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
  34 #include "internal.h"
  35 #include "nfs4filelayout.h"
  36
  37 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  38
  39 MODULE_LICENSE("GPL");
  40 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
  41 MODULE_DESCRIPTION("The NFSv4 file layout driver");
  42
  43 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
  44
  45 static loff_t
  46 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
  47                             loff_t offset)
  48 {
  49         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
  50         u64 tmp;
  51
  52         offset -= flseg->pattern_offset;
  53         tmp = offset;
  54         do_div(tmp, stripe_width);
  55
  56         return tmp * flseg->stripe_unit + do_div(offset, flseg->stripe_unit);
  57 }
  58
  59 /* This function is used by the layout driver to calculate the
  60  * offset of the file on the dserver based on whether the
  61  * layout type is STRIPE_DENSE or STRIPE_SPARSE
  62  */
  63 static loff_t
  64 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
  65 {
  66         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
  67
  68         switch (flseg->stripe_type) {
  69         case STRIPE_SPARSE:
  70                 return offset;
  71
  72         case STRIPE_DENSE:
  73                 return filelayout_get_dense_offset(flseg, offset);
  74         }
  75
  76         BUG();
  77 }
  78
  79 /* For data server errors we don't recover from */
  80 static void
  81 filelayout_set_lo_fail(struct pnfs_layout_segment *lseg)
  82 {
  83         if (lseg->pls_range.iomode == IOMODE_RW) {
  84                 dprintk("%s Setting layout IOMODE_RW fail bit\n", __func__);
  85                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
  86         } else {
  87                 dprintk("%s Setting layout IOMODE_READ fail bit\n", __func__);
  88                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
  89         }
  90 }
  91
  92 static int filelayout_async_handle_error(struct rpc_task *task,
  93                                          struct nfs4_state *state,
  94                                          struct nfs_client *clp,
  95                                          int *reset)
  96 {
  97         if (task->tk_status >= 0)
  98                 return 0;
  99
 100         *reset = 0;
 101
 102         switch (task->tk_status) {
 103         case -NFS4ERR_BADSESSION:
 104         case -NFS4ERR_BADSLOT:
 105         case -NFS4ERR_BAD_HIGH_SLOT:
 106         case -NFS4ERR_DEADSESSION:
 107         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
 108         case -NFS4ERR_SEQ_FALSE_RETRY:
 109         case -NFS4ERR_SEQ_MISORDERED:
 110                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 111                         "flags 0x%x\n", __func__, task->tk_status,
 112                         clp->cl_exchange_flags);
 113                 nfs4_schedule_session_recovery(clp->cl_session);
 114                 break;
 115         case -NFS4ERR_DELAY:
 116         case -NFS4ERR_GRACE:
 117         case -EKEYEXPIRED:
 118                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 119                 break;
 120         default:
 121                 dprintk("%s DS error. Retry through MDS %d\n", __func__,
 122                         task->tk_status);
 123                 *reset = 1;
 124                 break;
 125         }
 126         task->tk_status = 0;
 127         return -EAGAIN;
 128 }
 129
 130 /* NFS_PROTO call done callback routines */
 131
 132 static int filelayout_read_done_cb(struct rpc_task *task,
 133                                 struct nfs_read_data *data)
 134 {
 135         struct nfs_client *clp = data->ds_clp;
 136         int reset = 0;
 137
 138         dprintk("%s DS read\n", __func__);
 139
 140         if (filelayout_async_handle_error(task, data->args.context->state,
 141                                           data->ds_clp, &reset) == -EAGAIN) {
 142                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 143                         __func__, data->ds_clp, data->ds_clp->cl_session);
 144                 if (reset) {
 145                         filelayout_set_lo_fail(data->lseg);
 146                         nfs4_reset_read(task, data);
 147                         clp = NFS_SERVER(data->inode)->nfs_client;
 148                 }
 149                 nfs_restart_rpc(task, clp);
 150                 return -EAGAIN;
 151         }
 152
 153         return 0;
 154 }
 155
 156 /*
 157  * We reference the rpc_cred of the first WRITE that triggers the need for
 158  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 159  * rfc5661 is not clear about which credential should be used.
 160  */
 161 static void
 162 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
 163 {
 164         if (FILELAYOUT_LSEG(wdata->lseg)->commit_through_mds ||
 165             wdata->res.verf->committed == NFS_FILE_SYNC)
 166                 return;
 167
 168         pnfs_set_layoutcommit(wdata);
 169         dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, wdata->inode->i_ino,
 170                 (unsigned long) wdata->lseg->pls_end_pos);
 171 }
 172
 173 /*
 174  * Call ops for the async read/write cases
 175  * In the case of dense layouts, the offset needs to be reset to its
 176  * original value.
 177  */
 178 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 179 {
 180         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 181
 182         rdata->read_done_cb = filelayout_read_done_cb;
 183
 184         if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
 185                                 &rdata->args.seq_args, &rdata->res.seq_res,
 186                                 0, task))
 187                 return;
 188
 189         rpc_call_start(task);
 190 }
 191
 192 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 193 {
 194         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 195
 196         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 197
 198         /* Note this may cause RPC to be resent */
 199         rdata->mds_ops->rpc_call_done(task, data);
 200 }
 201
 202 static void filelayout_read_release(void *data)
 203 {
 204         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 205
 206         rdata->mds_ops->rpc_release(data);
 207 }
 208
 209 static int filelayout_write_done_cb(struct rpc_task *task,
 210                                 struct nfs_write_data *data)
 211 {
 212         int reset = 0;
 213
 214         if (filelayout_async_handle_error(task, data->args.context->state,
 215                                           data->ds_clp, &reset) == -EAGAIN) {
 216                 struct nfs_client *clp;
 217
 218                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 219                         __func__, data->ds_clp, data->ds_clp->cl_session);
 220                 if (reset) {
 221                         filelayout_set_lo_fail(data->lseg);
 222                         nfs4_reset_write(task, data);
 223                         clp = NFS_SERVER(data->inode)->nfs_client;
 224                 } else
 225                         clp = data->ds_clp;
 226                 nfs_restart_rpc(task, clp);
 227                 return -EAGAIN;
 228         }
 229
 230         filelayout_set_layoutcommit(data);
 231         return 0;
 232 }
 233
 234 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 235 static void prepare_to_resend_writes(struct nfs_write_data *data)
 236 {
 237         struct nfs_page *first = nfs_list_entry(data->pages.next);
 238
 239         data->task.tk_status = 0;
 240         memcpy(data->verf.verifier, first->wb_verf.verifier,
 241                sizeof(first->wb_verf.verifier));
 242         data->verf.verifier[0]++; /* ensure verifier mismatch */
 243 }
 244
 245 static int filelayout_commit_done_cb(struct rpc_task *task,
 246                                      struct nfs_write_data *data)
 247 {
 248         int reset = 0;
 249
 250         if (filelayout_async_handle_error(task, data->args.context->state,
 251                                           data->ds_clp, &reset) == -EAGAIN) {
 252                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 253                         __func__, data->ds_clp, data->ds_clp->cl_session);
 254                 if (reset) {
 255                         prepare_to_resend_writes(data);
 256                         filelayout_set_lo_fail(data->lseg);
 257                 } else
 258                         nfs_restart_rpc(task, data->ds_clp);
 259                 return -EAGAIN;
 260         }
 261
 262         return 0;
 263 }
 264
 265 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 266 {
 267         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 268
 269         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 270                                 &wdata->args.seq_args, &wdata->res.seq_res,
 271                                 0, task))
 272                 return;
 273
 274         rpc_call_start(task);
 275 }
 276
 277 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 278 {
 279         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 280
 281         /* Note this may cause RPC to be resent */
 282         wdata->mds_ops->rpc_call_done(task, data);
 283 }
 284
 285 static void filelayout_write_release(void *data)
 286 {
 287         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 288
 289         wdata->mds_ops->rpc_release(data);
 290 }
 291
 292 static void filelayout_commit_release(void *data)
 293 {
 294         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 295
 296         nfs_commit_release_pages(wdata);
 297         if (atomic_dec_and_test(&NFS_I(wdata->inode)->commits_outstanding))
 298                 nfs_commit_clear_lock(NFS_I(wdata->inode));
 299         nfs_commitdata_release(wdata);
 300 }
 301
 302 struct rpc_call_ops filelayout_read_call_ops = {
 303         .rpc_call_prepare = filelayout_read_prepare,
 304         .rpc_call_done = filelayout_read_call_done,
 305         .rpc_release = filelayout_read_release,
 306 };
 307
 308 struct rpc_call_ops filelayout_write_call_ops = {
 309         .rpc_call_prepare = filelayout_write_prepare,
 310         .rpc_call_done = filelayout_write_call_done,
 311         .rpc_release = filelayout_write_release,
 312 };
 313
 314 struct rpc_call_ops filelayout_commit_call_ops = {
 315         .rpc_call_prepare = filelayout_write_prepare,
 316         .rpc_call_done = filelayout_write_call_done,
 317         .rpc_release = filelayout_commit_release,
 318 };
 319
 320 static enum pnfs_try_status
 321 filelayout_read_pagelist(struct nfs_read_data *data)
 322 {
 323         struct pnfs_layout_segment *lseg = data->lseg;
 324         struct nfs4_pnfs_ds *ds;
 325         loff_t offset = data->args.offset;
 326         u32 j, idx;
 327         struct nfs_fh *fh;
 328         int status;
 329
 330         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 331                 __func__, data->inode->i_ino,
 332                 data->args.pgbase, (size_t)data->args.count, offset);
 333
 334         /* Retrieve the correct rpc_client for the byte range */
 335         j = nfs4_fl_calc_j_index(lseg, offset);
 336         idx = nfs4_fl_calc_ds_index(lseg, j);
 337         ds = nfs4_fl_prepare_ds(lseg, idx);
 338         if (!ds) {
 339                 /* Either layout fh index faulty, or ds connect failed */
 340                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 341                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 342                 return PNFS_NOT_ATTEMPTED;
 343         }
 344         dprintk("%s USE DS:ip %x %hu\n", __func__,
 345                 ntohl(ds->ds_ip_addr), ntohs(ds->ds_port));
 346
 347         /* No multipath support. Use first DS */
 348         data->ds_clp = ds->ds_clp;
 349         fh = nfs4_fl_select_ds_fh(lseg, j);
 350         if (fh)
 351                 data->args.fh = fh;
 352
 353         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 354         data->mds_offset = offset;
 355
 356         /* Perform an asynchronous read to ds */
 357         status = nfs_initiate_read(data, ds->ds_clp->cl_rpcclient,
 358                                    &filelayout_read_call_ops);
 359         BUG_ON(status != 0);
 360         return PNFS_ATTEMPTED;
 361 }
 362
 363 /* Perform async writes. */
 364 static enum pnfs_try_status
 365 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
 366 {
 367         struct pnfs_layout_segment *lseg = data->lseg;
 368         struct nfs4_pnfs_ds *ds;
 369         loff_t offset = data->args.offset;
 370         u32 j, idx;
 371         struct nfs_fh *fh;
 372         int status;
 373
 374         /* Retrieve the correct rpc_client for the byte range */
 375         j = nfs4_fl_calc_j_index(lseg, offset);
 376         idx = nfs4_fl_calc_ds_index(lseg, j);
 377         ds = nfs4_fl_prepare_ds(lseg, idx);
 378         if (!ds) {
 379                 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
 380                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 381                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 382                 return PNFS_NOT_ATTEMPTED;
 383         }
 384         dprintk("%s ino %lu sync %d req %Zu@%llu DS:%x:%hu\n", __func__,
 385                 data->inode->i_ino, sync, (size_t) data->args.count, offset,
 386                 ntohl(ds->ds_ip_addr), ntohs(ds->ds_port));
 387
 388         data->write_done_cb = filelayout_write_done_cb;
 389         data->ds_clp = ds->ds_clp;
 390         fh = nfs4_fl_select_ds_fh(lseg, j);
 391         if (fh)
 392                 data->args.fh = fh;
 393         /*
 394          * Get the file offset on the dserver. Set the write offset to
 395          * this offset and save the original offset.
 396          */
 397         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 398         data->mds_offset = offset;
 399
 400         /* Perform an asynchronous write */
 401         status = nfs_initiate_write(data, ds->ds_clp->cl_rpcclient,
 402                                     &filelayout_write_call_ops, sync);
 403         BUG_ON(status != 0);
 404         return PNFS_ATTEMPTED;
 405 }
 406
 407 /*
 408  * filelayout_check_layout()
 409  *
 410  * Make sure layout segment parameters are sane WRT the device.
 411  * At this point no generic layer initialization of the lseg has occurred,
 412  * and nothing has been added to the layout_hdr cache.
 413  *
 414  */
 415 static int
 416 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 417                         struct nfs4_filelayout_segment *fl,
 418                         struct nfs4_layoutget_res *lgr,
 419                         struct nfs4_deviceid *id)
 420 {
 421         struct nfs4_file_layout_dsaddr *dsaddr;
 422         int status = -EINVAL;
 423         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 424
 425         dprintk("--> %s\n", __func__);
 426
 427         if (fl->pattern_offset > lgr->range.offset) {
 428                 dprintk("%s pattern_offset %lld to large\n",
 429                                 __func__, fl->pattern_offset);
 430                 goto out;
 431         }
 432
 433         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 434                 dprintk("%s Invalid stripe unit (%u)\n",
 435                         __func__, fl->stripe_unit);
 436                 goto out;
 437         }
 438
 439         /* find and reference the deviceid */
 440         dsaddr = nfs4_fl_find_get_deviceid(id);
 441         if (dsaddr == NULL) {
 442                 dsaddr = get_device_info(lo->plh_inode, id);
 443                 if (dsaddr == NULL)
 444                         goto out;
 445         }
 446         fl->dsaddr = dsaddr;
 447
 448         if (fl->first_stripe_index < 0 ||
 449             fl->first_stripe_index >= dsaddr->stripe_count) {
 450                 dprintk("%s Bad first_stripe_index %d\n",
 451                                 __func__, fl->first_stripe_index);
 452                 goto out_put;
 453         }
 454
 455         if ((fl->stripe_type == STRIPE_SPARSE &&
 456             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 457             (fl->stripe_type == STRIPE_DENSE &&
 458             fl->num_fh != dsaddr->stripe_count)) {
 459                 dprintk("%s num_fh %u not valid for given packing\n",
 460                         __func__, fl->num_fh);
 461                 goto out_put;
 462         }
 463
 464         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
 465                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
 466                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
 467                         nfss->wsize);
 468         }
 469
 470         status = 0;
 471 out:
 472         dprintk("--> %s returns %d\n", __func__, status);
 473         return status;
 474 out_put:
 475         nfs4_fl_put_deviceid(dsaddr);
 476         goto out;
 477 }
 478
 479 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 480 {
 481         int i;
 482
 483         for (i = 0; i < fl->num_fh; i++) {
 484                 if (!fl->fh_array[i])
 485                         break;
 486                 kfree(fl->fh_array[i]);
 487         }
 488         kfree(fl->fh_array);
 489         fl->fh_array = NULL;
 490 }
 491
 492 static void
 493 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 494 {
 495         filelayout_free_fh_array(fl);
 496         kfree(fl);
 497 }
 498
 499 static int
 500 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 501                          struct nfs4_filelayout_segment *fl,
 502                          struct nfs4_layoutget_res *lgr,
 503                          struct nfs4_deviceid *id)
 504 {
 505         struct xdr_stream stream;
 506         struct xdr_buf buf = {
 507                 .pages =  lgr->layoutp->pages,
 508                 .page_len =  lgr->layoutp->len,
 509                 .buflen =  lgr->layoutp->len,
 510                 .len = lgr->layoutp->len,
 511         };
 512         struct page *scratch;
 513         __be32 *p;
 514         uint32_t nfl_util;
 515         int i;
 516
 517         dprintk("%s: set_layout_map Begin\n", __func__);
 518
 519         scratch = alloc_page(GFP_KERNEL);
 520         if (!scratch)
 521                 return -ENOMEM;
 522
 523         xdr_init_decode(&stream, &buf, NULL);
 524         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 525
 526         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 527          * num_fh (4) */
 528         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 529         if (unlikely(!p))
 530                 goto out_err;
 531
 532         memcpy(id, p, sizeof(*id));
 533         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 534         print_deviceid(id);
 535
 536         nfl_util = be32_to_cpup(p++);
 537         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 538                 fl->commit_through_mds = 1;
 539         if (nfl_util & NFL4_UFLG_DENSE)
 540                 fl->stripe_type = STRIPE_DENSE;
 541         else
 542                 fl->stripe_type = STRIPE_SPARSE;
 543         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 544
 545         fl->first_stripe_index = be32_to_cpup(p++);
 546         p = xdr_decode_hyper(p, &fl->pattern_offset);
 547         fl->num_fh = be32_to_cpup(p++);
 548
 549         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 550                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 551                 fl->pattern_offset);
 552
 553         if (!fl->num_fh)
 554                 goto out_err;
 555
 556         fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *),
 557                                GFP_KERNEL);
 558         if (!fl->fh_array)
 559                 goto out_err;
 560
 561         for (i = 0; i < fl->num_fh; i++) {
 562                 /* Do we want to use a mempool here? */
 563                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), GFP_KERNEL);
 564                 if (!fl->fh_array[i])
 565                         goto out_err_free;
 566
 567                 p = xdr_inline_decode(&stream, 4);
 568                 if (unlikely(!p))
 569                         goto out_err_free;
 570                 fl->fh_array[i]->size = be32_to_cpup(p++);
 571                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 572                         printk(KERN_ERR "Too big fh %d received %d\n",
 573                                i, fl->fh_array[i]->size);
 574                         goto out_err_free;
 575                 }
 576
 577                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 578                 if (unlikely(!p))
 579                         goto out_err_free;
 580                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 581                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 582                         fl->fh_array[i]->size);
 583         }
 584
 585         __free_page(scratch);
 586         return 0;
 587
 588 out_err_free:
 589         filelayout_free_fh_array(fl);
 590 out_err:
 591         __free_page(scratch);
 592         return -EIO;
 593 }
 594
 595 static void
 596 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 597 {
 598         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 599
 600         dprintk("--> %s\n", __func__);
 601         nfs4_fl_put_deviceid(fl->dsaddr);
 602         kfree(fl->commit_buckets);
 603         _filelayout_free_lseg(fl);
 604 }
 605
 606 static struct pnfs_layout_segment *
 607 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 608                       struct nfs4_layoutget_res *lgr)
 609 {
 610         struct nfs4_filelayout_segment *fl;
 611         int rc;
 612         struct nfs4_deviceid id;
 613
 614         dprintk("--> %s\n", __func__);
 615         fl = kzalloc(sizeof(*fl), GFP_KERNEL);
 616         if (!fl)
 617                 return NULL;
 618
 619         rc = filelayout_decode_layout(layoutid, fl, lgr, &id);
 620         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id)) {
 621                 _filelayout_free_lseg(fl);
 622                 return NULL;
 623         }
 624
 625         /* This assumes there is only one IOMODE_RW lseg.  What
 626          * we really want to do is have a layout_hdr level
 627          * dictionary of <multipath_list4, fh> keys, each
 628          * associated with a struct list_head, populated by calls
 629          * to filelayout_write_pagelist().
 630          * */
 631         if ((!fl->commit_through_mds) && (lgr->range.iomode == IOMODE_RW)) {
 632                 int i;
 633                 int size = (fl->stripe_type == STRIPE_SPARSE) ?
 634                         fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 635
 636                 fl->commit_buckets = kcalloc(size, sizeof(struct list_head), GFP_KERNEL);
 637                 if (!fl->commit_buckets) {
 638                         filelayout_free_lseg(&fl->generic_hdr);
 639                         return NULL;
 640                 }
 641                 fl->number_of_buckets = size;
 642                 for (i = 0; i < size; i++)
 643                         INIT_LIST_HEAD(&fl->commit_buckets[i]);
 644         }
 645         return &fl->generic_hdr;
 646 }
 647
 648 /*
 649  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 650  *
 651  * return 1 :  coalesce page
 652  * return 0 :  don't coalesce page
 653  */
 654 int
 655 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 656                    struct nfs_page *req)
 657 {
 658         u64 p_stripe, r_stripe;
 659         u32 stripe_unit;
 660
 661         if (!pgio->pg_lseg)
 662                 return 1;
 663         p_stripe = (u64)prev->wb_index << PAGE_CACHE_SHIFT;
 664         r_stripe = (u64)req->wb_index << PAGE_CACHE_SHIFT;
 665         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 666
 667         do_div(p_stripe, stripe_unit);
 668         do_div(r_stripe, stripe_unit);
 669
 670         return (p_stripe == r_stripe);
 671 }
 672
 673 static bool filelayout_mark_pnfs_commit(struct pnfs_layout_segment *lseg)
 674 {
 675         return !FILELAYOUT_LSEG(lseg)->commit_through_mds;
 676 }
 677
 678 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 679 {
 680         if (fl->stripe_type == STRIPE_SPARSE)
 681                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
 682         else
 683                 return j;
 684 }
 685
 686 struct list_head *filelayout_choose_commit_list(struct nfs_page *req)
 687 {
 688         struct pnfs_layout_segment *lseg = req->wb_commit_lseg;
 689         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 690         u32 i, j;
 691         struct list_head *list;
 692
 693         /* Note that we are calling nfs4_fl_calc_j_index on each page
 694          * that ends up being committed to a data server.  An attractive
 695          * alternative is to add a field to nfs_write_data and nfs_page
 696          * to store the value calculated in filelayout_write_pagelist
 697          * and just use that here.
 698          */
 699         j = nfs4_fl_calc_j_index(lseg,
 700                                  (loff_t)req->wb_index << PAGE_CACHE_SHIFT);
 701         i = select_bucket_index(fl, j);
 702         list = &fl->commit_buckets[i];
 703         if (list_empty(list)) {
 704                 /* Non-empty buckets hold a reference on the lseg */
 705                 get_lseg(lseg);
 706         }
 707         return list;
 708 }
 709
 710 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 711 {
 712         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 713
 714         if (flseg->stripe_type == STRIPE_SPARSE)
 715                 return i;
 716         else
 717                 return nfs4_fl_calc_ds_index(lseg, i);
 718 }
 719
 720 static struct nfs_fh *
 721 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 722 {
 723         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 724
 725         if (flseg->stripe_type == STRIPE_SPARSE) {
 726                 if (flseg->num_fh == 1)
 727                         i = 0;
 728                 else if (flseg->num_fh == 0)
 729                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
 730                         return NULL;
 731         }
 732         return flseg->fh_array[i];
 733 }
 734
 735 static int filelayout_initiate_commit(struct nfs_write_data *data, int how)
 736 {
 737         struct pnfs_layout_segment *lseg = data->lseg;
 738         struct nfs4_pnfs_ds *ds;
 739         u32 idx;
 740         struct nfs_fh *fh;
 741
 742         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
 743         ds = nfs4_fl_prepare_ds(lseg, idx);
 744         if (!ds) {
 745                 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
 746                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 747                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 748                 prepare_to_resend_writes(data);
 749                 data->mds_ops->rpc_release(data);
 750                 return -EAGAIN;
 751         }
 752         dprintk("%s ino %lu, how %d\n", __func__, data->inode->i_ino, how);
 753         data->write_done_cb = filelayout_commit_done_cb;
 754         data->ds_clp = ds->ds_clp;
 755         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
 756         if (fh)
 757                 data->args.fh = fh;
 758         return nfs_initiate_commit(data, ds->ds_clp->cl_rpcclient,
 759                                    &filelayout_commit_call_ops, how);
 760 }
 761
 762 /*
 763  * This is only useful while we are using whole file layouts.
 764  */
 765 static struct pnfs_layout_segment *find_only_write_lseg(struct inode *inode)
 766 {
 767         struct pnfs_layout_segment *lseg, *rv = NULL;
 768
 769         spin_lock(&inode->i_lock);
 770         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
 771                 if (lseg->pls_range.iomode == IOMODE_RW)
 772                         rv = get_lseg(lseg);
 773         spin_unlock(&inode->i_lock);
 774         return rv;
 775 }
 776
 777 static int alloc_ds_commits(struct inode *inode, struct list_head *list)
 778 {
 779         struct pnfs_layout_segment *lseg;
 780         struct nfs4_filelayout_segment *fl;
 781         struct nfs_write_data *data;
 782         int i, j;
 783
 784         /* Won't need this when non-whole file layout segments are supported
 785          * instead we will use a pnfs_layout_hdr structure */
 786         lseg = find_only_write_lseg(inode);
 787         if (!lseg)
 788                 return 0;
 789         fl = FILELAYOUT_LSEG(lseg);
 790         for (i = 0; i < fl->number_of_buckets; i++) {
 791                 if (list_empty(&fl->commit_buckets[i]))
 792                         continue;
 793                 data = nfs_commitdata_alloc();
 794                 if (!data)
 795                         goto out_bad;
 796                 data->ds_commit_index = i;
 797                 data->lseg = lseg;
 798                 list_add(&data->pages, list);
 799         }
 800         put_lseg(lseg);
 801         return 0;
 802
 803 out_bad:
 804         for (j = i; j < fl->number_of_buckets; j++) {
 805                 if (list_empty(&fl->commit_buckets[i]))
 806                         continue;
 807                 nfs_retry_commit(&fl->commit_buckets[i], lseg);
 808                 put_lseg(lseg);  /* associated with emptying bucket */
 809         }
 810         put_lseg(lseg);
 811         /* Caller will clean up entries put on list */
 812         return -ENOMEM;
 813 }
 814
 815 /* This follows nfs_commit_list pretty closely */
 816 static int
 817 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
 818                            int how)
 819 {
 820         struct nfs_write_data   *data, *tmp;
 821         LIST_HEAD(list);
 822
 823         if (!list_empty(mds_pages)) {
 824                 data = nfs_commitdata_alloc();
 825                 if (!data)
 826                         goto out_bad;
 827                 data->lseg = NULL;
 828                 list_add(&data->pages, &list);
 829         }
 830
 831         if (alloc_ds_commits(inode, &list))
 832                 goto out_bad;
 833
 834         list_for_each_entry_safe(data, tmp, &list, pages) {
 835                 list_del_init(&data->pages);
 836                 atomic_inc(&NFS_I(inode)->commits_outstanding);
 837                 if (!data->lseg) {
 838                         nfs_init_commit(data, mds_pages, NULL);
 839                         nfs_initiate_commit(data, NFS_CLIENT(inode),
 840                                             data->mds_ops, how);
 841                 } else {
 842                         nfs_init_commit(data, &FILELAYOUT_LSEG(data->lseg)->commit_buckets[data->ds_commit_index], data->lseg);
 843                         filelayout_initiate_commit(data, how);
 844                 }
 845         }
 846         return 0;
 847  out_bad:
 848         list_for_each_entry_safe(data, tmp, &list, pages) {
 849                 nfs_retry_commit(&data->pages, data->lseg);
 850                 list_del_init(&data->pages);
 851                 nfs_commit_free(data);
 852         }
 853         nfs_retry_commit(mds_pages, NULL);
 854         nfs_commit_clear_lock(NFS_I(inode));
 855         return -ENOMEM;
 856 }
 857
 858 static struct pnfs_layoutdriver_type filelayout_type = {
 859         .id                     = LAYOUT_NFSV4_1_FILES,
 860         .name                   = "LAYOUT_NFSV4_1_FILES",
 861         .owner                  = THIS_MODULE,
 862         .alloc_lseg             = filelayout_alloc_lseg,
 863         .free_lseg              = filelayout_free_lseg,
 864         .pg_test                = filelayout_pg_test,
 865         .mark_pnfs_commit       = filelayout_mark_pnfs_commit,
 866         .choose_commit_list     = filelayout_choose_commit_list,
 867         .commit_pagelist        = filelayout_commit_pagelist,
 868         .read_pagelist          = filelayout_read_pagelist,
 869         .write_pagelist         = filelayout_write_pagelist,
 870 };
 871
 872 static int __init nfs4filelayout_init(void)
 873 {
 874         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
 875                __func__);
 876         return pnfs_register_layoutdriver(&filelayout_type);
 877 }
 878
 879 static void __exit nfs4filelayout_exit(void)
 880 {
 881         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
 882                __func__);
 883         pnfs_unregister_layoutdriver(&filelayout_type);
 884 }
 885
 886 module_init(nfs4filelayout_init);
 887 module_exit(nfs4filelayout_exit);