fs/nfs/filelayout/filelayout.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 "filelayout.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_header *hdr)
  88 {
  89         struct rpc_task *task = &hdr->task;
  90
  91         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
  92                 dprintk("%s Reset task %5u for i/o through MDS "
  93                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
  94                         hdr->task.tk_pid,
  95                         hdr->inode->i_sb->s_id,
  96                         (unsigned long long)NFS_FILEID(hdr->inode),
  97                         hdr->args.count,
  98                         (unsigned long long)hdr->args.offset);
  99
 100                 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
 101         }
 102 }
 103
 104 static void filelayout_reset_read(struct nfs_pgio_header *hdr)
 105 {
 106         struct rpc_task *task = &hdr->task;
 107
 108         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 109                 dprintk("%s Reset task %5u for i/o through MDS "
 110                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
 111                         hdr->task.tk_pid,
 112                         hdr->inode->i_sb->s_id,
 113                         (unsigned long long)NFS_FILEID(hdr->inode),
 114                         hdr->args.count,
 115                         (unsigned long long)hdr->args.offset);
 116
 117                 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
 118         }
 119 }
 120
 121 static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo)
 122 {
 123         if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 124                 return;
 125         pnfs_return_layout(inode);
 126 }
 127
 128 static int filelayout_async_handle_error(struct rpc_task *task,
 129                                          struct nfs4_state *state,
 130                                          struct nfs_client *clp,
 131                                          struct pnfs_layout_segment *lseg)
 132 {
 133         struct pnfs_layout_hdr *lo = lseg->pls_layout;
 134         struct inode *inode = lo->plh_inode;
 135         struct nfs_server *mds_server = NFS_SERVER(inode);
 136         struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
 137         struct nfs_client *mds_client = mds_server->nfs_client;
 138         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
 139
 140         if (task->tk_status >= 0)
 141                 return 0;
 142
 143         switch (task->tk_status) {
 144         /* MDS state errors */
 145         case -NFS4ERR_DELEG_REVOKED:
 146         case -NFS4ERR_ADMIN_REVOKED:
 147         case -NFS4ERR_BAD_STATEID:
 148         case -NFS4ERR_OPENMODE:
 149                 if (state == NULL)
 150                         break;
 151                 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
 152                         goto out_bad_stateid;
 153                 goto wait_on_recovery;
 154         case -NFS4ERR_EXPIRED:
 155                 if (state != NULL) {
 156                         if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
 157                                 goto out_bad_stateid;
 158                 }
 159                 nfs4_schedule_lease_recovery(mds_client);
 160                 goto wait_on_recovery;
 161         /* DS session errors */
 162         case -NFS4ERR_BADSESSION:
 163         case -NFS4ERR_BADSLOT:
 164         case -NFS4ERR_BAD_HIGH_SLOT:
 165         case -NFS4ERR_DEADSESSION:
 166         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
 167         case -NFS4ERR_SEQ_FALSE_RETRY:
 168         case -NFS4ERR_SEQ_MISORDERED:
 169                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 170                         "flags 0x%x\n", __func__, task->tk_status,
 171                         clp->cl_exchange_flags);
 172                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
 173                 break;
 174         case -NFS4ERR_DELAY:
 175         case -NFS4ERR_GRACE:
 176                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 177                 break;
 178         case -NFS4ERR_RETRY_UNCACHED_REP:
 179                 break;
 180         /* Invalidate Layout errors */
 181         case -NFS4ERR_PNFS_NO_LAYOUT:
 182         case -ESTALE:           /* mapped NFS4ERR_STALE */
 183         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
 184         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
 185         case -NFS4ERR_FHEXPIRED:
 186         case -NFS4ERR_WRONG_TYPE:
 187                 dprintk("%s Invalid layout error %d\n", __func__,
 188                         task->tk_status);
 189                 /*
 190                  * Destroy layout so new i/o will get a new layout.
 191                  * Layout will not be destroyed until all current lseg
 192                  * references are put. Mark layout as invalid to resend failed
 193                  * i/o and all i/o waiting on the slot table to the MDS until
 194                  * layout is destroyed and a new valid layout is obtained.
 195                  */
 196                 pnfs_destroy_layout(NFS_I(inode));
 197                 rpc_wake_up(&tbl->slot_tbl_waitq);
 198                 goto reset;
 199         /* RPC connection errors */
 200         case -ECONNREFUSED:
 201         case -EHOSTDOWN:
 202         case -EHOSTUNREACH:
 203         case -ENETUNREACH:
 204         case -EIO:
 205         case -ETIMEDOUT:
 206         case -EPIPE:
 207                 dprintk("%s DS connection error %d\n", __func__,
 208                         task->tk_status);
 209                 nfs4_mark_deviceid_unavailable(devid);
 210                 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
 211                 rpc_wake_up(&tbl->slot_tbl_waitq);
 212                 /* fall through */
 213         default:
 214 reset:
 215                 dprintk("%s Retry through MDS. Error %d\n", __func__,
 216                         task->tk_status);
 217                 return -NFS4ERR_RESET_TO_MDS;
 218         }
 219 out:
 220         task->tk_status = 0;
 221         return -EAGAIN;
 222 out_bad_stateid:
 223         task->tk_status = -EIO;
 224         return 0;
 225 wait_on_recovery:
 226         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
 227         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
 228                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
 229         goto out;
 230 }
 231
 232 /* NFS_PROTO call done callback routines */
 233
 234 static int filelayout_read_done_cb(struct rpc_task *task,
 235                                 struct nfs_pgio_header *hdr)
 236 {
 237         int err;
 238
 239         trace_nfs4_pnfs_read(hdr, task->tk_status);
 240         err = filelayout_async_handle_error(task, hdr->args.context->state,
 241                                             hdr->ds_clp, hdr->lseg);
 242
 243         switch (err) {
 244         case -NFS4ERR_RESET_TO_MDS:
 245                 filelayout_reset_read(hdr);
 246                 return task->tk_status;
 247         case -EAGAIN:
 248                 rpc_restart_call_prepare(task);
 249                 return -EAGAIN;
 250         }
 251
 252         return 0;
 253 }
 254
 255 /*
 256  * We reference the rpc_cred of the first WRITE that triggers the need for
 257  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 258  * rfc5661 is not clear about which credential should be used.
 259  */
 260 static void
 261 filelayout_set_layoutcommit(struct nfs_pgio_header *hdr)
 262 {
 263
 264         if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
 265             hdr->res.verf->committed != NFS_DATA_SYNC)
 266                 return;
 267
 268         pnfs_set_layoutcommit(hdr);
 269         dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
 270                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
 271 }
 272
 273 bool
 274 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
 275 {
 276         return filelayout_test_devid_invalid(node) ||
 277                 nfs4_test_deviceid_unavailable(node);
 278 }
 279
 280 static bool
 281 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
 282 {
 283         struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
 284
 285         return filelayout_test_devid_unavailable(node);
 286 }
 287
 288 /*
 289  * Call ops for the async read/write cases
 290  * In the case of dense layouts, the offset needs to be reset to its
 291  * original value.
 292  */
 293 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 294 {
 295         struct nfs_pgio_header *hdr = data;
 296
 297         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
 298                 rpc_exit(task, -EIO);
 299                 return;
 300         }
 301         if (filelayout_reset_to_mds(hdr->lseg)) {
 302                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 303                 filelayout_reset_read(hdr);
 304                 rpc_exit(task, 0);
 305                 return;
 306         }
 307         hdr->pgio_done_cb = filelayout_read_done_cb;
 308
 309         if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
 310                         &hdr->args.seq_args,
 311                         &hdr->res.seq_res,
 312                         task))
 313                 return;
 314         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
 315                         hdr->args.lock_context, FMODE_READ) == -EIO)
 316                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
 317 }
 318
 319 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 320 {
 321         struct nfs_pgio_header *hdr = data;
 322
 323         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 324
 325         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
 326             task->tk_status == 0) {
 327                 nfs41_sequence_done(task, &hdr->res.seq_res);
 328                 return;
 329         }
 330
 331         /* Note this may cause RPC to be resent */
 332         hdr->mds_ops->rpc_call_done(task, data);
 333 }
 334
 335 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
 336 {
 337         struct nfs_pgio_header *hdr = data;
 338
 339         rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
 340 }
 341
 342 static void filelayout_read_release(void *data)
 343 {
 344         struct nfs_pgio_header *hdr = data;
 345         struct pnfs_layout_hdr *lo = hdr->lseg->pls_layout;
 346
 347         filelayout_fenceme(lo->plh_inode, lo);
 348         nfs_put_client(hdr->ds_clp);
 349         hdr->mds_ops->rpc_release(data);
 350 }
 351
 352 static int filelayout_write_done_cb(struct rpc_task *task,
 353                                 struct nfs_pgio_header *hdr)
 354 {
 355         int err;
 356
 357         trace_nfs4_pnfs_write(hdr, task->tk_status);
 358         err = filelayout_async_handle_error(task, hdr->args.context->state,
 359                                             hdr->ds_clp, hdr->lseg);
 360
 361         switch (err) {
 362         case -NFS4ERR_RESET_TO_MDS:
 363                 filelayout_reset_write(hdr);
 364                 return task->tk_status;
 365         case -EAGAIN:
 366                 rpc_restart_call_prepare(task);
 367                 return -EAGAIN;
 368         }
 369
 370         filelayout_set_layoutcommit(hdr);
 371         return 0;
 372 }
 373
 374 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 375 static void prepare_to_resend_writes(struct nfs_commit_data *data)
 376 {
 377         struct nfs_page *first = nfs_list_entry(data->pages.next);
 378
 379         data->task.tk_status = 0;
 380         memcpy(&data->verf.verifier, &first->wb_verf,
 381                sizeof(data->verf.verifier));
 382         data->verf.verifier.data[0]++; /* ensure verifier mismatch */
 383 }
 384
 385 static int filelayout_commit_done_cb(struct rpc_task *task,
 386                                      struct nfs_commit_data *data)
 387 {
 388         int err;
 389
 390         trace_nfs4_pnfs_commit_ds(data, task->tk_status);
 391         err = filelayout_async_handle_error(task, NULL, data->ds_clp,
 392                                             data->lseg);
 393
 394         switch (err) {
 395         case -NFS4ERR_RESET_TO_MDS:
 396                 prepare_to_resend_writes(data);
 397                 return -EAGAIN;
 398         case -EAGAIN:
 399                 rpc_restart_call_prepare(task);
 400                 return -EAGAIN;
 401         }
 402
 403         if (data->verf.committed == NFS_UNSTABLE)
 404                 pnfs_commit_set_layoutcommit(data);
 405
 406         return 0;
 407 }
 408
 409 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 410 {
 411         struct nfs_pgio_header *hdr = data;
 412
 413         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
 414                 rpc_exit(task, -EIO);
 415                 return;
 416         }
 417         if (filelayout_reset_to_mds(hdr->lseg)) {
 418                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 419                 filelayout_reset_write(hdr);
 420                 rpc_exit(task, 0);
 421                 return;
 422         }
 423         if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
 424                         &hdr->args.seq_args,
 425                         &hdr->res.seq_res,
 426                         task))
 427                 return;
 428         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
 429                         hdr->args.lock_context, FMODE_WRITE) == -EIO)
 430                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
 431 }
 432
 433 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 434 {
 435         struct nfs_pgio_header *hdr = data;
 436
 437         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
 438             task->tk_status == 0) {
 439                 nfs41_sequence_done(task, &hdr->res.seq_res);
 440                 return;
 441         }
 442
 443         /* Note this may cause RPC to be resent */
 444         hdr->mds_ops->rpc_call_done(task, data);
 445 }
 446
 447 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
 448 {
 449         struct nfs_pgio_header *hdr = data;
 450
 451         rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
 452 }
 453
 454 static void filelayout_write_release(void *data)
 455 {
 456         struct nfs_pgio_header *hdr = data;
 457         struct pnfs_layout_hdr *lo = hdr->lseg->pls_layout;
 458
 459         filelayout_fenceme(lo->plh_inode, lo);
 460         nfs_put_client(hdr->ds_clp);
 461         hdr->mds_ops->rpc_release(data);
 462 }
 463
 464 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
 465 {
 466         struct nfs_commit_data *wdata = data;
 467
 468         nfs41_setup_sequence(wdata->ds_clp->cl_session,
 469                         &wdata->args.seq_args,
 470                         &wdata->res.seq_res,
 471                         task);
 472 }
 473
 474 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
 475 {
 476         struct nfs_commit_data *wdata = data;
 477
 478         /* Note this may cause RPC to be resent */
 479         wdata->mds_ops->rpc_call_done(task, data);
 480 }
 481
 482 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
 483 {
 484         struct nfs_commit_data *cdata = data;
 485
 486         rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
 487 }
 488
 489 static void filelayout_commit_release(void *calldata)
 490 {
 491         struct nfs_commit_data *data = calldata;
 492
 493         data->completion_ops->completion(data);
 494         pnfs_put_lseg(data->lseg);
 495         nfs_put_client(data->ds_clp);
 496         nfs_commitdata_release(data);
 497 }
 498
 499 static const struct rpc_call_ops filelayout_read_call_ops = {
 500         .rpc_call_prepare = filelayout_read_prepare,
 501         .rpc_call_done = filelayout_read_call_done,
 502         .rpc_count_stats = filelayout_read_count_stats,
 503         .rpc_release = filelayout_read_release,
 504 };
 505
 506 static const struct rpc_call_ops filelayout_write_call_ops = {
 507         .rpc_call_prepare = filelayout_write_prepare,
 508         .rpc_call_done = filelayout_write_call_done,
 509         .rpc_count_stats = filelayout_write_count_stats,
 510         .rpc_release = filelayout_write_release,
 511 };
 512
 513 static const struct rpc_call_ops filelayout_commit_call_ops = {
 514         .rpc_call_prepare = filelayout_commit_prepare,
 515         .rpc_call_done = filelayout_write_commit_done,
 516         .rpc_count_stats = filelayout_commit_count_stats,
 517         .rpc_release = filelayout_commit_release,
 518 };
 519
 520 static enum pnfs_try_status
 521 filelayout_read_pagelist(struct nfs_pgio_header *hdr)
 522 {
 523         struct pnfs_layout_segment *lseg = hdr->lseg;
 524         struct nfs4_pnfs_ds *ds;
 525         struct rpc_clnt *ds_clnt;
 526         loff_t offset = hdr->args.offset;
 527         u32 j, idx;
 528         struct nfs_fh *fh;
 529
 530         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 531                 __func__, hdr->inode->i_ino,
 532                 hdr->args.pgbase, (size_t)hdr->args.count, offset);
 533
 534         /* Retrieve the correct rpc_client for the byte range */
 535         j = nfs4_fl_calc_j_index(lseg, offset);
 536         idx = nfs4_fl_calc_ds_index(lseg, j);
 537         ds = nfs4_fl_prepare_ds(lseg, idx);
 538         if (!ds)
 539                 return PNFS_NOT_ATTEMPTED;
 540
 541         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
 542         if (IS_ERR(ds_clnt))
 543                 return PNFS_NOT_ATTEMPTED;
 544
 545         dprintk("%s USE DS: %s cl_count %d\n", __func__,
 546                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 547
 548         /* No multipath support. Use first DS */
 549         atomic_inc(&ds->ds_clp->cl_count);
 550         hdr->ds_clp = ds->ds_clp;
 551         hdr->ds_idx = idx;
 552         fh = nfs4_fl_select_ds_fh(lseg, j);
 553         if (fh)
 554                 hdr->args.fh = fh;
 555
 556         hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
 557         hdr->mds_offset = offset;
 558
 559         /* Perform an asynchronous read to ds */
 560         nfs_initiate_pgio(ds_clnt, hdr,
 561                             &filelayout_read_call_ops, 0, RPC_TASK_SOFTCONN);
 562         return PNFS_ATTEMPTED;
 563 }
 564
 565 /* Perform async writes. */
 566 static enum pnfs_try_status
 567 filelayout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
 568 {
 569         struct pnfs_layout_segment *lseg = hdr->lseg;
 570         struct nfs4_pnfs_ds *ds;
 571         struct rpc_clnt *ds_clnt;
 572         loff_t offset = hdr->args.offset;
 573         u32 j, idx;
 574         struct nfs_fh *fh;
 575
 576         /* Retrieve the correct rpc_client for the byte range */
 577         j = nfs4_fl_calc_j_index(lseg, offset);
 578         idx = nfs4_fl_calc_ds_index(lseg, j);
 579         ds = nfs4_fl_prepare_ds(lseg, idx);
 580         if (!ds)
 581                 return PNFS_NOT_ATTEMPTED;
 582
 583         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
 584         if (IS_ERR(ds_clnt))
 585                 return PNFS_NOT_ATTEMPTED;
 586
 587         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
 588                 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
 589                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 590
 591         hdr->pgio_done_cb = filelayout_write_done_cb;
 592         atomic_inc(&ds->ds_clp->cl_count);
 593         hdr->ds_clp = ds->ds_clp;
 594         hdr->ds_idx = idx;
 595         fh = nfs4_fl_select_ds_fh(lseg, j);
 596         if (fh)
 597                 hdr->args.fh = fh;
 598         hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
 599
 600         /* Perform an asynchronous write */
 601         nfs_initiate_pgio(ds_clnt, hdr,
 602                                     &filelayout_write_call_ops, sync,
 603                                     RPC_TASK_SOFTCONN);
 604         return PNFS_ATTEMPTED;
 605 }
 606
 607 /*
 608  * filelayout_check_layout()
 609  *
 610  * Make sure layout segment parameters are sane WRT the device.
 611  * At this point no generic layer initialization of the lseg has occurred,
 612  * and nothing has been added to the layout_hdr cache.
 613  *
 614  */
 615 static int
 616 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 617                         struct nfs4_filelayout_segment *fl,
 618                         struct nfs4_layoutget_res *lgr,
 619                         struct nfs4_deviceid *id,
 620                         gfp_t gfp_flags)
 621 {
 622         struct nfs4_deviceid_node *d;
 623         struct nfs4_file_layout_dsaddr *dsaddr;
 624         int status = -EINVAL;
 625
 626         dprintk("--> %s\n", __func__);
 627
 628         /* FIXME: remove this check when layout segment support is added */
 629         if (lgr->range.offset != 0 ||
 630             lgr->range.length != NFS4_MAX_UINT64) {
 631                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 632                         __func__);
 633                 goto out;
 634         }
 635
 636         if (fl->pattern_offset > lgr->range.offset) {
 637                 dprintk("%s pattern_offset %lld too large\n",
 638                                 __func__, fl->pattern_offset);
 639                 goto out;
 640         }
 641
 642         if (!fl->stripe_unit) {
 643                 dprintk("%s Invalid stripe unit (%u)\n",
 644                         __func__, fl->stripe_unit);
 645                 goto out;
 646         }
 647
 648         /* find and reference the deviceid */
 649         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode), id,
 650                         lo->plh_lc_cred, gfp_flags);
 651         if (d == NULL)
 652                 goto out;
 653
 654         dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 655         /* Found deviceid is unavailable */
 656         if (filelayout_test_devid_unavailable(&dsaddr->id_node))
 657                 goto out_put;
 658
 659         fl->dsaddr = dsaddr;
 660
 661         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 662                 dprintk("%s Bad first_stripe_index %u\n",
 663                                 __func__, fl->first_stripe_index);
 664                 goto out_put;
 665         }
 666
 667         if ((fl->stripe_type == STRIPE_SPARSE &&
 668             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 669             (fl->stripe_type == STRIPE_DENSE &&
 670             fl->num_fh != dsaddr->stripe_count)) {
 671                 dprintk("%s num_fh %u not valid for given packing\n",
 672                         __func__, fl->num_fh);
 673                 goto out_put;
 674         }
 675
 676         status = 0;
 677 out:
 678         dprintk("--> %s returns %d\n", __func__, status);
 679         return status;
 680 out_put:
 681         nfs4_fl_put_deviceid(dsaddr);
 682         goto out;
 683 }
 684
 685 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 686 {
 687         int i;
 688
 689         for (i = 0; i < fl->num_fh; i++) {
 690                 if (!fl->fh_array[i])
 691                         break;
 692                 kfree(fl->fh_array[i]);
 693         }
 694         kfree(fl->fh_array);
 695         fl->fh_array = NULL;
 696 }
 697
 698 static void
 699 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 700 {
 701         filelayout_free_fh_array(fl);
 702         kfree(fl);
 703 }
 704
 705 static int
 706 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 707                          struct nfs4_filelayout_segment *fl,
 708                          struct nfs4_layoutget_res *lgr,
 709                          struct nfs4_deviceid *id,
 710                          gfp_t gfp_flags)
 711 {
 712         struct xdr_stream stream;
 713         struct xdr_buf buf;
 714         struct page *scratch;
 715         __be32 *p;
 716         uint32_t nfl_util;
 717         int i;
 718
 719         dprintk("%s: set_layout_map Begin\n", __func__);
 720
 721         scratch = alloc_page(gfp_flags);
 722         if (!scratch)
 723                 return -ENOMEM;
 724
 725         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 726         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 727
 728         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 729          * num_fh (4) */
 730         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 731         if (unlikely(!p))
 732                 goto out_err;
 733
 734         memcpy(id, p, sizeof(*id));
 735         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 736         nfs4_print_deviceid(id);
 737
 738         nfl_util = be32_to_cpup(p++);
 739         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 740                 fl->commit_through_mds = 1;
 741         if (nfl_util & NFL4_UFLG_DENSE)
 742                 fl->stripe_type = STRIPE_DENSE;
 743         else
 744                 fl->stripe_type = STRIPE_SPARSE;
 745         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 746
 747         fl->first_stripe_index = be32_to_cpup(p++);
 748         p = xdr_decode_hyper(p, &fl->pattern_offset);
 749         fl->num_fh = be32_to_cpup(p++);
 750
 751         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 752                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 753                 fl->pattern_offset);
 754
 755         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 756          * Futher checking is done in filelayout_check_layout */
 757         if (fl->num_fh >
 758             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 759                 goto out_err;
 760
 761         if (fl->num_fh > 0) {
 762                 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
 763                                        gfp_flags);
 764                 if (!fl->fh_array)
 765                         goto out_err;
 766         }
 767
 768         for (i = 0; i < fl->num_fh; i++) {
 769                 /* Do we want to use a mempool here? */
 770                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 771                 if (!fl->fh_array[i])
 772                         goto out_err_free;
 773
 774                 p = xdr_inline_decode(&stream, 4);
 775                 if (unlikely(!p))
 776                         goto out_err_free;
 777                 fl->fh_array[i]->size = be32_to_cpup(p++);
 778                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 779                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
 780                                i, fl->fh_array[i]->size);
 781                         goto out_err_free;
 782                 }
 783
 784                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 785                 if (unlikely(!p))
 786                         goto out_err_free;
 787                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 788                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 789                         fl->fh_array[i]->size);
 790         }
 791
 792         __free_page(scratch);
 793         return 0;
 794
 795 out_err_free:
 796         filelayout_free_fh_array(fl);
 797 out_err:
 798         __free_page(scratch);
 799         return -EIO;
 800 }
 801
 802 static void
 803 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 804 {
 805         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 806
 807         dprintk("--> %s\n", __func__);
 808         nfs4_fl_put_deviceid(fl->dsaddr);
 809         /* This assumes a single RW lseg */
 810         if (lseg->pls_range.iomode == IOMODE_RW) {
 811                 struct nfs4_filelayout *flo;
 812
 813                 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
 814                 flo->commit_info.nbuckets = 0;
 815                 kfree(flo->commit_info.buckets);
 816                 flo->commit_info.buckets = NULL;
 817         }
 818         _filelayout_free_lseg(fl);
 819 }
 820
 821 static int
 822 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 823                              struct nfs_commit_info *cinfo,
 824                              gfp_t gfp_flags)
 825 {
 826         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 827         struct pnfs_commit_bucket *buckets;
 828         int size, i;
 829
 830         if (fl->commit_through_mds)
 831                 return 0;
 832
 833         size = (fl->stripe_type == STRIPE_SPARSE) ?
 834                 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 835
 836         if (cinfo->ds->nbuckets >= size) {
 837                 /* This assumes there is only one IOMODE_RW lseg.  What
 838                  * we really want to do is have a layout_hdr level
 839                  * dictionary of <multipath_list4, fh> keys, each
 840                  * associated with a struct list_head, populated by calls
 841                  * to filelayout_write_pagelist().
 842                  * */
 843                 return 0;
 844         }
 845
 846         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 847                           gfp_flags);
 848         if (!buckets)
 849                 return -ENOMEM;
 850         for (i = 0; i < size; i++) {
 851                 INIT_LIST_HEAD(&buckets[i].written);
 852                 INIT_LIST_HEAD(&buckets[i].committing);
 853                 /* mark direct verifier as unset */
 854                 buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW;
 855         }
 856
 857         spin_lock(cinfo->lock);
 858         if (cinfo->ds->nbuckets >= size)
 859                 goto out;
 860         for (i = 0; i < cinfo->ds->nbuckets; i++) {
 861                 list_splice(&cinfo->ds->buckets[i].written,
 862                             &buckets[i].written);
 863                 list_splice(&cinfo->ds->buckets[i].committing,
 864                             &buckets[i].committing);
 865                 buckets[i].direct_verf.committed =
 866                         cinfo->ds->buckets[i].direct_verf.committed;
 867                 buckets[i].wlseg = cinfo->ds->buckets[i].wlseg;
 868                 buckets[i].clseg = cinfo->ds->buckets[i].clseg;
 869         }
 870         swap(cinfo->ds->buckets, buckets);
 871         cinfo->ds->nbuckets = size;
 872 out:
 873         spin_unlock(cinfo->lock);
 874         kfree(buckets);
 875         return 0;
 876 }
 877
 878 static struct pnfs_layout_segment *
 879 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 880                       struct nfs4_layoutget_res *lgr,
 881                       gfp_t gfp_flags)
 882 {
 883         struct nfs4_filelayout_segment *fl;
 884         int rc;
 885         struct nfs4_deviceid id;
 886
 887         dprintk("--> %s\n", __func__);
 888         fl = kzalloc(sizeof(*fl), gfp_flags);
 889         if (!fl)
 890                 return NULL;
 891
 892         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 893         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 894                 _filelayout_free_lseg(fl);
 895                 return NULL;
 896         }
 897         return &fl->generic_hdr;
 898 }
 899
 900 /*
 901  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 902  *
 903  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
 904  * of bytes (maximum @req->wb_bytes) that can be coalesced.
 905  */
 906 static size_t
 907 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 908                    struct nfs_page *req)
 909 {
 910         unsigned int size;
 911         u64 p_stripe, r_stripe;
 912         u32 stripe_offset;
 913         u64 segment_offset = pgio->pg_lseg->pls_range.offset;
 914         u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 915
 916         /* calls nfs_generic_pg_test */
 917         size = pnfs_generic_pg_test(pgio, prev, req);
 918         if (!size)
 919                 return 0;
 920
 921         /* see if req and prev are in the same stripe */
 922         if (prev) {
 923                 p_stripe = (u64)req_offset(prev) - segment_offset;
 924                 r_stripe = (u64)req_offset(req) - segment_offset;
 925                 do_div(p_stripe, stripe_unit);
 926                 do_div(r_stripe, stripe_unit);
 927
 928                 if (p_stripe != r_stripe)
 929                         return 0;
 930         }
 931
 932         /* calculate remaining bytes in the current stripe */
 933         div_u64_rem((u64)req_offset(req) - segment_offset,
 934                         stripe_unit,
 935                         &stripe_offset);
 936         WARN_ON_ONCE(stripe_offset > stripe_unit);
 937         if (stripe_offset >= stripe_unit)
 938                 return 0;
 939         return min(stripe_unit - (unsigned int)stripe_offset, size);
 940 }
 941
 942 static void
 943 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 944                         struct nfs_page *req)
 945 {
 946         if (!pgio->pg_lseg)
 947                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 948                                            req->wb_context,
 949                                            0,
 950                                            NFS4_MAX_UINT64,
 951                                            IOMODE_READ,
 952                                            GFP_KERNEL);
 953         /* If no lseg, fall back to read through mds */
 954         if (pgio->pg_lseg == NULL)
 955                 nfs_pageio_reset_read_mds(pgio);
 956 }
 957
 958 static void
 959 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 960                          struct nfs_page *req)
 961 {
 962         struct nfs_commit_info cinfo;
 963         int status;
 964
 965         if (!pgio->pg_lseg)
 966                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 967                                            req->wb_context,
 968                                            0,
 969                                            NFS4_MAX_UINT64,
 970                                            IOMODE_RW,
 971                                            GFP_NOFS);
 972         /* If no lseg, fall back to write through mds */
 973         if (pgio->pg_lseg == NULL)
 974                 goto out_mds;
 975         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 976         status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 977         if (status < 0) {
 978                 pnfs_put_lseg(pgio->pg_lseg);
 979                 pgio->pg_lseg = NULL;
 980                 goto out_mds;
 981         }
 982         return;
 983 out_mds:
 984         nfs_pageio_reset_write_mds(pgio);
 985 }
 986
 987 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 988         .pg_init = filelayout_pg_init_read,
 989         .pg_test = filelayout_pg_test,
 990         .pg_doio = pnfs_generic_pg_readpages,
 991 };
 992
 993 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 994         .pg_init = filelayout_pg_init_write,
 995         .pg_test = filelayout_pg_test,
 996         .pg_doio = pnfs_generic_pg_writepages,
 997 };
 998
 999 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
1000 {
1001         if (fl->stripe_type == STRIPE_SPARSE)
1002                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
1003         else
1004                 return j;
1005 }
1006
1007 /* The generic layer is about to remove the req from the commit list.
1008  * If this will make the bucket empty, it will need to put the lseg reference.
1009  * Note this is must be called holding the inode (/cinfo) lock
1010  */
1011 static void
1012 filelayout_clear_request_commit(struct nfs_page *req,
1013                                 struct nfs_commit_info *cinfo)
1014 {
1015         struct pnfs_layout_segment *freeme = NULL;
1016
1017         if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
1018                 goto out;
1019         cinfo->ds->nwritten--;
1020         if (list_is_singular(&req->wb_list)) {
1021                 struct pnfs_commit_bucket *bucket;
1022
1023                 bucket = list_first_entry(&req->wb_list,
1024                                           struct pnfs_commit_bucket,
1025                                           written);
1026                 freeme = bucket->wlseg;
1027                 bucket->wlseg = NULL;
1028         }
1029 out:
1030         nfs_request_remove_commit_list(req, cinfo);
1031         pnfs_put_lseg_locked(freeme);
1032 }
1033
1034 static void
1035 filelayout_mark_request_commit(struct nfs_page *req,
1036                                struct pnfs_layout_segment *lseg,
1037                                struct nfs_commit_info *cinfo)
1038
1039 {
1040         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1041         u32 i, j;
1042         struct list_head *list;
1043         struct pnfs_commit_bucket *buckets;
1044
1045         if (fl->commit_through_mds) {
1046                 list = &cinfo->mds->list;
1047                 spin_lock(cinfo->lock);
1048                 goto mds_commit;
1049         }
1050
1051         /* Note that we are calling nfs4_fl_calc_j_index on each page
1052          * that ends up being committed to a data server.  An attractive
1053          * alternative is to add a field to nfs_write_data and nfs_page
1054          * to store the value calculated in filelayout_write_pagelist
1055          * and just use that here.
1056          */
1057         j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1058         i = select_bucket_index(fl, j);
1059         spin_lock(cinfo->lock);
1060         buckets = cinfo->ds->buckets;
1061         list = &buckets[i].written;
1062         if (list_empty(list)) {
1063                 /* Non-empty buckets hold a reference on the lseg.  That ref
1064                  * is normally transferred to the COMMIT call and released
1065                  * there.  It could also be released if the last req is pulled
1066                  * off due to a rewrite, in which case it will be done in
1067                  * filelayout_clear_request_commit
1068                  */
1069                 buckets[i].wlseg = pnfs_get_lseg(lseg);
1070         }
1071         set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1072         cinfo->ds->nwritten++;
1073
1074 mds_commit:
1075         /* nfs_request_add_commit_list(). We need to add req to list without
1076          * dropping cinfo lock.
1077          */
1078         set_bit(PG_CLEAN, &(req)->wb_flags);
1079         nfs_list_add_request(req, list);
1080         cinfo->mds->ncommit++;
1081         spin_unlock(cinfo->lock);
1082         if (!cinfo->dreq) {
1083                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1084                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1085                              BDI_RECLAIMABLE);
1086                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
1087                                    I_DIRTY_DATASYNC);
1088         }
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 /* Note called with cinfo->lock held. */
1174 static int
1175 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1176                                struct nfs_commit_info *cinfo,
1177                                int max)
1178 {
1179         struct list_head *src = &bucket->written;
1180         struct list_head *dst = &bucket->committing;
1181         int ret;
1182
1183         ret = transfer_commit_list(src, dst, cinfo, max);
1184         if (ret) {
1185                 cinfo->ds->nwritten -= ret;
1186                 cinfo->ds->ncommitting += ret;
1187                 bucket->clseg = bucket->wlseg;
1188                 if (list_empty(src))
1189                         bucket->wlseg = NULL;
1190                 else
1191                         pnfs_get_lseg(bucket->clseg);
1192         }
1193         return ret;
1194 }
1195
1196 /* Move reqs from written to committing lists, returning count of number moved.
1197  * Note called with cinfo->lock held.
1198  */
1199 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1200                                         int max)
1201 {
1202         int i, rv = 0, cnt;
1203
1204         for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1205                 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1206                                                      cinfo, max);
1207                 max -= cnt;
1208                 rv += cnt;
1209         }
1210         return rv;
1211 }
1212
1213 /* Pull everything off the committing lists and dump into @dst */
1214 static void filelayout_recover_commit_reqs(struct list_head *dst,
1215                                            struct nfs_commit_info *cinfo)
1216 {
1217         struct pnfs_commit_bucket *b;
1218         struct pnfs_layout_segment *freeme;
1219         int i;
1220
1221 restart:
1222         spin_lock(cinfo->lock);
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                         freeme = b->wlseg;
1226                         b->wlseg = NULL;
1227                         spin_unlock(cinfo->lock);
1228                         pnfs_put_lseg(freeme);
1229                         goto restart;
1230                 }
1231         }
1232         cinfo->ds->nwritten = 0;
1233         spin_unlock(cinfo->lock);
1234 }
1235
1236 /* filelayout_search_commit_reqs - Search lists in @cinfo for the head reqest
1237  *                                 for @page
1238  * @cinfo - commit info for current inode
1239  * @page - page to search for matching head request
1240  *
1241  * Returns a the head request if one is found, otherwise returns NULL.
1242  */
1243 static struct nfs_page *
1244 filelayout_search_commit_reqs(struct nfs_commit_info *cinfo, struct page *page)
1245 {
1246         struct nfs_page *freq, *t;
1247         struct pnfs_commit_bucket *b;
1248         int i;
1249
1250         /* Linearly search the commit lists for each bucket until a matching
1251          * request is found */
1252         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1253                 list_for_each_entry_safe(freq, t, &b->written, wb_list) {
1254                         if (freq->wb_page == page)
1255                                 return freq->wb_head;
1256                 }
1257                 list_for_each_entry_safe(freq, t, &b->committing, wb_list) {
1258                         if (freq->wb_page == page)
1259                                 return freq->wb_head;
1260                 }
1261         }
1262
1263         return NULL;
1264 }
1265
1266 static void filelayout_retry_commit(struct nfs_commit_info *cinfo, int idx)
1267 {
1268         struct pnfs_ds_commit_info *fl_cinfo = cinfo->ds;
1269         struct pnfs_commit_bucket *bucket;
1270         struct pnfs_layout_segment *freeme;
1271         int i;
1272
1273         for (i = idx; i < fl_cinfo->nbuckets; i++) {
1274                 bucket = &fl_cinfo->buckets[i];
1275                 if (list_empty(&bucket->committing))
1276                         continue;
1277                 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1278                 spin_lock(cinfo->lock);
1279                 freeme = bucket->clseg;
1280                 bucket->clseg = NULL;
1281                 spin_unlock(cinfo->lock);
1282                 pnfs_put_lseg(freeme);
1283         }
1284 }
1285
1286 static unsigned int
1287 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1288 {
1289         struct pnfs_ds_commit_info *fl_cinfo;
1290         struct pnfs_commit_bucket *bucket;
1291         struct nfs_commit_data *data;
1292         int i;
1293         unsigned int nreq = 0;
1294
1295         fl_cinfo = cinfo->ds;
1296         bucket = fl_cinfo->buckets;
1297         for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1298                 if (list_empty(&bucket->committing))
1299                         continue;
1300                 data = nfs_commitdata_alloc();
1301                 if (!data)
1302                         break;
1303                 data->ds_commit_index = i;
1304                 spin_lock(cinfo->lock);
1305                 data->lseg = bucket->clseg;
1306                 bucket->clseg = NULL;
1307                 spin_unlock(cinfo->lock);
1308                 list_add(&data->pages, list);
1309                 nreq++;
1310         }
1311
1312         /* Clean up on error */
1313         filelayout_retry_commit(cinfo, i);
1314         /* Caller will clean up entries put on list */
1315         return nreq;
1316 }
1317
1318 /* This follows nfs_commit_list pretty closely */
1319 static int
1320 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1321                            int how, struct nfs_commit_info *cinfo)
1322 {
1323         struct nfs_commit_data *data, *tmp;
1324         LIST_HEAD(list);
1325         unsigned int nreq = 0;
1326
1327         if (!list_empty(mds_pages)) {
1328                 data = nfs_commitdata_alloc();
1329                 if (data != NULL) {
1330                         data->lseg = NULL;
1331                         list_add(&data->pages, &list);
1332                         nreq++;
1333                 } else {
1334                         nfs_retry_commit(mds_pages, NULL, cinfo);
1335                         filelayout_retry_commit(cinfo, 0);
1336                         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1337                         return -ENOMEM;
1338                 }
1339         }
1340
1341         nreq += alloc_ds_commits(cinfo, &list);
1342
1343         if (nreq == 0) {
1344                 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1345                 goto out;
1346         }
1347
1348         atomic_add(nreq, &cinfo->mds->rpcs_out);
1349
1350         list_for_each_entry_safe(data, tmp, &list, pages) {
1351                 list_del_init(&data->pages);
1352                 if (!data->lseg) {
1353                         nfs_init_commit(data, mds_pages, NULL, cinfo);
1354                         nfs_initiate_commit(NFS_CLIENT(inode), data,
1355                                             data->mds_ops, how, 0);
1356                 } else {
1357                         struct pnfs_commit_bucket *buckets;
1358
1359                         buckets = cinfo->ds->buckets;
1360                         nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1361                         filelayout_initiate_commit(data, how);
1362                 }
1363         }
1364 out:
1365         cinfo->ds->ncommitting = 0;
1366         return PNFS_ATTEMPTED;
1367 }
1368 static struct nfs4_deviceid_node *
1369 filelayout_alloc_deviceid_node(struct nfs_server *server,
1370                 struct pnfs_device *pdev, gfp_t gfp_flags)
1371 {
1372         struct nfs4_file_layout_dsaddr *dsaddr;
1373
1374         dsaddr = nfs4_fl_alloc_deviceid_node(server, pdev, gfp_flags);
1375         if (!dsaddr)
1376                 return NULL;
1377         return &dsaddr->id_node;
1378 }
1379
1380 static void
1381 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1382 {
1383         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1384 }
1385
1386 static struct pnfs_layout_hdr *
1387 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1388 {
1389         struct nfs4_filelayout *flo;
1390
1391         flo = kzalloc(sizeof(*flo), gfp_flags);
1392         return flo != NULL ? &flo->generic_hdr : NULL;
1393 }
1394
1395 static void
1396 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1397 {
1398         kfree(FILELAYOUT_FROM_HDR(lo));
1399 }
1400
1401 static struct pnfs_ds_commit_info *
1402 filelayout_get_ds_info(struct inode *inode)
1403 {
1404         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1405
1406         if (layout == NULL)
1407                 return NULL;
1408         else
1409                 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1410 }
1411
1412 static struct pnfs_layoutdriver_type filelayout_type = {
1413         .id                     = LAYOUT_NFSV4_1_FILES,
1414         .name                   = "LAYOUT_NFSV4_1_FILES",
1415         .owner                  = THIS_MODULE,
1416         .alloc_layout_hdr       = filelayout_alloc_layout_hdr,
1417         .free_layout_hdr        = filelayout_free_layout_hdr,
1418         .alloc_lseg             = filelayout_alloc_lseg,
1419         .free_lseg              = filelayout_free_lseg,
1420         .pg_read_ops            = &filelayout_pg_read_ops,
1421         .pg_write_ops           = &filelayout_pg_write_ops,
1422         .get_ds_info            = &filelayout_get_ds_info,
1423         .mark_request_commit    = filelayout_mark_request_commit,
1424         .clear_request_commit   = filelayout_clear_request_commit,
1425         .scan_commit_lists      = filelayout_scan_commit_lists,
1426         .recover_commit_reqs    = filelayout_recover_commit_reqs,
1427         .search_commit_reqs     = filelayout_search_commit_reqs,
1428         .commit_pagelist        = filelayout_commit_pagelist,
1429         .read_pagelist          = filelayout_read_pagelist,
1430         .write_pagelist         = filelayout_write_pagelist,
1431         .alloc_deviceid_node    = filelayout_alloc_deviceid_node,
1432         .free_deviceid_node     = filelayout_free_deveiceid_node,
1433 };
1434
1435 static int __init nfs4filelayout_init(void)
1436 {
1437         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1438                __func__);
1439         return pnfs_register_layoutdriver(&filelayout_type);
1440 }
1441
1442 static void __exit nfs4filelayout_exit(void)
1443 {
1444         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1445                __func__);
1446         pnfs_unregister_layoutdriver(&filelayout_type);
1447 }
1448
1449 MODULE_ALIAS("nfs-layouttype4-1");
1450
1451 module_init(nfs4filelayout_init);
1452 module_exit(nfs4filelayout_exit);