lustre/ptlrpc/nrs_orr.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License version 2 for more details.  A copy is
  14  * included in the COPYING file that accompanied this code.
  15
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2013, Intel Corporation.
  24  *
  25  * Copyright 2012 Xyratex Technology Limited
  26  */
  27 /*
  28  * lustre/ptlrpc/nrs_orr.c
  29  *
  30  * Network Request Scheduler (NRS) ORR and TRR policies
  31  *
  32  * Request scheduling in a Round-Robin manner over backend-fs objects and OSTs
  33  * respectively
  34  *
  35  * Author: Liang Zhen <liang@whamcloud.com>
  36  * Author: Nikitas Angelinas <nikitas_angelinas@xyratex.com>
  37  */
  38 #ifdef HAVE_SERVER_SUPPORT
  39
  40 /**
  41  * \addtogoup nrs
  42  * @{
  43  */
  44 #define DEBUG_SUBSYSTEM S_RPC
  45 #include <obd_support.h>
  46 #include <obd_class.h>
  47 #include <lustre_net.h>
  48 #include <lustre/lustre_idl.h>
  49 #include <lustre_req_layout.h>
  50 #include "ptlrpc_internal.h"
  51
  52 /**
  53  * \name ORR/TRR policy
  54  *
  55  * ORR/TRR (Object-based Round Robin/Target-based Round Robin) NRS policies
  56  *
  57  * ORR performs batched Round Robin shceduling of brw RPCs, based on the FID of
  58  * the backend-fs object that the brw RPC pertains to; the TRR policy performs
  59  * batched Round Robin scheduling of brw RPCs, based on the OST index that the
  60  * RPC pertains to. Both policies also order RPCs in each batch in ascending
  61  * offset order, which is lprocfs-tunable between logical file offsets, and
  62  * physical disk offsets, as reported by fiemap.
  63  *
  64  * The TRR policy reuses much of the functionality of ORR. These two scheduling
  65  * algorithms could alternatively be implemented under a single NRS policy, that
  66  * uses an lprocfs tunable in order to switch between the two types of
  67  * scheduling behaviour. The two algorithms have been implemented as separate
  68  * policies for reasons of clarity to the user, and to avoid issues that would
  69  * otherwise arise at the point of switching between behaviours in the case of
  70  * having a single policy, such as resource cleanup for nrs_orr_object
  71  * instances. It is possible that this may need to be re-examined in the future,
  72  * along with potentially coalescing other policies that perform batched request
  73  * scheduling in a Round-Robin manner, all into one policy.
  74  *
  75  * @{
  76  */
  77
  78 #define NRS_POL_NAME_ORR        "orr"
  79 #define NRS_POL_NAME_TRR        "trr"
  80
  81 /**
  82  * Checks if the RPC type of \a nrq is currently handled by an ORR/TRR policy
  83  *
  84  * \param[in]  orrd   the ORR/TRR policy scheduler instance
  85  * \param[in]  nrq    the request
  86  * \param[out] opcode the opcode is saved here, just in order to avoid calling
  87  *                    lustre_msg_get_opc() again later
  88  *
  89  * \retval true  request type is supported by the policy instance
  90  * \retval false request type is not supported by the policy instance
  91  */
  92 static bool nrs_orr_req_supported(struct nrs_orr_data *orrd,
  93                                   struct ptlrpc_nrs_request *nrq, __u32 *opcode)
  94 {
  95         struct ptlrpc_request  *req = container_of(nrq, struct ptlrpc_request,
  96                                                    rq_nrq);
  97         __u32                   opc = lustre_msg_get_opc(req->rq_reqmsg);
  98         bool                    rc = false;
  99
 100         /**
 101          * XXX: nrs_orr_data::od_supp accessed unlocked.
 102          */
 103         switch (opc) {
 104         case OST_READ:
 105                 rc = orrd->od_supp & NOS_OST_READ;
 106                 break;
 107         case OST_WRITE:
 108                 rc = orrd->od_supp & NOS_OST_WRITE;
 109                 break;
 110         }
 111
 112         if (rc)
 113                 *opcode = opc;
 114
 115         return rc;
 116 }
 117
 118 /**
 119  * Returns the ORR/TRR key fields for the request \a nrq in \a key.
 120  *
 121  * \param[in]  orrd the ORR/TRR policy scheduler instance
 122  * \param[in]  nrq  the request
 123  * \param[in]  opc  the request's opcode
 124  * \param[in]  name the policy name
 125  * \param[out] key  fields of the key are returned here.
 126  *
 127  * \retval 0   key filled successfully
 128  * \retval < 0 error
 129  */
 130 static int nrs_orr_key_fill(struct nrs_orr_data *orrd,
 131                             struct ptlrpc_nrs_request *nrq, __u32 opc,
 132                             char *name, struct nrs_orr_key *key)
 133 {
 134         struct ptlrpc_request  *req = container_of(nrq, struct ptlrpc_request,
 135                                                    rq_nrq);
 136         struct ost_body        *body;
 137         __u32                   ost_idx;
 138         bool                    is_orr = strncmp(name, NRS_POL_NAME_ORR,
 139                                                  NRS_POL_NAME_MAX) == 0;
 140
 141         LASSERT(req != NULL);
 142
 143         /**
 144          * This is an attempt to fill in the request key fields while
 145          * moving a request from the regular to the high-priority NRS
 146          * head (via ldlm_lock_reorder_req()), but the request key has
 147          * been adequately filled when nrs_orr_res_get() was called through
 148          * ptlrpc_nrs_req_initialize() for the regular NRS head's ORR/TRR
 149          * policy, so there is nothing to do.
 150          */
 151         if ((is_orr && nrq->nr_u.orr.or_orr_set) ||
 152             (!is_orr && nrq->nr_u.orr.or_trr_set)) {
 153                 *key = nrq->nr_u.orr.or_key;
 154                 return 0;
 155         }
 156
 157         if (nrq->nr_u.orr.or_orr_set || nrq->nr_u.orr.or_trr_set)
 158                 memset(&nrq->nr_u.orr.or_key, 0, sizeof(nrq->nr_u.orr.or_key));
 159
 160         ost_idx = class_server_data(req->rq_export->exp_obd)->lsd_osd_index;
 161
 162         if (is_orr) {
 163                 int     rc;
 164                 /**
 165                  * The request pill for OST_READ and OST_WRITE requests is
 166                  * initialized in the ost_io service's
 167                  * ptlrpc_service_ops::so_hpreq_handler, ost_io_hpreq_handler(),
 168                  * so no need to redo it here.
 169                  */
 170                 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 171                 if (body == NULL)
 172                         RETURN(-EFAULT);
 173
 174                 rc = ostid_to_fid(&key->ok_fid, &body->oa.o_oi, ost_idx);
 175                 if (rc < 0)
 176                         return rc;
 177
 178                 nrq->nr_u.orr.or_orr_set = 1;
 179         } else {
 180                 key->ok_idx = ost_idx;
 181                 nrq->nr_u.orr.or_trr_set = 1;
 182         }
 183
 184         return 0;
 185 }
 186
 187 /**
 188  * Populates the range values in \a range with logical offsets obtained via
 189  * \a nb.
 190  *
 191  * \param[in]  nb       niobuf_remote struct array for this request
 192  * \param[in]  niocount count of niobuf_remote structs for this request
 193  * \param[out] range    the offset range is returned here
 194  */
 195 static void nrs_orr_range_fill_logical(struct niobuf_remote *nb, int niocount,
 196                                        struct nrs_orr_req_range *range)
 197 {
 198         /* Should we do this at page boundaries ? */
 199         range->or_start = nb[0].offset & CFS_PAGE_MASK;
 200         range->or_end = (nb[niocount - 1].offset +
 201                          nb[niocount - 1].len - 1) | ~CFS_PAGE_MASK;
 202 }
 203
 204 /**
 205  * We obtain information just for a single extent, as the request can only be in
 206  * a single place in the binary heap anyway.
 207  */
 208 #define ORR_NUM_EXTENTS 1
 209
 210 /**
 211  * Converts the logical file offset range in \a range, to a physical disk offset
 212  * range in \a range, for a request. Uses obd_get_info() in order to carry out a
 213  * fiemap call and obtain backend-fs extent information. The returned range is
 214  * in physical block numbers.
 215  *
 216  * \param[in]     nrq   the request
 217  * \param[in]     oa    obdo struct for this request
 218  * \param[in,out] range the offset range in bytes; logical range in, physical
 219  *                      range out
 220  *
 221  * \retval 0    physical offsets obtained successfully
 222  * \retvall < 0 error
 223  */
 224 static int nrs_orr_range_fill_physical(struct ptlrpc_nrs_request *nrq,
 225                                        struct obdo *oa,
 226                                        struct nrs_orr_req_range *range)
 227 {
 228         struct ptlrpc_request     *req = container_of(nrq,
 229                                                       struct ptlrpc_request,
 230                                                       rq_nrq);
 231         char                       fiemap_buf[offsetof(struct ll_user_fiemap,
 232                                                   fm_extents[ORR_NUM_EXTENTS])];
 233         struct ll_user_fiemap     *fiemap = (struct ll_user_fiemap *)fiemap_buf;
 234         struct ll_fiemap_info_key  key;
 235         loff_t                     start;
 236         loff_t                     end;
 237         int                        rc;
 238
 239         key = (typeof(key)) {
 240                 .name = KEY_FIEMAP,
 241                 .oa = *oa,
 242                 .fiemap = {
 243                         .fm_start = range->or_start,
 244                         .fm_length = range->or_end - range->or_start,
 245                         .fm_extent_count = ORR_NUM_EXTENTS
 246                 }
 247         };
 248
 249         rc = obd_get_info(req->rq_svc_thread->t_env, req->rq_export,
 250                           sizeof(key), &key, NULL, fiemap, NULL);
 251         if (rc < 0)
 252                 GOTO(out, rc);
 253
 254         if (fiemap->fm_mapped_extents == 0 ||
 255             fiemap->fm_mapped_extents > ORR_NUM_EXTENTS)
 256                 GOTO(out, rc = -EFAULT);
 257
 258         /**
 259          * Calculate the physical offset ranges for the request from the extent
 260          * information and the logical request offsets.
 261          */
 262         start = fiemap->fm_extents[0].fe_physical + range->or_start -
 263                 fiemap->fm_extents[0].fe_logical;
 264         end = start + range->or_end - range->or_start;
 265
 266         range->or_start = start;
 267         range->or_end = end;
 268
 269         nrq->nr_u.orr.or_physical_set = 1;
 270 out:
 271         return rc;
 272 }
 273
 274 /**
 275  * Sets the offset range the request covers; either in logical file
 276  * offsets or in physical disk offsets.
 277  *
 278  * \param[in] nrq        the request
 279  * \param[in] orrd       the ORR/TRR policy scheduler instance
 280  * \param[in] opc        the request's opcode
 281  * \param[in] moving_req is the request in the process of moving onto the
 282  *                       high-priority NRS head?
 283  *
 284  * \retval 0    range filled successfully
 285  * \retval != 0 error
 286  */
 287 static int nrs_orr_range_fill(struct ptlrpc_nrs_request *nrq,
 288                               struct nrs_orr_data *orrd, __u32 opc,
 289                               bool moving_req)
 290 {
 291         struct ptlrpc_request       *req = container_of(nrq,
 292                                                         struct ptlrpc_request,
 293                                                         rq_nrq);
 294         struct obd_ioobj            *ioo;
 295         struct niobuf_remote        *nb;
 296         struct ost_body             *body;
 297         struct nrs_orr_req_range     range;
 298         int                          niocount;
 299         int                          rc = 0;
 300
 301         /**
 302          * If we are scheduling using physical disk offsets, but we have filled
 303          * the offset information in the request previously
 304          * (i.e. ldlm_lock_reorder_req() is moving the request to the
 305          * high-priority NRS head), there is no need to do anything, and we can
 306          * exit. Moreover than the lack of need, we would be unable to perform
 307          * the obd_get_info() call required in nrs_orr_range_fill_physical(),
 308          * because ldlm_lock_reorder_lock() calls into here while holding a
 309          * spinlock, and retrieving fiemap information via obd_get_info() is a
 310          * potentially sleeping operation.
 311          */
 312         if (orrd->od_physical && nrq->nr_u.orr.or_physical_set)
 313                 return 0;
 314
 315         ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
 316         if (ioo == NULL)
 317                 GOTO(out, rc = -EFAULT);
 318
 319         niocount = ioo->ioo_bufcnt;
 320
 321         nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
 322         if (nb == NULL)
 323                 GOTO(out, rc = -EFAULT);
 324
 325         /**
 326          * Use logical information from niobuf_remote structures.
 327          */
 328         nrs_orr_range_fill_logical(nb, niocount, &range);
 329
 330         /**
 331          * Obtain physical offsets if selected, and this is an OST_READ RPC
 332          * RPC. We do not enter this block if moving_req is set which indicates
 333          * that the request is being moved to the high-priority NRS head by
 334          * ldlm_lock_reorder_req(), as that function calls in here while holding
 335          * a spinlock, and nrs_orr_range_physical() can sleep, so we just use
 336          * logical file offsets for the range values for such requests.
 337          */
 338         if (orrd->od_physical && opc == OST_READ && !moving_req) {
 339                 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 340                 if (body == NULL)
 341                         GOTO(out, rc = -EFAULT);
 342
 343                 /**
 344                  * Translate to physical block offsets from backend filesystem
 345                  * extents.
 346                  * Ignore return values; if obtaining the physical offsets
 347                  * fails, use the logical offsets.
 348                  */
 349                 nrs_orr_range_fill_physical(nrq, &body->oa, &range);
 350         }
 351
 352         nrq->nr_u.orr.or_range = range;
 353 out:
 354         return rc;
 355 }
 356
 357 /**
 358  * Generates a character string that can be used in order to register uniquely
 359  * named libcfs_hash and slab objects for ORR/TRR policy instances. The
 360  * character string is unique per policy instance, as it includes the policy's
 361  * name, the CPT number, and a {reg|hp} token, and there is one policy instance
 362  * per NRS head on each CPT, and the policy is only compatible with the ost_io
 363  * service.
 364  *
 365  * \param[in] policy the policy instance
 366  * \param[out] name  the character array that will hold the generated name
 367  */
 368 static void nrs_orr_genobjname(struct ptlrpc_nrs_policy *policy, char *name)
 369 {
 370         snprintf(name, NRS_ORR_OBJ_NAME_MAX, "%s%s%s%d",
 371                  "nrs_", policy->pol_desc->pd_name,
 372                  policy->pol_nrs->nrs_queue_type == PTLRPC_NRS_QUEUE_REG ?
 373                  "_reg_" : "_hp_", nrs_pol2cptid(policy));
 374 }
 375
 376 /**
 377  * ORR/TRR hash operations
 378  */
 379 #define NRS_ORR_BITS            24
 380 #define NRS_ORR_BKT_BITS        12
 381 #define NRS_ORR_HASH_FLAGS      (CFS_HASH_RW_BKTLOCK | CFS_HASH_ASSERT_EMPTY)
 382
 383 #define NRS_TRR_BITS            4
 384 #define NRS_TRR_BKT_BITS        2
 385 #define NRS_TRR_HASH_FLAGS      CFS_HASH_RW_BKTLOCK
 386
 387 static unsigned nrs_orr_hop_hash(cfs_hash_t *hs, const void *key, unsigned mask)
 388 {
 389         return cfs_hash_djb2_hash(key, sizeof(struct nrs_orr_key), mask);
 390 }
 391
 392 static void *nrs_orr_hop_key(cfs_hlist_node_t *hnode)
 393 {
 394         struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
 395                                                       struct nrs_orr_object,
 396                                                       oo_hnode);
 397         return &orro->oo_key;
 398 }
 399
 400 static int nrs_orr_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 401 {
 402         struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
 403                                                       struct nrs_orr_object,
 404                                                       oo_hnode);
 405
 406         return lu_fid_eq(&orro->oo_key.ok_fid,
 407                          &((struct nrs_orr_key *)key)->ok_fid);
 408 }
 409
 410 static void *nrs_orr_hop_object(cfs_hlist_node_t *hnode)
 411 {
 412         return cfs_hlist_entry(hnode, struct nrs_orr_object, oo_hnode);
 413 }
 414
 415 static void nrs_orr_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 416 {
 417         struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
 418                                                       struct nrs_orr_object,
 419                                                       oo_hnode);
 420         cfs_atomic_inc(&orro->oo_ref);
 421 }
 422
 423 /**
 424  * Removes an nrs_orr_object the hash and frees its memory, if the object has
 425  * no active users.
 426  */
 427 static void nrs_orr_hop_put_free(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 428 {
 429         struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
 430                                                       struct nrs_orr_object,
 431                                                       oo_hnode);
 432         struct nrs_orr_data   *orrd = container_of(orro->oo_res.res_parent,
 433                                                    struct nrs_orr_data, od_res);
 434         cfs_hash_bd_t          bds[2];
 435
 436         if (cfs_atomic_dec_return(&orro->oo_ref) > 1)
 437                 return;
 438
 439         cfs_hash_lock(hs, 0);
 440         cfs_hash_dual_bd_get_and_lock(hs, &orro->oo_key, bds, 1);
 441
 442         /**
 443          * Another thread may have won the race and taken a reference on the
 444          * nrs_orr_object.
 445          */
 446         if (cfs_atomic_read(&orro->oo_ref) > 1)
 447                 goto lost_race;
 448
 449         if (bds[1].bd_bucket == NULL)
 450                 cfs_hash_bd_del_locked(hs, &bds[0], hnode);
 451         else
 452                 hnode = cfs_hash_dual_bd_finddel_locked(hs, bds, &orro->oo_key,
 453                                                         hnode);
 454         LASSERT(hnode != NULL);
 455
 456         OBD_SLAB_FREE_PTR(orro, orrd->od_cache);
 457
 458 lost_race:
 459
 460         cfs_hash_dual_bd_unlock(hs, bds, 1);
 461         cfs_hash_unlock(hs, 0);
 462 }
 463
 464 static void nrs_orr_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 465 {
 466         struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
 467                                                       struct nrs_orr_object,
 468                                                       oo_hnode);
 469         cfs_atomic_dec(&orro->oo_ref);
 470 }
 471
 472 static int nrs_trr_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 473 {
 474         struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
 475                                                       struct nrs_orr_object,
 476                                                       oo_hnode);
 477
 478         return orro->oo_key.ok_idx == ((struct nrs_orr_key *)key)->ok_idx;
 479 }
 480
 481 static void nrs_trr_hop_exit(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 482 {
 483         struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
 484                                                       struct nrs_orr_object,
 485                                                       oo_hnode);
 486         struct nrs_orr_data   *orrd = container_of(orro->oo_res.res_parent,
 487                                                    struct nrs_orr_data, od_res);
 488
 489         LASSERTF(cfs_atomic_read(&orro->oo_ref) == 0,
 490                  "Busy NRS TRR policy object for OST with index %u, with %d "
 491                  "refs\n", orro->oo_key.ok_idx, cfs_atomic_read(&orro->oo_ref));
 492
 493         OBD_SLAB_FREE_PTR(orro, orrd->od_cache);
 494 }
 495
 496 static cfs_hash_ops_t nrs_orr_hash_ops = {
 497         .hs_hash        = nrs_orr_hop_hash,
 498         .hs_key         = nrs_orr_hop_key,
 499         .hs_keycmp      = nrs_orr_hop_keycmp,
 500         .hs_object      = nrs_orr_hop_object,
 501         .hs_get         = nrs_orr_hop_get,
 502         .hs_put         = nrs_orr_hop_put_free,
 503         .hs_put_locked  = nrs_orr_hop_put,
 504 };
 505
 506 static cfs_hash_ops_t nrs_trr_hash_ops = {
 507         .hs_hash        = nrs_orr_hop_hash,
 508         .hs_key         = nrs_orr_hop_key,
 509         .hs_keycmp      = nrs_trr_hop_keycmp,
 510         .hs_object      = nrs_orr_hop_object,
 511         .hs_get         = nrs_orr_hop_get,
 512         .hs_put         = nrs_orr_hop_put,
 513         .hs_put_locked  = nrs_orr_hop_put,
 514         .hs_exit        = nrs_trr_hop_exit,
 515 };
 516
 517 #define NRS_ORR_QUANTUM_DFLT    256
 518
 519 /**
 520  * Binary heap predicate.
 521  *
 522  * Uses
 523  * ptlrpc_nrs_request::nr_u::orr::or_round,
 524  * ptlrpc_nrs_request::nr_u::orr::or_sequence, and
 525  * ptlrpc_nrs_request::nr_u::orr::or_range to compare two binheap nodes and
 526  * produce a binary predicate that indicates their relative priority, so that
 527  * the binary heap can perform the necessary sorting operations.
 528  *
 529  * \param[in] e1 the first binheap node to compare
 530  * \param[in] e2 the second binheap node to compare
 531  *
 532  * \retval 0 e1 > e2
 533  * \retval 1 e1 < e2
 534  */
 535 static int orr_req_compare(cfs_binheap_node_t *e1, cfs_binheap_node_t *e2)
 536 {
 537         struct ptlrpc_nrs_request *nrq1;
 538         struct ptlrpc_nrs_request *nrq2;
 539
 540         nrq1 = container_of(e1, struct ptlrpc_nrs_request, nr_node);
 541         nrq2 = container_of(e2, struct ptlrpc_nrs_request, nr_node);
 542
 543         /**
 544          * Requests have been scheduled against a different scheduling round.
 545          */
 546         if (nrq1->nr_u.orr.or_round < nrq2->nr_u.orr.or_round)
 547                 return 1;
 548         else if (nrq1->nr_u.orr.or_round > nrq2->nr_u.orr.or_round)
 549                 return 0;
 550
 551         /**
 552          * Requests have been scheduled against the same scheduling round, but
 553          * belong to a different batch, i.e. they pertain to a different
 554          * backend-fs object (for ORR policy instances) or OST (for TRR policy
 555          * instances).
 556          */
 557         if (nrq1->nr_u.orr.or_sequence < nrq2->nr_u.crr.cr_sequence)
 558                 return 1;
 559         else if (nrq1->nr_u.orr.or_sequence > nrq2->nr_u.crr.cr_sequence)
 560                 return 0;
 561
 562         /**
 563          * If round numbers and sequence numbers are equal, the two requests
 564          * have been scheduled on the same round, and belong to the same batch,
 565          * which means they pertain to the same backend-fs object (if this is an
 566          * ORR policy instance), or to the same OST (if this is a TRR policy
 567          * instance), so these requests should be sorted by ascending offset
 568          * order.
 569          */
 570         if (nrq1->nr_u.orr.or_range.or_start <
 571             nrq2->nr_u.orr.or_range.or_start) {
 572                 return 1;
 573         } else if (nrq1->nr_u.orr.or_range.or_start >
 574                  nrq2->nr_u.orr.or_range.or_start) {
 575                 return 0;
 576         } else {
 577                 /**
 578                  * Requests start from the same offset; Dispatch the shorter one
 579                  * first; perhaps slightly more chances of hitting caches like
 580                  * this.
 581                  */
 582                 return nrq1->nr_u.orr.or_range.or_end <
 583                        nrq2->nr_u.orr.or_range.or_end;
 584         }
 585 }
 586
 587 /**
 588  * ORR binary heap operations
 589  */
 590 static cfs_binheap_ops_t nrs_orr_heap_ops = {
 591         .hop_enter      = NULL,
 592         .hop_exit       = NULL,
 593         .hop_compare    = orr_req_compare,
 594 };
 595
 596 /**
 597  * Prints a warning message if an ORR/TRR policy is started on a service with
 598  * more than one CPT.  Not printed on the console for now, since we don't
 599  * have any performance metrics in the first place, and it is annoying.
 600  *
 601  * \param[in] policy the policy instance
 602  *
 603  * \retval 0 success
 604  */
 605 static int nrs_orr_init(struct ptlrpc_nrs_policy *policy)
 606 {
 607         if (policy->pol_nrs->nrs_svcpt->scp_service->srv_ncpts > 1)
 608                 CDEBUG(D_CONFIG, "%s: The %s NRS policy was registered on a "
 609                       "service with multiple service partitions. This policy "
 610                       "may perform better with a single partition.\n",
 611                       policy->pol_nrs->nrs_svcpt->scp_service->srv_name,
 612                       policy->pol_desc->pd_name);
 613
 614         return 0;
 615 }
 616
 617 /**
 618  * Called when an ORR policy instance is started.
 619  *
 620  * \param[in] policy the policy
 621  *
 622  * \retval -ENOMEM OOM error
 623  * \retval 0       success
 624  */
 625 static int nrs_orr_start(struct ptlrpc_nrs_policy *policy)
 626 {
 627         struct nrs_orr_data    *orrd;
 628         cfs_hash_ops_t         *ops;
 629         unsigned                cur_bits;
 630         unsigned                max_bits;
 631         unsigned                bkt_bits;
 632         unsigned                flags;
 633         int                     rc = 0;
 634         ENTRY;
 635
 636         OBD_CPT_ALLOC_PTR(orrd, nrs_pol2cptab(policy), nrs_pol2cptid(policy));
 637         if (orrd == NULL)
 638                 RETURN(-ENOMEM);
 639
 640         /*
 641          * Binary heap instance for sorted incoming requests.
 642          */
 643         orrd->od_binheap = cfs_binheap_create(&nrs_orr_heap_ops,
 644                                               CBH_FLAG_ATOMIC_GROW, 4096, NULL,
 645                                               nrs_pol2cptab(policy),
 646                                               nrs_pol2cptid(policy));
 647         if (orrd->od_binheap == NULL)
 648                 GOTO(failed, rc = -ENOMEM);
 649
 650         nrs_orr_genobjname(policy, orrd->od_objname);
 651
 652         /**
 653          * Slab cache for NRS ORR/TRR objects.
 654          */
 655         orrd->od_cache = cfs_mem_cache_create(orrd->od_objname,
 656                                               sizeof(struct nrs_orr_object),
 657                                               0, 0);
 658         if (orrd->od_cache == NULL)
 659                 GOTO(failed, rc = -ENOMEM);
 660
 661         if (strncmp(policy->pol_desc->pd_name, NRS_POL_NAME_ORR,
 662                     NRS_POL_NAME_MAX) == 0) {
 663                 ops = &nrs_orr_hash_ops;
 664                 cur_bits = NRS_ORR_BITS;
 665                 max_bits = NRS_ORR_BITS;
 666                 bkt_bits = NRS_ORR_BKT_BITS;
 667                 flags = NRS_ORR_HASH_FLAGS;
 668         } else {
 669                 ops = &nrs_trr_hash_ops;
 670                 cur_bits = NRS_TRR_BITS;
 671                 max_bits = NRS_TRR_BITS;
 672                 bkt_bits = NRS_TRR_BKT_BITS;
 673                 flags = NRS_TRR_HASH_FLAGS;
 674         }
 675
 676         /**
 677          * Hash for finding objects by struct nrs_orr_key.
 678          * XXX: For TRR, it might be better to avoid using libcfs_hash?
 679          * All that needs to be resolved are OST indices, and they
 680          * will stay relatively stable during an OSS node's lifetime.
 681          */
 682         orrd->od_obj_hash = cfs_hash_create(orrd->od_objname, cur_bits,
 683                                             max_bits, bkt_bits, 0,
 684                                             CFS_HASH_MIN_THETA,
 685                                             CFS_HASH_MAX_THETA, ops, flags);
 686         if (orrd->od_obj_hash == NULL)
 687                 GOTO(failed, rc = -ENOMEM);
 688
 689         /* XXX: Fields accessed unlocked */
 690         orrd->od_quantum = NRS_ORR_QUANTUM_DFLT;
 691         orrd->od_supp = NOS_DFLT;
 692         orrd->od_physical = true;
 693         /**
 694          * Set to 1 so that the test inside nrs_orr_req_add() can evaluate to
 695          * true.
 696          */
 697         orrd->od_sequence = 1;
 698
 699         policy->pol_private = orrd;
 700
 701         RETURN(rc);
 702
 703 failed:
 704         if (orrd->od_cache) {
 705                 rc = cfs_mem_cache_destroy(orrd->od_cache);
 706                 LASSERTF(rc == 0, "Could not destroy od_cache slab\n");
 707         }
 708         if (orrd->od_binheap != NULL)
 709                 cfs_binheap_destroy(orrd->od_binheap);
 710
 711         OBD_FREE_PTR(orrd);
 712
 713         RETURN(rc);
 714 }
 715
 716 /**
 717  * Called when an ORR/TRR policy instance is stopped.
 718  *
 719  * Called when the policy has been instructed to transition to the
 720  * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state and has no more
 721  * pending requests to serve.
 722  *
 723  * \param[in] policy the policy
 724  */
 725 static void nrs_orr_stop(struct ptlrpc_nrs_policy *policy)
 726 {
 727         struct nrs_orr_data *orrd = policy->pol_private;
 728         ENTRY;
 729
 730         LASSERT(orrd != NULL);
 731         LASSERT(orrd->od_binheap != NULL);
 732         LASSERT(orrd->od_obj_hash != NULL);
 733         LASSERT(orrd->od_cache != NULL);
 734         LASSERT(cfs_binheap_is_empty(orrd->od_binheap));
 735
 736         cfs_binheap_destroy(orrd->od_binheap);
 737         cfs_hash_putref(orrd->od_obj_hash);
 738         cfs_mem_cache_destroy(orrd->od_cache);
 739
 740         OBD_FREE_PTR(orrd);
 741 }
 742
 743 /**
 744  * Performs a policy-specific ctl function on ORR/TRR policy instances; similar
 745  * to ioctl.
 746  *
 747  * \param[in]     policy the policy instance
 748  * \param[in]     opc    the opcode
 749  * \param[in,out] arg    used for passing parameters and information
 750  *
 751  * \pre spin_is_locked(&policy->pol_nrs->->nrs_lock)
 752  * \post spin_is_locked(&policy->pol_nrs->->nrs_lock)
 753  *
 754  * \retval 0   operation carried successfully
 755  * \retval -ve error
 756  */
 757 int nrs_orr_ctl(struct ptlrpc_nrs_policy *policy, enum ptlrpc_nrs_ctl opc,
 758                 void *arg)
 759 {
 760         LASSERT(spin_is_locked(&policy->pol_nrs->nrs_lock));
 761
 762         switch(opc) {
 763         default:
 764                 RETURN(-EINVAL);
 765
 766         case NRS_CTL_ORR_RD_QUANTUM: {
 767                 struct nrs_orr_data     *orrd = policy->pol_private;
 768
 769                 *(__u16 *)arg = orrd->od_quantum;
 770                 }
 771                 break;
 772
 773         case NRS_CTL_ORR_WR_QUANTUM: {
 774                 struct nrs_orr_data     *orrd = policy->pol_private;
 775
 776                 orrd->od_quantum = *(__u16 *)arg;
 777                 LASSERT(orrd->od_quantum != 0);
 778                 }
 779                 break;
 780
 781         case NRS_CTL_ORR_RD_OFF_TYPE: {
 782                 struct nrs_orr_data     *orrd = policy->pol_private;
 783
 784                 *(bool *)arg = orrd->od_physical;
 785                 }
 786                 break;
 787
 788         case NRS_CTL_ORR_WR_OFF_TYPE: {
 789                 struct nrs_orr_data     *orrd = policy->pol_private;
 790
 791                 orrd->od_physical = *(bool *)arg;
 792                 }
 793                 break;
 794
 795         case NRS_CTL_ORR_RD_SUPP_REQ: {
 796                 struct nrs_orr_data     *orrd = policy->pol_private;
 797
 798                 *(enum nrs_orr_supp *)arg = orrd->od_supp;
 799                 }
 800                 break;
 801
 802         case NRS_CTL_ORR_WR_SUPP_REQ: {
 803                 struct nrs_orr_data     *orrd = policy->pol_private;
 804
 805                 orrd->od_supp = *(enum nrs_orr_supp *)arg;
 806                 LASSERT((orrd->od_supp & NOS_OST_RW) != 0);
 807                 }
 808                 break;
 809         }
 810         RETURN(0);
 811 }
 812
 813 /**
 814  * Obtains resources for ORR/TRR policy instances. The top-level resource lives
 815  * inside \e nrs_orr_data and the second-level resource inside
 816  * \e nrs_orr_object instances.
 817  *
 818  * \param[in]  policy     the policy for which resources are being taken for
 819  *                        request \a nrq
 820  * \param[in]  nrq        the request for which resources are being taken
 821  * \param[in]  parent     parent resource, embedded in nrs_orr_data for the
 822  *                        ORR/TRR policies
 823  * \param[out] resp       used to return resource references
 824  * \param[in]  moving_req signifies limited caller context; used to perform
 825  *                        memory allocations in an atomic context in this
 826  *                        policy
 827  *
 828  * \retval 0   we are returning a top-level, parent resource, one that is
 829  *             embedded in an nrs_orr_data object
 830  * \retval 1   we are returning a bottom-level resource, one that is embedded
 831  *             in an nrs_orr_object object
 832  *
 833  * \see nrs_resource_get_safe()
 834  */
 835 int nrs_orr_res_get(struct ptlrpc_nrs_policy *policy,
 836                     struct ptlrpc_nrs_request *nrq,
 837                     const struct ptlrpc_nrs_resource *parent,
 838                     struct ptlrpc_nrs_resource **resp, bool moving_req)
 839 {
 840         struct nrs_orr_data            *orrd;
 841         struct nrs_orr_object          *orro;
 842         struct nrs_orr_object          *tmp;
 843         struct nrs_orr_key              key = { { { 0 } } };
 844         __u32                           opc;
 845         int                             rc = 0;
 846
 847         /**
 848          * struct nrs_orr_data is requested.
 849          */
 850         if (parent == NULL) {
 851                 *resp = &((struct nrs_orr_data *)policy->pol_private)->od_res;
 852                 return 0;
 853         }
 854
 855         orrd = container_of(parent, struct nrs_orr_data, od_res);
 856
 857         /**
 858          * If the request type is not supported, fail the enqueuing; the RPC
 859          * will be handled by the fallback NRS policy.
 860          */
 861         if (!nrs_orr_req_supported(orrd, nrq, &opc))
 862                 return -1;
 863
 864         /**
 865          * Fill in the key for the request; OST FID for ORR policy instances,
 866          * and OST index for TRR policy instances.
 867          */
 868         rc = nrs_orr_key_fill(orrd, nrq, opc, policy->pol_desc->pd_name, &key);
 869         if (rc < 0)
 870                 RETURN(rc);
 871
 872         /**
 873          * Set the offset range the request covers
 874          */
 875         rc = nrs_orr_range_fill(nrq, orrd, opc, moving_req);
 876         if (rc < 0)
 877                 RETURN(rc);
 878
 879         orro = cfs_hash_lookup(orrd->od_obj_hash, &key);
 880         if (orro != NULL)
 881                 goto out;
 882
 883         OBD_SLAB_CPT_ALLOC_PTR_GFP(orro, orrd->od_cache,
 884                                    nrs_pol2cptab(policy), nrs_pol2cptid(policy),
 885                                    (moving_req ? CFS_ALLOC_ATOMIC :
 886                                     CFS_ALLOC_IO));
 887         if (orro == NULL)
 888                 RETURN(-ENOMEM);
 889
 890         orro->oo_key = key;
 891         cfs_atomic_set(&orro->oo_ref, 1);
 892
 893         tmp = cfs_hash_findadd_unique(orrd->od_obj_hash, &orro->oo_key,
 894                                       &orro->oo_hnode);
 895         if (tmp != orro) {
 896                 OBD_SLAB_FREE_PTR(orro, orrd->od_cache);
 897                 orro = tmp;
 898         }
 899 out:
 900         /**
 901          * For debugging purposes
 902          */
 903         nrq->nr_u.orr.or_key = orro->oo_key;
 904
 905         *resp = &orro->oo_res;
 906
 907         return 1;
 908 }
 909
 910 /**
 911  * Called when releasing references to the resource hierachy obtained for a
 912  * request for scheduling using ORR/TRR policy instances
 913  *
 914  * \param[in] policy   the policy the resource belongs to
 915  * \param[in] res      the resource to be released
 916  */
 917 static void nrs_orr_res_put(struct ptlrpc_nrs_policy *policy,
 918                             const struct ptlrpc_nrs_resource *res)
 919 {
 920         struct nrs_orr_data     *orrd;
 921         struct nrs_orr_object   *orro;
 922
 923         /**
 924          * Do nothing for freeing parent, nrs_orr_data resources.
 925          */
 926         if (res->res_parent == NULL)
 927                 return;
 928
 929         orro = container_of(res, struct nrs_orr_object, oo_res);
 930         orrd = container_of(res->res_parent, struct nrs_orr_data, od_res);
 931
 932         cfs_hash_put(orrd->od_obj_hash, &orro->oo_hnode);
 933 }
 934
 935 /**
 936  * Called when polling an ORR/TRR policy instance for a request so that it can
 937  * be served. Returns the request that is at the root of the binary heap, as
 938  * that is the lowest priority one (i.e. libcfs_heap is an implementation of a
 939  * min-heap)
 940  *
 941  * \param[in] policy the policy instance being polled
 942  * \param[in] peek   when set, signifies that we just want to examine the
 943  *                   request, and not handle it, so the request is not removed
 944  *                   from the policy.
 945  * \param[in] force  force the policy to return a request; unused in this policy
 946  *
 947  * \retval the request to be handled
 948  * \retval NULL no request available
 949  *
 950  * \see ptlrpc_nrs_req_get_nolock()
 951  * \see nrs_request_get()
 952  */
 953 static
 954 struct ptlrpc_nrs_request *nrs_orr_req_get(struct ptlrpc_nrs_policy *policy,
 955                                            bool peek, bool force)
 956 {
 957         struct nrs_orr_data       *orrd = policy->pol_private;
 958         cfs_binheap_node_t        *node = cfs_binheap_root(orrd->od_binheap);
 959         struct ptlrpc_nrs_request *nrq;
 960
 961         nrq = unlikely(node == NULL) ? NULL :
 962               container_of(node, struct ptlrpc_nrs_request, nr_node);
 963
 964         if (likely(!peek && nrq != NULL)) {
 965                 struct nrs_orr_object *orro;
 966
 967                 orro = container_of(nrs_request_resource(nrq),
 968                                     struct nrs_orr_object, oo_res);
 969
 970                 LASSERT(nrq->nr_u.orr.or_round <= orro->oo_round);
 971
 972                 cfs_binheap_remove(orrd->od_binheap, &nrq->nr_node);
 973                 orro->oo_active--;
 974
 975                 if (strncmp(policy->pol_desc->pd_name, NRS_POL_NAME_ORR,
 976                                  NRS_POL_NAME_MAX) == 0)
 977                         CDEBUG(D_RPCTRACE,
 978                                "NRS: starting to handle %s request for object "
 979                                "with FID "DFID", from OST with index %u, with "
 980                                "round "LPU64"\n", NRS_POL_NAME_ORR,
 981                                PFID(&orro->oo_key.ok_fid),
 982                                nrq->nr_u.orr.or_key.ok_idx,
 983                                nrq->nr_u.orr.or_round);
 984                 else
 985                         CDEBUG(D_RPCTRACE,
 986                                "NRS: starting to handle %s request from OST "
 987                                "with index %u, with round "LPU64"\n",
 988                                NRS_POL_NAME_TRR, nrq->nr_u.orr.or_key.ok_idx,
 989                                nrq->nr_u.orr.or_round);
 990
 991                 /** Peek at the next request to be served */
 992                 node = cfs_binheap_root(orrd->od_binheap);
 993
 994                 /** No more requests */
 995                 if (unlikely(node == NULL)) {
 996                         orrd->od_round++;
 997                 } else {
 998                         struct ptlrpc_nrs_request *next;
 999
1000                         next = container_of(node, struct ptlrpc_nrs_request,
1001                                             nr_node);
1002
1003                         if (orrd->od_round < next->nr_u.orr.or_round)
1004                                 orrd->od_round = next->nr_u.orr.or_round;
1005                 }
1006         }
1007
1008         return nrq;
1009 }
1010
1011 /**
1012  * Sort-adds request \a nrq to an ORR/TRR \a policy instance's set of queued
1013  * requests in the policy's binary heap.
1014  *
1015  * A scheduling round is a stream of requests that have been sorted in batches
1016  * according to the backend-fs object (for ORR policy instances) or OST (for TRR
1017  * policy instances) that they pertain to (as identified by its IDIF FID or OST
1018  * index respectively); there can be only one batch for each object or OST in
1019  * each round. The batches are of maximum size nrs_orr_data:od_quantum. When a
1020  * new request arrives for scheduling for an object or OST that has exhausted
1021  * its quantum in its current round, the request will be scheduled on the next
1022  * scheduling round. Requests are allowed to be scheduled against a round until
1023  * all requests for the round are serviced, so an object or OST might miss a
1024  * round if requests are not scheduled for it for a long enough period of time.
1025  * Objects or OSTs that miss a round will continue with having their next
1026  * request scheduled, starting at the round that requests are being dispatched
1027  * for, at the time of arrival of this request.
1028  *
1029  * Requests are tagged with the round number and a sequence number; the sequence
1030  * number indicates the relative ordering amongst the batches of requests in a
1031  * round, and is identical for all requests in a batch, as is the round number.
1032  * The round and sequence numbers are used by orr_req_compare() in order to use
1033  * nrs_orr_data::od_binheap in order to maintain an ordered set of rounds, with
1034  * each round consisting of an ordered set of batches of requests, and each
1035  * batch consisting of an ordered set of requests according to their logical
1036  * file or physical disk offsets.
1037  *
1038  * \param[in] policy the policy
1039  * \param[in] nrq    the request to add
1040  *
1041  * \retval 0    request successfully added
1042  * \retval != 0 error
1043  */
1044 static int nrs_orr_req_add(struct ptlrpc_nrs_policy *policy,
1045                            struct ptlrpc_nrs_request *nrq)
1046 {
1047         struct nrs_orr_data     *orrd;
1048         struct nrs_orr_object   *orro;
1049         int                      rc;
1050
1051         orro = container_of(nrs_request_resource(nrq),
1052                             struct nrs_orr_object, oo_res);
1053         orrd = container_of(nrs_request_resource(nrq)->res_parent,
1054                             struct nrs_orr_data, od_res);
1055
1056         if (orro->oo_quantum == 0 || orro->oo_round < orrd->od_round ||
1057             (orro->oo_active == 0 && orro->oo_quantum > 0)) {
1058
1059                 /**
1060                  * If there are no pending requests for the object/OST, but some
1061                  * of its quantum still remains unused, which implies we did not
1062                  * get a chance to schedule up to its maximum allowed batch size
1063                  * of requests in the previous round this object/OST
1064                  * participated in, schedule this next request on a new round;
1065                  * this avoids fragmentation of request batches caused by
1066                  * intermittent inactivity on the object/OST, at the expense of
1067                  * potentially slightly increased service time for the request
1068                  * batch this request will be a part of.
1069                  */
1070                 if (orro->oo_active == 0 && orro->oo_quantum > 0)
1071                         orro->oo_round++;
1072
1073                 /** A new scheduling round has commenced */
1074                 if (orro->oo_round < orrd->od_round)
1075                         orro->oo_round = orrd->od_round;
1076
1077                 /** I was not the last object/OST that scheduled a request */
1078                 if (orro->oo_sequence < orrd->od_sequence)
1079                         orro->oo_sequence = ++orrd->od_sequence;
1080                 /**
1081                  * Reset the quantum if we have reached the maximum quantum
1082                  * size for this batch, or even if we have not managed to
1083                  * complete a batch size up to its maximum allowed size.
1084                  * XXX: Accessed unlocked
1085                  */
1086                 orro->oo_quantum = orrd->od_quantum;
1087         }
1088
1089         nrq->nr_u.crr.cr_round = orro->oo_round;
1090         nrq->nr_u.crr.cr_sequence = orro->oo_sequence;
1091
1092         rc = cfs_binheap_insert(orrd->od_binheap, &nrq->nr_node);
1093         if (rc == 0) {
1094                 orro->oo_active++;
1095                 if (--orro->oo_quantum == 0)
1096                         orro->oo_round++;
1097         }
1098         return rc;
1099 }
1100
1101 /**
1102  * Removes request \a nrq from an ORR/TRR \a policy instance's set of queued
1103  * requests.
1104  *
1105  * \param[in] policy the policy
1106  * \param[in] nrq    the request to remove
1107  */
1108 static void nrs_orr_req_del(struct ptlrpc_nrs_policy *policy,
1109                             struct ptlrpc_nrs_request *nrq)
1110 {
1111         struct nrs_orr_data     *orrd;
1112         struct nrs_orr_object   *orro;
1113         bool                     is_root;
1114
1115         orro = container_of(nrs_request_resource(nrq),
1116                             struct nrs_orr_object, oo_res);
1117         orrd = container_of(nrs_request_resource(nrq)->res_parent,
1118                             struct nrs_orr_data, od_res);
1119
1120         LASSERT(nrq->nr_u.orr.or_round <= orro->oo_round);
1121
1122         is_root = &nrq->nr_node == cfs_binheap_root(orrd->od_binheap);
1123
1124         cfs_binheap_remove(orrd->od_binheap, &nrq->nr_node);
1125         orro->oo_active--;
1126
1127         /**
1128          * If we just deleted the node at the root of the binheap, we may have
1129          * to adjust round numbers.
1130          */
1131         if (unlikely(is_root)) {
1132                 /** Peek at the next request to be served */
1133                 cfs_binheap_node_t *node = cfs_binheap_root(orrd->od_binheap);
1134
1135                 /** No more requests */
1136                 if (unlikely(node == NULL)) {
1137                         orrd->od_round++;
1138                 } else {
1139                         nrq = container_of(node, struct ptlrpc_nrs_request,
1140                                            nr_node);
1141
1142                         if (orrd->od_round < nrq->nr_u.orr.or_round)
1143                                 orrd->od_round = nrq->nr_u.orr.or_round;
1144                 }
1145         }
1146 }
1147
1148 /**
1149  * Called right after the request \a nrq finishes being handled by ORR policy
1150  * instance \a policy.
1151  *
1152  * \param[in] policy the policy that handled the request
1153  * \param[in] nrq    the request that was handled
1154  */
1155 static void nrs_orr_req_stop(struct ptlrpc_nrs_policy *policy,
1156                              struct ptlrpc_nrs_request *nrq)
1157 {
1158         /** NB: resource control, credits etc can be added here */
1159         if (strncmp(policy->pol_desc->pd_name, NRS_POL_NAME_ORR,
1160                     NRS_POL_NAME_MAX) == 0)
1161                 CDEBUG(D_RPCTRACE,
1162                        "NRS: finished handling %s request for object with FID "
1163                        DFID", from OST with index %u, with round "LPU64"\n",
1164                        NRS_POL_NAME_ORR, PFID(&nrq->nr_u.orr.or_key.ok_fid),
1165                        nrq->nr_u.orr.or_key.ok_idx, nrq->nr_u.orr.or_round);
1166         else
1167                 CDEBUG(D_RPCTRACE,
1168                        "NRS: finished handling %s request from OST with index %u,"
1169                        " with round "LPU64"\n",
1170                        NRS_POL_NAME_TRR, nrq->nr_u.orr.or_key.ok_idx,
1171                        nrq->nr_u.orr.or_round);
1172 }
1173
1174 /**
1175  * lprocfs interface
1176  */
1177
1178 #ifdef LPROCFS
1179
1180 /**
1181  * This allows to bundle the policy name into the lprocfs_vars::data pointer
1182  * so that lprocfs read/write functions can be used by both the ORR and TRR
1183  * policies.
1184  */
1185 struct nrs_lprocfs_orr_data {
1186         struct ptlrpc_service   *svc;
1187         char                    *name;
1188 } lprocfs_orr_data = {
1189         .name = NRS_POL_NAME_ORR
1190 }, lprocfs_trr_data = {
1191         .name = NRS_POL_NAME_TRR
1192 };
1193
1194 /**
1195  * Retrieves the value of the Round Robin quantum (i.e. the maximum batch size)
1196  * for ORR/TRR policy instances on both the regular and high-priority NRS head
1197  * of a service, as long as a policy instance is not in the
1198  * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state; policy instances in this
1199  * state are skipped later by nrs_orr_ctl().
1200  *
1201  * Quantum values are in # of RPCs, and the output is in YAML format.
1202  *
1203  * For example:
1204  *
1205  *      reg_quantum:256
1206  *      hp_quantum:8
1207  *
1208  * XXX: the CRR-N version of this, ptlrpc_lprocfs_rd_nrs_crrn_quantum() is
1209  * almost identical; it can be reworked and then reused for ORR/TRR.
1210  */
1211 static int ptlrpc_lprocfs_rd_nrs_orr_quantum(char *page, char **start,
1212                                              off_t off, int count, int *eof,
1213                                              void *data)
1214 {
1215         struct nrs_lprocfs_orr_data *orr_data = data;
1216         struct ptlrpc_service       *svc = orr_data->svc;
1217         __u16                        quantum;
1218         int                          rc;
1219         int                          rc2 = 0;
1220
1221         /**
1222          * Perform two separate calls to this as only one of the NRS heads'
1223          * policies may be in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
1224          * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING state.
1225          */
1226         rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1227                                        orr_data->name,
1228                                        NRS_CTL_ORR_RD_QUANTUM,
1229                                        true, &quantum);
1230         if (rc == 0) {
1231                 *eof = 1;
1232                 rc2 = snprintf(page, count, NRS_LPROCFS_QUANTUM_NAME_REG
1233                                "%-5d\n", quantum);
1234                 /**
1235                  * Ignore -ENODEV as the regular NRS head's policy may be in the
1236                  * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1237                  */
1238         } else if (rc != -ENODEV) {
1239                 return rc;
1240         }
1241
1242         /**
1243          * We know the ost_io service which is the only one ORR/TRR policies are
1244          * compatible with, do have an HP NRS head, but it may be best to guard
1245          * against a possible change of this in the future.
1246          */
1247         if (!nrs_svc_has_hp(svc))
1248                 goto no_hp;
1249
1250         rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1251                                        orr_data->name, NRS_CTL_ORR_RD_QUANTUM,
1252                                        true, &quantum);
1253         if (rc == 0) {
1254                 *eof = 1;
1255                 rc2 += snprintf(page + rc2, count - rc2,
1256                                 NRS_LPROCFS_QUANTUM_NAME_HP"%-5d\n", quantum);
1257                 /**
1258                  * Ignore -ENODEV as the high priority NRS head's policy may be
1259                  * in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1260                  */
1261         } else if (rc != -ENODEV) {
1262                 return rc;
1263         }
1264
1265 no_hp:
1266
1267         return rc2 ? : rc;
1268 }
1269
1270 /**
1271  * Sets the value of the Round Robin quantum (i.e. the maximum batch size)
1272  * for ORR/TRR policy instances of a service. The user can set the quantum size
1273  * for the regular and high priority NRS head separately by specifying each
1274  * value, or both together in a single invocation.
1275  *
1276  * For example:
1277  *
1278  * lctl set_param ost.OSS.ost_io.nrs_orr_quantum=req_quantum:64, to set the
1279  * request quantum size of the ORR policy instance on the regular NRS head of
1280  * the ost_io service to 64
1281  *
1282  * lctl set_param ost.OSS.ost_io.nrs_trr_quantum=hp_quantum:8 to set the request
1283  * quantum size of the TRR policy instance on the high priority NRS head of the
1284  * ost_io service to 8
1285  *
1286  * lctl set_param ost.OSS.ost_io.nrs_orr_quantum=32, to set both the request
1287  * quantum size of the ORR policy instance on both the regular and the high
1288  * priority NRS head of the ost_io service to 32
1289  *
1290  * policy instances in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state
1291  * are skipped later by nrs_orr_ctl().
1292  *
1293  * XXX: the CRR-N version of this, ptlrpc_lprocfs_wr_nrs_crrn_quantum() is
1294  * almost identical; it can be reworked and then reused for ORR/TRR.
1295  */
1296 static int ptlrpc_lprocfs_wr_nrs_orr_quantum(struct file *file,
1297                                              const char *buffer,
1298                                              unsigned long count, void *data)
1299 {
1300         struct nrs_lprocfs_orr_data *orr_data = data;
1301         struct ptlrpc_service       *svc = orr_data->svc;
1302         enum ptlrpc_nrs_queue_type   queue = 0;
1303         char                         kernbuf[LPROCFS_NRS_WR_QUANTUM_MAX_CMD];
1304         char                        *val;
1305         long                         quantum_reg;
1306         long                         quantum_hp;
1307         /** lprocfs_find_named_value() modifies its argument, so keep a copy */
1308         unsigned long                count_copy;
1309         int                          rc = 0;
1310         int                          rc2 = 0;
1311
1312         if (count > (sizeof(kernbuf) - 1))
1313                 return -EINVAL;
1314
1315         if (cfs_copy_from_user(kernbuf, buffer, count))
1316                 return -EFAULT;
1317
1318         kernbuf[count] = '\0';
1319
1320         count_copy = count;
1321
1322         /**
1323          * Check if the regular quantum value has been specified
1324          */
1325         val = lprocfs_find_named_value(kernbuf, NRS_LPROCFS_QUANTUM_NAME_REG,
1326                                        &count_copy);
1327         if (val != kernbuf) {
1328                 quantum_reg = simple_strtol(val, NULL, 10);
1329
1330                 queue |= PTLRPC_NRS_QUEUE_REG;
1331         }
1332
1333         count_copy = count;
1334
1335         /**
1336          * Check if the high priority quantum value has been specified
1337          */
1338         val = lprocfs_find_named_value(kernbuf, NRS_LPROCFS_QUANTUM_NAME_HP,
1339                                        &count_copy);
1340         if (val != kernbuf) {
1341                 if (!nrs_svc_has_hp(svc))
1342                         return -ENODEV;
1343
1344                 quantum_hp = simple_strtol(val, NULL, 10);
1345
1346                 queue |= PTLRPC_NRS_QUEUE_HP;
1347         }
1348
1349         /**
1350          * If none of the queues has been specified, look for a valid numerical
1351          * value
1352          */
1353         if (queue == 0) {
1354                 if (!isdigit(kernbuf[0]))
1355                         return -EINVAL;
1356
1357                 quantum_reg = simple_strtol(kernbuf, NULL, 10);
1358
1359                 queue = PTLRPC_NRS_QUEUE_REG;
1360
1361                 if (nrs_svc_has_hp(svc)) {
1362                         queue |= PTLRPC_NRS_QUEUE_HP;
1363                         quantum_hp = quantum_reg;
1364                 }
1365         }
1366
1367         if ((((queue & PTLRPC_NRS_QUEUE_REG) != 0) &&
1368             ((quantum_reg > LPROCFS_NRS_QUANTUM_MAX || quantum_reg <= 0))) ||
1369             (((queue & PTLRPC_NRS_QUEUE_HP) != 0) &&
1370             ((quantum_hp > LPROCFS_NRS_QUANTUM_MAX || quantum_hp <= 0))))
1371                 return -EINVAL;
1372
1373         /**
1374          * We change the values on regular and HP NRS heads separately, so that
1375          * we do not exit early from ptlrpc_nrs_policy_control() with an error
1376          * returned by nrs_policy_ctl_locked(), in cases where the user has not
1377          * started the policy on either the regular or HP NRS head; i.e. we are
1378          * ignoring -ENODEV within nrs_policy_ctl_locked(). -ENODEV is returned
1379          * only if the operation fails with -ENODEV on all heads that have been
1380          * specified by the command; if at least one operation succeeds,
1381          * success is returned.
1382          */
1383         if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1384                 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1385                                                orr_data->name,
1386                                                NRS_CTL_ORR_WR_QUANTUM, false,
1387                                                &quantum_reg);
1388                 if ((rc < 0 && rc != -ENODEV) ||
1389                     (rc == -ENODEV && queue == PTLRPC_NRS_QUEUE_REG))
1390                         return rc;
1391         }
1392
1393         if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1394                 rc2 = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1395                                                 orr_data->name,
1396                                                 NRS_CTL_ORR_WR_QUANTUM, false,
1397                                                 &quantum_hp);
1398                 if ((rc2 < 0 && rc2 != -ENODEV) ||
1399                     (rc2 == -ENODEV && queue == PTLRPC_NRS_QUEUE_HP))
1400                         return rc2;
1401         }
1402
1403         return rc == -ENODEV && rc2 == -ENODEV ? -ENODEV : count;
1404 }
1405
1406 #define LPROCFS_NRS_OFF_NAME_REG                "reg_offset_type:"
1407 #define LPROCFS_NRS_OFF_NAME_HP                 "hp_offset_type:"
1408
1409 #define LPROCFS_NRS_OFF_NAME_PHYSICAL           "physical"
1410 #define LPROCFS_NRS_OFF_NAME_LOGICAL            "logical"
1411
1412 /**
1413  * Retrieves the offset type used by ORR/TRR policy instances on both the
1414  * regular and high-priority NRS head of a service, as long as a policy
1415  * instance is not in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state;
1416  * policy instances in this state are skipped later by nrs_orr_ctl().
1417  *
1418  * Offset type information is a (physical|logical) string, and output is
1419  * in YAML format.
1420  *
1421  * For example:
1422  *
1423  *      reg_offset_type:physical
1424  *      hp_offset_type:logical
1425  */
1426 static int ptlrpc_lprocfs_rd_nrs_orr_offset_type(char *page, char **start,
1427                                                  off_t off, int count, int *eof,
1428                                                  void *data)
1429 {
1430         struct nrs_lprocfs_orr_data *orr_data = data;
1431         struct ptlrpc_service       *svc = orr_data->svc;
1432         bool                         physical;
1433         int                          rc;
1434         int                          rc2 = 0;
1435
1436         /**
1437          * Perform two separate calls to this as only one of the NRS heads'
1438          * policies may be in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED
1439          * or ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING state.
1440          */
1441         rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1442                                        orr_data->name, NRS_CTL_ORR_RD_OFF_TYPE,
1443                                        true, &physical);
1444         if (rc == 0) {
1445                 *eof = 1;
1446                 rc2 = snprintf(page, count,
1447                                LPROCFS_NRS_OFF_NAME_REG"%s\n",
1448                                physical ? LPROCFS_NRS_OFF_NAME_PHYSICAL :
1449                                LPROCFS_NRS_OFF_NAME_LOGICAL);
1450                 /**
1451                  * Ignore -ENODEV as the regular NRS head's policy may be in the
1452                  * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1453                  */
1454         } else if (rc != -ENODEV) {
1455                 return rc;
1456         }
1457
1458         /**
1459          * We know the ost_io service which is the only one ORR/TRR policies are
1460          * compatible with, do have an HP NRS head, but it may be best to guard
1461          * against a possible change of this in the future.
1462          */
1463         if (!nrs_svc_has_hp(svc))
1464                 goto no_hp;
1465
1466         rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1467                                        orr_data->name, NRS_CTL_ORR_RD_OFF_TYPE,
1468                                        true, &physical);
1469         if (rc == 0) {
1470                 *eof = 1;
1471                 rc2 += snprintf(page + rc2, count - rc2,
1472                                 LPROCFS_NRS_OFF_NAME_HP"%s\n",
1473                                 physical ? LPROCFS_NRS_OFF_NAME_PHYSICAL :
1474                                 LPROCFS_NRS_OFF_NAME_LOGICAL);
1475                 /**
1476                  * Ignore -ENODEV as the high priority NRS head's policy may be
1477                  * in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1478                  */
1479         } else if (rc != -ENODEV) {
1480                 return rc;
1481         }
1482
1483 no_hp:
1484
1485         return rc2 ? : rc;
1486 }
1487
1488 /**
1489  * Max valid command string is the size of the labels, plus "physical" twice.
1490  * plus a separating ' '
1491  */
1492 #define LPROCFS_NRS_WR_OFF_TYPE_MAX_CMD                                        \
1493         sizeof(LPROCFS_NRS_OFF_NAME_REG LPROCFS_NRS_OFF_NAME_PHYSICAL " "      \
1494                LPROCFS_NRS_OFF_NAME_HP LPROCFS_NRS_OFF_NAME_PHYSICAL)
1495
1496 /**
1497  * Sets the type of offsets used to order RPCs in ORR/TRR policy instances. The
1498  * user can set offset type for the regular or high priority NRS head
1499  * separately by specifying each value, or both together in a single invocation.
1500  *
1501  * For example:
1502  *
1503  * lctl set_param ost.OSS.ost_io.nrs_orr_offset_type=
1504  * reg_offset_type:physical, to enable the ORR policy instance on the regular
1505  * NRS head of the ost_io service to use physical disk offset ordering.
1506  *
1507  * lctl set_param ost.OSS.ost_io.nrs_trr_offset_type=logical, to enable the TRR
1508  * policy instances on both the regular ang high priority NRS heads of the
1509  * ost_io service to use logical file offset ordering.
1510  *
1511  * policy instances in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state are
1512  * are skipped later by nrs_orr_ctl().
1513  */
1514 static int ptlrpc_lprocfs_wr_nrs_orr_offset_type(struct file *file,
1515                                                  const char *buffer,
1516                                                  unsigned long count,
1517                                                  void *data)
1518 {
1519         struct nrs_lprocfs_orr_data *orr_data = data;
1520         struct ptlrpc_service       *svc = orr_data->svc;
1521         enum ptlrpc_nrs_queue_type   queue = 0;
1522         char                         kernbuf[LPROCFS_NRS_WR_OFF_TYPE_MAX_CMD];
1523         char                        *val_reg;
1524         char                        *val_hp;
1525         bool                         physical_reg;
1526         bool                         physical_hp;
1527         unsigned long                count_copy;
1528         int                          rc = 0;
1529         int                          rc2 = 0;
1530
1531         if (count > (sizeof(kernbuf) - 1))
1532                 return -EINVAL;
1533
1534         if (cfs_copy_from_user(kernbuf, buffer, count))
1535                 return -EFAULT;
1536
1537         kernbuf[count] = '\0';
1538
1539         count_copy = count;
1540
1541         /**
1542          * Check if the regular offset type has been specified
1543          */
1544         val_reg = lprocfs_find_named_value(kernbuf,
1545                                            LPROCFS_NRS_OFF_NAME_REG,
1546                                            &count_copy);
1547         if (val_reg != kernbuf)
1548                 queue |= PTLRPC_NRS_QUEUE_REG;
1549
1550         count_copy = count;
1551
1552         /**
1553          * Check if the high priority offset type has been specified
1554          */
1555         val_hp = lprocfs_find_named_value(kernbuf, LPROCFS_NRS_OFF_NAME_HP,
1556                                           &count_copy);
1557         if (val_hp != kernbuf) {
1558                 if (!nrs_svc_has_hp(svc))
1559                         return -ENODEV;
1560
1561                 queue |= PTLRPC_NRS_QUEUE_HP;
1562         }
1563
1564         /**
1565          * If none of the queues has been specified, there may be a valid
1566          * command string at the start of the buffer.
1567          */
1568         if (queue == 0) {
1569                 queue = PTLRPC_NRS_QUEUE_REG;
1570
1571                 if (nrs_svc_has_hp(svc))
1572                         queue |= PTLRPC_NRS_QUEUE_HP;
1573         }
1574
1575         if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1576                 if (strncmp(val_reg, LPROCFS_NRS_OFF_NAME_PHYSICAL,
1577                             sizeof(LPROCFS_NRS_OFF_NAME_PHYSICAL) - 1) == 0)
1578                         physical_reg = true;
1579                 else if (strncmp(val_reg, LPROCFS_NRS_OFF_NAME_LOGICAL,
1580                          sizeof(LPROCFS_NRS_OFF_NAME_LOGICAL) - 1) == 0)
1581                         physical_reg = false;
1582                 else
1583                         return -EINVAL;
1584         }
1585
1586         if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1587                 if (strncmp(val_hp, LPROCFS_NRS_OFF_NAME_PHYSICAL,
1588                             sizeof(LPROCFS_NRS_OFF_NAME_PHYSICAL) - 1) == 0)
1589                         physical_hp = true;
1590                 else if (strncmp(val_hp, LPROCFS_NRS_OFF_NAME_LOGICAL,
1591                                  sizeof(LPROCFS_NRS_OFF_NAME_LOGICAL) - 1) == 0)
1592                         physical_hp = false;
1593                 else
1594                         return -EINVAL;
1595         }
1596
1597         /**
1598          * We change the values on regular and HP NRS heads separately, so that
1599          * we do not exit early from ptlrpc_nrs_policy_control() with an error
1600          * returned by nrs_policy_ctl_locked(), in cases where the user has not
1601          * started the policy on either the regular or HP NRS head; i.e. we are
1602          * ignoring -ENODEV within nrs_policy_ctl_locked(). -ENODEV is returned
1603          * only if the operation fails with -ENODEV on all heads that have been
1604          * specified by the command; if at least one operation succeeds,
1605          * success is returned.
1606          */
1607         if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1608                 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1609                                                orr_data->name,
1610                                                NRS_CTL_ORR_WR_OFF_TYPE, false,
1611                                                &physical_reg);
1612                 if ((rc < 0 && rc != -ENODEV) ||
1613                     (rc == -ENODEV && queue == PTLRPC_NRS_QUEUE_REG))
1614                         return rc;
1615         }
1616
1617         if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1618                 rc2 = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1619                                                 orr_data->name,
1620                                                 NRS_CTL_ORR_WR_OFF_TYPE, false,
1621                                                 &physical_hp);
1622                 if ((rc2 < 0 && rc2 != -ENODEV) ||
1623                     (rc2 == -ENODEV && queue == PTLRPC_NRS_QUEUE_HP))
1624                         return rc2;
1625         }
1626
1627         return rc == -ENODEV && rc2 == -ENODEV ? -ENODEV : count;
1628 }
1629
1630 #define NRS_LPROCFS_REQ_SUPP_NAME_REG           "reg_supported:"
1631 #define NRS_LPROCFS_REQ_SUPP_NAME_HP            "hp_supported:"
1632
1633 #define LPROCFS_NRS_SUPP_NAME_READS             "reads"
1634 #define LPROCFS_NRS_SUPP_NAME_WRITES            "writes"
1635 #define LPROCFS_NRS_SUPP_NAME_READWRITES        "reads_and_writes"
1636
1637 /**
1638  * Translates enum nrs_orr_supp values to a corresponding string.
1639  */
1640 static const char *nrs_orr_supp2str(enum nrs_orr_supp supp)
1641 {
1642         switch(supp) {
1643         default:
1644                 LBUG();
1645         case NOS_OST_READ:
1646                 return LPROCFS_NRS_SUPP_NAME_READS;
1647         case NOS_OST_WRITE:
1648                 return LPROCFS_NRS_SUPP_NAME_WRITES;
1649         case NOS_OST_RW:
1650                 return LPROCFS_NRS_SUPP_NAME_READWRITES;
1651         }
1652 }
1653
1654 /**
1655  * Translates strings to the corresponding enum nrs_orr_supp value
1656  */
1657 static enum nrs_orr_supp nrs_orr_str2supp(const char *val)
1658 {
1659         if (strncmp(val, LPROCFS_NRS_SUPP_NAME_READWRITES,
1660                     sizeof(LPROCFS_NRS_SUPP_NAME_READWRITES) - 1) == 0)
1661                 return NOS_OST_RW;
1662         else if (strncmp(val, LPROCFS_NRS_SUPP_NAME_READS,
1663                          sizeof(LPROCFS_NRS_SUPP_NAME_READS) - 1) == 0)
1664                 return NOS_OST_READ;
1665         else if (strncmp(val, LPROCFS_NRS_SUPP_NAME_WRITES,
1666                          sizeof(LPROCFS_NRS_SUPP_NAME_WRITES) - 1) == 0)
1667                 return NOS_OST_WRITE;
1668         else
1669                 return -EINVAL;
1670 }
1671
1672 /**
1673  * Retrieves the type of RPCs handled at the point of invocation by ORR/TRR
1674  * policy instances on both the regular and high-priority NRS head of a service,
1675  * as long as a policy instance is not in the
1676  * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state; policy instances in this
1677  * state are skipped later by nrs_orr_ctl().
1678  *
1679  * Supported RPC type information is a (reads|writes|reads_and_writes) string,
1680  * and output is in YAML format.
1681  *
1682  * For example:
1683  *
1684  *      reg_supported:reads
1685  *      hp_supported:reads_and_writes
1686  */
1687 static int ptlrpc_lprocfs_rd_nrs_orr_supported(char *page, char **start,
1688                                                off_t off, int count, int *eof,
1689                                                void *data)
1690 {
1691         struct nrs_lprocfs_orr_data *orr_data = data;
1692         struct ptlrpc_service       *svc = orr_data->svc;
1693         enum nrs_orr_supp            supported;
1694         int                          rc;
1695         int                          rc2 = 0;
1696
1697         /**
1698          * Perform two separate calls to this as only one of the NRS heads'
1699          * policies may be in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED
1700          * or ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING state.
1701          */
1702         rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1703                                        orr_data->name,
1704                                        NRS_CTL_ORR_RD_SUPP_REQ, true,
1705                                        &supported);
1706
1707         if (rc == 0) {
1708                 *eof = 1;
1709                 rc2 = snprintf(page, count,
1710                                NRS_LPROCFS_REQ_SUPP_NAME_REG"%s\n",
1711                                nrs_orr_supp2str(supported));
1712                 /**
1713                  * Ignore -ENODEV as the regular NRS head's policy may be in the
1714                  * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1715                  */
1716         } else if (rc != -ENODEV) {
1717                 return rc;
1718         }
1719
1720         /**
1721          * We know the ost_io service which is the only one ORR/TRR policies are
1722          * compatible with, do have an HP NRS head, but it may be best to guard
1723          * against a possible change of this in the future.
1724          */
1725         if (!nrs_svc_has_hp(svc))
1726                 goto no_hp;
1727
1728         rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1729                                        orr_data->name,
1730                                        NRS_CTL_ORR_RD_SUPP_REQ, true,
1731                                        &supported);
1732         if (rc == 0) {
1733                 *eof = 1;
1734                 rc2 += snprintf(page + rc2, count - rc2,
1735                                NRS_LPROCFS_REQ_SUPP_NAME_HP"%s\n",
1736                                nrs_orr_supp2str(supported));
1737                 /**
1738                  * Ignore -ENODEV as the high priority NRS head's policy may be
1739                  * in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1740                  */
1741         } else if (rc != -ENODEV) {
1742                 return rc;
1743         }
1744
1745 no_hp:
1746
1747         return rc2 ? : rc;
1748 }
1749
1750 /**
1751  * Max valid command string is the size of the labels, plus "reads_and_writes"
1752  * twice, plus a separating ' '
1753  */
1754 #define LPROCFS_NRS_WR_REQ_SUPP_MAX_CMD                                        \
1755         sizeof(NRS_LPROCFS_REQ_SUPP_NAME_REG LPROCFS_NRS_SUPP_NAME_READWRITES  \
1756                NRS_LPROCFS_REQ_SUPP_NAME_HP LPROCFS_NRS_SUPP_NAME_READWRITES   \
1757                " ")
1758
1759 /**
1760  * Sets the type of RPCs handled by ORR/TRR policy instances. The user can
1761  * modify this setting for the regular or high priority NRS heads separately, or
1762  * both together in a single invocation.
1763  *
1764  * For example:
1765  *
1766  * lctl set_param ost.OSS.ost_io.nrs_orr_supported=
1767  * "reg_supported:reads", to enable the ORR policy instance on the regular NRS
1768  * head of the ost_io service to handle OST_READ RPCs.
1769  *
1770  * lctl set_param ost.OSS.ost_io.nrs_trr_supported=reads_and_writes, to enable
1771  * the TRR policy instances on both the regular ang high priority NRS heads of
1772  * the ost_io service to use handle OST_READ and OST_WRITE RPCs.
1773  *
1774  * policy instances in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state are
1775  * are skipped later by nrs_orr_ctl().
1776  */
1777 static int ptlrpc_lprocfs_wr_nrs_orr_supported(struct file *file,
1778                                                const char *buffer,
1779                                                unsigned long count, void *data)
1780 {
1781         struct nrs_lprocfs_orr_data *orr_data = data;
1782         struct ptlrpc_service       *svc = orr_data->svc;
1783         enum ptlrpc_nrs_queue_type   queue = 0;
1784         char                         kernbuf[LPROCFS_NRS_WR_REQ_SUPP_MAX_CMD];
1785         char                        *val_reg;
1786         char                        *val_hp;
1787         enum nrs_orr_supp            supp_reg;
1788         enum nrs_orr_supp            supp_hp;
1789         unsigned long                count_copy;
1790         int                          rc = 0;
1791         int                          rc2 = 0;
1792
1793         if (count > (sizeof(kernbuf) - 1))
1794                 return -EINVAL;
1795
1796         if (cfs_copy_from_user(kernbuf, buffer, count))
1797                 return -EFAULT;
1798
1799         kernbuf[count] = '\0';
1800
1801         count_copy = count;
1802
1803         /**
1804          * Check if the regular supported requests setting has been specified
1805          */
1806         val_reg = lprocfs_find_named_value(kernbuf,
1807                                            NRS_LPROCFS_REQ_SUPP_NAME_REG,
1808                                            &count_copy);
1809         if (val_reg != kernbuf)
1810                 queue |= PTLRPC_NRS_QUEUE_REG;
1811
1812         count_copy = count;
1813
1814         /**
1815          * Check if the high priority supported requests setting has been
1816          * specified
1817          */
1818         val_hp = lprocfs_find_named_value(kernbuf, NRS_LPROCFS_REQ_SUPP_NAME_HP,
1819                                           &count_copy);
1820         if (val_hp != kernbuf) {
1821                 if (!nrs_svc_has_hp(svc))
1822                         return -ENODEV;
1823
1824                 queue |= PTLRPC_NRS_QUEUE_HP;
1825         }
1826
1827         /**
1828          * If none of the queues has been specified, there may be a valid
1829          * command string at the start of the buffer.
1830          */
1831         if (queue == 0) {
1832                 queue = PTLRPC_NRS_QUEUE_REG;
1833
1834                 if (nrs_svc_has_hp(svc))
1835                         queue |= PTLRPC_NRS_QUEUE_HP;
1836         }
1837
1838         if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1839                 supp_reg = nrs_orr_str2supp(val_reg);
1840                 if (supp_reg == -EINVAL)
1841                         return -EINVAL;
1842         }
1843
1844         if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1845                 supp_hp = nrs_orr_str2supp(val_hp);
1846                 if (supp_hp == -EINVAL)
1847                         return -EINVAL;
1848         }
1849
1850         /**
1851          * We change the values on regular and HP NRS heads separately, so that
1852          * we do not exit early from ptlrpc_nrs_policy_control() with an error
1853          * returned by nrs_policy_ctl_locked(), in cases where the user has not
1854          * started the policy on either the regular or HP NRS head; i.e. we are
1855          * ignoring -ENODEV within nrs_policy_ctl_locked(). -ENODEV is returned
1856          * only if the operation fails with -ENODEV on all heads that have been
1857          * specified by the command; if at least one operation succeeds,
1858          * success is returned.
1859          */
1860         if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1861                 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1862                                                orr_data->name,
1863                                                NRS_CTL_ORR_WR_SUPP_REQ, false,
1864                                                &supp_reg);
1865                 if ((rc < 0 && rc != -ENODEV) ||
1866                     (rc == -ENODEV && queue == PTLRPC_NRS_QUEUE_REG))
1867                         return rc;
1868         }
1869
1870         if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1871                 rc2 = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1872                                                 orr_data->name,
1873                                                 NRS_CTL_ORR_WR_SUPP_REQ, false,
1874                                                 &supp_hp);
1875                 if ((rc2 < 0 && rc2 != -ENODEV) ||
1876                     (rc2 == -ENODEV && queue == PTLRPC_NRS_QUEUE_HP))
1877                         return rc2;
1878         }
1879
1880         return rc == -ENODEV && rc2 == -ENODEV ? -ENODEV : count;
1881 }
1882
1883 int nrs_orr_lprocfs_init(struct ptlrpc_service *svc)
1884 {
1885         int     rc;
1886         int     i;
1887
1888         struct lprocfs_vars nrs_orr_lprocfs_vars[] = {
1889                 { .name         = "nrs_orr_quantum",
1890                   .read_fptr    = ptlrpc_lprocfs_rd_nrs_orr_quantum,
1891                   .write_fptr   = ptlrpc_lprocfs_wr_nrs_orr_quantum },
1892                 { .name         = "nrs_orr_offset_type",
1893                   .read_fptr    = ptlrpc_lprocfs_rd_nrs_orr_offset_type,
1894                   .write_fptr   = ptlrpc_lprocfs_wr_nrs_orr_offset_type },
1895                 { .name         = "nrs_orr_supported",
1896                   .read_fptr    = ptlrpc_lprocfs_rd_nrs_orr_supported,
1897                   .write_fptr   = ptlrpc_lprocfs_wr_nrs_orr_supported },
1898                 { NULL }
1899         };
1900
1901         if (svc->srv_procroot == NULL)
1902                 return 0;
1903
1904         lprocfs_orr_data.svc = svc;
1905
1906         for (i = 0; i < ARRAY_SIZE(nrs_orr_lprocfs_vars); i++)
1907                 nrs_orr_lprocfs_vars[i].data = &lprocfs_orr_data;
1908
1909         rc = lprocfs_add_vars(svc->srv_procroot, nrs_orr_lprocfs_vars, NULL);
1910
1911         return rc;
1912 }
1913
1914 void nrs_orr_lprocfs_fini(struct ptlrpc_service *svc)
1915 {
1916         if (svc->srv_procroot == NULL)
1917                 return;
1918
1919         lprocfs_remove_proc_entry("nrs_orr_quantum", svc->srv_procroot);
1920         lprocfs_remove_proc_entry("nrs_orr_offset_type", svc->srv_procroot);
1921         lprocfs_remove_proc_entry("nrs_orr_supported", svc->srv_procroot);
1922 }
1923
1924 #endif /* LPROCFS */
1925
1926 static const struct ptlrpc_nrs_pol_ops nrs_orr_ops = {
1927         .op_policy_init         = nrs_orr_init,
1928         .op_policy_start        = nrs_orr_start,
1929         .op_policy_stop         = nrs_orr_stop,
1930         .op_policy_ctl          = nrs_orr_ctl,
1931         .op_res_get             = nrs_orr_res_get,
1932         .op_res_put             = nrs_orr_res_put,
1933         .op_req_get             = nrs_orr_req_get,
1934         .op_req_enqueue         = nrs_orr_req_add,
1935         .op_req_dequeue         = nrs_orr_req_del,
1936         .op_req_stop            = nrs_orr_req_stop,
1937 #ifdef LPROCFS
1938         .op_lprocfs_init        = nrs_orr_lprocfs_init,
1939         .op_lprocfs_fini        = nrs_orr_lprocfs_fini,
1940 #endif
1941 };
1942
1943 struct ptlrpc_nrs_pol_conf nrs_conf_orr = {
1944         .nc_name                = NRS_POL_NAME_ORR,
1945         .nc_ops                 = &nrs_orr_ops,
1946         .nc_compat              = nrs_policy_compat_one,
1947         .nc_compat_svc_name     = "ost_io",
1948 };
1949
1950 /**
1951  * TRR, Target-based Round Robin policy
1952  *
1953  * TRR reuses much of the functions and data structures of ORR
1954  */
1955
1956 #ifdef LPROCFS
1957
1958 int nrs_trr_lprocfs_init(struct ptlrpc_service *svc)
1959 {
1960         int     rc;
1961         int     i;
1962
1963         struct lprocfs_vars nrs_trr_lprocfs_vars[] = {
1964                 { .name         = "nrs_trr_quantum",
1965                   .read_fptr    = ptlrpc_lprocfs_rd_nrs_orr_quantum,
1966                   .write_fptr   = ptlrpc_lprocfs_wr_nrs_orr_quantum },
1967                 { .name         = "nrs_trr_offset_type",
1968                   .read_fptr    = ptlrpc_lprocfs_rd_nrs_orr_offset_type,
1969                   .write_fptr   = ptlrpc_lprocfs_wr_nrs_orr_offset_type },
1970                 { .name         = "nrs_trr_supported",
1971                   .read_fptr    = ptlrpc_lprocfs_rd_nrs_orr_supported,
1972                   .write_fptr   = ptlrpc_lprocfs_wr_nrs_orr_supported },
1973                 { NULL }
1974         };
1975
1976         if (svc->srv_procroot == NULL)
1977                 return 0;
1978
1979         lprocfs_trr_data.svc = svc;
1980
1981         for (i = 0; i < ARRAY_SIZE(nrs_trr_lprocfs_vars); i++)
1982                 nrs_trr_lprocfs_vars[i].data = &lprocfs_trr_data;
1983
1984         rc = lprocfs_add_vars(svc->srv_procroot, nrs_trr_lprocfs_vars, NULL);
1985
1986         return rc;
1987 }
1988
1989 void nrs_trr_lprocfs_fini(struct ptlrpc_service *svc)
1990 {
1991         if (svc->srv_procroot == NULL)
1992                 return;
1993
1994         lprocfs_remove_proc_entry("nrs_trr_quantum", svc->srv_procroot);
1995         lprocfs_remove_proc_entry("nrs_trr_offset_type", svc->srv_procroot);
1996         lprocfs_remove_proc_entry("nrs_trr_supported", svc->srv_procroot);
1997 }
1998
1999 #endif /* LPROCFS */
2000
2001 /**
2002  * Reuse much of the ORR functionality for TRR.
2003  */
2004 static const struct ptlrpc_nrs_pol_ops nrs_trr_ops = {
2005         .op_policy_init         = nrs_orr_init,
2006         .op_policy_start        = nrs_orr_start,
2007         .op_policy_stop         = nrs_orr_stop,
2008         .op_policy_ctl          = nrs_orr_ctl,
2009         .op_res_get             = nrs_orr_res_get,
2010         .op_res_put             = nrs_orr_res_put,
2011         .op_req_get             = nrs_orr_req_get,
2012         .op_req_enqueue         = nrs_orr_req_add,
2013         .op_req_dequeue         = nrs_orr_req_del,
2014         .op_req_stop            = nrs_orr_req_stop,
2015 #ifdef LPROCFS
2016         .op_lprocfs_init        = nrs_trr_lprocfs_init,
2017         .op_lprocfs_fini        = nrs_trr_lprocfs_fini,
2018 #endif
2019 };
2020
2021 struct ptlrpc_nrs_pol_conf nrs_conf_trr = {
2022         .nc_name                = NRS_POL_NAME_TRR,
2023         .nc_ops                 = &nrs_trr_ops,
2024         .nc_compat              = nrs_policy_compat_one,
2025         .nc_compat_svc_name     = "ost_io",
2026 };
2027
2028 /** @} ORR/TRR policy */
2029
2030 /** @} nrs */
2031
2032 #endif /* HAVE_SERVER_SUPPORT */