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 "internal.h"
  37 #include "nfs4filelayout.h"
  38
  39 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  40
  41 MODULE_LICENSE("GPL");
  42 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
  43 MODULE_DESCRIPTION("The NFSv4 file layout driver");
  44
  45 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
  46
  47 static loff_t
  48 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
  49                             loff_t offset)
  50 {
  51         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
  52         u64 tmp;
  53
  54         offset -= flseg->pattern_offset;
  55         tmp = offset;
  56         do_div(tmp, stripe_width);
  57
  58         return tmp * flseg->stripe_unit + do_div(offset, flseg->stripe_unit);
  59 }
  60
  61 /* This function is used by the layout driver to calculate the
  62  * offset of the file on the dserver based on whether the
  63  * layout type is STRIPE_DENSE or STRIPE_SPARSE
  64  */
  65 static loff_t
  66 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
  67 {
  68         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
  69
  70         switch (flseg->stripe_type) {
  71         case STRIPE_SPARSE:
  72                 return offset;
  73
  74         case STRIPE_DENSE:
  75                 return filelayout_get_dense_offset(flseg, offset);
  76         }
  77
  78         BUG();
  79 }
  80
  81 static int filelayout_async_handle_error(struct rpc_task *task,
  82                                          struct nfs4_state *state,
  83                                          struct nfs_client *clp,
  84                                          int *reset)
  85 {
  86         if (task->tk_status >= 0)
  87                 return 0;
  88
  89         *reset = 0;
  90
  91         switch (task->tk_status) {
  92         case -NFS4ERR_BADSESSION:
  93         case -NFS4ERR_BADSLOT:
  94         case -NFS4ERR_BAD_HIGH_SLOT:
  95         case -NFS4ERR_DEADSESSION:
  96         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
  97         case -NFS4ERR_SEQ_FALSE_RETRY:
  98         case -NFS4ERR_SEQ_MISORDERED:
  99                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 100                         "flags 0x%x\n", __func__, task->tk_status,
 101                         clp->cl_exchange_flags);
 102                 nfs4_schedule_session_recovery(clp->cl_session);
 103                 break;
 104         case -NFS4ERR_DELAY:
 105         case -NFS4ERR_GRACE:
 106         case -EKEYEXPIRED:
 107                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 108                 break;
 109         case -NFS4ERR_RETRY_UNCACHED_REP:
 110                 break;
 111         default:
 112                 dprintk("%s DS error. Retry through MDS %d\n", __func__,
 113                         task->tk_status);
 114                 *reset = 1;
 115                 break;
 116         }
 117         task->tk_status = 0;
 118         return -EAGAIN;
 119 }
 120
 121 /* NFS_PROTO call done callback routines */
 122
 123 static int filelayout_read_done_cb(struct rpc_task *task,
 124                                 struct nfs_read_data *data)
 125 {
 126         int reset = 0;
 127
 128         dprintk("%s DS read\n", __func__);
 129
 130         if (filelayout_async_handle_error(task, data->args.context->state,
 131                                           data->ds_clp, &reset) == -EAGAIN) {
 132                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 133                         __func__, data->ds_clp, data->ds_clp->cl_session);
 134                 if (reset) {
 135                         pnfs_set_lo_fail(data->lseg);
 136                         nfs4_reset_read(task, data);
 137                 }
 138                 rpc_restart_call_prepare(task);
 139                 return -EAGAIN;
 140         }
 141
 142         return 0;
 143 }
 144
 145 /*
 146  * We reference the rpc_cred of the first WRITE that triggers the need for
 147  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 148  * rfc5661 is not clear about which credential should be used.
 149  */
 150 static void
 151 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
 152 {
 153         if (FILELAYOUT_LSEG(wdata->lseg)->commit_through_mds ||
 154             wdata->res.verf->committed == NFS_FILE_SYNC)
 155                 return;
 156
 157         pnfs_set_layoutcommit(wdata);
 158         dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, wdata->inode->i_ino,
 159                 (unsigned long) NFS_I(wdata->inode)->layout->plh_lwb);
 160 }
 161
 162 /*
 163  * Call ops for the async read/write cases
 164  * In the case of dense layouts, the offset needs to be reset to its
 165  * original value.
 166  */
 167 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 168 {
 169         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 170
 171         rdata->read_done_cb = filelayout_read_done_cb;
 172
 173         if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
 174                                 &rdata->args.seq_args, &rdata->res.seq_res,
 175                                 0, task))
 176                 return;
 177
 178         rpc_call_start(task);
 179 }
 180
 181 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 182 {
 183         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 184
 185         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 186
 187         /* Note this may cause RPC to be resent */
 188         rdata->mds_ops->rpc_call_done(task, data);
 189 }
 190
 191 static void filelayout_read_release(void *data)
 192 {
 193         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 194
 195         rdata->mds_ops->rpc_release(data);
 196 }
 197
 198 static int filelayout_write_done_cb(struct rpc_task *task,
 199                                 struct nfs_write_data *data)
 200 {
 201         int reset = 0;
 202
 203         if (filelayout_async_handle_error(task, data->args.context->state,
 204                                           data->ds_clp, &reset) == -EAGAIN) {
 205                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 206                         __func__, data->ds_clp, data->ds_clp->cl_session);
 207                 if (reset) {
 208                         pnfs_set_lo_fail(data->lseg);
 209                         nfs4_reset_write(task, data);
 210                 }
 211                 rpc_restart_call_prepare(task);
 212                 return -EAGAIN;
 213         }
 214
 215         filelayout_set_layoutcommit(data);
 216         return 0;
 217 }
 218
 219 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 220 static void prepare_to_resend_writes(struct nfs_write_data *data)
 221 {
 222         struct nfs_page *first = nfs_list_entry(data->pages.next);
 223
 224         data->task.tk_status = 0;
 225         memcpy(data->verf.verifier, first->wb_verf.verifier,
 226                sizeof(first->wb_verf.verifier));
 227         data->verf.verifier[0]++; /* ensure verifier mismatch */
 228 }
 229
 230 static int filelayout_commit_done_cb(struct rpc_task *task,
 231                                      struct nfs_write_data *data)
 232 {
 233         int reset = 0;
 234
 235         if (filelayout_async_handle_error(task, data->args.context->state,
 236                                           data->ds_clp, &reset) == -EAGAIN) {
 237                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 238                         __func__, data->ds_clp, data->ds_clp->cl_session);
 239                 if (reset) {
 240                         prepare_to_resend_writes(data);
 241                         pnfs_set_lo_fail(data->lseg);
 242                 } else
 243                         rpc_restart_call_prepare(task);
 244                 return -EAGAIN;
 245         }
 246
 247         return 0;
 248 }
 249
 250 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 251 {
 252         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 253
 254         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 255                                 &wdata->args.seq_args, &wdata->res.seq_res,
 256                                 0, task))
 257                 return;
 258
 259         rpc_call_start(task);
 260 }
 261
 262 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 263 {
 264         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 265
 266         /* Note this may cause RPC to be resent */
 267         wdata->mds_ops->rpc_call_done(task, data);
 268 }
 269
 270 static void filelayout_write_release(void *data)
 271 {
 272         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 273
 274         wdata->mds_ops->rpc_release(data);
 275 }
 276
 277 static void filelayout_commit_release(void *data)
 278 {
 279         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 280
 281         nfs_commit_release_pages(wdata);
 282         if (atomic_dec_and_test(&NFS_I(wdata->inode)->commits_outstanding))
 283                 nfs_commit_clear_lock(NFS_I(wdata->inode));
 284         nfs_commitdata_release(wdata);
 285 }
 286
 287 struct rpc_call_ops filelayout_read_call_ops = {
 288         .rpc_call_prepare = filelayout_read_prepare,
 289         .rpc_call_done = filelayout_read_call_done,
 290         .rpc_release = filelayout_read_release,
 291 };
 292
 293 struct rpc_call_ops filelayout_write_call_ops = {
 294         .rpc_call_prepare = filelayout_write_prepare,
 295         .rpc_call_done = filelayout_write_call_done,
 296         .rpc_release = filelayout_write_release,
 297 };
 298
 299 struct rpc_call_ops filelayout_commit_call_ops = {
 300         .rpc_call_prepare = filelayout_write_prepare,
 301         .rpc_call_done = filelayout_write_call_done,
 302         .rpc_release = filelayout_commit_release,
 303 };
 304
 305 static enum pnfs_try_status
 306 filelayout_read_pagelist(struct nfs_read_data *data)
 307 {
 308         struct pnfs_layout_segment *lseg = data->lseg;
 309         struct nfs4_pnfs_ds *ds;
 310         loff_t offset = data->args.offset;
 311         u32 j, idx;
 312         struct nfs_fh *fh;
 313         int status;
 314
 315         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 316                 __func__, data->inode->i_ino,
 317                 data->args.pgbase, (size_t)data->args.count, offset);
 318
 319         if (test_bit(NFS_DEVICEID_INVALID, &FILELAYOUT_DEVID_NODE(lseg)->flags))
 320                 return PNFS_NOT_ATTEMPTED;
 321
 322         /* Retrieve the correct rpc_client for the byte range */
 323         j = nfs4_fl_calc_j_index(lseg, offset);
 324         idx = nfs4_fl_calc_ds_index(lseg, j);
 325         ds = nfs4_fl_prepare_ds(lseg, idx);
 326         if (!ds) {
 327                 /* Either layout fh index faulty, or ds connect failed */
 328                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 329                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 330                 return PNFS_NOT_ATTEMPTED;
 331         }
 332         dprintk("%s USE DS: %s\n", __func__, ds->ds_remotestr);
 333
 334         /* No multipath support. Use first DS */
 335         data->ds_clp = ds->ds_clp;
 336         fh = nfs4_fl_select_ds_fh(lseg, j);
 337         if (fh)
 338                 data->args.fh = fh;
 339
 340         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 341         data->mds_offset = offset;
 342
 343         /* Perform an asynchronous read to ds */
 344         status = nfs_initiate_read(data, ds->ds_clp->cl_rpcclient,
 345                                    &filelayout_read_call_ops);
 346         BUG_ON(status != 0);
 347         return PNFS_ATTEMPTED;
 348 }
 349
 350 /* Perform async writes. */
 351 static enum pnfs_try_status
 352 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
 353 {
 354         struct pnfs_layout_segment *lseg = data->lseg;
 355         struct nfs4_pnfs_ds *ds;
 356         loff_t offset = data->args.offset;
 357         u32 j, idx;
 358         struct nfs_fh *fh;
 359         int status;
 360
 361         if (test_bit(NFS_DEVICEID_INVALID, &FILELAYOUT_DEVID_NODE(lseg)->flags))
 362                 return PNFS_NOT_ATTEMPTED;
 363
 364         /* Retrieve the correct rpc_client for the byte range */
 365         j = nfs4_fl_calc_j_index(lseg, offset);
 366         idx = nfs4_fl_calc_ds_index(lseg, j);
 367         ds = nfs4_fl_prepare_ds(lseg, idx);
 368         if (!ds) {
 369                 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
 370                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 371                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 372                 return PNFS_NOT_ATTEMPTED;
 373         }
 374         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s\n", __func__,
 375                 data->inode->i_ino, sync, (size_t) data->args.count, offset,
 376                 ds->ds_remotestr);
 377
 378         data->write_done_cb = filelayout_write_done_cb;
 379         data->ds_clp = ds->ds_clp;
 380         fh = nfs4_fl_select_ds_fh(lseg, j);
 381         if (fh)
 382                 data->args.fh = fh;
 383         /*
 384          * Get the file offset on the dserver. Set the write offset to
 385          * this offset and save the original offset.
 386          */
 387         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 388
 389         /* Perform an asynchronous write */
 390         status = nfs_initiate_write(data, ds->ds_clp->cl_rpcclient,
 391                                     &filelayout_write_call_ops, sync);
 392         BUG_ON(status != 0);
 393         return PNFS_ATTEMPTED;
 394 }
 395
 396 /*
 397  * filelayout_check_layout()
 398  *
 399  * Make sure layout segment parameters are sane WRT the device.
 400  * At this point no generic layer initialization of the lseg has occurred,
 401  * and nothing has been added to the layout_hdr cache.
 402  *
 403  */
 404 static int
 405 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 406                         struct nfs4_filelayout_segment *fl,
 407                         struct nfs4_layoutget_res *lgr,
 408                         struct nfs4_deviceid *id,
 409                         gfp_t gfp_flags)
 410 {
 411         struct nfs4_deviceid_node *d;
 412         struct nfs4_file_layout_dsaddr *dsaddr;
 413         int status = -EINVAL;
 414         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 415
 416         dprintk("--> %s\n", __func__);
 417
 418         /* FIXME: remove this check when layout segment support is added */
 419         if (lgr->range.offset != 0 ||
 420             lgr->range.length != NFS4_MAX_UINT64) {
 421                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 422                         __func__);
 423                 goto out;
 424         }
 425
 426         if (fl->pattern_offset > lgr->range.offset) {
 427                 dprintk("%s pattern_offset %lld too large\n",
 428                                 __func__, fl->pattern_offset);
 429                 goto out;
 430         }
 431
 432         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 433                 dprintk("%s Invalid stripe unit (%u)\n",
 434                         __func__, fl->stripe_unit);
 435                 goto out;
 436         }
 437
 438         /* find and reference the deviceid */
 439         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
 440                                    NFS_SERVER(lo->plh_inode)->nfs_client, id);
 441         if (d == NULL) {
 442                 dsaddr = get_device_info(lo->plh_inode, id, gfp_flags);
 443                 if (dsaddr == NULL)
 444                         goto out;
 445         } else
 446                 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 447         /* Found deviceid is being reaped */
 448         if (test_bit(NFS_DEVICEID_INVALID, &dsaddr->id_node.flags))
 449                         goto out_put;
 450
 451         fl->dsaddr = dsaddr;
 452
 453         if (fl->first_stripe_index < 0 ||
 454             fl->first_stripe_index >= dsaddr->stripe_count) {
 455                 dprintk("%s Bad first_stripe_index %d\n",
 456                                 __func__, fl->first_stripe_index);
 457                 goto out_put;
 458         }
 459
 460         if ((fl->stripe_type == STRIPE_SPARSE &&
 461             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 462             (fl->stripe_type == STRIPE_DENSE &&
 463             fl->num_fh != dsaddr->stripe_count)) {
 464                 dprintk("%s num_fh %u not valid for given packing\n",
 465                         __func__, fl->num_fh);
 466                 goto out_put;
 467         }
 468
 469         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
 470                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
 471                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
 472                         nfss->wsize);
 473         }
 474
 475         status = 0;
 476 out:
 477         dprintk("--> %s returns %d\n", __func__, status);
 478         return status;
 479 out_put:
 480         nfs4_fl_put_deviceid(dsaddr);
 481         goto out;
 482 }
 483
 484 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 485 {
 486         int i;
 487
 488         for (i = 0; i < fl->num_fh; i++) {
 489                 if (!fl->fh_array[i])
 490                         break;
 491                 kfree(fl->fh_array[i]);
 492         }
 493         kfree(fl->fh_array);
 494         fl->fh_array = NULL;
 495 }
 496
 497 static void
 498 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 499 {
 500         filelayout_free_fh_array(fl);
 501         kfree(fl);
 502 }
 503
 504 static int
 505 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 506                          struct nfs4_filelayout_segment *fl,
 507                          struct nfs4_layoutget_res *lgr,
 508                          struct nfs4_deviceid *id,
 509                          gfp_t gfp_flags)
 510 {
 511         struct xdr_stream stream;
 512         struct xdr_buf buf;
 513         struct page *scratch;
 514         __be32 *p;
 515         uint32_t nfl_util;
 516         int i;
 517
 518         dprintk("%s: set_layout_map Begin\n", __func__);
 519
 520         scratch = alloc_page(gfp_flags);
 521         if (!scratch)
 522                 return -ENOMEM;
 523
 524         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 525         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 526
 527         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 528          * num_fh (4) */
 529         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 530         if (unlikely(!p))
 531                 goto out_err;
 532
 533         memcpy(id, p, sizeof(*id));
 534         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 535         nfs4_print_deviceid(id);
 536
 537         nfl_util = be32_to_cpup(p++);
 538         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 539                 fl->commit_through_mds = 1;
 540         if (nfl_util & NFL4_UFLG_DENSE)
 541                 fl->stripe_type = STRIPE_DENSE;
 542         else
 543                 fl->stripe_type = STRIPE_SPARSE;
 544         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 545
 546         fl->first_stripe_index = be32_to_cpup(p++);
 547         p = xdr_decode_hyper(p, &fl->pattern_offset);
 548         fl->num_fh = be32_to_cpup(p++);
 549
 550         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 551                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 552                 fl->pattern_offset);
 553
 554         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 555          * Futher checking is done in filelayout_check_layout */
 556         if (fl->num_fh < 0 || fl->num_fh >
 557             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 558                 goto out_err;
 559
 560         if (fl->num_fh > 0) {
 561                 fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *),
 562                                        gfp_flags);
 563                 if (!fl->fh_array)
 564                         goto out_err;
 565         }
 566
 567         for (i = 0; i < fl->num_fh; i++) {
 568                 /* Do we want to use a mempool here? */
 569                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 570                 if (!fl->fh_array[i])
 571                         goto out_err_free;
 572
 573                 p = xdr_inline_decode(&stream, 4);
 574                 if (unlikely(!p))
 575                         goto out_err_free;
 576                 fl->fh_array[i]->size = be32_to_cpup(p++);
 577                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 578                         printk(KERN_ERR "Too big fh %d received %d\n",
 579                                i, fl->fh_array[i]->size);
 580                         goto out_err_free;
 581                 }
 582
 583                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 584                 if (unlikely(!p))
 585                         goto out_err_free;
 586                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 587                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 588                         fl->fh_array[i]->size);
 589         }
 590
 591         __free_page(scratch);
 592         return 0;
 593
 594 out_err_free:
 595         filelayout_free_fh_array(fl);
 596 out_err:
 597         __free_page(scratch);
 598         return -EIO;
 599 }
 600
 601 static void
 602 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 603 {
 604         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 605
 606         dprintk("--> %s\n", __func__);
 607         nfs4_fl_put_deviceid(fl->dsaddr);
 608         kfree(fl->commit_buckets);
 609         _filelayout_free_lseg(fl);
 610 }
 611
 612 static struct pnfs_layout_segment *
 613 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 614                       struct nfs4_layoutget_res *lgr,
 615                       gfp_t gfp_flags)
 616 {
 617         struct nfs4_filelayout_segment *fl;
 618         int rc;
 619         struct nfs4_deviceid id;
 620
 621         dprintk("--> %s\n", __func__);
 622         fl = kzalloc(sizeof(*fl), gfp_flags);
 623         if (!fl)
 624                 return NULL;
 625
 626         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 627         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 628                 _filelayout_free_lseg(fl);
 629                 return NULL;
 630         }
 631
 632         /* This assumes there is only one IOMODE_RW lseg.  What
 633          * we really want to do is have a layout_hdr level
 634          * dictionary of <multipath_list4, fh> keys, each
 635          * associated with a struct list_head, populated by calls
 636          * to filelayout_write_pagelist().
 637          * */
 638         if ((!fl->commit_through_mds) && (lgr->range.iomode == IOMODE_RW)) {
 639                 int i;
 640                 int size = (fl->stripe_type == STRIPE_SPARSE) ?
 641                         fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 642
 643                 fl->commit_buckets = kcalloc(size, sizeof(struct list_head), gfp_flags);
 644                 if (!fl->commit_buckets) {
 645                         filelayout_free_lseg(&fl->generic_hdr);
 646                         return NULL;
 647                 }
 648                 fl->number_of_buckets = size;
 649                 for (i = 0; i < size; i++)
 650                         INIT_LIST_HEAD(&fl->commit_buckets[i]);
 651         }
 652         return &fl->generic_hdr;
 653 }
 654
 655 /*
 656  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 657  *
 658  * return true  : coalesce page
 659  * return false : don't coalesce page
 660  */
 661 static bool
 662 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 663                    struct nfs_page *req)
 664 {
 665         u64 p_stripe, r_stripe;
 666         u32 stripe_unit;
 667
 668         if (!pnfs_generic_pg_test(pgio, prev, req) ||
 669             !nfs_generic_pg_test(pgio, prev, req))
 670                 return false;
 671
 672         p_stripe = (u64)prev->wb_index << PAGE_CACHE_SHIFT;
 673         r_stripe = (u64)req->wb_index << PAGE_CACHE_SHIFT;
 674         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 675
 676         do_div(p_stripe, stripe_unit);
 677         do_div(r_stripe, stripe_unit);
 678
 679         return (p_stripe == r_stripe);
 680 }
 681
 682 void
 683 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 684                         struct nfs_page *req)
 685 {
 686         BUG_ON(pgio->pg_lseg != NULL);
 687
 688         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 689                                            req->wb_context,
 690                                            0,
 691                                            NFS4_MAX_UINT64,
 692                                            IOMODE_READ,
 693                                            GFP_KERNEL);
 694         /* If no lseg, fall back to read through mds */
 695         if (pgio->pg_lseg == NULL)
 696                 nfs_pageio_reset_read_mds(pgio);
 697 }
 698
 699 void
 700 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 701                          struct nfs_page *req)
 702 {
 703         BUG_ON(pgio->pg_lseg != NULL);
 704
 705         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 706                                            req->wb_context,
 707                                            0,
 708                                            NFS4_MAX_UINT64,
 709                                            IOMODE_RW,
 710                                            GFP_NOFS);
 711         /* If no lseg, fall back to write through mds */
 712         if (pgio->pg_lseg == NULL)
 713                 nfs_pageio_reset_write_mds(pgio);
 714 }
 715
 716 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 717         .pg_init = filelayout_pg_init_read,
 718         .pg_test = filelayout_pg_test,
 719         .pg_doio = pnfs_generic_pg_readpages,
 720 };
 721
 722 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 723         .pg_init = filelayout_pg_init_write,
 724         .pg_test = filelayout_pg_test,
 725         .pg_doio = pnfs_generic_pg_writepages,
 726 };
 727
 728 static bool filelayout_mark_pnfs_commit(struct pnfs_layout_segment *lseg)
 729 {
 730         return !FILELAYOUT_LSEG(lseg)->commit_through_mds;
 731 }
 732
 733 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 734 {
 735         if (fl->stripe_type == STRIPE_SPARSE)
 736                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
 737         else
 738                 return j;
 739 }
 740
 741 struct list_head *filelayout_choose_commit_list(struct nfs_page *req)
 742 {
 743         struct pnfs_layout_segment *lseg = req->wb_commit_lseg;
 744         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 745         u32 i, j;
 746         struct list_head *list;
 747
 748         /* Note that we are calling nfs4_fl_calc_j_index on each page
 749          * that ends up being committed to a data server.  An attractive
 750          * alternative is to add a field to nfs_write_data and nfs_page
 751          * to store the value calculated in filelayout_write_pagelist
 752          * and just use that here.
 753          */
 754         j = nfs4_fl_calc_j_index(lseg,
 755                                  (loff_t)req->wb_index << PAGE_CACHE_SHIFT);
 756         i = select_bucket_index(fl, j);
 757         list = &fl->commit_buckets[i];
 758         if (list_empty(list)) {
 759                 /* Non-empty buckets hold a reference on the lseg */
 760                 get_lseg(lseg);
 761         }
 762         return list;
 763 }
 764
 765 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 766 {
 767         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 768
 769         if (flseg->stripe_type == STRIPE_SPARSE)
 770                 return i;
 771         else
 772                 return nfs4_fl_calc_ds_index(lseg, i);
 773 }
 774
 775 static struct nfs_fh *
 776 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 777 {
 778         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 779
 780         if (flseg->stripe_type == STRIPE_SPARSE) {
 781                 if (flseg->num_fh == 1)
 782                         i = 0;
 783                 else if (flseg->num_fh == 0)
 784                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
 785                         return NULL;
 786         }
 787         return flseg->fh_array[i];
 788 }
 789
 790 static int filelayout_initiate_commit(struct nfs_write_data *data, int how)
 791 {
 792         struct pnfs_layout_segment *lseg = data->lseg;
 793         struct nfs4_pnfs_ds *ds;
 794         u32 idx;
 795         struct nfs_fh *fh;
 796
 797         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
 798         ds = nfs4_fl_prepare_ds(lseg, idx);
 799         if (!ds) {
 800                 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
 801                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 802                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 803                 prepare_to_resend_writes(data);
 804                 data->mds_ops->rpc_release(data);
 805                 return -EAGAIN;
 806         }
 807         dprintk("%s ino %lu, how %d\n", __func__, data->inode->i_ino, how);
 808         data->write_done_cb = filelayout_commit_done_cb;
 809         data->ds_clp = ds->ds_clp;
 810         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
 811         if (fh)
 812                 data->args.fh = fh;
 813         return nfs_initiate_commit(data, ds->ds_clp->cl_rpcclient,
 814                                    &filelayout_commit_call_ops, how);
 815 }
 816
 817 /*
 818  * This is only useful while we are using whole file layouts.
 819  */
 820 static struct pnfs_layout_segment *find_only_write_lseg(struct inode *inode)
 821 {
 822         struct pnfs_layout_segment *lseg, *rv = NULL;
 823
 824         spin_lock(&inode->i_lock);
 825         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
 826                 if (lseg->pls_range.iomode == IOMODE_RW)
 827                         rv = get_lseg(lseg);
 828         spin_unlock(&inode->i_lock);
 829         return rv;
 830 }
 831
 832 static int alloc_ds_commits(struct inode *inode, struct list_head *list)
 833 {
 834         struct pnfs_layout_segment *lseg;
 835         struct nfs4_filelayout_segment *fl;
 836         struct nfs_write_data *data;
 837         int i, j;
 838
 839         /* Won't need this when non-whole file layout segments are supported
 840          * instead we will use a pnfs_layout_hdr structure */
 841         lseg = find_only_write_lseg(inode);
 842         if (!lseg)
 843                 return 0;
 844         fl = FILELAYOUT_LSEG(lseg);
 845         for (i = 0; i < fl->number_of_buckets; i++) {
 846                 if (list_empty(&fl->commit_buckets[i]))
 847                         continue;
 848                 data = nfs_commitdata_alloc();
 849                 if (!data)
 850                         goto out_bad;
 851                 data->ds_commit_index = i;
 852                 data->lseg = lseg;
 853                 list_add(&data->pages, list);
 854         }
 855         put_lseg(lseg);
 856         return 0;
 857
 858 out_bad:
 859         for (j = i; j < fl->number_of_buckets; j++) {
 860                 if (list_empty(&fl->commit_buckets[i]))
 861                         continue;
 862                 nfs_retry_commit(&fl->commit_buckets[i], lseg);
 863                 put_lseg(lseg);  /* associated with emptying bucket */
 864         }
 865         put_lseg(lseg);
 866         /* Caller will clean up entries put on list */
 867         return -ENOMEM;
 868 }
 869
 870 /* This follows nfs_commit_list pretty closely */
 871 static int
 872 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
 873                            int how)
 874 {
 875         struct nfs_write_data   *data, *tmp;
 876         LIST_HEAD(list);
 877
 878         if (!list_empty(mds_pages)) {
 879                 data = nfs_commitdata_alloc();
 880                 if (!data)
 881                         goto out_bad;
 882                 data->lseg = NULL;
 883                 list_add(&data->pages, &list);
 884         }
 885
 886         if (alloc_ds_commits(inode, &list))
 887                 goto out_bad;
 888
 889         list_for_each_entry_safe(data, tmp, &list, pages) {
 890                 list_del_init(&data->pages);
 891                 atomic_inc(&NFS_I(inode)->commits_outstanding);
 892                 if (!data->lseg) {
 893                         nfs_init_commit(data, mds_pages, NULL);
 894                         nfs_initiate_commit(data, NFS_CLIENT(inode),
 895                                             data->mds_ops, how);
 896                 } else {
 897                         nfs_init_commit(data, &FILELAYOUT_LSEG(data->lseg)->commit_buckets[data->ds_commit_index], data->lseg);
 898                         filelayout_initiate_commit(data, how);
 899                 }
 900         }
 901         return 0;
 902  out_bad:
 903         list_for_each_entry_safe(data, tmp, &list, pages) {
 904                 nfs_retry_commit(&data->pages, data->lseg);
 905                 list_del_init(&data->pages);
 906                 nfs_commit_free(data);
 907         }
 908         nfs_retry_commit(mds_pages, NULL);
 909         nfs_commit_clear_lock(NFS_I(inode));
 910         return -ENOMEM;
 911 }
 912
 913 static void
 914 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
 915 {
 916         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
 917 }
 918
 919 static struct pnfs_layoutdriver_type filelayout_type = {
 920         .id                     = LAYOUT_NFSV4_1_FILES,
 921         .name                   = "LAYOUT_NFSV4_1_FILES",
 922         .owner                  = THIS_MODULE,
 923         .alloc_lseg             = filelayout_alloc_lseg,
 924         .free_lseg              = filelayout_free_lseg,
 925         .pg_read_ops            = &filelayout_pg_read_ops,
 926         .pg_write_ops           = &filelayout_pg_write_ops,
 927         .mark_pnfs_commit       = filelayout_mark_pnfs_commit,
 928         .choose_commit_list     = filelayout_choose_commit_list,
 929         .commit_pagelist        = filelayout_commit_pagelist,
 930         .read_pagelist          = filelayout_read_pagelist,
 931         .write_pagelist         = filelayout_write_pagelist,
 932         .free_deviceid_node     = filelayout_free_deveiceid_node,
 933 };
 934
 935 static int __init nfs4filelayout_init(void)
 936 {
 937         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
 938                __func__);
 939         return pnfs_register_layoutdriver(&filelayout_type);
 940 }
 941
 942 static void __exit nfs4filelayout_exit(void)
 943 {
 944         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
 945                __func__);
 946         pnfs_unregister_layoutdriver(&filelayout_type);
 947 }
 948
 949 MODULE_ALIAS("nfs-layouttype4-1");
 950
 951 module_init(nfs4filelayout_init);
 952 module_exit(nfs4filelayout_exit);