fs/nfs/filelayout/filelayout.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 <[email protected]>
  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 #include <linux/backing-dev.h>
  36
  37 #include <linux/sunrpc/metrics.h>
  38
  39 #include "../nfs4session.h"
  40 #include "../internal.h"
  41 #include "../delegation.h"
  42 #include "filelayout.h"
  43 #include "../nfs4trace.h"
  44
  45 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  46
  47 MODULE_LICENSE("GPL");
  48 MODULE_AUTHOR("Dean Hildebrand <[email protected]>");
  49 MODULE_DESCRIPTION("The NFSv4 file layout driver");
  50
  51 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
  52 static const struct pnfs_commit_ops filelayout_commit_ops;
  53
  54 static loff_t
  55 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
  56                             loff_t offset)
  57 {
  58         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
  59         u64 stripe_no;
  60         u32 rem;
  61
  62         offset -= flseg->pattern_offset;
  63         stripe_no = div_u64(offset, stripe_width);
  64         div_u64_rem(offset, flseg->stripe_unit, &rem);
  65
  66         return stripe_no * flseg->stripe_unit + rem;
  67 }
  68
  69 /* This function is used by the layout driver to calculate the
  70  * offset of the file on the dserver based on whether the
  71  * layout type is STRIPE_DENSE or STRIPE_SPARSE
  72  */
  73 static loff_t
  74 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
  75 {
  76         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
  77
  78         switch (flseg->stripe_type) {
  79         case STRIPE_SPARSE:
  80                 return offset;
  81
  82         case STRIPE_DENSE:
  83                 return filelayout_get_dense_offset(flseg, offset);
  84         }
  85
  86         BUG();
  87 }
  88
  89 static void filelayout_reset_write(struct nfs_pgio_header *hdr)
  90 {
  91         struct rpc_task *task = &hdr->task;
  92
  93         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
  94                 dprintk("%s Reset task %5u for i/o through MDS "
  95                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
  96                         hdr->task.tk_pid,
  97                         hdr->inode->i_sb->s_id,
  98                         (unsigned long long)NFS_FILEID(hdr->inode),
  99                         hdr->args.count,
 100                         (unsigned long long)hdr->args.offset);
 101
 102                 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
 103         }
 104 }
 105
 106 static void filelayout_reset_read(struct nfs_pgio_header *hdr)
 107 {
 108         struct rpc_task *task = &hdr->task;
 109
 110         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 111                 dprintk("%s Reset task %5u for i/o through MDS "
 112                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
 113                         hdr->task.tk_pid,
 114                         hdr->inode->i_sb->s_id,
 115                         (unsigned long long)NFS_FILEID(hdr->inode),
 116                         hdr->args.count,
 117                         (unsigned long long)hdr->args.offset);
 118
 119                 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
 120         }
 121 }
 122
 123 static int filelayout_async_handle_error(struct rpc_task *task,
 124                                          struct nfs4_state *state,
 125                                          struct nfs_client *clp,
 126                                          struct pnfs_layout_segment *lseg)
 127 {
 128         struct pnfs_layout_hdr *lo = lseg->pls_layout;
 129         struct inode *inode = lo->plh_inode;
 130         struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
 131         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
 132
 133         if (task->tk_status >= 0)
 134                 return 0;
 135
 136         switch (task->tk_status) {
 137         /* DS session errors */
 138         case -NFS4ERR_BADSESSION:
 139         case -NFS4ERR_BADSLOT:
 140         case -NFS4ERR_BAD_HIGH_SLOT:
 141         case -NFS4ERR_DEADSESSION:
 142         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
 143         case -NFS4ERR_SEQ_FALSE_RETRY:
 144         case -NFS4ERR_SEQ_MISORDERED:
 145                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 146                         "flags 0x%x\n", __func__, task->tk_status,
 147                         clp->cl_exchange_flags);
 148                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
 149                 break;
 150         case -NFS4ERR_DELAY:
 151         case -NFS4ERR_GRACE:
 152                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 153                 break;
 154         case -NFS4ERR_RETRY_UNCACHED_REP:
 155                 break;
 156         /* Invalidate Layout errors */
 157         case -NFS4ERR_ACCESS:
 158         case -NFS4ERR_PNFS_NO_LAYOUT:
 159         case -ESTALE:           /* mapped NFS4ERR_STALE */
 160         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
 161         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
 162         case -NFS4ERR_FHEXPIRED:
 163         case -NFS4ERR_WRONG_TYPE:
 164                 dprintk("%s Invalid layout error %d\n", __func__,
 165                         task->tk_status);
 166                 /*
 167                  * Destroy layout so new i/o will get a new layout.
 168                  * Layout will not be destroyed until all current lseg
 169                  * references are put. Mark layout as invalid to resend failed
 170                  * i/o and all i/o waiting on the slot table to the MDS until
 171                  * layout is destroyed and a new valid layout is obtained.
 172                  */
 173                 pnfs_destroy_layout(NFS_I(inode));
 174                 rpc_wake_up(&tbl->slot_tbl_waitq);
 175                 goto reset;
 176         /* RPC connection errors */
 177         case -ECONNREFUSED:
 178         case -EHOSTDOWN:
 179         case -EHOSTUNREACH:
 180         case -ENETUNREACH:
 181         case -EIO:
 182         case -ETIMEDOUT:
 183         case -EPIPE:
 184                 dprintk("%s DS connection error %d\n", __func__,
 185                         task->tk_status);
 186                 nfs4_mark_deviceid_unavailable(devid);
 187                 pnfs_error_mark_layout_for_return(inode, lseg);
 188                 pnfs_set_lo_fail(lseg);
 189                 rpc_wake_up(&tbl->slot_tbl_waitq);
 190                 fallthrough;
 191         default:
 192 reset:
 193                 dprintk("%s Retry through MDS. Error %d\n", __func__,
 194                         task->tk_status);
 195                 return -NFS4ERR_RESET_TO_MDS;
 196         }
 197         task->tk_status = 0;
 198         return -EAGAIN;
 199 }
 200
 201 /* NFS_PROTO call done callback routines */
 202
 203 static int filelayout_read_done_cb(struct rpc_task *task,
 204                                 struct nfs_pgio_header *hdr)
 205 {
 206         int err;
 207
 208         trace_nfs4_pnfs_read(hdr, task->tk_status);
 209         err = filelayout_async_handle_error(task, hdr->args.context->state,
 210                                             hdr->ds_clp, hdr->lseg);
 211
 212         switch (err) {
 213         case -NFS4ERR_RESET_TO_MDS:
 214                 filelayout_reset_read(hdr);
 215                 return task->tk_status;
 216         case -EAGAIN:
 217                 rpc_restart_call_prepare(task);
 218                 return -EAGAIN;
 219         }
 220
 221         return 0;
 222 }
 223
 224 /*
 225  * We reference the rpc_cred of the first WRITE that triggers the need for
 226  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 227  * rfc5661 is not clear about which credential should be used.
 228  */
 229 static void
 230 filelayout_set_layoutcommit(struct nfs_pgio_header *hdr)
 231 {
 232         loff_t end_offs = 0;
 233
 234         if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
 235             hdr->res.verf->committed == NFS_FILE_SYNC)
 236                 return;
 237         if (hdr->res.verf->committed == NFS_DATA_SYNC)
 238                 end_offs = hdr->mds_offset + (loff_t)hdr->res.count;
 239
 240         /* Note: if the write is unstable, don't set end_offs until commit */
 241         pnfs_set_layoutcommit(hdr->inode, hdr->lseg, end_offs);
 242         dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
 243                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
 244 }
 245
 246 bool
 247 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
 248 {
 249         return filelayout_test_devid_invalid(node) ||
 250                 nfs4_test_deviceid_unavailable(node);
 251 }
 252
 253 static bool
 254 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
 255 {
 256         struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
 257
 258         return filelayout_test_devid_unavailable(node);
 259 }
 260
 261 /*
 262  * Call ops for the async read/write cases
 263  * In the case of dense layouts, the offset needs to be reset to its
 264  * original value.
 265  */
 266 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 267 {
 268         struct nfs_pgio_header *hdr = data;
 269
 270         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
 271                 rpc_exit(task, -EIO);
 272                 return;
 273         }
 274         if (filelayout_reset_to_mds(hdr->lseg)) {
 275                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 276                 filelayout_reset_read(hdr);
 277                 rpc_exit(task, 0);
 278                 return;
 279         }
 280         hdr->pgio_done_cb = filelayout_read_done_cb;
 281
 282         if (nfs4_setup_sequence(hdr->ds_clp,
 283                         &hdr->args.seq_args,
 284                         &hdr->res.seq_res,
 285                         task))
 286                 return;
 287         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
 288                         hdr->args.lock_context, FMODE_READ) == -EIO)
 289                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
 290 }
 291
 292 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 293 {
 294         struct nfs_pgio_header *hdr = data;
 295
 296         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
 297             task->tk_status == 0) {
 298                 nfs41_sequence_done(task, &hdr->res.seq_res);
 299                 return;
 300         }
 301
 302         /* Note this may cause RPC to be resent */
 303         hdr->mds_ops->rpc_call_done(task, data);
 304 }
 305
 306 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
 307 {
 308         struct nfs_pgio_header *hdr = data;
 309
 310         rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
 311 }
 312
 313 static int filelayout_write_done_cb(struct rpc_task *task,
 314                                 struct nfs_pgio_header *hdr)
 315 {
 316         int err;
 317
 318         trace_nfs4_pnfs_write(hdr, task->tk_status);
 319         err = filelayout_async_handle_error(task, hdr->args.context->state,
 320                                             hdr->ds_clp, hdr->lseg);
 321
 322         switch (err) {
 323         case -NFS4ERR_RESET_TO_MDS:
 324                 filelayout_reset_write(hdr);
 325                 return task->tk_status;
 326         case -EAGAIN:
 327                 rpc_restart_call_prepare(task);
 328                 return -EAGAIN;
 329         }
 330
 331         filelayout_set_layoutcommit(hdr);
 332
 333         /* zero out the fattr */
 334         hdr->fattr.valid = 0;
 335         if (task->tk_status >= 0)
 336                 nfs_writeback_update_inode(hdr);
 337
 338         return 0;
 339 }
 340
 341 static int filelayout_commit_done_cb(struct rpc_task *task,
 342                                      struct nfs_commit_data *data)
 343 {
 344         int err;
 345
 346         trace_nfs4_pnfs_commit_ds(data, task->tk_status);
 347         err = filelayout_async_handle_error(task, NULL, data->ds_clp,
 348                                             data->lseg);
 349
 350         switch (err) {
 351         case -NFS4ERR_RESET_TO_MDS:
 352                 pnfs_generic_prepare_to_resend_writes(data);
 353                 return -EAGAIN;
 354         case -EAGAIN:
 355                 rpc_restart_call_prepare(task);
 356                 return -EAGAIN;
 357         }
 358
 359         pnfs_set_layoutcommit(data->inode, data->lseg, data->lwb);
 360
 361         return 0;
 362 }
 363
 364 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 365 {
 366         struct nfs_pgio_header *hdr = data;
 367
 368         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
 369                 rpc_exit(task, -EIO);
 370                 return;
 371         }
 372         if (filelayout_reset_to_mds(hdr->lseg)) {
 373                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 374                 filelayout_reset_write(hdr);
 375                 rpc_exit(task, 0);
 376                 return;
 377         }
 378         if (nfs4_setup_sequence(hdr->ds_clp,
 379                         &hdr->args.seq_args,
 380                         &hdr->res.seq_res,
 381                         task))
 382                 return;
 383         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
 384                         hdr->args.lock_context, FMODE_WRITE) == -EIO)
 385                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
 386 }
 387
 388 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 389 {
 390         struct nfs_pgio_header *hdr = data;
 391
 392         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
 393             task->tk_status == 0) {
 394                 nfs41_sequence_done(task, &hdr->res.seq_res);
 395                 return;
 396         }
 397
 398         /* Note this may cause RPC to be resent */
 399         hdr->mds_ops->rpc_call_done(task, data);
 400 }
 401
 402 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
 403 {
 404         struct nfs_pgio_header *hdr = data;
 405
 406         rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
 407 }
 408
 409 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
 410 {
 411         struct nfs_commit_data *wdata = data;
 412
 413         nfs4_setup_sequence(wdata->ds_clp,
 414                         &wdata->args.seq_args,
 415                         &wdata->res.seq_res,
 416                         task);
 417 }
 418
 419 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
 420 {
 421         struct nfs_commit_data *cdata = data;
 422
 423         rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
 424 }
 425
 426 static const struct rpc_call_ops filelayout_read_call_ops = {
 427         .rpc_call_prepare = filelayout_read_prepare,
 428         .rpc_call_done = filelayout_read_call_done,
 429         .rpc_count_stats = filelayout_read_count_stats,
 430         .rpc_release = pnfs_generic_rw_release,
 431 };
 432
 433 static const struct rpc_call_ops filelayout_write_call_ops = {
 434         .rpc_call_prepare = filelayout_write_prepare,
 435         .rpc_call_done = filelayout_write_call_done,
 436         .rpc_count_stats = filelayout_write_count_stats,
 437         .rpc_release = pnfs_generic_rw_release,
 438 };
 439
 440 static const struct rpc_call_ops filelayout_commit_call_ops = {
 441         .rpc_call_prepare = filelayout_commit_prepare,
 442         .rpc_call_done = pnfs_generic_write_commit_done,
 443         .rpc_count_stats = filelayout_commit_count_stats,
 444         .rpc_release = pnfs_generic_commit_release,
 445 };
 446
 447 static enum pnfs_try_status
 448 filelayout_read_pagelist(struct nfs_pgio_header *hdr)
 449 {
 450         struct pnfs_layout_segment *lseg = hdr->lseg;
 451         struct nfs4_pnfs_ds *ds;
 452         struct rpc_clnt *ds_clnt;
 453         loff_t offset = hdr->args.offset;
 454         u32 j, idx;
 455         struct nfs_fh *fh;
 456
 457         dprintk("--> %s ino %lu pgbase %u req %zu@%llu\n",
 458                 __func__, hdr->inode->i_ino,
 459                 hdr->args.pgbase, (size_t)hdr->args.count, offset);
 460
 461         /* Retrieve the correct rpc_client for the byte range */
 462         j = nfs4_fl_calc_j_index(lseg, offset);
 463         idx = nfs4_fl_calc_ds_index(lseg, j);
 464         ds = nfs4_fl_prepare_ds(lseg, idx);
 465         if (!ds)
 466                 return PNFS_NOT_ATTEMPTED;
 467
 468         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
 469         if (IS_ERR(ds_clnt))
 470                 return PNFS_NOT_ATTEMPTED;
 471
 472         dprintk("%s USE DS: %s cl_count %d\n", __func__,
 473                 ds->ds_remotestr, refcount_read(&ds->ds_clp->cl_count));
 474
 475         /* No multipath support. Use first DS */
 476         refcount_inc(&ds->ds_clp->cl_count);
 477         hdr->ds_clp = ds->ds_clp;
 478         hdr->ds_commit_idx = idx;
 479         fh = nfs4_fl_select_ds_fh(lseg, j);
 480         if (fh)
 481                 hdr->args.fh = fh;
 482
 483         hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
 484         hdr->mds_offset = offset;
 485
 486         /* Perform an asynchronous read to ds */
 487         nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
 488                           NFS_PROTO(hdr->inode), &filelayout_read_call_ops,
 489                           0, RPC_TASK_SOFTCONN);
 490         return PNFS_ATTEMPTED;
 491 }
 492
 493 /* Perform async writes. */
 494 static enum pnfs_try_status
 495 filelayout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
 496 {
 497         struct pnfs_layout_segment *lseg = hdr->lseg;
 498         struct nfs4_pnfs_ds *ds;
 499         struct rpc_clnt *ds_clnt;
 500         loff_t offset = hdr->args.offset;
 501         u32 j, idx;
 502         struct nfs_fh *fh;
 503
 504         /* Retrieve the correct rpc_client for the byte range */
 505         j = nfs4_fl_calc_j_index(lseg, offset);
 506         idx = nfs4_fl_calc_ds_index(lseg, j);
 507         ds = nfs4_fl_prepare_ds(lseg, idx);
 508         if (!ds)
 509                 return PNFS_NOT_ATTEMPTED;
 510
 511         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
 512         if (IS_ERR(ds_clnt))
 513                 return PNFS_NOT_ATTEMPTED;
 514
 515         dprintk("%s ino %lu sync %d req %zu@%llu DS: %s cl_count %d\n",
 516                 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
 517                 offset, ds->ds_remotestr, refcount_read(&ds->ds_clp->cl_count));
 518
 519         hdr->pgio_done_cb = filelayout_write_done_cb;
 520         refcount_inc(&ds->ds_clp->cl_count);
 521         hdr->ds_clp = ds->ds_clp;
 522         hdr->ds_commit_idx = idx;
 523         fh = nfs4_fl_select_ds_fh(lseg, j);
 524         if (fh)
 525                 hdr->args.fh = fh;
 526         hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
 527
 528         /* Perform an asynchronous write */
 529         nfs_initiate_pgio(ds_clnt, hdr, hdr->cred,
 530                           NFS_PROTO(hdr->inode), &filelayout_write_call_ops,
 531                           sync, RPC_TASK_SOFTCONN);
 532         return PNFS_ATTEMPTED;
 533 }
 534
 535 static int
 536 filelayout_check_deviceid(struct pnfs_layout_hdr *lo,
 537                           struct nfs4_filelayout_segment *fl,
 538                           gfp_t gfp_flags)
 539 {
 540         struct nfs4_deviceid_node *d;
 541         struct nfs4_file_layout_dsaddr *dsaddr;
 542         int status = -EINVAL;
 543
 544         /* Is the deviceid already set? If so, we're good. */
 545         if (fl->dsaddr != NULL)
 546                 return 0;
 547
 548         /* find and reference the deviceid */
 549         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode), &fl->deviceid,
 550                         lo->plh_lc_cred, gfp_flags);
 551         if (d == NULL)
 552                 goto out;
 553
 554         dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 555         /* Found deviceid is unavailable */
 556         if (filelayout_test_devid_unavailable(&dsaddr->id_node))
 557                 goto out_put;
 558
 559         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 560                 dprintk("%s Bad first_stripe_index %u\n",
 561                                 __func__, fl->first_stripe_index);
 562                 goto out_put;
 563         }
 564
 565         if ((fl->stripe_type == STRIPE_SPARSE &&
 566             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 567             (fl->stripe_type == STRIPE_DENSE &&
 568             fl->num_fh != dsaddr->stripe_count)) {
 569                 dprintk("%s num_fh %u not valid for given packing\n",
 570                         __func__, fl->num_fh);
 571                 goto out_put;
 572         }
 573         status = 0;
 574
 575         /*
 576          * Atomic compare and xchange to ensure we don't scribble
 577          * over a non-NULL pointer.
 578          */
 579         if (cmpxchg(&fl->dsaddr, NULL, dsaddr) != NULL)
 580                 goto out_put;
 581 out:
 582         return status;
 583 out_put:
 584         nfs4_fl_put_deviceid(dsaddr);
 585         goto out;
 586 }
 587
 588 /*
 589  * filelayout_check_layout()
 590  *
 591  * Make sure layout segment parameters are sane WRT the device.
 592  * At this point no generic layer initialization of the lseg has occurred,
 593  * and nothing has been added to the layout_hdr cache.
 594  *
 595  */
 596 static int
 597 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 598                         struct nfs4_filelayout_segment *fl,
 599                         struct nfs4_layoutget_res *lgr,
 600                         gfp_t gfp_flags)
 601 {
 602         int status = -EINVAL;
 603
 604         dprintk("--> %s\n", __func__);
 605
 606         /* FIXME: remove this check when layout segment support is added */
 607         if (lgr->range.offset != 0 ||
 608             lgr->range.length != NFS4_MAX_UINT64) {
 609                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 610                         __func__);
 611                 goto out;
 612         }
 613
 614         if (fl->pattern_offset > lgr->range.offset) {
 615                 dprintk("%s pattern_offset %lld too large\n",
 616                                 __func__, fl->pattern_offset);
 617                 goto out;
 618         }
 619
 620         if (!fl->stripe_unit) {
 621                 dprintk("%s Invalid stripe unit (%u)\n",
 622                         __func__, fl->stripe_unit);
 623                 goto out;
 624         }
 625
 626         status = 0;
 627 out:
 628         dprintk("--> %s returns %d\n", __func__, status);
 629         return status;
 630 }
 631
 632 static void _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 633 {
 634         int i;
 635
 636         if (fl->fh_array) {
 637                 for (i = 0; i < fl->num_fh; i++) {
 638                         if (!fl->fh_array[i])
 639                                 break;
 640                         kfree(fl->fh_array[i]);
 641                 }
 642                 kfree(fl->fh_array);
 643         }
 644         kfree(fl);
 645 }
 646
 647 static int
 648 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 649                          struct nfs4_filelayout_segment *fl,
 650                          struct nfs4_layoutget_res *lgr,
 651                          gfp_t gfp_flags)
 652 {
 653         struct xdr_stream stream;
 654         struct xdr_buf buf;
 655         struct page *scratch;
 656         __be32 *p;
 657         uint32_t nfl_util;
 658         int i;
 659
 660         dprintk("%s: set_layout_map Begin\n", __func__);
 661
 662         scratch = alloc_page(gfp_flags);
 663         if (!scratch)
 664                 return -ENOMEM;
 665
 666         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 667         xdr_set_scratch_page(&stream, scratch);
 668
 669         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 670          * num_fh (4) */
 671         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 672         if (unlikely(!p))
 673                 goto out_err;
 674
 675         memcpy(&fl->deviceid, p, sizeof(fl->deviceid));
 676         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 677         nfs4_print_deviceid(&fl->deviceid);
 678
 679         nfl_util = be32_to_cpup(p++);
 680         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 681                 fl->commit_through_mds = 1;
 682         if (nfl_util & NFL4_UFLG_DENSE)
 683                 fl->stripe_type = STRIPE_DENSE;
 684         else
 685                 fl->stripe_type = STRIPE_SPARSE;
 686         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 687
 688         fl->first_stripe_index = be32_to_cpup(p++);
 689         p = xdr_decode_hyper(p, &fl->pattern_offset);
 690         fl->num_fh = be32_to_cpup(p++);
 691
 692         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 693                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 694                 fl->pattern_offset);
 695
 696         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 697          * Futher checking is done in filelayout_check_layout */
 698         if (fl->num_fh >
 699             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 700                 goto out_err;
 701
 702         if (fl->num_fh > 0) {
 703                 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
 704                                        gfp_flags);
 705                 if (!fl->fh_array)
 706                         goto out_err;
 707         }
 708
 709         for (i = 0; i < fl->num_fh; i++) {
 710                 /* Do we want to use a mempool here? */
 711                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 712                 if (!fl->fh_array[i])
 713                         goto out_err;
 714
 715                 p = xdr_inline_decode(&stream, 4);
 716                 if (unlikely(!p))
 717                         goto out_err;
 718                 fl->fh_array[i]->size = be32_to_cpup(p++);
 719                 if (fl->fh_array[i]->size > NFS_MAXFHSIZE) {
 720                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
 721                                i, fl->fh_array[i]->size);
 722                         goto out_err;
 723                 }
 724
 725                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 726                 if (unlikely(!p))
 727                         goto out_err;
 728                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 729                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 730                         fl->fh_array[i]->size);
 731         }
 732
 733         __free_page(scratch);
 734         return 0;
 735
 736 out_err:
 737         __free_page(scratch);
 738         return -EIO;
 739 }
 740
 741 static void
 742 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 743 {
 744         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 745
 746         dprintk("--> %s\n", __func__);
 747         if (fl->dsaddr != NULL)
 748                 nfs4_fl_put_deviceid(fl->dsaddr);
 749         /* This assumes a single RW lseg */
 750         if (lseg->pls_range.iomode == IOMODE_RW) {
 751                 struct nfs4_filelayout *flo;
 752                 struct inode *inode;
 753
 754                 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
 755                 inode = flo->generic_hdr.plh_inode;
 756                 spin_lock(&inode->i_lock);
 757                 pnfs_generic_ds_cinfo_release_lseg(&flo->commit_info, lseg);
 758                 spin_unlock(&inode->i_lock);
 759         }
 760         _filelayout_free_lseg(fl);
 761 }
 762
 763 static struct pnfs_layout_segment *
 764 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 765                       struct nfs4_layoutget_res *lgr,
 766                       gfp_t gfp_flags)
 767 {
 768         struct nfs4_filelayout_segment *fl;
 769         int rc;
 770
 771         dprintk("--> %s\n", __func__);
 772         fl = kzalloc(sizeof(*fl), gfp_flags);
 773         if (!fl)
 774                 return NULL;
 775
 776         rc = filelayout_decode_layout(layoutid, fl, lgr, gfp_flags);
 777         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, gfp_flags)) {
 778                 _filelayout_free_lseg(fl);
 779                 return NULL;
 780         }
 781         return &fl->generic_hdr;
 782 }
 783
 784 /*
 785  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 786  *
 787  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
 788  * of bytes (maximum @req->wb_bytes) that can be coalesced.
 789  */
 790 static size_t
 791 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 792                    struct nfs_page *req)
 793 {
 794         unsigned int size;
 795         u64 p_stripe, r_stripe;
 796         u32 stripe_offset;
 797         u64 segment_offset = pgio->pg_lseg->pls_range.offset;
 798         u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 799
 800         /* calls nfs_generic_pg_test */
 801         size = pnfs_generic_pg_test(pgio, prev, req);
 802         if (!size)
 803                 return 0;
 804
 805         /* see if req and prev are in the same stripe */
 806         if (prev) {
 807                 p_stripe = (u64)req_offset(prev) - segment_offset;
 808                 r_stripe = (u64)req_offset(req) - segment_offset;
 809                 do_div(p_stripe, stripe_unit);
 810                 do_div(r_stripe, stripe_unit);
 811
 812                 if (p_stripe != r_stripe)
 813                         return 0;
 814         }
 815
 816         /* calculate remaining bytes in the current stripe */
 817         div_u64_rem((u64)req_offset(req) - segment_offset,
 818                         stripe_unit,
 819                         &stripe_offset);
 820         WARN_ON_ONCE(stripe_offset > stripe_unit);
 821         if (stripe_offset >= stripe_unit)
 822                 return 0;
 823         return min(stripe_unit - (unsigned int)stripe_offset, size);
 824 }
 825
 826 static struct pnfs_layout_segment *
 827 fl_pnfs_update_layout(struct inode *ino,
 828                       struct nfs_open_context *ctx,
 829                       loff_t pos,
 830                       u64 count,
 831                       enum pnfs_iomode iomode,
 832                       bool strict_iomode,
 833                       gfp_t gfp_flags)
 834 {
 835         struct pnfs_layout_segment *lseg = NULL;
 836         struct pnfs_layout_hdr *lo;
 837         struct nfs4_filelayout_segment *fl;
 838         int status;
 839
 840         lseg = pnfs_update_layout(ino, ctx, pos, count, iomode, strict_iomode,
 841                                   gfp_flags);
 842         if (IS_ERR_OR_NULL(lseg))
 843                 goto out;
 844
 845         lo = NFS_I(ino)->layout;
 846         fl = FILELAYOUT_LSEG(lseg);
 847
 848         status = filelayout_check_deviceid(lo, fl, gfp_flags);
 849         if (status) {
 850                 pnfs_put_lseg(lseg);
 851                 lseg = NULL;
 852         }
 853 out:
 854         return lseg;
 855 }
 856
 857 static void
 858 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 859                         struct nfs_page *req)
 860 {
 861         pnfs_generic_pg_check_layout(pgio);
 862         if (!pgio->pg_lseg) {
 863                 pgio->pg_lseg = fl_pnfs_update_layout(pgio->pg_inode,
 864                                                       nfs_req_openctx(req),
 865                                                       0,
 866                                                       NFS4_MAX_UINT64,
 867                                                       IOMODE_READ,
 868                                                       false,
 869                                                       GFP_KERNEL);
 870                 if (IS_ERR(pgio->pg_lseg)) {
 871                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 872                         pgio->pg_lseg = NULL;
 873                         return;
 874                 }
 875         }
 876         /* If no lseg, fall back to read through mds */
 877         if (pgio->pg_lseg == NULL)
 878                 nfs_pageio_reset_read_mds(pgio);
 879 }
 880
 881 static void
 882 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 883                          struct nfs_page *req)
 884 {
 885         pnfs_generic_pg_check_layout(pgio);
 886         if (!pgio->pg_lseg) {
 887                 pgio->pg_lseg = fl_pnfs_update_layout(pgio->pg_inode,
 888                                                       nfs_req_openctx(req),
 889                                                       0,
 890                                                       NFS4_MAX_UINT64,
 891                                                       IOMODE_RW,
 892                                                       false,
 893                                                       GFP_NOFS);
 894                 if (IS_ERR(pgio->pg_lseg)) {
 895                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
 896                         pgio->pg_lseg = NULL;
 897                         return;
 898                 }
 899         }
 900
 901         /* If no lseg, fall back to write through mds */
 902         if (pgio->pg_lseg == NULL)
 903                 nfs_pageio_reset_write_mds(pgio);
 904 }
 905
 906 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 907         .pg_init = filelayout_pg_init_read,
 908         .pg_test = filelayout_pg_test,
 909         .pg_doio = pnfs_generic_pg_readpages,
 910         .pg_cleanup = pnfs_generic_pg_cleanup,
 911 };
 912
 913 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 914         .pg_init = filelayout_pg_init_write,
 915         .pg_test = filelayout_pg_test,
 916         .pg_doio = pnfs_generic_pg_writepages,
 917         .pg_cleanup = pnfs_generic_pg_cleanup,
 918 };
 919
 920 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 921 {
 922         if (fl->stripe_type == STRIPE_SPARSE)
 923                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
 924         else
 925                 return j;
 926 }
 927
 928 static void
 929 filelayout_mark_request_commit(struct nfs_page *req,
 930                                struct pnfs_layout_segment *lseg,
 931                                struct nfs_commit_info *cinfo,
 932                                u32 ds_commit_idx)
 933
 934 {
 935         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 936         u32 i, j;
 937
 938         if (fl->commit_through_mds) {
 939                 nfs_request_add_commit_list(req, cinfo);
 940         } else {
 941                 /* Note that we are calling nfs4_fl_calc_j_index on each page
 942                  * that ends up being committed to a data server.  An attractive
 943                  * alternative is to add a field to nfs_write_data and nfs_page
 944                  * to store the value calculated in filelayout_write_pagelist
 945                  * and just use that here.
 946                  */
 947                 j = nfs4_fl_calc_j_index(lseg, req_offset(req));
 948                 i = select_bucket_index(fl, j);
 949                 pnfs_layout_mark_request_commit(req, lseg, cinfo, i);
 950         }
 951 }
 952
 953 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 954 {
 955         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 956
 957         if (flseg->stripe_type == STRIPE_SPARSE)
 958                 return i;
 959         else
 960                 return nfs4_fl_calc_ds_index(lseg, i);
 961 }
 962
 963 static struct nfs_fh *
 964 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
 965 {
 966         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
 967
 968         if (flseg->stripe_type == STRIPE_SPARSE) {
 969                 if (flseg->num_fh == 1)
 970                         i = 0;
 971                 else if (flseg->num_fh == 0)
 972                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
 973                         return NULL;
 974         }
 975         return flseg->fh_array[i];
 976 }
 977
 978 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
 979 {
 980         struct pnfs_layout_segment *lseg = data->lseg;
 981         struct nfs4_pnfs_ds *ds;
 982         struct rpc_clnt *ds_clnt;
 983         u32 idx;
 984         struct nfs_fh *fh;
 985
 986         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
 987         ds = nfs4_fl_prepare_ds(lseg, idx);
 988         if (!ds)
 989                 goto out_err;
 990
 991         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
 992         if (IS_ERR(ds_clnt))
 993                 goto out_err;
 994
 995         dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
 996                 data->inode->i_ino, how, refcount_read(&ds->ds_clp->cl_count));
 997         data->commit_done_cb = filelayout_commit_done_cb;
 998         refcount_inc(&ds->ds_clp->cl_count);
 999         data->ds_clp = ds->ds_clp;
1000         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1001         if (fh)
1002                 data->args.fh = fh;
1003         return nfs_initiate_commit(ds_clnt, data, NFS_PROTO(data->inode),
1004                                    &filelayout_commit_call_ops, how,
1005                                    RPC_TASK_SOFTCONN);
1006 out_err:
1007         pnfs_generic_prepare_to_resend_writes(data);
1008         pnfs_generic_commit_release(data);
1009         return -EAGAIN;
1010 }
1011
1012 static int
1013 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1014                            int how, struct nfs_commit_info *cinfo)
1015 {
1016         return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1017                                             filelayout_initiate_commit);
1018 }
1019
1020 static struct nfs4_deviceid_node *
1021 filelayout_alloc_deviceid_node(struct nfs_server *server,
1022                 struct pnfs_device *pdev, gfp_t gfp_flags)
1023 {
1024         struct nfs4_file_layout_dsaddr *dsaddr;
1025
1026         dsaddr = nfs4_fl_alloc_deviceid_node(server, pdev, gfp_flags);
1027         if (!dsaddr)
1028                 return NULL;
1029         return &dsaddr->id_node;
1030 }
1031
1032 static void
1033 filelayout_free_deviceid_node(struct nfs4_deviceid_node *d)
1034 {
1035         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1036 }
1037
1038 static struct pnfs_layout_hdr *
1039 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1040 {
1041         struct nfs4_filelayout *flo;
1042
1043         flo = kzalloc(sizeof(*flo), gfp_flags);
1044         if (flo == NULL)
1045                 return NULL;
1046         pnfs_init_ds_commit_info(&flo->commit_info);
1047         flo->commit_info.ops = &filelayout_commit_ops;
1048         return &flo->generic_hdr;
1049 }
1050
1051 static void
1052 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1053 {
1054         kfree_rcu(FILELAYOUT_FROM_HDR(lo), generic_hdr.plh_rcu);
1055 }
1056
1057 static struct pnfs_ds_commit_info *
1058 filelayout_get_ds_info(struct inode *inode)
1059 {
1060         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1061
1062         if (layout == NULL)
1063                 return NULL;
1064         else
1065                 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1066 }
1067
1068 static void
1069 filelayout_setup_ds_info(struct pnfs_ds_commit_info *fl_cinfo,
1070                 struct pnfs_layout_segment *lseg)
1071 {
1072         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1073         struct inode *inode = lseg->pls_layout->plh_inode;
1074         struct pnfs_commit_array *array, *new;
1075         unsigned int size = (fl->stripe_type == STRIPE_SPARSE) ?
1076                 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
1077
1078         new = pnfs_alloc_commit_array(size, nfs_io_gfp_mask());
1079         if (new) {
1080                 spin_lock(&inode->i_lock);
1081                 array = pnfs_add_commit_array(fl_cinfo, new, lseg);
1082                 spin_unlock(&inode->i_lock);
1083                 if (array != new)
1084                         pnfs_free_commit_array(new);
1085         }
1086 }
1087
1088 static void
1089 filelayout_release_ds_info(struct pnfs_ds_commit_info *fl_cinfo,
1090                 struct inode *inode)
1091 {
1092         spin_lock(&inode->i_lock);
1093         pnfs_generic_ds_cinfo_destroy(fl_cinfo);
1094         spin_unlock(&inode->i_lock);
1095 }
1096
1097 static const struct pnfs_commit_ops filelayout_commit_ops = {
1098         .setup_ds_info          = filelayout_setup_ds_info,
1099         .release_ds_info        = filelayout_release_ds_info,
1100         .mark_request_commit    = filelayout_mark_request_commit,
1101         .clear_request_commit   = pnfs_generic_clear_request_commit,
1102         .scan_commit_lists      = pnfs_generic_scan_commit_lists,
1103         .recover_commit_reqs    = pnfs_generic_recover_commit_reqs,
1104         .search_commit_reqs     = pnfs_generic_search_commit_reqs,
1105         .commit_pagelist        = filelayout_commit_pagelist,
1106 };
1107
1108 static struct pnfs_layoutdriver_type filelayout_type = {
1109         .id                     = LAYOUT_NFSV4_1_FILES,
1110         .name                   = "LAYOUT_NFSV4_1_FILES",
1111         .owner                  = THIS_MODULE,
1112         .flags                  = PNFS_LAYOUTGET_ON_OPEN,
1113         .max_layoutget_response = 4096, /* 1 page or so... */
1114         .alloc_layout_hdr       = filelayout_alloc_layout_hdr,
1115         .free_layout_hdr        = filelayout_free_layout_hdr,
1116         .alloc_lseg             = filelayout_alloc_lseg,
1117         .free_lseg              = filelayout_free_lseg,
1118         .pg_read_ops            = &filelayout_pg_read_ops,
1119         .pg_write_ops           = &filelayout_pg_write_ops,
1120         .get_ds_info            = &filelayout_get_ds_info,
1121         .read_pagelist          = filelayout_read_pagelist,
1122         .write_pagelist         = filelayout_write_pagelist,
1123         .alloc_deviceid_node    = filelayout_alloc_deviceid_node,
1124         .free_deviceid_node     = filelayout_free_deviceid_node,
1125         .sync                   = pnfs_nfs_generic_sync,
1126 };
1127
1128 static int __init nfs4filelayout_init(void)
1129 {
1130         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1131                __func__);
1132         return pnfs_register_layoutdriver(&filelayout_type);
1133 }
1134
1135 static void __exit nfs4filelayout_exit(void)
1136 {
1137         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1138                __func__);
1139         pnfs_unregister_layoutdriver(&filelayout_type);
1140 }
1141
1142 MODULE_ALIAS("nfs-layouttype4-1");
1143
1144 module_init(nfs4filelayout_init);
1145 module_exit(nfs4filelayout_exit);