);
 378                 return -EAGAIN;
 379         }
 380
 381         filelayout_set_layoutcommit(data);
 382         return 0;
 383 }
 384
 385 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 386 static void prepare_to_resend_writes(struct nfs_commit_data *data)
 387 {
 388         struct nfs_page *first = nfs_list_entry(data->pages.next);
 389
 390         data->task.tk_status = 0;
 391         memcpy(&data->verf.verifier, &first->wb_verf,
 392                sizeof(data->verf.verifier));
 393         data->verf.verifier.data[0]++; /* ensure verifier mismatch */
 394 }
 395
 396 static int filelayout_commit_done_cb(struct rpc_task *task,
 397                                      struct nfs_commit_data *data)
 398 {
 399         int err;
 400
 401         trace_nfs4_pnfs_commit_ds(data, task->tk_status);
 402         err = filelayout_async_handle_error(task, NULL, data->ds_clp,
 403                                             data->lseg);
 404
 405         switch (err) {
 406         case -NFS4ERR_RESET_TO_MDS:
 407                 prepare_to_resend_writes(data);
 408                 return -EAGAIN;
 409         case -EAGAIN:
 410                 rpc_restart_call_prepare(task);
 411                 return -EAGAIN;
 412         }
 413
 414         return 0;
 415 }
 416
 417 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 418 {
 419         struct nfs_write_data *wdata = data;
 420
 421         if (unlikely(test_bit(NFS_CONTEXT_BAD, &wdata->args.context->flags))) {
 422                 rpc_exit(task, -EIO);
 423                 return;
 424         }
 425         if (filelayout_reset_to_mds(wdata->header->lseg)) {
 426                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 427                 filelayout_reset_write(wdata);
 428                 rpc_exit(task, 0);
 429                 return;
 430         }
 431         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 432                         &wdata->args.seq_args,
 433                         &wdata->res.seq_res,
 434                         task))
 435                 return;
 436         nfs4_set_rw_stateid(&wdata->args.stateid, wdata->args.context,
 437                         wdata->args.lock_context, FMODE_WRITE);
 438 }
 439
 440 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 441 {
 442         struct nfs_write_data *wdata = data;
 443
 444         if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) &&
 445             task->tk_status == 0)
 446                 return;
 447
 448         /* Note this may cause RPC to be resent */
 449         wdata->header->mds_ops->rpc_call_done(task, data);
 450 }
 451
 452 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
 453 {
 454         struct nfs_write_data *wdata = data;
 455
 456         rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
 457 }
 458
 459 static void filelayout_write_release(void *data)
 460 {
 461         struct nfs_write_data *wdata = data;
 462         struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout;
 463
 464         filelayout_fenceme(lo->plh_inode, lo);
 465         nfs_put_client(wdata->ds_clp);
 466         wdata->header->mds_ops->rpc_release(data);
 467 }
 468
 469 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
 470 {
 471         struct nfs_commit_data *wdata = data;
 472
 473         nfs41_setup_sequence(wdata->ds_clp->cl_session,
 474                         &wdata->args.seq_args,
 475                         &wdata->res.seq_res,
 476                         task);
 477 }
 478
 479 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
 480 {
 481         struct nfs_commit_data *wdata = data;
 482
 483         /* Note this may cause RPC to be resent */
 484         wdata->mds_ops->rpc_call_done(task, data);
 485 }
 486
 487 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
 488 {
 489         struct nfs_commit_data *cdata = data;
 490
 491         rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
 492 }
 493
 494 static void filelayout_commit_release(void *calldata)
 495 {
 496         struct nfs_commit_data *data = calldata;
 497
 498         data->completion_ops->completion(data);
 499         pnfs_put_lseg(data->lseg);
 500         nfs_put_client(data->ds_clp);
 501         nfs_commitdata_release(data);
 502 }
 503
 504 static const struct rpc_call_ops filelayout_read_call_ops = {
 505         .rpc_call_prepare = filelayout_read_prepare,
 506         .rpc_call_done = filelayout_read_call_done,
 507         .rpc_count_stats = filelayout_read_count_stats,
 508         .rpc_release = filelayout_read_release,
 509 };
 510
 511 static const struct rpc_call_ops filelayout_write_call_ops = {
 512         .rpc_call_prepare = filelayout_write_prepare,
 513         .rpc_call_done = filelayout_write_call_done,
 514         .rpc_count_stats = filelayout_write_count_stats,
 515         .rpc_release = filelayout_write_release,
 516 };
 517
 518 static const struct rpc_call_ops filelayout_commit_call_ops = {
 519         .rpc_call_prepare = filelayout_commit_prepare,
 520         .rpc_call_done = filelayout_write_commit_done,
 521         .rpc_count_stats = filelayout_commit_count_stats,
 522         .rpc_release = filelayout_commit_release,
 523 };
 524
 525 static enum pnfs_try_status
 526 filelayout_read_pagelist(struct nfs_read_data *data)
 527 {
 528         struct nfs_pgio_header *hdr = data->header;
 529         struct pnfs_layout_segment *lseg = hdr->lseg;
 530         struct nfs4_pnfs_ds *ds;
 531         struct rpc_clnt *ds_clnt;
 532         loff_t offset = data->args.offset;
 533         u32 j, idx;
 534         struct nfs_fh *fh;
 535
 536         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 537                 __func__, hdr->inode->i_ino,
 538                 data->args.pgbase, (size_t)data->args.count, offset);
 539
 540         /* Retrieve the correct rpc_client for the byte range */
 541         j = nfs4_fl_calc_j_index(lseg, offset);
 542         idx = nfs4_fl_calc_ds_index(lseg, j);
 543         ds = nfs4_fl_prepare_ds(lseg, idx);
 544         if (!ds)
 545                 return PNFS_NOT_ATTEMPTED;
 546
 547         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
 548         if (IS_ERR(ds_clnt))
 549                 return PNFS_NOT_ATTEMPTED;
 550
 551         dprintk("%s USE DS: %s cl_count %d\n", __func__,
 552                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 553
 554         /* No multipath support. Use first DS */
 555         atomic_inc(&ds->ds_clp->cl_count);
 556         data->ds_clp = ds->ds_clp;
 557         fh = nfs4_fl_select_ds_fh(lseg, j);
 558         if (fh)
 559                 data->args.fh = fh;
 560
 561         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 562         data->mds_offset = offset;
 563
 564         /* Perform an asynchronous read to ds */
 565         nfs_initiate_read(ds_clnt, data,
 566                                   &filelayout_read_call_ops, RPC_TASK_SOFTCONN);
 567         return PNFS_ATTEMPTED;
 568 }
 569
 570 /* Perform async writes. */
 571 static enum pnfs_try_status
 572 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
 573 {
 574         struct nfs_pgio_header *hdr = data->header;
 575         struct pnfs_layout_segment *lseg = hdr->lseg;
 576         struct nfs4_pnfs_ds *ds;
 577         struct rpc_clnt *ds_clnt;
 578         loff_t offset = data->args.offset;
 579         u32 j, idx;
 580         struct nfs_fh *fh;
 581
 582         /* Retrieve the correct rpc_client for the byte range */
 583         j = nfs4_fl_calc_j_index(lseg, offset);
 584         idx = nfs4_fl_calc_ds_index(lseg, j);
 585         ds = nfs4_fl_prepare_ds(lseg, idx);
 586         if (!ds)
 587                 return PNFS_NOT_ATTEMPTED;
 588
 589         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
 590         if (IS_ERR(ds_clnt))
 591                 return PNFS_NOT_ATTEMPTED;
 592
 593         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
 594                 __func__, hdr->inode->i_ino, sync, (size_t) data->args.count,
 595                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 596
 597         data->write_done_cb = filelayout_write_done_cb;
 598         atomic_inc(&ds->ds_clp->cl_count);
 599         data->ds_clp = ds->ds_clp;
 600         fh = nfs4_fl_select_ds_fh(lseg, j);
 601         if (fh)
 602                 data->args.fh = fh;
 603         /*
 604          * Get the file offset on the dserver. Set the write offset to
 605          * this offset and save the original offset.
 606          */
 607         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 608
 609         /* Perform an asynchronous write */
 610         nfs_initiate_write(ds_clnt, data,
 611                                     &filelayout_write_call_ops, sync,
 612                                     RPC_TASK_SOFTCONN);
 613         return PNFS_ATTEMPTED;
 614 }
 615
 616 /*
 617  * filelayout_check_layout()
 618  *
 619  * Make sure layout segment parameters are sane WRT the device.
 620  * At this point no generic layer initialization of the lseg has occurred,
 621  * and nothing has been added to the layout_hdr cache.
 622  *
 623  */
 624 static int
 625 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 626                         struct nfs4_filelayout_segment *fl,
 627                         struct nfs4_layoutget_res *lgr,
 628                         struct nfs4_deviceid *id,
 629                         gfp_t gfp_flags)
 630 {
 631         struct nfs4_deviceid_node *d;
 632         struct nfs4_file_layout_dsaddr *dsaddr;
 633         int status = -EINVAL;
 634         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 635
 636         dprintk("--> %s\n", __func__);
 637
 638         /* FIXME: remove this check when layout segment support is added */
 639         if (lgr->range.offset != 0 ||
 640             lgr->range.length != NFS4_MAX_UINT64) {
 641                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 642                         __func__);
 643                 goto out;
 644         }
 645
 646         if (fl->pattern_offset > lgr->range.offset) {
 647                 dprintk("%s pattern_offset %lld too large\n",
 648                                 __func__, fl->pattern_offset);
 649                 goto out;
 650         }
 651
 652         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 653                 dprintk("%s Invalid stripe unit (%u)\n",
 654                         __func__, fl->stripe_unit);
 655                 goto out;
 656         }
 657
 658         /* find and reference the deviceid */
 659         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
 660                                    NFS_SERVER(lo->plh_inode)->nfs_client, id);
 661         if (d == NULL) {
 662                 dsaddr = filelayout_get_device_info(lo->plh_inode, id,
 663                                 lo->plh_lc_cred, gfp_flags);
 664                 if (dsaddr == NULL)
 665                         goto out;
 666         } else
 667                 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 668         /* Found deviceid is unavailable */
 669         if (filelayout_test_devid_unavailable(&dsaddr->id_node))
 670                         goto out_put;
 671
 672         fl->dsaddr = dsaddr;
 673
 674         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 675                 dprintk("%s Bad first_stripe_index %u\n",
 676                                 __func__, fl->first_stripe_index);
 677                 goto out_put;
 678         }
 679
 680         if ((fl->stripe_type == STRIPE_SPARSE &&
 681             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 682             (fl->stripe_type == STRIPE_DENSE &&
 683             fl->num_fh != dsaddr->stripe_count)) {
 684                 dprintk("%s num_fh %u not valid for given packing\n",
 685                         __func__, fl->num_fh);
 686                 goto out_put;
 687         }
 688
 689         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
 690                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
 691                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
 692                         nfss->wsize);
 693         }
 694
 695         status = 0;
 696 out:
 697         dprintk("--> %s returns %d\n", __func__, status);
 698         return status;
 699 out_put:
 700         nfs4_fl_put_deviceid(dsaddr);
 701         goto out;
 702 }
 703
 704 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 705 {
 706         int i;
 707
 708         for (i = 0; i < fl->num_fh; i++) {
 709                 if (!fl->fh_array[i])
 710                         break;
 711                 kfree(fl->fh_array[i]);
 712         }
 713         kfree(fl->fh_array);
 714         fl->fh_array = NULL;
 715 }
 716
 717 static void
 718 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 719 {
 720         filelayout_free_fh_array(fl);
 721         kfree(fl);
 722 }
 723
 724 static int
 725 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 726                          struct nfs4_filelayout_segment *fl,
 727                          struct nfs4_layoutget_res *lgr,
 728                          struct nfs4_deviceid *id,
 729                          gfp_t gfp_flags)
 730 {
 731         struct xdr_stream stream;
 732         struct xdr_buf buf;
 733         struct page *scratch;
 734         __be32 *p;
 735         uint32_t nfl_util;
 736         int i;
 737
 738         dprintk("%s: set_layout_map Begin\n", __func__);
 739
 740         scratch = alloc_page(gfp_flags);
 741         if (!scratch)
 742                 return -ENOMEM;
 743
 744         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 745         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 746
 747         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 748          * num_fh (4) */
 749         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 750         if (unlikely(!p))
 751                 goto out_err;
 752
 753         memcpy(id, p, sizeof(*id));
 754         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 755         nfs4_print_deviceid(id);
 756
 757         nfl_util = be32_to_cpup(p++);
 758         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 759                 fl->commit_through_mds = 1;
 760         if (nfl_util & NFL4_UFLG_DENSE)
 761                 fl->stripe_type = STRIPE_DENSE;
 762         else
 763                 fl->stripe_type = STRIPE_SPARSE;
 764         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 765
 766         fl->first_stripe_index = be32_to_cpup(p++);
 767         p = xdr_decode_hyper(p, &fl->pattern_offset);
 768         fl->num_fh = be32_to_cpup(p++);
 769
 770         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 771                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 772                 fl->pattern_offset);
 773
 774         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 775          * Futher checking is done in filelayout_check_layout */
 776         if (fl->num_fh >
 777             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 778                 goto out_err;
 779
 780         if (fl->num_fh > 0) {
 781                 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
 782                                        gfp_flags);
 783                 if (!fl->fh_array)
 784                         goto out_err;
 785         }
 786
 787         for (i = 0; i < fl->num_fh; i++) {
 788                 /* Do we want to use a mempool here? */
 789                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 790                 if (!fl->fh_array[i])
 791                         goto out_err_free;
 792
 793                 p = xdr_inline_decode(&stream, 4);
 794                 if (unlikely(!p))
 795                         goto out_err_free;
 796                 fl->fh_array[i]->size = be32_to_cpup(p++);
 797                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 798                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
 799                                i, fl->fh_array[i]->size);
 800                         goto out_err_free;
 801                 }
 802
 803                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 804                 if (unlikely(!p))
 805                         goto out_err_free;
 806                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 807                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 808                         fl->fh_array[i]->size);
 809         }
 810
 811         __free_page(scratch);
 812         return 0;
 813
 814 out_err_free:
 815         filelayout_free_fh_array(fl);
 816 out_err:
 817         __free_page(scratch);
 818         return -EIO;
 819 }
 820
 821 static void
 822 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 823 {
 824         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 825
 826         dprintk("--> %s\n", __func__);
 827         nfs4_fl_put_deviceid(fl->dsaddr);
 828         /* This assumes a single RW lseg */
 829         if (lseg->pls_range.iomode == IOMODE_RW) {
 830                 struct nfs4_filelayout *flo;
 831
 832                 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
 833                 flo->commit_info.nbuckets = 0;
 834                 kfree(flo->commit_info.buckets);
 835                 flo->commit_info.buckets = NULL;
 836         }
 837         _filelayout_free_lseg(fl);
 838 }
 839
 840 static int
 841 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 842                              struct nfs_commit_info *cinfo,
 843                              gfp_t gfp_flags)
 844 {
 845         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 846         struct pnfs_commit_bucket *buckets;
 847         int size;
 848
 849         if (fl->commit_through_mds)
 850                 return 0;
 851         if (cinfo->ds->nbuckets != 0) {
 852                 /* This assumes there is only one IOMODE_RW lseg.  What
 853                  * we really want to do is have a layout_hdr level
 854                  * dictionary of <multipath_list4, fh> keys, each
 855                  * associated with a struct list_head, populated by calls
 856                  * to filelayout_write_pagelist().
 857                  * */
 858                 return 0;
 859         }
 860
 861         size = (fl->stripe_type == STRIPE_SPARSE) ?
 862                 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 863
 864         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 865                           gfp_flags);
 866         if (!buckets)
 867                 return -ENOMEM;
 868         else {
 869                 int i;
 870
 871                 spin_lock(cinfo->lock);
 872                 if (cinfo->ds->nbuckets != 0)
 873                         kfree(buckets);
 874                 else {
 875                         cinfo->ds->buckets = buckets;
 876                         cinfo->ds->nbuckets = size;
 877                         for (i = 0; i < size; i++) {
 878                                 INIT_LIST_HEAD(&buckets[i].written);
 879                                 INIT_LIST_HEAD(&buckets[i].committing);
 880                         }
 881                 }
 882                 spin_unlock(cinfo->lock);
 883                 return 0;
 884         }
 885 }
 886
 887 static struct pnfs_layout_segment *
 888 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 889                       struct nfs4_layoutget_res *lgr,
 890                       gfp_t gfp_flags)
 891 {
 892         struct nfs4_filelayout_segment *fl;
 893         int rc;
 894         struct nfs4_deviceid id;
 895
 896         dprintk("--> %s\n", __func__);
 897         fl = kzalloc(sizeof(*fl), gfp_flags);
 898         if (!fl)
 899                 return NULL;
 900
 901         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 902         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 903                 _filelayout_free_lseg(fl);
 904                 return NULL;
 905         }
 906         return &fl->generic_hdr;
 907 }
 908
 909 /*
 910  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 911  *
 912  * return true  : coalesce page
 913  * return false : don't coalesce page
 914  */
 915 static bool
 916 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 917                    struct nfs_page *req)
 918 {
 919         u64 p_stripe, r_stripe;
 920         u32 stripe_unit;
 921
 922         if (!pnfs_generic_pg_test(pgio, prev, req) ||
 923             !nfs_generic_pg_test(pgio, prev, req))
 924                 return false;
 925
 926         p_stripe = (u64)req_offset(prev);
 927         r_stripe = (u64)req_offset(req);
 928         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 929
 930         do_div(p_stripe, stripe_unit);
 931         do_div(r_stripe, stripe_unit);
 932
 933         return (p_stripe == r_stripe);
 934 }
 935
 936 static void
 937 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 938                         struct nfs_page *req)
 939 {
 940         WARN_ON_ONCE(pgio->pg_lseg != NULL);
 941
 942         if (req->wb_offset != req->wb_pgbase) {
 943                 /*
 944                  * Handling unaligned pages is difficult, because have to
 945                  * somehow split a req in two in certain cases in the
 946                  * pg.test code.  Avoid this by just not using pnfs
 947                  * in this case.
 948                  */
 949                 nfs_pageio_reset_read_mds(pgio);
 950                 return;
 951         }
 952         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 953                                            req->wb_context,
 954                                            0,
 955                                            NFS4_MAX_UINT64,
 956                                            IOMODE_READ,
 957                                            GFP_KERNEL);
 958         /* If no lseg, fall back to read through mds */
 959         if (pgio->pg_lseg == NULL)
 960                 nfs_pageio_reset_read_mds(pgio);
 961 }
 962
 963 static void
 964 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 965                          struct nfs_page *req)
 966 {
 967         struct nfs_commit_info cinfo;
 968         int status;
 969
 970         WARN_ON_ONCE(pgio->pg_lseg != NULL);
 971
 972         if (req->wb_offset != req->wb_pgbase)
 973                 goto out_mds;
 974         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 975                                            req->wb_context,
 976                                            0,
 977                                            NFS4_MAX_UINT64,
 978                                            IOMODE_RW,
 979                                            GFP_NOFS);
 980         /* If no lseg, fall back to write through mds */
 981         if (pgio->pg_lseg == NULL)
 982                 goto out_mds;
 983         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 984         status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 985         if (status < 0) {
 986                 pnfs_put_lseg(pgio->pg_lseg);
 987                 pgio->pg_lseg = NULL;
 988                 goto out_mds;
 989         }
 990         return;
 991 out_mds:
 992         nfs_pageio_reset_write_mds(pgio);
 993 }
 994
 995 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 996         .pg_init = filelayout_pg_init_read,
 997         .pg_test = filelayout_pg_test,
 998         .pg_doio = pnfs_generic_pg_readpages,
 999 };
1000
1001 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
1002         .pg_init = filelayout_pg_init_write,
1003         .pg_test = filelayout_pg_test,
1004         .pg_doio = pnfs_generic_pg_writepages,
1005 };
1006
1007 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
1008 {
1009         if (fl->stripe_type == STRIPE_SPARSE)
1010                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
1011         else
1012                 return j;
1013 }
1014
1015 /* The generic layer is about to remove the req from the commit list.
1016  * If this will make the bucket empty, it will need to put the lseg reference.
1017  */
1018 static void
1019 filelayout_clear_request_commit(struct nfs_page *req,
1020                                 struct nfs_commit_info *cinfo)
1021 {
1022         struct pnfs_layout_segment *freeme = NULL;
1023
1024         spin_lock(cinfo->lock);
1025         if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
1026                 goto out;
1027         cinfo->ds->nwritten--;
1028         if (list_is_singular(&req->wb_list)) {
1029                 struct pnfs_commit_bucket *bucket;
1030
1031                 bucket = list_first_entry(&req->wb_list,
1032                                           struct pnfs_commit_bucket,
1033                                           written);
1034                 freeme = bucket->wlseg;
1035                 bucket->wlseg = NULL;
1036         }
1037 out:
1038         nfs_request_remove_commit_list(req, cinfo);
1039         spin_unlock(cinfo->lock);
1040         pnfs_put_lseg(freeme);
1041 }
1042
1043 static struct list_head *
1044 filelayout_choose_commit_list(struct nfs_page *req,
1045                               struct pnfs_layout_segment *lseg,
1046                               struct nfs_commit_info *cinfo)
1047 {
1048         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1049         u32 i, j;
1050         struct list_head *list;
1051         struct pnfs_commit_bucket *buckets;
1052
1053         if (fl->commit_through_mds)
1054                 return &cinfo->mds->list;
1055
1056         /* Note that we are calling nfs4_fl_calc_j_index on each page
1057          * that ends up being committed to a data server.  An attractive
1058          * alternative is to add a field to nfs_write_data and nfs_page
1059          * to store the value calculated in filelayout_write_pagelist
1060          * and just use that here.
1061          */
1062         j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1063         i = select_bucket_index(fl, j);
1064         buckets = cinfo->ds->buckets;
1065         list = &buckets[i].written;
1066         if (list_empty(list)) {
1067                 /* Non-empty buckets hold a reference on the lseg.  That ref
1068                  * is normally transferred to the COMMIT call and released
1069                  * there.  It could also be released if the last req is pulled
1070                  * off due to a rewrite, in which case it will be done in
1071                  * filelayout_clear_request_commit
1072                  */
1073                 buckets[i].wlseg = pnfs_get_lseg(lseg);
1074         }
1075         set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1076         cinfo->ds->nwritten++;
1077         return list;
1078 }
1079
1080 static void
1081 filelayout_mark_request_commit(struct nfs_page *req,
1082                                struct pnfs_layout_segment *lseg,
1083                                struct nfs_commit_info *cinfo)
1084 {
1085         struct list_head *list;
1086
1087         list = filelayout_choose_commit_list(req, lseg, cinfo);
1088         nfs_request_add_commit_list(req, list, cinfo);
1089 }
1090
1091 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1092 {
1093         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1094
1095         if (flseg->stripe_type == STRIPE_SPARSE)
1096                 return i;
1097         else
1098                 return nfs4_fl_calc_ds_index(lseg, i);
1099 }
1100
1101 static struct nfs_fh *
1102 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1103 {
1104         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1105
1106         if (flseg->stripe_type == STRIPE_SPARSE) {
1107                 if (flseg->num_fh == 1)
1108                         i = 0;
1109                 else if (flseg->num_fh == 0)
1110                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1111                         return NULL;
1112         }
1113         return flseg->fh_array[i];
1114 }
1115
1116 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1117 {
1118         struct pnfs_layout_segment *lseg = data->lseg;
1119         struct nfs4_pnfs_ds *ds;
1120         struct rpc_clnt *ds_clnt;
1121         u32 idx;
1122         struct nfs_fh *fh;
1123
1124         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1125         ds = nfs4_fl_prepare_ds(lseg, idx);
1126         if (!ds)
1127                 goto out_err;
1128
1129         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1130         if (IS_ERR(ds_clnt))
1131                 goto out_err;
1132
1133         dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1134                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1135         data->commit_done_cb = filelayout_commit_done_cb;
1136         atomic_inc(&ds->ds_clp->cl_count);
1137         data->ds_clp = ds->ds_clp;
1138         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1139         if (fh)
1140                 data->args.fh = fh;
1141         return nfs_initiate_commit(ds_clnt, data,
1142                                    &filelayout_commit_call_ops, how,
1143                                    RPC_TASK_SOFTCONN);
1144 out_err:
1145         prepare_to_resend_writes(data);
1146         filelayout_commit_release(data);
1147         return -EAGAIN;
1148 }
1149
1150 static int
1151 transfer_commit_list(struct list_head *src, struct list_head *dst,
1152                      struct nfs_commit_info *cinfo, int max)
1153 {
1154         struct nfs_page *req, *tmp;
1155         int ret = 0;
1156
1157         list_for_each_entry_safe(req, tmp, src, wb_list) {
1158                 if (!nfs_lock_request(req))
1159                         continue;
1160                 kref_get(&req->wb_kref);
1161                 if (cond_resched_lock(cinfo->lock))
1162                         list_safe_reset_next(req, tmp, wb_list);
1163                 nfs_request_remove_commit_list(req, cinfo);
1164                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1165                 nfs_list_add_request(req, dst);
1166                 ret++;
1167                 if ((ret == max) && !cinfo->dreq)
1168                         break;
1169         }
1170         return ret;
1171 }
1172
1173 static int
1174 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1175                                struct nfs_commit_info *cinfo,
1176                                int max)
1177 {
1178         struct list_head *src = &bucket->written;
1179         struct list_head *dst = &bucket->committing;
1180         int ret;
1181
1182         ret = transfer_commit_list(src, dst, cinfo, max);
1183         if (ret) {
1184                 cinfo->ds->nwritten -= ret;
1185                 cinfo->ds->ncommitting += ret;
1186                 bucket->clseg = bucket->wlseg;
1187                 if (list_empty(src))
1188                         bucket->wlseg = NULL;
1189                 else
1190                         pnfs_get_lseg(bucket->clseg);
1191         }
1192         return ret;
1193 }
1194
1195 /* Move reqs from written to committing lists, returning count of number moved.
1196  * Note called with cinfo->lock held.
1197  */
1198 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1199                                         int max)
1200 {
1201         int i, rv = 0, cnt;
1202
1203         for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1204                 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1205                                                      cinfo, max);
1206                 max -= cnt;
1207                 rv += cnt;
1208         }
1209         return rv;
1210 }
1211
1212 /* Pull everything off the committing lists and dump into @dst */
1213 static void filelayout_recover_commit_reqs(struct list_head *dst,
1214                                            struct nfs_commit_info *cinfo)
1215 {
1216         struct pnfs_commit_bucket *b;
1217         int i;
1218
1219         /* NOTE cinfo->lock is NOT held, relying on fact that this is
1220          * only called on single thread per dreq.
1221          * Can't take the lock because need to do pnfs_put_lseg
1222          */
1223         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1224                 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1225                         pnfs_put_lseg(b->wlseg);
1226                         b->wlseg = NULL;
1227                 }
1228         }
1229         cinfo->ds->nwritten = 0;
1230 }
1231
1232 static unsigned int
1233 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1234 {
1235         struct pnfs_ds_commit_info *fl_cinfo;
1236         struct pnfs_commit_bucket *bucket;
1237         struct nfs_commit_data *data;
1238         int i, j;
1239         unsigned int nreq = 0;
1240
1241         fl_cinfo = cinfo->ds;
1242         bucket = fl_cinfo->buckets;
1243         for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1244                 if (list_empty(&bucket->committing))
1245                         continue;
1246                 data = nfs_commitdata_alloc();
1247                 if (!data)
1248                         break;
1249                 data->ds_commit_index = i;
1250                 data->lseg = bucket->clseg;
1251                 bucket->clseg = NULL;
1252                 list_add(&data->pages, list);
1253                 nreq++;
1254         }
1255
1256         /* Clean up on error */
1257         for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1258                 if (list_empty(&bucket->committing))
1259                         continue;
1260                 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1261                 pnfs_put_lseg(bucket->clseg);
1262                 bucket->clseg = NULL;
1263         }
1264         /* Caller will clean up entries put on list */
1265         return nreq;
1266 }
1267
1268 /* This follows nfs_commit_list pretty closely */
1269 static int
1270 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1271                            int how, struct nfs_commit_info *cinfo)
1272 {
1273         struct nfs_commit_data *data, *tmp;
1274         LIST_HEAD(list);
1275         unsigned int nreq = 0;
1276
1277         if (!list_empty(mds_pages)) {
1278                 data = nfs_commitdata_alloc();
1279                 if (data != NULL) {
1280                         data->lseg = NULL;
1281                         list_add(&data->pages, &list);
1282                         nreq++;
1283                 } else
1284                         nfs_retry_commit(mds_pages, NULL, cinfo);
1285         }
1286
1287         nreq += alloc_ds_commits(cinfo, &list);
1288
1289         if (nreq == 0) {
1290                 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1291                 goto out;
1292         }
1293
1294         atomic_add(nreq, &cinfo->mds->rpcs_out);
1295
1296         list_for_each_entry_safe(data, tmp, &list, pages) {
1297                 list_del_init(&data->pages);
1298                 if (!data->lseg) {
1299                         nfs_init_commit(data, mds_pages, NULL, cinfo);
1300                         nfs_initiate_commit(NFS_CLIENT(inode), data,
1301                                             data->mds_ops, how, 0);
1302                 } else {
1303                         struct pnfs_commit_bucket *buckets;
1304
1305                         buckets = cinfo->ds->buckets;
1306                         nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1307                         filelayout_initiate_commit(data, how);
1308                 }
1309         }
1310 out:
1311         cinfo->ds->ncommitting = 0;
1312         return PNFS_ATTEMPTED;
1313 }
1314
1315 static void
1316 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1317 {
1318         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1319 }
1320
1321 static struct pnfs_layout_hdr *
1322 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1323 {
1324         struct nfs4_filelayout *flo;
1325
1326         flo = kzalloc(sizeof(*flo), gfp_flags);
1327         return &flo->generic_hdr;
1328 }
1329
1330 static void
1331 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1332 {
1333         kfree(FILELAYOUT_FROM_HDR(lo));
1334 }
1335
1336 static struct pnfs_ds_commit_info *
1337 filelayout_get_ds_info(struct inode *inode)
1338 {
1339         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1340
1341         if (layout == NULL)
1342                 return NULL;
1343         else
1344                 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1345 }
1346
1347 static struct pnfs_layoutdriver_type filelayout_type = {
1348         .id                     = LAYOUT_NFSV4_1_FILES,
1349         .name                   = "LAYOUT_NFSV4_1_FILES",
1350         .owner                  = THIS_MODULE,
1351         .alloc_layout_hdr       = filelayout_alloc_layout_hdr,
1352         .free_layout_hdr        = filelayout_free_layout_hdr,
1353         .alloc_lseg             = filelayout_alloc_lseg,
1354         .free_lseg              = filelayout_free_lseg,
1355         .pg_read_ops            = &filelayout_pg_read_ops,
1356         .pg_write_ops           = &filelayout_pg_write_ops,
1357         .get_ds_info            = &filelayout_get_ds_info,
1358         .mark_request_commit    = filelayout_mark_request_commit,
1359         .clear_request_commit   = filelayout_clear_request_commit,
1360         .scan_commit_lists      = filelayout_scan_commit_lists,
1361         .recover_commit_reqs    = filelayout_recover_commit_reqs,
1362         .commit_pagelist        = filelayout_commit_pagelist,
1363         .read_pagelist          = filelayout_read_pagelist,
1364         .write_pagelist         = filelayout_write_pagelist,
1365         .free_deviceid_node     = filelayout_free_deveiceid_node,
1366 };
1367
1368 static int __init nfs4filelayout_init(void)
1369 {
1370         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1371                __func__);
1372         return pnfs_register_layoutdriver(&filelayout_type);
1373 }
1374
1375 static void __exit nfs4filelayout_exit(void)
1376 {
1377         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1378                __func__);
1379         pnfs_unregister_layoutdriver(&filelayout_type);
1380 }
1381
1382 MODULE_ALIAS("nfs-layouttype4-1");
1383
1384 module_init(nfs4filelayout_init);
1385 module_exit(nfs4filelayout_exit);