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         case -NFS4ERR_RETRY_UNCACHED_REP:
 121                 break;
 122         default:
 123                 dprintk("%s DS error. Retry through MDS %d\n", __func__,
 124                         task->tk_status);
 125                 *reset = 1;
 126                 break;
 127         }
 128         task->tk_status = 0;
 129         return -EAGAIN;
 130 }
 131
 132 /* NFS_PROTO call done callback routines */
 133
 134 static int filelayout_read_done_cb(struct rpc_task *task,
 135                                 struct nfs_read_data *data)
 136 {
 137         struct nfs_client *clp = data->ds_clp;
 138         int reset = 0;
 139
 140         dprintk("%s DS read\n", __func__);
 141
 142         if (filelayout_async_handle_error(task, data->args.context->state,
 143                                           data->ds_clp, &reset) == -EAGAIN) {
 144                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 145                         __func__, data->ds_clp, data->ds_clp->cl_session);
 146                 if (reset) {
 147                         filelayout_set_lo_fail(data->lseg);
 148                         nfs4_reset_read(task, data);
 149                         clp = NFS_SERVER(data->inode)->nfs_client;
 150                 }
 151                 nfs_restart_rpc(task, clp);
 152                 return -EAGAIN;
 153         }
 154
 155         return 0;
 156 }
 157
 158 /*
 159  * We reference the rpc_cred of the first WRITE that triggers the need for
 160  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 161  * rfc5661 is not clear about which credential should be used.
 162  */
 163 static void
 164 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
 165 {
 166         if (FILELAYOUT_LSEG(wdata->lseg)->commit_through_mds ||
 167             wdata->res.verf->committed == NFS_FILE_SYNC)
 168                 return;
 169
 170         pnfs_set_layoutcommit(wdata);
 171         dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, wdata->inode->i_ino,
 172                 (unsigned long) wdata->lseg->pls_end_pos);
 173 }
 174
 175 /*
 176  * Call ops for the async read/write cases
 177  * In the case of dense layouts, the offset needs to be reset to its
 178  * original value.
 179  */
 180 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 181 {
 182         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 183
 184         rdata->read_done_cb = filelayout_read_done_cb;
 185
 186         if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
 187                                 &rdata->args.seq_args, &rdata->res.seq_res,
 188                                 0, task))
 189                 return;
 190
 191         rpc_call_start(task);
 192 }
 193
 194 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 195 {
 196         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 197
 198         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 199
 200         /* Note this may cause RPC to be resent */
 201         rdata->mds_ops->rpc_call_done(task, data);
 202 }
 203
 204 static void filelayout_read_release(void *data)
 205 {
 206         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
 207
 208         rdata->mds_ops->rpc_release(data);
 209 }
 210
 211 static int filelayout_write_done_cb(struct rpc_task *task,
 212                                 struct nfs_write_data *data)
 213 {
 214         int reset = 0;
 215
 216         if (filelayout_async_handle_error(task, data->args.context->state,
 217                                           data->ds_clp, &reset) == -EAGAIN) {
 218                 struct nfs_client *clp;
 219
 220                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 221                         __func__, data->ds_clp, data->ds_clp->cl_session);
 222                 if (reset) {
 223                         filelayout_set_lo_fail(data->lseg);
 224                         nfs4_reset_write(task, data);
 225                         clp = NFS_SERVER(data->inode)->nfs_client;
 226                 } else
 227                         clp = data->ds_clp;
 228                 nfs_restart_rpc(task, clp);
 229                 return -EAGAIN;
 230         }
 231
 232         filelayout_set_layoutcommit(data);
 233         return 0;
 234 }
 235
 236 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 237 static void prepare_to_resend_writes(struct nfs_write_data *data)
 238 {
 239         struct nfs_page *first = nfs_list_entry(data->pages.next);
 240
 241         data->task.tk_status = 0;
 242         memcpy(data->verf.verifier, first->wb_verf.verifier,
 243                sizeof(first->wb_verf.verifier));
 244         data->verf.verifier[0]++; /* ensure verifier mismatch */
 245 }
 246
 247 static int filelayout_commit_done_cb(struct rpc_task *task,
 248                                      struct nfs_write_data *data)
 249 {
 250         int reset = 0;
 251
 252         if (filelayout_async_handle_error(task, data->args.context->state,
 253                                           data->ds_clp, &reset) == -EAGAIN) {
 254                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
 255                         __func__, data->ds_clp, data->ds_clp->cl_session);
 256                 if (reset) {
 257                         prepare_to_resend_writes(data);
 258                         filelayout_set_lo_fail(data->lseg);
 259                 } else
 260                         nfs_restart_rpc(task, data->ds_clp);
 261                 return -EAGAIN;
 262         }
 263
 264         return 0;
 265 }
 266
 267 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 268 {
 269         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 270
 271         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
 272                                 &wdata->args.seq_args, &wdata->res.seq_res,
 273                                 0, task))
 274                 return;
 275
 276         rpc_call_start(task);
 277 }
 278
 279 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 280 {
 281         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 282
 283         /* Note this may cause RPC to be resent */
 284         wdata->mds_ops->rpc_call_done(task, data);
 285 }
 286
 287 static void filelayout_write_release(void *data)
 288 {
 289         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 290
 291         wdata->mds_ops->rpc_release(data);
 292 }
 293
 294 static void filelayout_commit_release(void *data)
 295 {
 296         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
 297
 298         nfs_commit_release_pages(wdata);
 299         if (atomic_dec_and_test(&NFS_I(wdata->inode)->commits_outstanding))
 300                 nfs_commit_clear_lock(NFS_I(wdata->inode));
 301         nfs_commitdata_release(wdata);
 302 }
 303
 304 struct rpc_call_ops filelayout_read_call_ops = {
 305         .rpc_call_prepare = filelayout_read_prepare,
 306         .rpc_call_done = filelayout_read_call_done,
 307         .rpc_release = filelayout_read_release,
 308 };
 309
 310 struct rpc_call_ops filelayout_write_call_ops = {
 311         .rpc_call_prepare = filelayout_write_prepare,
 312         .rpc_call_done = filelayout_write_call_done,
 313         .rpc_release = filelayout_write_release,
 314 };
 315
 316 struct rpc_call_ops filelayout_commit_call_ops = {
 317         .rpc_call_prepare = filelayout_write_prepare,
 318         .rpc_call_done = filelayout_write_call_done,
 319         .rpc_release = filelayout_commit_release,
 320 };
 321
 322 static enum pnfs_try_status
 323 filelayout_read_pagelist(struct nfs_read_data *data)
 324 {
 325         struct pnfs_layout_segment *lseg = data->lseg;
 326         struct nfs4_pnfs_ds *ds;
 327         loff_t offset = data->args.offset;
 328         u32 j, idx;
 329         struct nfs_fh *fh;
 330         int status;
 331
 332         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 333                 __func__, data->inode->i_ino,
 334                 data->args.pgbase, (size_t)data->args.count, offset);
 335
 336         /* Retrieve the correct rpc_client for the byte range */
 337         j = nfs4_fl_calc_j_index(lseg, offset);
 338         idx = nfs4_fl_calc_ds_index(lseg, j);
 339         ds = nfs4_fl_prepare_ds(lseg, idx);
 340         if (!ds) {
 341                 /* Either layout fh index faulty, or ds connect failed */
 342                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 343                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 344                 return PNFS_NOT_ATTEMPTED;
 345         }
 346         dprintk("%s USE DS:ip %x %hu\n", __func__,
 347                 ntohl(ds->ds_ip_addr), ntohs(ds->ds_port));
 348
 349         /* No multipath support. Use first DS */
 350         data->ds_clp = ds->ds_clp;
 351         fh = nfs4_fl_select_ds_fh(lseg, j);
 352         if (fh)
 353                 data->args.fh = fh;
 354
 355         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 356         data->mds_offset = offset;
 357
 358         /* Perform an asynchronous read to ds */
 359         status = nfs_initiate_read(data, ds->ds_clp->cl_rpcclient,
 360                                    &filelayout_read_call_ops);
 361         BUG_ON(status != 0);
 362         return PNFS_ATTEMPTED;
 363 }
 364
 365 /* Perform async writes. */
 366 static enum pnfs_try_status
 367 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
 368 {
 369         struct pnfs_layout_segment *lseg = data->lseg;
 370         struct nfs4_pnfs_ds *ds;
 371         loff_t offset = data->args.offset;
 372         u32 j, idx;
 373         struct nfs_fh *fh;
 374         int status;
 375
 376         /* Retrieve the correct rpc_client for the byte range */
 377         j = nfs4_fl_calc_j_index(lseg, offset);
 378         idx = nfs4_fl_calc_ds_index(lseg, j);
 379         ds = nfs4_fl_prepare_ds(lseg, idx);
 380         if (!ds) {
 381                 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
 382                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 383                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 384                 return PNFS_NOT_ATTEMPTED;
 385         }
 386         dprintk("%s ino %lu sync %d req %Zu@%llu DS:%x:%hu\n", __func__,
 387                 data->inode->i_ino, sync, (size_t) data->args.count, offset,
 388                 ntohl(ds->ds_ip_addr), ntohs(ds->ds_port));
 389
 390         data->write_done_cb = filelayout_write_done_cb;
 391         data->ds_clp = ds->ds_clp;
 392         fh = nfs4_fl_select_ds_fh(lseg, j);
 393         if (fh)
 394                 data->args.fh = fh;
 395         /*
 396          * Get the file offset on the dserver. Set the write offset to
 397          * this offset and save the original offset.
 398          */
 399         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
 400         data->mds_offset = offset;
 401
 402         /* Perform an asynchronous write */
 403         status = nfs_initiate_write(data, ds->ds_clp->cl_rpcclient,
 404                                     &filelayout_write_call_ops, sync);
 405         BUG_ON(status != 0);
 406         return PNFS_ATTEMPTED;
 407 }
 408
 409 /*
 410  * filelayout_check_layout()
 411  *
 412  * Make sure layout segment parameters are sane WRT the device.
 413  * At this point no generic layer initialization of the lseg has occurred,
 414  * and nothing has been added to the layout_hdr cache.
 415  *
 416  */
 417 static int
 418 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 419                         struct nfs4_filelayout_segment *fl,
 420                         struct nfs4_layoutget_res *lgr,
 421                         struct nfs4_deviceid *id,
 422                         gfp_t gfp_flags)
 423 {
 424         struct nfs4_file_layout_dsaddr *dsaddr;
 425         int status = -EINVAL;
 426         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
 427
 428         dprintk("--> %s\n", __func__);
 429
 430         if (fl->pattern_offset > lgr->range.offset) {
 431                 dprintk("%s pattern_offset %lld to large\n",
 432                                 __func__, fl->pattern_offset);
 433                 goto out;
 434         }
 435
 436         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
 437                 dprintk("%s Invalid stripe unit (%u)\n",
 438                         __func__, fl->stripe_unit);
 439                 goto out;
 440         }
 441
 442         /* find and reference the deviceid */
 443         dsaddr = nfs4_fl_find_get_deviceid(id);
 444         if (dsaddr == NULL) {
 445                 dsaddr = get_device_info(lo->plh_inode, id, gfp_flags);
 446                 if (dsaddr == NULL)
 447                         goto out;
 448         }
 449         fl->dsaddr = dsaddr;
 450
 451         if (fl->first_stripe_index < 0 ||
 452             fl->first_stripe_index >= dsaddr->stripe_count) {
 453                 dprintk("%s Bad first_stripe_index %d\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                 .pages =  lgr->layoutp->pages,
 512                 .page_len =  lgr->layoutp->len,
 513                 .buflen =  lgr->layoutp->len,
 514                 .len = lgr->layoutp->len,
 515         };
 516         struct page *scratch;
 517         __be32 *p;
 518         uint32_t nfl_util;
 519         int i;
 520
 521         dprintk("%s: set_layout_map Begin\n", __func__);
 522
 523         scratch = alloc_page(gfp_flags);
 524         if (!scratch)
 525                 return -ENOMEM;
 526
 527         xdr_init_decode(&stream, &buf, NULL);
 528         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 529
 530         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 531          * num_fh (4) */
 532         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 533         if (unlikely(!p))
 534                 goto out_err;
 535
 536         memcpy(id, p, sizeof(*id));
 537         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 538         print_deviceid(id);
 539
 540         nfl_util = be32_to_cpup(p++);
 541         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 542                 fl->commit_through_mds = 1;
 543         if (nfl_util & NFL4_UFLG_DENSE)
 544                 fl->stripe_type = STRIPE_DENSE;
 545         else
 546                 fl->stripe_type = STRIPE_SPARSE;
 547         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 548
 549         fl->first_stripe_index = be32_to_cpup(p++);
 550         p = xdr_decode_hyper(p, &fl->pattern_offset);
 551         fl->num_fh = be32_to_cpup(p++);
 552
 553         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 554                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 555                 fl->pattern_offset);
 556
 557         if (!fl->num_fh)
 558                 goto out_err;
 559
 560         fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *),
 561                                gfp_flags);
 562         if (!fl->fh_array)
 563                 goto out_err;
 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 1 :  coalesce page
 657  * return 0 :  don't coalesce page
 658  */
 659 int
 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 (!pgio->pg_lseg)
 667                 return 1;
 668         p_stripe = (u64)prev->wb_index << PAGE_CACHE_SHIFT;
 669         r_stripe = (u64)req->wb_index << PAGE_CACHE_SHIFT;
 670         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 671
 672         do_div(p_stripe, stripe_unit);
 673         do_div(r_stripe, stripe_unit);
 674
 675         return (p_stripe == r_stripe);
 676 }
 677
 678 static bool filelayout_mark_pnfs_commit(struct pnfs_layout_segment *lseg)
 679 {
 680         return !FILELAYOUT_LSEG(lseg)->commit_through_mds;
 681 }
 682
 683 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 684 {
 685         if (fl->stripe_type == STRIPE_SPARSE)
 686                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
 687         else
 688                 return j;
 689 }
 690
 691 struct list_head *filelayout_choose_commit_list(struct nfs_page *req)
 692 {
 693         struct pnfs_layout_segment *lseg = req->wb_commit_lseg;
 694         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 695         u32 i, j;
 696         struct list_head *list;
 697
 698         /* Note that we are calling nfs4_fl_calc_j_index on each page
 699          * that ends up being committed to a data server.  An attractive
 700          * alternative is to add a field to nfs_write_data and nfs_page
 701          * to store the value calculated in filelayout_write_pagelist
 702          * and just use that here.
 703          */
 704         j = nfs4_fl_calc_j_index(lseg,
 705                                  (loff_t)req->wb_index << PAGE_CACHE_SHIFT);
 706         i = select_bucket_index(fl, j);
 707         list = &fl->commit_buckets[i];
 708         if (list_empty(list)) {
 709                 /* Non-empty buckets hold a reference on the lseg */
 710                 get_lseg(lseg);
 711         }
 712         return list;
 713 }
 714
 715 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 716 {
 717         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 718
 719         if (flseg->stripe_type == STRIPE_SPARSE)
 720                 return i;
 721         else
 722                 return nfs4_fl_calc_ds_index(lseg, i);
 723 }
 724
 725 static struct nfs_fh *
 726 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 727 {
 728         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 729
 730         if (flseg->stripe_type == STRIPE_SPARSE) {
 731                 if (flseg->num_fh == 1)
 732                         i = 0;
 733                 else if (flseg->num_fh == 0)
 734                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
 735                         return NULL;
 736         }
 737         return flseg->fh_array[i];
 738 }
 739
 740 static int filelayout_initiate_commit(struct nfs_write_data *data, int how)
 741 {
 742         struct pnfs_layout_segment *lseg = data->lseg;
 743         struct nfs4_pnfs_ds *ds;
 744         u32 idx;
 745         struct nfs_fh *fh;
 746
 747         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
 748         ds = nfs4_fl_prepare_ds(lseg, idx);
 749         if (!ds) {
 750                 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
 751                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
 752                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
 753                 prepare_to_resend_writes(data);
 754                 data->mds_ops->rpc_release(data);
 755                 return -EAGAIN;
 756         }
 757         dprintk("%s ino %lu, how %d\n", __func__, data->inode->i_ino, how);
 758         data->write_done_cb = filelayout_commit_done_cb;
 759         data->ds_clp = ds->ds_clp;
 760         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
 761         if (fh)
 762                 data->args.fh = fh;
 763         return nfs_initiate_commit(data, ds->ds_clp->cl_rpcclient,
 764                                    &filelayout_commit_call_ops, how);
 765 }
 766
 767 /*
 768  * This is only useful while we are using whole file layouts.
 769  */
 770 static struct pnfs_layout_segment *find_only_write_lseg(struct inode *inode)
 771 {
 772         struct pnfs_layout_segment *lseg, *rv = NULL;
 773
 774         spin_lock(&inode->i_lock);
 775         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
 776                 if (lseg->pls_range.iomode == IOMODE_RW)
 777                         rv = get_lseg(lseg);
 778         spin_unlock(&inode->i_lock);
 779         return rv;
 780 }
 781
 782 static int alloc_ds_commits(struct inode *inode, struct list_head *list)
 783 {
 784         struct pnfs_layout_segment *lseg;
 785         struct nfs4_filelayout_segment *fl;
 786         struct nfs_write_data *data;
 787         int i, j;
 788
 789         /* Won't need this when non-whole file layout segments are supported
 790          * instead we will use a pnfs_layout_hdr structure */
 791         lseg = find_only_write_lseg(inode);
 792         if (!lseg)
 793                 return 0;
 794         fl = FILELAYOUT_LSEG(lseg);
 795         for (i = 0; i < fl->number_of_buckets; i++) {
 796                 if (list_empty(&fl->commit_buckets[i]))
 797                         continue;
 798                 data = nfs_commitdata_alloc();
 799                 if (!data)
 800                         goto out_bad;
 801                 data->ds_commit_index = i;
 802                 data->lseg = lseg;
 803                 list_add(&data->pages, list);
 804         }
 805         put_lseg(lseg);
 806         return 0;
 807
 808 out_bad:
 809         for (j = i; j < fl->number_of_buckets; j++) {
 810                 if (list_empty(&fl->commit_buckets[i]))
 811                         continue;
 812                 nfs_retry_commit(&fl->commit_buckets[i], lseg);
 813                 put_lseg(lseg);  /* associated with emptying bucket */
 814         }
 815         put_lseg(lseg);
 816         /* Caller will clean up entries put on list */
 817         return -ENOMEM;
 818 }
 819
 820 /* This follows nfs_commit_list pretty closely */
 821 static int
 822 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
 823                            int how)
 824 {
 825         struct nfs_write_data   *data, *tmp;
 826         LIST_HEAD(list);
 827
 828         if (!list_empty(mds_pages)) {
 829                 data = nfs_commitdata_alloc();
 830                 if (!data)
 831                         goto out_bad;
 832                 data->lseg = NULL;
 833                 list_add(&data->pages, &list);
 834         }
 835
 836         if (alloc_ds_commits(inode, &list))
 837                 goto out_bad;
 838
 839         list_for_each_entry_safe(data, tmp, &list, pages) {
 840                 list_del_init(&data->pages);
 841                 atomic_inc(&NFS_I(inode)->commits_outstanding);
 842                 if (!data->lseg) {
 843                         nfs_init_commit(data, mds_pages, NULL);
 844                         nfs_initiate_commit(data, NFS_CLIENT(inode),
 845                                             data->mds_ops, how);
 846                 } else {
 847                         nfs_init_commit(data, &FILELAYOUT_LSEG(data->lseg)->commit_buckets[data->ds_commit_index], data->lseg);
 848                         filelayout_initiate_commit(data, how);
 849                 }
 850         }
 851         return 0;
 852  out_bad:
 853         list_for_each_entry_safe(data, tmp, &list, pages) {
 854                 nfs_retry_commit(&data->pages, data->lseg);
 855                 list_del_init(&data->pages);
 856                 nfs_commit_free(data);
 857         }
 858         nfs_retry_commit(mds_pages, NULL);
 859         nfs_commit_clear_lock(NFS_I(inode));
 860         return -ENOMEM;
 861 }
 862
 863 static struct pnfs_layoutdriver_type filelayout_type = {
 864         .id                     = LAYOUT_NFSV4_1_FILES,
 865         .name                   = "LAYOUT_NFSV4_1_FILES",
 866         .owner                  = THIS_MODULE,
 867         .alloc_lseg             = filelayout_alloc_lseg,
 868         .free_lseg              = filelayout_free_lseg,
 869         .pg_test                = filelayout_pg_test,
 870         .mark_pnfs_commit       = filelayout_mark_pnfs_commit,
 871         .choose_commit_list     = filelayout_choose_commit_list,
 872         .commit_pagelist        = filelayout_commit_pagelist,
 873         .read_pagelist          = filelayout_read_pagelist,
 874         .write_pagelist         = filelayout_write_pagelist,
 875 };
 876
 877 static int __init nfs4filelayout_init(void)
 878 {
 879         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
 880                __func__);
 881         return pnfs_register_layoutdriver(&filelayout_type);
 882 }
 883
 884 static void __exit nfs4filelayout_exit(void)
 885 {
 886         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
 887                __func__);
 888         pnfs_unregister_layoutdriver(&filelayout_type);
 889 }
 890
 891 module_init(nfs4filelayout_init);
 892 module_exit(nfs4filelayout_exit);