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