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 "internal.h"
  39 #include "delegation.h"
  40 #include "nfs4filelayout.h"
  41
  42 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  43
  44 MODULE_LICENSE("GPL");
  45 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
  46 MODULE_DESCRIPTION("The NFSv4 file layout driver");
  47
  48 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
  49
  50 static loff_t
  51 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
  52                             loff_t offset)
  53 {
  54         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
  55         u64 stripe_no;
  56         u32 rem;
  57
  58         offset -= flseg->pattern_offset;
  59         stripe_no = div_u64(offset, stripe_width);
  60         div_u64_rem(offset, flseg->stripe_unit, &rem);
  61
  62         return stripe_no * flseg->stripe_unit + rem;
  63 }
  64
  65 /* This function is used by the layout driver to calculate the
  66  * offset of the file on the dserver based on whether the
  67  * layout type is STRIPE_DENSE or STRIPE_SPARSE
  68  */
  69 static loff_t
  70 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
  71 {
  72         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
  73
  74         switch (flseg->stripe_type) {
  75         case STRIPE_SPARSE:
  76                 return offset;
  77
  78         case STRIPE_DENSE:
  79                 return filelayout_get_dense_offset(flseg, offset);
  80         }
  81
  82         BUG();
  83 }
  84
  85 static int filelayout_async_handle_error(struct rpc_task *task,
  86                                          struct nfs4_state *state,
  87                                          struct nfs_client *clp,
  88                                          int *reset)
  89 {
  90         struct nfs_server *mds_server = NFS_SERVER(state->inode);
  91         struct nfs_client *mds_client = mds_server->nfs_client;
  92
  93         if (task->tk_status >= 0)
  94                 return 0;
  95         *reset = 0;
  96
  97         switch (task->tk_status) {
  98         /* MDS state errors */
  99         case -NFS4ERR_DELEG_REVOKED:
 100         case -NFS4ERR_ADMIN_REVOKED:
 101         case -NFS4ERR_BAD_STATEID:
 102                 nfs_remove_bad_delegation(state->inode);
 103         case -NFS4ERR_OPENMODE:
 104                 nfs4_schedule_stateid_recovery(mds_server, state);
 105                 goto wait_on_recovery;
 106         case -NFS4ERR_EXPIRED:
 107                 nfs4_schedule_stateid_recovery(mds_server, state);
 108                 nfs4_schedule_lease_recovery(mds_client);
 109                 goto wait_on_recovery;
 110         /* DS session errors */
 111         case -NFS4ERR_BADSESSION:
 112         case -NFS4ERR_BADSLOT:
 113         case -NFS4ERR_BAD_HIGH_SLOT:
 114         case -NFS4ERR_DEADSESSION:
 115         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
 116         case -NFS4ERR_SEQ_FALSE_RETRY:
 117         case -NFS4ERR_SEQ_MISORDERED:
 118                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 119                         "flags 0x%x\n", __func__, task->tk_status,
 120                         clp->cl_exchange_flags);
 121                 nfs4_schedule_session_recovery(clp->cl_session);
 122                 break;
 123         case -NFS4ERR_DELAY:
 124         case -NFS4ERR_GRACE:
 125         case -EKEYEXPIRED:
 126                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 127                 break;
 128         case -NFS4ERR_RETRY_UNCACHED_REP:
 129                 break;
 130         default:
 131                 dprintk("%s DS error. Retry through MDS %d\n", __func__,
 132                         task->tk_status);
 133                 *reset = 1;
 134                 break;
 135         }
 136 out:
 137         task->tk_status = 0;
 138         return -EAGAIN;
 139 wait_on_recovery:
 140         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
 141         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
 142                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
 143         goto out;
 144 }
 145
 146 /* NFS_PROTO call done callback routines */
 147
 148 static int filelayout_read_done_cb(struct rpc_task *task,
 149                                 struct nfs_read_data *data)
 150 {
 151         struct nfs_pgio_header *hdr = data->header;
 152         int reset = 0;
 153
 154         dprintk("%s DS read\n", __func__);
 155
 156         if (filelayout_async_handle_error(task, data->args.context->state,
 157                                           data->ds_clp, &reset) == -EAGAIN) {
 158                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 159                         __func__, data->ds_clp, data->ds_clp->cl_session);
 160                 if (reset) {
 161                         pnfs_set_lo_fail(hdr->lseg);
 162                         nfs4_reset_read(task, data);
 163                 }
 164                 rpc_restart_call_prepare(task);
 165                 return -EAGAIN;
 166         }
 167
 168         return 0;
 169 }
 170
 171 /*
 172  * We reference the rpc_cred of the first WRITE that triggers the need for
 173  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 174  * rfc5661 is not clear about which credential should be used.
 175  */
 176 static void
 177 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
 178 {
 179         struct nfs_pgio_header *hdr = wdata->header;
 180
 181         if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
 182             wdata->res.verf->committed == NFS_FILE_SYNC)
 183                 return;
 184
 185         pnfs_set_layoutcommit(wdata);
 186         dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
 187                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
 188 }
 189
 190 /*
 191  * Call ops for the async read/write cases
 192  * In the case of dense layouts, the offset needs to be reset to its
 193  * original value.
 194  */
 195 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 196 {
 197         struct nfs_read_data *rdata = data;
 198
 199         rdata->read_done_cb = filelayout_read_done_cb;
 200
 201         if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
 202                                 &rdata->args.seq_args, &rdata->res.seq_res,
 203                                 task))
 204                 return;
 205
 206         rpc_call_start(task);
 207 }
 208
 209 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 210 {
 211         struct nfs_read_data *rdata = data;
 212
 213         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 214
 215         /* Note this may cause RPC to be resent */
 216         rdata->header->mds_ops->rpc_call_done(task, data);
 217 }
 218
 219 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
 220 {
 221         struct nfs_read_data *rdata = data;
 222
 223         rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
 224 }
 225
 226 static void filelayout_read_release(void *data)
 227 {
 228         struct nfs_read_data *rdata = data;
 229
 230         rdata->header->mds_ops->rpc_release(data);
 231 }
 232
 233 static int filelayout_write_done_cb(struct rpc_task *task,
 234                                 struct nfs_write_data *data)
 235 {
 236         struct nfs_pgio_header *hdr = data->header;
 237         int reset = 0;
 238
 239         if (filelayout_async_handle_error(task, data->args.context->state,
 240                                           data->ds_clp, &reset) == -EAGAIN) {
 241                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 242                         __func__, data->ds_clp, data->ds_clp->cl_session);
 243                 if (reset) {
 244                         pnfs_set_lo_fail(hdr->lseg);
 245                         nfs4_reset_write(task, data);
 246                 }
 247                 rpc_restart_call_prepare(task);
 248                 return -EAGAIN;
 249         }
 250
 251         filelayout_set_layoutcommit(data);
 252         return 0;
 253 }
 254
 255 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 256 static void prepare_to_resend_writes(struct nfs_commit_data *data)
 257 {
 258         struct nfs_page *first = nfs_list_entry(data->pages.next);
 259
 260         data->task.tk_status = 0;
 261         memcpy(data->verf.verifier, first->wb_verf.verifier,
 262                sizeof(first->wb_verf.verifier));
 263         data->verf.verifier[0]++; /* ensure verifier mismatch */
 264 }
 265
 266 static int filelayout_commit_done_cb(struct rpc_task *task,
 267                                      struct nfs_commit_data *data)
 268 {
 269         int reset = 0;
 270
 271         if (filelayout_async_handle_error(task, data->context->state,
 272                                           data->ds_clp, &reset) == -EAGAIN) {
 273                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 274                         __func__, data->ds_clp, data->ds_clp->cl_session);
 275                 if (reset) {
 276                         prepare_to_resend_writes(data);
 277                         pnfs_set_lo_fail(data->lseg);
 278                 } else
 279                         rpc_restart_call_prepare(task);
 280                 return -EAGAIN;
 281         }
 282
 283         return 0;
 284 }
 285
 286 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 287 {
 288         struct nfs_write_data *wdata = data;
 289
 290         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 291                                 &wdata->args.seq_args, &wdata->res.seq_res,
 292                                 task))
 293                 return;
 294
 295         rpc_call_start(task);
 296 }
 297
 298 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 299 {
 300         struct nfs_write_data *wdata = data;
 301
 302         /* Note this may cause RPC to be resent */
 303         wdata->header->mds_ops->rpc_call_done(task, data);
 304 }
 305
 306 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
 307 {
 308         struct nfs_write_data *wdata = data;
 309
 310         rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
 311 }
 312
 313 static void filelayout_write_release(void *data)
 314 {
 315         struct nfs_write_data *wdata = data;
 316
 317         wdata->header->mds_ops->rpc_release(data);
 318 }
 319
 320 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
 321 {
 322         struct nfs_commit_data *wdata = data;
 323
 324         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 325                                 &wdata->args.seq_args, &wdata->res.seq_res,
 326                                 task))
 327                 return;
 328
 329         rpc_call_start(task);
 330 }
 331
 332 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
 333 {
 334         struct nfs_commit_data *wdata = data;
 335
 336         /* Note this may cause RPC to be resent */
 337         wdata->mds_ops->rpc_call_done(task, data);
 338 }
 339
 340 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
 341 {
 342         struct nfs_commit_data *cdata = data;
 343
 344         rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
 345 }
 346
 347 static void filelayout_commit_release(void *calldata)
 348 {
 349         struct nfs_commit_data *data = calldata;
 350
 351         nfs_commit_release_pages(data);
 352         if (atomic_dec_and_test(&NFS_I(data->inode)->commits_outstanding))
 353                 nfs_commit_clear_lock(NFS_I(data->inode));
 354         put_lseg(data->lseg);
 355         nfs_commitdata_release(data);
 356 }
 357
 358 static const struct rpc_call_ops filelayout_read_call_ops = {
 359         .rpc_call_prepare = filelayout_read_prepare,
 360         .rpc_call_done = filelayout_read_call_done,
 361         .rpc_count_stats = filelayout_read_count_stats,
 362         .rpc_release = filelayout_read_release,
 363 };
 364
 365 static const struct rpc_call_ops filelayout_write_call_ops = {
 366         .rpc_call_prepare = filelayout_write_prepare,
 367         .rpc_call_done = filelayout_write_call_done,
 368         .rpc_count_stats = filelayout_write_count_stats,
 369         .rpc_release = filelayout_write_release,
 370 };
 371
 372 static const struct rpc_call_ops filelayout_commit_call_ops = {
 373         .rpc_call_prepare = filelayout_commit_prepare,
 374         .rpc_call_done = filelayout_write_commit_done,
 375         .rpc_count_stats = filelayout_commit_count_stats,
 376         .rpc_release = filelayout_commit_release,
 377 };
 378
 379 static enum pnfs_try_status
 380 filelayout_read_pagelist(struct nfs_read_data *data)
 381 {
 382         struct nfs_pgio_header *hdr = data->header;
 383         struct pnfs_layout_segment *lseg = hdr->lseg;
 384         struct nfs4_pnfs_ds *ds;
 385         loff_t offset = data->args.offset;
 386         u32 j, idx;
 387         struct nfs_fh *fh;
 388         int status;
 389
 390         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 391                 __func__, hdr->inode->i_ino,
 392                 data->args.pgbase, (size_t)data->args.count, offset);
 393
 394         if (test_bit(NFS_DEVICEID_INVALID, &FILELAYOUT_DEVID_NODE(lseg)->flags))
 395                 return PNFS_NOT_ATTEMPTED;
 396
 397         /* Retrieve the correct rpc_client for the byte range */
 398         j = nfs4_fl_calc_j_index(lseg, offset);
 399         idx = nfs4_fl_calc_ds_index(lseg, j);
 400         ds = nfs4_fl_prepare_ds(lseg, idx);
 401         if (!ds) {
 402                 /* Either layout fh index faulty, or ds connect failed */
 403                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 404                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 405                 return PNFS_NOT_ATTEMPTED;
 406         }
 407         dprintk("%s USE DS: %s\n", __func__, ds->ds_remotestr);
 408
 409         /* No multipath support. Use first DS */
 410         data->ds_clp = ds->ds_clp;
 411         fh = nfs4_fl_select_ds_fh(lseg, j);
 412         if (fh)
 413                 data->args.fh = fh;
 414
 415         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 416         data->mds_offset = offset;
 417
 418         /* Perform an asynchronous read to ds */
 419         status = nfs_initiate_read(ds->ds_clp->cl_rpcclient, data,
 420                                    &filelayout_read_call_ops);
 421         BUG_ON(status != 0);
 422         return PNFS_ATTEMPTED;
 423 }
 424
 425 /* Perform async writes. */
 426 static enum pnfs_try_status
 427 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
 428 {
 429         struct nfs_pgio_header *hdr = data->header;
 430         struct pnfs_layout_segment *lseg = hdr->lseg;
 431         struct nfs4_pnfs_ds *ds;
 432         loff_t offset = data->args.offset;
 433         u32 j, idx;
 434         struct nfs_fh *fh;
 435         int status;
 436
 437         if (test_bit(NFS_DEVICEID_INVALID, &FILELAYOUT_DEVID_NODE(lseg)->flags))
 438                 return PNFS_NOT_ATTEMPTED;
 439
 440         /* Retrieve the correct rpc_client for the byte range */
 441         j = nfs4_fl_calc_j_index(lseg, offset);
 442         idx = nfs4_fl_calc_ds_index(lseg, j);
 443         ds = nfs4_fl_prepare_ds(lseg, idx);
 444         if (!ds) {
 445                 printk(KERN_ERR "NFS: %s: prepare_ds failed, use MDS\n",
 446                         __func__);
 447                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 448                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 449                 return PNFS_NOT_ATTEMPTED;
 450         }
 451         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s\n", __func__,
 452                 hdr->inode->i_ino, sync, (size_t) data->args.count, offset,
 453                 ds->ds_remotestr);
 454
 455         data->write_done_cb = filelayout_write_done_cb;
 456         data->ds_clp = ds->ds_clp;
 457         fh = nfs4_fl_select_ds_fh(lseg, j);
 458         if (fh)
 459                 data->args.fh = fh;
 460         /*
 461          * Get the file offset on the dserver. Set the write offset to
 462          * this offset and save the original offset.
 463          */
 464         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 465
 466         /* Perform an asynchronous write */
 467         status = nfs_initiate_write(ds->ds_clp->cl_rpcclient, data,
 468                                     &filelayout_write_call_ops, sync);
 469         BUG_ON(status != 0);
 470         return PNFS_ATTEMPTED;
 471 }
 472
 473 /*
 474  * filelayout_check_layout()
 475  *
 476  * Make sure layout segment parameters are sane WRT the device.
 477  * At this point no generic layer initialization of the lseg has occurred,
 478  * and nothing has been added to the layout_hdr cache.
 479  *
 480  */
 481 static int
 482 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 483                         struct nfs4_filelayout_segment *fl,
 484                         struct nfs4_layoutget_res *lgr,
 485                         struct nfs4_deviceid *id,
 486                         gfp_t gfp_flags)
 487 {
 488         struct nfs4_deviceid_node *d;
 489         struct nfs4_file_layout_dsaddr *dsaddr;
 490         int status = -EINVAL;
 491         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 492
 493         dprintk("--> %s\n", __func__);
 494
 495         /* FIXME: remove this check when layout segment support is added */
 496         if (lgr->range.offset != 0 ||
 497             lgr->range.length != NFS4_MAX_UINT64) {
 498                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 499                         __func__);
 500                 goto out;
 501         }
 502
 503         if (fl->pattern_offset > lgr->range.offset) {
 504                 dprintk("%s pattern_offset %lld too large\n",
 505                                 __func__, fl->pattern_offset);
 506                 goto out;
 507         }
 508
 509         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 510                 dprintk("%s Invalid stripe unit (%u)\n",
 511                         __func__, fl->stripe_unit);
 512                 goto out;
 513         }
 514
 515         /* find and reference the deviceid */
 516         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
 517                                    NFS_SERVER(lo->plh_inode)->nfs_client, id);
 518         if (d == NULL) {
 519                 dsaddr = get_device_info(lo->plh_inode, id, gfp_flags);
 520                 if (dsaddr == NULL)
 521                         goto out;
 522         } else
 523                 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 524         /* Found deviceid is being reaped */
 525         if (test_bit(NFS_DEVICEID_INVALID, &dsaddr->id_node.flags))
 526                         goto out_put;
 527
 528         fl->dsaddr = dsaddr;
 529
 530         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 531                 dprintk("%s Bad first_stripe_index %u\n",
 532                                 __func__, fl->first_stripe_index);
 533                 goto out_put;
 534         }
 535
 536         if ((fl->stripe_type == STRIPE_SPARSE &&
 537             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 538             (fl->stripe_type == STRIPE_DENSE &&
 539             fl->num_fh != dsaddr->stripe_count)) {
 540                 dprintk("%s num_fh %u not valid for given packing\n",
 541                         __func__, fl->num_fh);
 542                 goto out_put;
 543         }
 544
 545         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
 546                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
 547                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
 548                         nfss->wsize);
 549         }
 550
 551         status = 0;
 552 out:
 553         dprintk("--> %s returns %d\n", __func__, status);
 554         return status;
 555 out_put:
 556         nfs4_fl_put_deviceid(dsaddr);
 557         goto out;
 558 }
 559
 560 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 561 {
 562         int i;
 563
 564         for (i = 0; i < fl->num_fh; i++) {
 565                 if (!fl->fh_array[i])
 566                         break;
 567                 kfree(fl->fh_array[i]);
 568         }
 569         kfree(fl->fh_array);
 570         fl->fh_array = NULL;
 571 }
 572
 573 static void
 574 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 575 {
 576         filelayout_free_fh_array(fl);
 577         kfree(fl);
 578 }
 579
 580 static int
 581 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 582                          struct nfs4_filelayout_segment *fl,
 583                          struct nfs4_layoutget_res *lgr,
 584                          struct nfs4_deviceid *id,
 585                          gfp_t gfp_flags)
 586 {
 587         struct xdr_stream stream;
 588         struct xdr_buf buf;
 589         struct page *scratch;
 590         __be32 *p;
 591         uint32_t nfl_util;
 592         int i;
 593
 594         dprintk("%s: set_layout_map Begin\n", __func__);
 595
 596         scratch = alloc_page(gfp_flags);
 597         if (!scratch)
 598                 return -ENOMEM;
 599
 600         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 601         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 602
 603         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 604          * num_fh (4) */
 605         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 606         if (unlikely(!p))
 607                 goto out_err;
 608
 609         memcpy(id, p, sizeof(*id));
 610         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 611         nfs4_print_deviceid(id);
 612
 613         nfl_util = be32_to_cpup(p++);
 614         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 615                 fl->commit_through_mds = 1;
 616         if (nfl_util & NFL4_UFLG_DENSE)
 617                 fl->stripe_type = STRIPE_DENSE;
 618         else
 619                 fl->stripe_type = STRIPE_SPARSE;
 620         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 621
 622         fl->first_stripe_index = be32_to_cpup(p++);
 623         p = xdr_decode_hyper(p, &fl->pattern_offset);
 624         fl->num_fh = be32_to_cpup(p++);
 625
 626         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 627                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 628                 fl->pattern_offset);
 629
 630         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 631          * Futher checking is done in filelayout_check_layout */
 632         if (fl->num_fh >
 633             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 634                 goto out_err;
 635
 636         if (fl->num_fh > 0) {
 637                 fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *),
 638                                        gfp_flags);
 639                 if (!fl->fh_array)
 640                         goto out_err;
 641         }
 642
 643         for (i = 0; i < fl->num_fh; i++) {
 644                 /* Do we want to use a mempool here? */
 645                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 646                 if (!fl->fh_array[i])
 647                         goto out_err_free;
 648
 649                 p = xdr_inline_decode(&stream, 4);
 650                 if (unlikely(!p))
 651                         goto out_err_free;
 652                 fl->fh_array[i]->size = be32_to_cpup(p++);
 653                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 654                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
 655                                i, fl->fh_array[i]->size);
 656                         goto out_err_free;
 657                 }
 658
 659                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 660                 if (unlikely(!p))
 661                         goto out_err_free;
 662                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 663                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 664                         fl->fh_array[i]->size);
 665         }
 666
 667         __free_page(scratch);
 668         return 0;
 669
 670 out_err_free:
 671         filelayout_free_fh_array(fl);
 672 out_err:
 673         __free_page(scratch);
 674         return -EIO;
 675 }
 676
 677 static void
 678 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 679 {
 680         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 681
 682         dprintk("--> %s\n", __func__);
 683         nfs4_fl_put_deviceid(fl->dsaddr);
 684         /* This assumes a single RW lseg */
 685         if (lseg->pls_range.iomode == IOMODE_RW) {
 686                 struct nfs4_filelayout *flo;
 687
 688                 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
 689                 flo->commit_info.nbuckets = 0;
 690                 kfree(flo->commit_info.buckets);
 691                 flo->commit_info.buckets = NULL;
 692         }
 693         _filelayout_free_lseg(fl);
 694 }
 695
 696 static int
 697 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 698                              gfp_t gfp_flags)
 699 {
 700         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 701         struct nfs4_filelayout *flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
 702
 703         struct nfs4_fl_commit_bucket *buckets;
 704         int size;
 705
 706         if (fl->commit_through_mds)
 707                 return 0;
 708         if (flo->commit_info.nbuckets != 0) {
 709                 /* This assumes there is only one IOMODE_RW lseg.  What
 710                  * we really want to do is have a layout_hdr level
 711                  * dictionary of <multipath_list4, fh> keys, each
 712                  * associated with a struct list_head, populated by calls
 713                  * to filelayout_write_pagelist().
 714                  * */
 715                 return 0;
 716         }
 717
 718         size = (fl->stripe_type == STRIPE_SPARSE) ?
 719                 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 720
 721         buckets = kcalloc(size, sizeof(struct nfs4_fl_commit_bucket),
 722                           gfp_flags);
 723         if (!buckets)
 724                 return -ENOMEM;
 725         else {
 726                 int i;
 727
 728                 spin_lock(&lseg->pls_layout->plh_inode->i_lock);
 729                 if (flo->commit_info.nbuckets != 0)
 730                         kfree(buckets);
 731                 else {
 732                         flo->commit_info.buckets = buckets;
 733                         flo->commit_info.nbuckets = size;
 734                         for (i = 0; i < size; i++) {
 735                                 INIT_LIST_HEAD(&buckets[i].written);
 736                                 INIT_LIST_HEAD(&buckets[i].committing);
 737                         }
 738                 }
 739                 spin_unlock(&lseg->pls_layout->plh_inode->i_lock);
 740                 return 0;
 741         }
 742 }
 743
 744 static struct pnfs_layout_segment *
 745 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 746                       struct nfs4_layoutget_res *lgr,
 747                       gfp_t gfp_flags)
 748 {
 749         struct nfs4_filelayout_segment *fl;
 750         int rc;
 751         struct nfs4_deviceid id;
 752
 753         dprintk("--> %s\n", __func__);
 754         fl = kzalloc(sizeof(*fl), gfp_flags);
 755         if (!fl)
 756                 return NULL;
 757
 758         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 759         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 760                 _filelayout_free_lseg(fl);
 761                 return NULL;
 762         }
 763         return &fl->generic_hdr;
 764 }
 765
 766 /*
 767  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 768  *
 769  * return true  : coalesce page
 770  * return false : don't coalesce page
 771  */
 772 static bool
 773 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 774                    struct nfs_page *req)
 775 {
 776         u64 p_stripe, r_stripe;
 777         u32 stripe_unit;
 778
 779         if (!pnfs_generic_pg_test(pgio, prev, req) ||
 780             !nfs_generic_pg_test(pgio, prev, req))
 781                 return false;
 782
 783         p_stripe = (u64)req_offset(prev);
 784         r_stripe = (u64)req_offset(req);
 785         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 786
 787         do_div(p_stripe, stripe_unit);
 788         do_div(r_stripe, stripe_unit);
 789
 790         return (p_stripe == r_stripe);
 791 }
 792
 793 static void
 794 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 795                         struct nfs_page *req)
 796 {
 797         BUG_ON(pgio->pg_lseg != NULL);
 798
 799         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 800                                            req->wb_context,
 801                                            0,
 802                                            NFS4_MAX_UINT64,
 803                                            IOMODE_READ,
 804                                            GFP_KERNEL);
 805         /* If no lseg, fall back to read through mds */
 806         if (pgio->pg_lseg == NULL)
 807                 nfs_pageio_reset_read_mds(pgio);
 808 }
 809
 810 static void
 811 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 812                          struct nfs_page *req)
 813 {
 814         int status;
 815
 816         BUG_ON(pgio->pg_lseg != NULL);
 817
 818         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 819                                            req->wb_context,
 820                                            0,
 821                                            NFS4_MAX_UINT64,
 822                                            IOMODE_RW,
 823                                            GFP_NOFS);
 824         /* If no lseg, fall back to write through mds */
 825         if (pgio->pg_lseg == NULL)
 826                 goto out_mds;
 827         status = filelayout_alloc_commit_info(pgio->pg_lseg, GFP_NOFS);
 828         if (status < 0) {
 829                 put_lseg(pgio->pg_lseg);
 830                 pgio->pg_lseg = NULL;
 831                 goto out_mds;
 832         }
 833         return;
 834 out_mds:
 835         nfs_pageio_reset_write_mds(pgio);
 836 }
 837
 838 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 839         .pg_init = filelayout_pg_init_read,
 840         .pg_test = filelayout_pg_test,
 841         .pg_doio = pnfs_generic_pg_readpages,
 842 };
 843
 844 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 845         .pg_init = filelayout_pg_init_write,
 846         .pg_test = filelayout_pg_test,
 847         .pg_doio = pnfs_generic_pg_writepages,
 848 };
 849
 850 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 851 {
 852         if (fl->stripe_type == STRIPE_SPARSE)
 853                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
 854         else
 855                 return j;
 856 }
 857
 858 /* The generic layer is about to remove the req from the commit list.
 859  * If this will make the bucket empty, it will need to put the lseg reference.
 860  */
 861 static void
 862 filelayout_clear_request_commit(struct nfs_page *req)
 863 {
 864         struct pnfs_layout_segment *freeme = NULL;
 865         struct inode *inode = req->wb_context->dentry->d_inode;
 866
 867         spin_lock(&inode->i_lock);
 868         if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
 869                 goto out;
 870         if (list_is_singular(&req->wb_list)) {
 871                 struct nfs4_fl_commit_bucket *bucket;
 872
 873                 bucket = list_first_entry(&req->wb_list,
 874                                           struct nfs4_fl_commit_bucket,
 875                                           written);
 876                 freeme = bucket->wlseg;
 877                 bucket->wlseg = NULL;
 878         }
 879 out:
 880         nfs_request_remove_commit_list(req);
 881         spin_unlock(&inode->i_lock);
 882         put_lseg(freeme);
 883 }
 884
 885 static struct list_head *
 886 filelayout_choose_commit_list(struct nfs_page *req,
 887                               struct pnfs_layout_segment *lseg)
 888 {
 889         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 890         u32 i, j;
 891         struct list_head *list;
 892         struct nfs4_fl_commit_bucket *buckets;
 893
 894         if (fl->commit_through_mds)
 895                 return &NFS_I(req->wb_context->dentry->d_inode)->commit_list;
 896
 897         /* Note that we are calling nfs4_fl_calc_j_index on each page
 898          * that ends up being committed to a data server.  An attractive
 899          * alternative is to add a field to nfs_write_data and nfs_page
 900          * to store the value calculated in filelayout_write_pagelist
 901          * and just use that here.
 902          */
 903         j = nfs4_fl_calc_j_index(lseg, req_offset(req));
 904         i = select_bucket_index(fl, j);
 905         buckets = FILELAYOUT_FROM_HDR(lseg->pls_layout)->commit_info.buckets;
 906         list = &buckets[i].written;
 907         if (list_empty(list)) {
 908                 /* Non-empty buckets hold a reference on the lseg.  That ref
 909                  * is normally transferred to the COMMIT call and released
 910                  * there.  It could also be released if the last req is pulled
 911                  * off due to a rewrite, in which case it will be done in
 912                  * filelayout_clear_request_commit
 913                  */
 914                 buckets[i].wlseg = get_lseg(lseg);
 915         }
 916         set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
 917         return list;
 918 }
 919
 920 static void
 921 filelayout_mark_request_commit(struct nfs_page *req,
 922                 struct pnfs_layout_segment *lseg)
 923 {
 924         struct list_head *list;
 925
 926         list = filelayout_choose_commit_list(req, lseg);
 927         nfs_request_add_commit_list(req, list);
 928 }
 929
 930 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 931 {
 932         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 933
 934         if (flseg->stripe_type == STRIPE_SPARSE)
 935                 return i;
 936         else
 937                 return nfs4_fl_calc_ds_index(lseg, i);
 938 }
 939
 940 static struct nfs_fh *
 941 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 942 {
 943         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 944
 945         if (flseg->stripe_type == STRIPE_SPARSE) {
 946                 if (flseg->num_fh == 1)
 947                         i = 0;
 948                 else if (flseg->num_fh == 0)
 949                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
 950                         return NULL;
 951         }
 952         return flseg->fh_array[i];
 953 }
 954
 955 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
 956 {
 957         struct pnfs_layout_segment *lseg = data->lseg;
 958         struct nfs4_pnfs_ds *ds;
 959         u32 idx;
 960         struct nfs_fh *fh;
 961
 962         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
 963         ds = nfs4_fl_prepare_ds(lseg, idx);
 964         if (!ds) {
 965                 printk(KERN_ERR "NFS: %s: prepare_ds failed, use MDS\n",
 966                         __func__);
 967                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 968                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 969                 prepare_to_resend_writes(data);
 970                 filelayout_commit_release(data);
 971                 return -EAGAIN;
 972         }
 973         dprintk("%s ino %lu, how %d\n", __func__, data->inode->i_ino, how);
 974         data->commit_done_cb = filelayout_commit_done_cb;
 975         data->ds_clp = ds->ds_clp;
 976         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
 977         if (fh)
 978                 data->args.fh = fh;
 979         return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data,
 980                                    &filelayout_commit_call_ops, how);
 981 }
 982
 983 static int
 984 filelayout_scan_ds_commit_list(struct nfs4_fl_commit_bucket *bucket, int max,
 985                 spinlock_t *lock)
 986 {
 987         struct list_head *src = &bucket->written;
 988         struct list_head *dst = &bucket->committing;
 989         struct nfs_page *req, *tmp;
 990         int ret = 0;
 991
 992         list_for_each_entry_safe(req, tmp, src, wb_list) {
 993                 if (!nfs_lock_request(req))
 994                         continue;
 995                 if (cond_resched_lock(lock))
 996                         list_safe_reset_next(req, tmp, wb_list);
 997                 nfs_request_remove_commit_list(req);
 998                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
 999                 nfs_list_add_request(req, dst);
1000                 ret++;
1001                 if (ret == max)
1002                         break;
1003         }
1004         if (ret) {
1005                 bucket->clseg = bucket->wlseg;
1006                 if (list_empty(src))
1007                         bucket->wlseg = NULL;
1008                 else
1009                         get_lseg(bucket->clseg);
1010         }
1011         return ret;
1012 }
1013
1014 /* Move reqs from written to committing lists, returning count of number moved.
1015  * Note called with i_lock held.
1016  */
1017 static int filelayout_scan_commit_lists(struct inode *inode, int max,
1018                 spinlock_t *lock)
1019 {
1020         struct nfs4_fl_commit_info *fl_cinfo;
1021         int i, rv = 0, cnt;
1022
1023         fl_cinfo = &FILELAYOUT_FROM_HDR(NFS_I(inode)->layout)->commit_info;
1024         if (fl_cinfo->nbuckets == 0)
1025                 goto out_done;
1026         for (i = 0; i < fl_cinfo->nbuckets && max != 0; i++) {
1027                 cnt = filelayout_scan_ds_commit_list(&fl_cinfo->buckets[i],
1028                                 max, lock);
1029                 max -= cnt;
1030                 rv += cnt;
1031         }
1032 out_done:
1033         return rv;
1034 }
1035
1036 static unsigned int
1037 alloc_ds_commits(struct inode *inode, struct list_head *list)
1038 {
1039         struct nfs4_fl_commit_info *fl_cinfo;
1040         struct nfs4_fl_commit_bucket *bucket;
1041         struct nfs_commit_data *data;
1042         int i, j;
1043         unsigned int nreq = 0;
1044
1045         fl_cinfo = &FILELAYOUT_FROM_HDR(NFS_I(inode)->layout)->commit_info;
1046         bucket = fl_cinfo->buckets;
1047         for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1048                 if (list_empty(&bucket->committing))
1049                         continue;
1050                 data = nfs_commitdata_alloc();
1051                 if (!data)
1052                         break;
1053                 data->ds_commit_index = i;
1054                 data->lseg = bucket->clseg;
1055                 bucket->clseg = NULL;
1056                 list_add(&data->pages, list);
1057                 nreq++;
1058         }
1059
1060         /* Clean up on error */
1061         for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1062                 if (list_empty(&bucket->committing))
1063                         continue;
1064                 nfs_retry_commit(&bucket->committing, bucket->clseg);
1065                 put_lseg(bucket->clseg);
1066                 bucket->clseg = NULL;
1067         }
1068         /* Caller will clean up entries put on list */
1069         return nreq;
1070 }
1071
1072 /* This follows nfs_commit_list pretty closely */
1073 static int
1074 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1075                            int how)
1076 {
1077         struct nfs_commit_data *data, *tmp;
1078         LIST_HEAD(list);
1079         unsigned int nreq = 0;
1080
1081         if (!list_empty(mds_pages)) {
1082                 data = nfs_commitdata_alloc();
1083                 if (data != NULL) {
1084                         data->lseg = NULL;
1085                         list_add(&data->pages, &list);
1086                         nreq++;
1087                 } else
1088                         nfs_retry_commit(mds_pages, NULL);
1089         }
1090
1091         nreq += alloc_ds_commits(inode, &list);
1092
1093         if (nreq == 0) {
1094                 nfs_commit_clear_lock(NFS_I(inode));
1095                 goto out;
1096         }
1097
1098         atomic_add(nreq, &NFS_I(inode)->commits_outstanding);
1099
1100         list_for_each_entry_safe(data, tmp, &list, pages) {
1101                 list_del_init(&data->pages);
1102                 if (!data->lseg) {
1103                         nfs_init_commit(data, mds_pages, NULL);
1104                         nfs_initiate_commit(NFS_CLIENT(inode), data,
1105                                             data->mds_ops, how);
1106                 } else {
1107                         struct nfs4_fl_commit_info *fl_cinfo;
1108
1109                         fl_cinfo = &FILELAYOUT_FROM_HDR(data->lseg->pls_layout)->commit_info;
1110                         nfs_init_commit(data, &fl_cinfo->buckets[data->ds_commit_index].committing, data->lseg);
1111                         filelayout_initiate_commit(data, how);
1112                 }
1113         }
1114 out:
1115         return PNFS_ATTEMPTED;
1116 }
1117
1118 static void
1119 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1120 {
1121         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1122 }
1123
1124 static struct pnfs_layout_hdr *
1125 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1126 {
1127         struct nfs4_filelayout *flo;
1128
1129         flo = kzalloc(sizeof(*flo), gfp_flags);
1130         return &flo->generic_hdr;
1131 }
1132
1133 static void
1134 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1135 {
1136         kfree(FILELAYOUT_FROM_HDR(lo));
1137 }
1138
1139 static struct pnfs_layoutdriver_type filelayout_type = {
1140         .id                     = LAYOUT_NFSV4_1_FILES,
1141         .name                   = "LAYOUT_NFSV4_1_FILES",
1142         .owner                  = THIS_MODULE,
1143         .alloc_layout_hdr       = filelayout_alloc_layout_hdr,
1144         .free_layout_hdr        = filelayout_free_layout_hdr,
1145         .alloc_lseg             = filelayout_alloc_lseg,
1146         .free_lseg              = filelayout_free_lseg,
1147         .pg_read_ops            = &filelayout_pg_read_ops,
1148         .pg_write_ops           = &filelayout_pg_write_ops,
1149         .mark_request_commit    = filelayout_mark_request_commit,
1150         .clear_request_commit   = filelayout_clear_request_commit,
1151         .scan_commit_lists      = filelayout_scan_commit_lists,
1152         .commit_pagelist        = filelayout_commit_pagelist,
1153         .read_pagelist          = filelayout_read_pagelist,
1154         .write_pagelist         = filelayout_write_pagelist,
1155         .free_deviceid_node     = filelayout_free_deveiceid_node,
1156 };
1157
1158 static int __init nfs4filelayout_init(void)
1159 {
1160         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1161                __func__);
1162         return pnfs_register_layoutdriver(&filelayout_type);
1163 }
1164
1165 static void __exit nfs4filelayout_exit(void)
1166 {
1167         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1168                __func__);
1169         pnfs_unregister_layoutdriver(&filelayout_type);
1170 }
1171
1172 MODULE_ALIAS("nfs-layouttype4-1");
1173
1174 module_init(nfs4filelayout_init);
1175 module_exit(nfs4filelayout_exit);