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