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
  35 #include "internal.h"
  36 #include "nfs4filelayout.h"
  37
  38 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  39
  40 MODULE_LICENSE("GPL");
  41 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
  42 MODULE_DESCRIPTION("The NFSv4 file layout driver");
  43
  44 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
  45
  46 static loff_t
  47 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
  48                             loff_t offset)
  49 {
  50         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
  51         u64 tmp;
  52
  53         offset -= flseg->pattern_offset;
  54         tmp = offset;
  55         do_div(tmp, stripe_width);
  56
  57         return tmp * flseg->stripe_unit + do_div(offset, flseg->stripe_unit);
  58 }
  59
  60 /* This function is used by the layout driver to calculate the
  61  * offset of the file on the dserver based on whether the
  62  * layout type is STRIPE_DENSE or STRIPE_SPARSE
  63  */
  64 static loff_t
  65 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
  66 {
  67         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
  68
  69         switch (flseg->stripe_type) {
  70         case STRIPE_SPARSE:
  71                 return offset;
  72
  73         case STRIPE_DENSE:
  74                 return filelayout_get_dense_offset(flseg, offset);
  75         }
  76
  77         BUG();
  78 }
  79
  80 static int filelayout_async_handle_error(struct rpc_task *task,
  81                                          struct nfs4_state *state,
  82                                          struct nfs_client *clp,
  83                                          int *reset)
  84 {
  85         if (task->tk_status >= 0)
  86                 return 0;
  87
  88         *reset = 0;
  89
  90         switch (task->tk_status) {
  91         case -NFS4ERR_BADSESSION:
  92         case -NFS4ERR_BADSLOT:
  93         case -NFS4ERR_BAD_HIGH_SLOT:
  94         case -NFS4ERR_DEADSESSION:
  95         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
  96         case -NFS4ERR_SEQ_FALSE_RETRY:
  97         case -NFS4ERR_SEQ_MISORDERED:
  98                 dprintk("%s ERROR %d, Reset session. Exchangeid "
  99                         "flags 0x%x\n", __func__, task->tk_status,
 100                         clp->cl_exchange_flags);
 101                 nfs4_schedule_session_recovery(clp->cl_session);
 102                 break;
 103         case -NFS4ERR_DELAY:
 104         case -NFS4ERR_GRACE:
 105         case -EKEYEXPIRED:
 106                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 107                 break;
 108         case -NFS4ERR_RETRY_UNCACHED_REP:
 109                 break;
 110         default:
 111                 dprintk("%s DS error. Retry through MDS %d\n", __func__,
 112                         task->tk_status);
 113                 *reset = 1;
 114                 break;
 115         }
 116         task->tk_status = 0;
 117         return -EAGAIN;
 118 }
 119
 120 /* NFS_PROTO call done callback routines */
 121
 122 static int filelayout_read_done_cb(struct rpc_task *task,
 123                                 struct nfs_read_data *data)
 124 {
 125         int reset = 0;
 126
 127         dprintk("%s DS read\n", __func__);
 128
 129         if (filelayout_async_handle_error(task, data->args.context->state,
 130                                           data->ds_clp, &reset) == -EAGAIN) {
 131                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 132                         __func__, data->ds_clp, data->ds_clp->cl_session);
 133                 if (reset) {
 134                         pnfs_set_lo_fail(data->lseg);
 135                         nfs4_reset_read(task, data);
 136                 }
 137                 rpc_restart_call_prepare(task);
 138                 return -EAGAIN;
 139         }
 140
 141         return 0;
 142 }
 143
 144 /*
 145  * We reference the rpc_cred of the first WRITE that triggers the need for
 146  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 147  * rfc5661 is not clear about which credential should be used.
 148  */
 149 static void
 150 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
 151 {
 152         if (FILELAYOUT_LSEG(wdata->lseg)->commit_through_mds ||
 153             wdata->res.verf->committed == NFS_FILE_SYNC)
 154                 return;
 155
 156         pnfs_set_layoutcommit(wdata);
 157         dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, wdata->inode->i_ino,
 158                 (unsigned long) NFS_I(wdata->inode)->layout->plh_lwb);
 159 }
 160
 161 /*
 162  * Call ops for the async read/write cases
 163  * In the case of dense layouts, the offset needs to be reset to its
 164  * original value.
 165  */
 166 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 167 {
 168         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 169
 170         rdata->read_done_cb = filelayout_read_done_cb;
 171
 172         if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
 173                                 &rdata->args.seq_args, &rdata->res.seq_res,
 174                                 0, task))
 175                 return;
 176
 177         rpc_call_start(task);
 178 }
 179
 180 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 181 {
 182         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 183
 184         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 185
 186         /* Note this may cause RPC to be resent */
 187         rdata->mds_ops->rpc_call_done(task, data);
 188 }
 189
 190 static void filelayout_read_release(void *data)
 191 {
 192         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 193
 194         rdata->mds_ops->rpc_release(data);
 195 }
 196
 197 static int filelayout_write_done_cb(struct rpc_task *task,
 198                                 struct nfs_write_data *data)
 199 {
 200         int reset = 0;
 201
 202         if (filelayout_async_handle_error(task, data->args.context->state,
 203                                           data->ds_clp, &reset) == -EAGAIN) {
 204                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 205                         __func__, data->ds_clp, data->ds_clp->cl_session);
 206                 if (reset) {
 207                         pnfs_set_lo_fail(data->lseg);
 208                         nfs4_reset_write(task, data);
 209                 }
 210                 rpc_restart_call_prepare(task);
 211                 return -EAGAIN;
 212         }
 213
 214         filelayout_set_layoutcommit(data);
 215         return 0;
 216 }
 217
 218 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 219 static void prepare_to_resend_writes(struct nfs_write_data *data)
 220 {
 221         struct nfs_page *first = nfs_list_entry(data->pages.next);
 222
 223         data->task.tk_status = 0;
 224         memcpy(data->verf.verifier, first->wb_verf.verifier,
 225                sizeof(first->wb_verf.verifier));
 226         data->verf.verifier[0]++; /* ensure verifier mismatch */
 227 }
 228
 229 static int filelayout_commit_done_cb(struct rpc_task *task,
 230                                      struct nfs_write_data *data)
 231 {
 232         int reset = 0;
 233
 234         if (filelayout_async_handle_error(task, data->args.context->state,
 235                                           data->ds_clp, &reset) == -EAGAIN) {
 236                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 237                         __func__, data->ds_clp, data->ds_clp->cl_session);
 238                 if (reset) {
 239                         prepare_to_resend_writes(data);
 240                         pnfs_set_lo_fail(data->lseg);
 241                 } else
 242                         rpc_restart_call_prepare(task);
 243                 return -EAGAIN;
 244         }
 245
 246         return 0;
 247 }
 248
 249 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 250 {
 251         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 252
 253         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 254                                 &wdata->args.seq_args, &wdata->res.seq_res,
 255                                 0, task))
 256                 return;
 257
 258         rpc_call_start(task);
 259 }
 260
 261 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 262 {
 263         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 264
 265         /* Note this may cause RPC to be resent */
 266         wdata->mds_ops->rpc_call_done(task, data);
 267 }
 268
 269 static void filelayout_write_release(void *data)
 270 {
 271         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 272
 273         wdata->mds_ops->rpc_release(data);
 274 }
 275
 276 static void filelayout_commit_release(void *data)
 277 {
 278         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 279
 280         nfs_commit_release_pages(wdata);
 281         if (atomic_dec_and_test(&NFS_I(wdata->inode)->commits_outstanding))
 282                 nfs_commit_clear_lock(NFS_I(wdata->inode));
 283         nfs_commitdata_release(wdata);
 284 }
 285
 286 struct rpc_call_ops filelayout_read_call_ops = {
 287         .rpc_call_prepare = filelayout_read_prepare,
 288         .rpc_call_done = filelayout_read_call_done,
 289         .rpc_release = filelayout_read_release,
 290 };
 291
 292 struct rpc_call_ops filelayout_write_call_ops = {
 293         .rpc_call_prepare = filelayout_write_prepare,
 294         .rpc_call_done = filelayout_write_call_done,
 295         .rpc_release = filelayout_write_release,
 296 };
 297
 298 struct rpc_call_ops filelayout_commit_call_ops = {
 299         .rpc_call_prepare = filelayout_write_prepare,
 300         .rpc_call_done = filelayout_write_call_done,
 301         .rpc_release = filelayout_commit_release,
 302 };
 303
 304 static enum pnfs_try_status
 305 filelayout_read_pagelist(struct nfs_read_data *data)
 306 {
 307         struct pnfs_layout_segment *lseg = data->lseg;
 308         struct nfs4_pnfs_ds *ds;
 309         loff_t offset = data->args.offset;
 310         u32 j, idx;
 311         struct nfs_fh *fh;
 312         int status;
 313
 314         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 315                 __func__, data->inode->i_ino,
 316                 data->args.pgbase, (size_t)data->args.count, offset);
 317
 318         if (test_bit(NFS_DEVICEID_INVALID, &FILELAYOUT_DEVID_NODE(lseg)->flags))
 319                 return PNFS_NOT_ATTEMPTED;
 320
 321         /* Retrieve the correct rpc_client for the byte range */
 322         j = nfs4_fl_calc_j_index(lseg, offset);
 323         idx = nfs4_fl_calc_ds_index(lseg, j);
 324         ds = nfs4_fl_prepare_ds(lseg, idx);
 325         if (!ds) {
 326                 /* Either layout fh index faulty, or ds connect failed */
 327                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 328                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 329                 return PNFS_NOT_ATTEMPTED;
 330         }
 331         dprintk("%s USE DS: %s\n", __func__, ds->ds_remotestr);
 332
 333         /* No multipath support. Use first DS */
 334         data->ds_clp = ds->ds_clp;
 335         fh = nfs4_fl_select_ds_fh(lseg, j);
 336         if (fh)
 337                 data->args.fh = fh;
 338
 339         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 340         data->mds_offset = offset;
 341
 342         /* Perform an asynchronous read to ds */
 343         status = nfs_initiate_read(data, ds->ds_clp->cl_rpcclient,
 344                                    &filelayout_read_call_ops);
 345         BUG_ON(status != 0);
 346         return PNFS_ATTEMPTED;
 347 }
 348
 349 /* Perform async writes. */
 350 static enum pnfs_try_status
 351 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
 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         if (test_bit(NFS_DEVICEID_INVALID, &FILELAYOUT_DEVID_NODE(lseg)->flags))
 361                 return PNFS_NOT_ATTEMPTED;
 362
 363         /* Retrieve the correct rpc_client for the byte range */
 364         j = nfs4_fl_calc_j_index(lseg, offset);
 365         idx = nfs4_fl_calc_ds_index(lseg, j);
 366         ds = nfs4_fl_prepare_ds(lseg, idx);
 367         if (!ds) {
 368                 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
 369                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 370                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 371                 return PNFS_NOT_ATTEMPTED;
 372         }
 373         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s\n", __func__,
 374                 data->inode->i_ino, sync, (size_t) data->args.count, offset,
 375                 ds->ds_remotestr);
 376
 377         data->write_done_cb = filelayout_write_done_cb;
 378         data->ds_clp = ds->ds_clp;
 379         fh = nfs4_fl_select_ds_fh(lseg, j);
 380         if (fh)
 381                 data->args.fh = fh;
 382         /*
 383          * Get the file offset on the dserver. Set the write offset to
 384          * this offset and save the original offset.
 385          */
 386         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 387
 388         /* Perform an asynchronous write */
 389         status = nfs_initiate_write(data, ds->ds_clp->cl_rpcclient,
 390                                     &filelayout_write_call_ops, sync);
 391         BUG_ON(status != 0);
 392         return PNFS_ATTEMPTED;
 393 }
 394
 395 /*
 396  * filelayout_check_layout()
 397  *
 398  * Make sure layout segment parameters are sane WRT the device.
 399  * At this point no generic layer initialization of the lseg has occurred,
 400  * and nothing has been added to the layout_hdr cache.
 401  *
 402  */
 403 static int
 404 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 405                         struct nfs4_filelayout_segment *fl,
 406                         struct nfs4_layoutget_res *lgr,
 407                         struct nfs4_deviceid *id,
 408                         gfp_t gfp_flags)
 409 {
 410         struct nfs4_deviceid_node *d;
 411         struct nfs4_file_layout_dsaddr *dsaddr;
 412         int status = -EINVAL;
 413         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 414
 415         dprintk("--> %s\n", __func__);
 416
 417         /* FIXME: remove this check when layout segment support is added */
 418         if (lgr->range.offset != 0 ||
 419             lgr->range.length != NFS4_MAX_UINT64) {
 420                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 421                         __func__);
 422                 goto out;
 423         }
 424
 425         if (fl->pattern_offset > lgr->range.offset) {
 426                 dprintk("%s pattern_offset %lld too large\n",
 427                                 __func__, fl->pattern_offset);
 428                 goto out;
 429         }
 430
 431         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 432                 dprintk("%s Invalid stripe unit (%u)\n",
 433                         __func__, fl->stripe_unit);
 434                 goto out;
 435         }
 436
 437         /* find and reference the deviceid */
 438         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
 439                                    NFS_SERVER(lo->plh_inode)->nfs_client, id);
 440         if (d == NULL) {
 441                 dsaddr = get_device_info(lo->plh_inode, id, gfp_flags);
 442                 if (dsaddr == NULL)
 443                         goto out;
 444         } else
 445                 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 446         /* Found deviceid is being reaped */
 447         if (test_bit(NFS_DEVICEID_INVALID, &dsaddr->id_node.flags))
 448                         goto out_put;
 449
 450         fl->dsaddr = dsaddr;
 451
 452         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 453                 dprintk("%s Bad first_stripe_index %u\n",
 454                                 __func__, fl->first_stripe_index);
 455                 goto out_put;
 456         }
 457
 458         if ((fl->stripe_type == STRIPE_SPARSE &&
 459             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 460             (fl->stripe_type == STRIPE_DENSE &&
 461             fl->num_fh != dsaddr->stripe_count)) {
 462                 dprintk("%s num_fh %u not valid for given packing\n",
 463                         __func__, fl->num_fh);
 464                 goto out_put;
 465         }
 466
 467         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
 468                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
 469                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
 470                         nfss->wsize);
 471         }
 472
 473         status = 0;
 474 out:
 475         dprintk("--> %s returns %d\n", __func__, status);
 476         return status;
 477 out_put:
 478         nfs4_fl_put_deviceid(dsaddr);
 479         goto out;
 480 }
 481
 482 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 483 {
 484         int i;
 485
 486         for (i = 0; i < fl->num_fh; i++) {
 487                 if (!fl->fh_array[i])
 488                         break;
 489                 kfree(fl->fh_array[i]);
 490         }
 491         kfree(fl->fh_array);
 492         fl->fh_array = NULL;
 493 }
 494
 495 static void
 496 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 497 {
 498         filelayout_free_fh_array(fl);
 499         kfree(fl);
 500 }
 501
 502 static int
 503 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 504                          struct nfs4_filelayout_segment *fl,
 505                          struct nfs4_layoutget_res *lgr,
 506                          struct nfs4_deviceid *id,
 507                          gfp_t gfp_flags)
 508 {
 509         struct xdr_stream stream;
 510         struct xdr_buf buf;
 511         struct page *scratch;
 512         __be32 *p;
 513         uint32_t nfl_util;
 514         int i;
 515
 516         dprintk("%s: set_layout_map Begin\n", __func__);
 517
 518         scratch = alloc_page(gfp_flags);
 519         if (!scratch)
 520                 return -ENOMEM;
 521
 522         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 523         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 524
 525         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 526          * num_fh (4) */
 527         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 528         if (unlikely(!p))
 529                 goto out_err;
 530
 531         memcpy(id, p, sizeof(*id));
 532         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 533         nfs4_print_deviceid(id);
 534
 535         nfl_util = be32_to_cpup(p++);
 536         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 537                 fl->commit_through_mds = 1;
 538         if (nfl_util & NFL4_UFLG_DENSE)
 539                 fl->stripe_type = STRIPE_DENSE;
 540         else
 541                 fl->stripe_type = STRIPE_SPARSE;
 542         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 543
 544         fl->first_stripe_index = be32_to_cpup(p++);
 545         p = xdr_decode_hyper(p, &fl->pattern_offset);
 546         fl->num_fh = be32_to_cpup(p++);
 547
 548         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 549                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 550                 fl->pattern_offset);
 551
 552         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 553          * Futher checking is done in filelayout_check_layout */
 554         if (fl->num_fh >
 555             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 556                 goto out_err;
 557
 558         if (fl->num_fh > 0) {
 559                 fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *),
 560                                        gfp_flags);
 561                 if (!fl->fh_array)
 562                         goto out_err;
 563         }
 564
 565         for (i = 0; i < fl->num_fh; i++) {
 566                 /* Do we want to use a mempool here? */
 567                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 568                 if (!fl->fh_array[i])
 569                         goto out_err_free;
 570
 571                 p = xdr_inline_decode(&stream, 4);
 572                 if (unlikely(!p))
 573                         goto out_err_free;
 574                 fl->fh_array[i]->size = be32_to_cpup(p++);
 575                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 576                         printk(KERN_ERR "Too big fh %d received %d\n",
 577                                i, fl->fh_array[i]->size);
 578                         goto out_err_free;
 579                 }
 580
 581                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 582                 if (unlikely(!p))
 583                         goto out_err_free;
 584                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 585                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 586                         fl->fh_array[i]->size);
 587         }
 588
 589         __free_page(scratch);
 590         return 0;
 591
 592 out_err_free:
 593         filelayout_free_fh_array(fl);
 594 out_err:
 595         __free_page(scratch);
 596         return -EIO;
 597 }
 598
 599 static void
 600 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 601 {
 602         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 603
 604         dprintk("--> %s\n", __func__);
 605         nfs4_fl_put_deviceid(fl->dsaddr);
 606         kfree(fl->commit_buckets);
 607         _filelayout_free_lseg(fl);
 608 }
 609
 610 static struct pnfs_layout_segment *
 611 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 612                       struct nfs4_layoutget_res *lgr,
 613                       gfp_t gfp_flags)
 614 {
 615         struct nfs4_filelayout_segment *fl;
 616         int rc;
 617         struct nfs4_deviceid id;
 618
 619         dprintk("--> %s\n", __func__);
 620         fl = kzalloc(sizeof(*fl), gfp_flags);
 621         if (!fl)
 622                 return NULL;
 623
 624         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 625         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 626                 _filelayout_free_lseg(fl);
 627                 return NULL;
 628         }
 629
 630         /* This assumes there is only one IOMODE_RW lseg.  What
 631          * we really want to do is have a layout_hdr level
 632          * dictionary of <multipath_list4, fh> keys, each
 633          * associated with a struct list_head, populated by calls
 634          * to filelayout_write_pagelist().
 635          * */
 636         if ((!fl->commit_through_mds) && (lgr->range.iomode == IOMODE_RW)) {
 637                 int i;
 638                 int size = (fl->stripe_type == STRIPE_SPARSE) ?
 639                         fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 640
 641                 fl->commit_buckets = kcalloc(size, sizeof(struct list_head), gfp_flags);
 642                 if (!fl->commit_buckets) {
 643                         filelayout_free_lseg(&fl->generic_hdr);
 644                         return NULL;
 645                 }
 646                 fl->number_of_buckets = size;
 647                 for (i = 0; i < size; i++)
 648                         INIT_LIST_HEAD(&fl->commit_buckets[i]);
 649         }
 650         return &fl->generic_hdr;
 651 }
 652
 653 /*
 654  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 655  *
 656  * return true  : coalesce page
 657  * return false : don't coalesce page
 658  */
 659 static bool
 660 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 661                    struct nfs_page *req)
 662 {
 663         u64 p_stripe, r_stripe;
 664         u32 stripe_unit;
 665
 666         if (!pnfs_generic_pg_test(pgio, prev, req) ||
 667             !nfs_generic_pg_test(pgio, prev, req))
 668                 return false;
 669
 670         p_stripe = (u64)prev->wb_index << PAGE_CACHE_SHIFT;
 671         r_stripe = (u64)req->wb_index << PAGE_CACHE_SHIFT;
 672         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 673
 674         do_div(p_stripe, stripe_unit);
 675         do_div(r_stripe, stripe_unit);
 676
 677         return (p_stripe == r_stripe);
 678 }
 679
 680 void
 681 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 682                         struct nfs_page *req)
 683 {
 684         BUG_ON(pgio->pg_lseg != NULL);
 685
 686         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 687                                            req->wb_context,
 688                                            0,
 689                                            NFS4_MAX_UINT64,
 690                                            IOMODE_READ,
 691                                            GFP_KERNEL);
 692         /* If no lseg, fall back to read through mds */
 693         if (pgio->pg_lseg == NULL)
 694                 nfs_pageio_reset_read_mds(pgio);
 695 }
 696
 697 void
 698 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 699                          struct nfs_page *req)
 700 {
 701         BUG_ON(pgio->pg_lseg != NULL);
 702
 703         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 704                                            req->wb_context,
 705                                            0,
 706                                            NFS4_MAX_UINT64,
 707                                            IOMODE_RW,
 708                                            GFP_NOFS);
 709         /* If no lseg, fall back to write through mds */
 710         if (pgio->pg_lseg == NULL)
 711                 nfs_pageio_reset_write_mds(pgio);
 712 }
 713
 714 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 715         .pg_init = filelayout_pg_init_read,
 716         .pg_test = filelayout_pg_test,
 717         .pg_doio = pnfs_generic_pg_readpages,
 718 };
 719
 720 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 721         .pg_init = filelayout_pg_init_write,
 722         .pg_test = filelayout_pg_test,
 723         .pg_doio = pnfs_generic_pg_writepages,
 724 };
 725
 726 static bool filelayout_mark_pnfs_commit(struct pnfs_layout_segment *lseg)
 727 {
 728         return !FILELAYOUT_LSEG(lseg)->commit_through_mds;
 729 }
 730
 731 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 732 {
 733         if (fl->stripe_type == STRIPE_SPARSE)
 734                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
 735         else
 736                 return j;
 737 }
 738
 739 struct list_head *filelayout_choose_commit_list(struct nfs_page *req)
 740 {
 741         struct pnfs_layout_segment *lseg = req->wb_commit_lseg;
 742         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 743         u32 i, j;
 744         struct list_head *list;
 745
 746         /* Note that we are calling nfs4_fl_calc_j_index on each page
 747          * that ends up being committed to a data server.  An attractive
 748          * alternative is to add a field to nfs_write_data and nfs_page
 749          * to store the value calculated in filelayout_write_pagelist
 750          * and just use that here.
 751          */
 752         j = nfs4_fl_calc_j_index(lseg,
 753                                  (loff_t)req->wb_index << PAGE_CACHE_SHIFT);
 754         i = select_bucket_index(fl, j);
 755         list = &fl->commit_buckets[i];
 756         if (list_empty(list)) {
 757                 /* Non-empty buckets hold a reference on the lseg */
 758                 get_lseg(lseg);
 759         }
 760         return list;
 761 }
 762
 763 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 764 {
 765         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 766
 767         if (flseg->stripe_type == STRIPE_SPARSE)
 768                 return i;
 769         else
 770                 return nfs4_fl_calc_ds_index(lseg, i);
 771 }
 772
 773 static struct nfs_fh *
 774 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 775 {
 776         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 777
 778         if (flseg->stripe_type == STRIPE_SPARSE) {
 779                 if (flseg->num_fh == 1)
 780                         i = 0;
 781                 else if (flseg->num_fh == 0)
 782                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
 783                         return NULL;
 784         }
 785         return flseg->fh_array[i];
 786 }
 787
 788 static int filelayout_initiate_commit(struct nfs_write_data *data, int how)
 789 {
 790         struct pnfs_layout_segment *lseg = data->lseg;
 791         struct nfs4_pnfs_ds *ds;
 792         u32 idx;
 793         struct nfs_fh *fh;
 794
 795         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
 796         ds = nfs4_fl_prepare_ds(lseg, idx);
 797         if (!ds) {
 798                 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
 799                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 800                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 801                 prepare_to_resend_writes(data);
 802                 data->mds_ops->rpc_release(data);
 803                 return -EAGAIN;
 804         }
 805         dprintk("%s ino %lu, how %d\n", __func__, data->inode->i_ino, how);
 806         data->write_done_cb = filelayout_commit_done_cb;
 807         data->ds_clp = ds->ds_clp;
 808         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
 809         if (fh)
 810                 data->args.fh = fh;
 811         return nfs_initiate_commit(data, ds->ds_clp->cl_rpcclient,
 812                                    &filelayout_commit_call_ops, how);
 813 }
 814
 815 /*
 816  * This is only useful while we are using whole file layouts.
 817  */
 818 static struct pnfs_layout_segment *find_only_write_lseg(struct inode *inode)
 819 {
 820         struct pnfs_layout_segment *lseg, *rv = NULL;
 821
 822         spin_lock(&inode->i_lock);
 823         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
 824                 if (lseg->pls_range.iomode == IOMODE_RW)
 825                         rv = get_lseg(lseg);
 826         spin_unlock(&inode->i_lock);
 827         return rv;
 828 }
 829
 830 static int alloc_ds_commits(struct inode *inode, struct list_head *list)
 831 {
 832         struct pnfs_layout_segment *lseg;
 833         struct nfs4_filelayout_segment *fl;
 834         struct nfs_write_data *data;
 835         int i, j;
 836
 837         /* Won't need this when non-whole file layout segments are supported
 838          * instead we will use a pnfs_layout_hdr structure */
 839         lseg = find_only_write_lseg(inode);
 840         if (!lseg)
 841                 return 0;
 842         fl = FILELAYOUT_LSEG(lseg);
 843         for (i = 0; i < fl->number_of_buckets; i++) {
 844                 if (list_empty(&fl->commit_buckets[i]))
 845                         continue;
 846                 data = nfs_commitdata_alloc();
 847                 if (!data)
 848                         goto out_bad;
 849                 data->ds_commit_index = i;
 850                 data->lseg = lseg;
 851                 list_add(&data->pages, list);
 852         }
 853         put_lseg(lseg);
 854         return 0;
 855
 856 out_bad:
 857         for (j = i; j < fl->number_of_buckets; j++) {
 858                 if (list_empty(&fl->commit_buckets[i]))
 859                         continue;
 860                 nfs_retry_commit(&fl->commit_buckets[i], lseg);
 861                 put_lseg(lseg);  /* associated with emptying bucket */
 862         }
 863         put_lseg(lseg);
 864         /* Caller will clean up entries put on list */
 865         return -ENOMEM;
 866 }
 867
 868 /* This follows nfs_commit_list pretty closely */
 869 static int
 870 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
 871                            int how)
 872 {
 873         struct nfs_write_data   *data, *tmp;
 874         LIST_HEAD(list);
 875
 876         if (!list_empty(mds_pages)) {
 877                 data = nfs_commitdata_alloc();
 878                 if (!data)
 879                         goto out_bad;
 880                 data->lseg = NULL;
 881                 list_add(&data->pages, &list);
 882         }
 883
 884         if (alloc_ds_commits(inode, &list))
 885                 goto out_bad;
 886
 887         list_for_each_entry_safe(data, tmp, &list, pages) {
 888                 list_del_init(&data->pages);
 889                 atomic_inc(&NFS_I(inode)->commits_outstanding);
 890                 if (!data->lseg) {
 891                         nfs_init_commit(data, mds_pages, NULL);
 892                         nfs_initiate_commit(data, NFS_CLIENT(inode),
 893                                             data->mds_ops, how);
 894                 } else {
 895                         nfs_init_commit(data, &FILELAYOUT_LSEG(data->lseg)->commit_buckets[data->ds_commit_index], data->lseg);
 896                         filelayout_initiate_commit(data, how);
 897                 }
 898         }
 899         return 0;
 900  out_bad:
 901         list_for_each_entry_safe(data, tmp, &list, pages) {
 902                 nfs_retry_commit(&data->pages, data->lseg);
 903                 list_del_init(&data->pages);
 904                 nfs_commit_free(data);
 905         }
 906         nfs_retry_commit(mds_pages, NULL);
 907         nfs_commit_clear_lock(NFS_I(inode));
 908         return -ENOMEM;
 909 }
 910
 911 static void
 912 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
 913 {
 914         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
 915 }
 916
 917 static struct pnfs_layoutdriver_type filelayout_type = {
 918         .id                     = LAYOUT_NFSV4_1_FILES,
 919         .name                   = "LAYOUT_NFSV4_1_FILES",
 920         .owner                  = THIS_MODULE,
 921         .alloc_lseg             = filelayout_alloc_lseg,
 922         .free_lseg              = filelayout_free_lseg,
 923         .pg_read_ops            = &filelayout_pg_read_ops,
 924         .pg_write_ops           = &filelayout_pg_write_ops,
 925         .mark_pnfs_commit       = filelayout_mark_pnfs_commit,
 926         .choose_commit_list     = filelayout_choose_commit_list,
 927         .commit_pagelist        = filelayout_commit_pagelist,
 928         .read_pagelist          = filelayout_read_pagelist,
 929         .write_pagelist         = filelayout_write_pagelist,
 930         .free_deviceid_node     = filelayout_free_deveiceid_node,
 931 };
 932
 933 static int __init nfs4filelayout_init(void)
 934 {
 935         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
 936                __func__);
 937         return pnfs_register_layoutdriver(&filelayout_type);
 938 }
 939
 940 static void __exit nfs4filelayout_exit(void)
 941 {
 942         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
 943                __func__);
 944         pnfs_unregister_layoutdriver(&filelayout_type);
 945 }
 946
 947 MODULE_ALIAS("nfs-layouttype4-1");
 948
 949 module_init(nfs4filelayout_init);
 950 module_exit(nfs4filelayout_exit);