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