4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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.
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
23 * Copyright (c) 2013, Intel Corporation.
25 * Copyright 2012 Xyratex Technology Limited
28 * lustre/ptlrpc/nrs_orr.c
30 * Network Request Scheduler (NRS) ORR and TRR policies
32 * Request scheduling in a Round-Robin manner over backend-fs objects and OSTs
35 * Author: Liang Zhen <liang@whamcloud.com>
36 * Author: Nikitas Angelinas <nikitas_angelinas@xyratex.com>
38 #ifdef HAVE_SERVER_SUPPORT
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"
53 * \name ORR/TRR policy
55 * ORR/TRR (Object-based Round Robin/Target-based Round Robin) NRS policies
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.
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.
78 #define NRS_POL_NAME_ORR "orr"
79 #define NRS_POL_NAME_TRR "trr"
82 * Checks if the RPC type of \a nrq is currently handled by an ORR/TRR policy
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
89 * \retval true request type is supported by the policy instance
90 * \retval false request type is not supported by the policy instance
92 static bool nrs_orr_req_supported(struct nrs_orr_data *orrd,
93 struct ptlrpc_nrs_request *nrq, __u32 *opcode)
95 struct ptlrpc_request *req = container_of(nrq, struct ptlrpc_request,
97 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
101 * XXX: nrs_orr_data::od_supp accessed unlocked.
105 rc = orrd->od_supp & NOS_OST_READ;
108 rc = orrd->od_supp & NOS_OST_WRITE;
119 * Returns the ORR/TRR key fields for the request \a nrq in \a key.
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.
127 * \retval 0 key filled successfully
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)
134 struct ptlrpc_request *req = container_of(nrq, struct ptlrpc_request,
136 struct ost_body *body;
138 bool is_orr = strncmp(name, NRS_POL_NAME_ORR,
139 NRS_POL_NAME_MAX) == 0;
141 LASSERT(req != NULL);
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.
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;
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));
160 ost_idx = class_server_data(req->rq_export->exp_obd)->lsd_osd_index;
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.
170 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
174 rc = ostid_to_fid(&key->ok_fid, &body->oa.o_oi, ost_idx);
178 nrq->nr_u.orr.or_orr_set = 1;
180 key->ok_idx = ost_idx;
181 nrq->nr_u.orr.or_trr_set = 1;
188 * Populates the range values in \a range with logical offsets obtained via
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
195 static void nrs_orr_range_fill_logical(struct niobuf_remote *nb, int niocount,
196 struct nrs_orr_req_range *range)
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;
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.
208 #define ORR_NUM_EXTENTS 1
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.
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
221 * \retval 0 physical offsets obtained successfully
224 static int nrs_orr_range_fill_physical(struct ptlrpc_nrs_request *nrq,
226 struct nrs_orr_req_range *range)
228 struct ptlrpc_request *req = container_of(nrq,
229 struct ptlrpc_request,
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;
239 key = (typeof(key)) {
243 .fm_start = range->or_start,
244 .fm_length = range->or_end - range->or_start,
245 .fm_extent_count = ORR_NUM_EXTENTS
249 rc = obd_get_info(req->rq_svc_thread->t_env, req->rq_export,
250 sizeof(key), &key, NULL, fiemap, NULL);
254 if (fiemap->fm_mapped_extents == 0 ||
255 fiemap->fm_mapped_extents > ORR_NUM_EXTENTS)
256 GOTO(out, rc = -EFAULT);
259 * Calculate the physical offset ranges for the request from the extent
260 * information and the logical request offsets.
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;
266 range->or_start = start;
269 nrq->nr_u.orr.or_physical_set = 1;
275 * Sets the offset range the request covers; either in logical file
276 * offsets or in physical disk offsets.
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?
284 * \retval 0 range filled successfully
287 static int nrs_orr_range_fill(struct ptlrpc_nrs_request *nrq,
288 struct nrs_orr_data *orrd, __u32 opc,
291 struct ptlrpc_request *req = container_of(nrq,
292 struct ptlrpc_request,
294 struct obd_ioobj *ioo;
295 struct niobuf_remote *nb;
296 struct ost_body *body;
297 struct nrs_orr_req_range range;
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.
312 if (orrd->od_physical && nrq->nr_u.orr.or_physical_set)
315 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
317 GOTO(out, rc = -EFAULT);
319 niocount = ioo->ioo_bufcnt;
321 nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
323 GOTO(out, rc = -EFAULT);
326 * Use logical information from niobuf_remote structures.
328 nrs_orr_range_fill_logical(nb, niocount, &range);
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.
338 if (orrd->od_physical && opc == OST_READ && !moving_req) {
339 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
341 GOTO(out, rc = -EFAULT);
344 * Translate to physical block offsets from backend filesystem
346 * Ignore return values; if obtaining the physical offsets
347 * fails, use the logical offsets.
349 nrs_orr_range_fill_physical(nrq, &body->oa, &range);
352 nrq->nr_u.orr.or_range = range;
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
365 * \param[in] policy the policy instance
366 * \param[out] name the character array that will hold the generated name
368 static void nrs_orr_genobjname(struct ptlrpc_nrs_policy *policy, char *name)
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));
377 * ORR/TRR hash operations
379 #define NRS_ORR_BITS 24
380 #define NRS_ORR_BKT_BITS 12
381 #define NRS_ORR_HASH_FLAGS (CFS_HASH_SPIN_BKTLOCK | CFS_HASH_ASSERT_EMPTY)
383 #define NRS_TRR_BITS 4
384 #define NRS_TRR_BKT_BITS 2
385 #define NRS_TRR_HASH_FLAGS CFS_HASH_SPIN_BKTLOCK
387 static unsigned nrs_orr_hop_hash(cfs_hash_t *hs, const void *key, unsigned mask)
389 return cfs_hash_djb2_hash(key, sizeof(struct nrs_orr_key), mask);
392 static void *nrs_orr_hop_key(cfs_hlist_node_t *hnode)
394 struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
395 struct nrs_orr_object,
397 return &orro->oo_key;
400 static int nrs_orr_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
402 struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
403 struct nrs_orr_object,
406 return lu_fid_eq(&orro->oo_key.ok_fid,
407 &((struct nrs_orr_key *)key)->ok_fid);
410 static void *nrs_orr_hop_object(cfs_hlist_node_t *hnode)
412 return cfs_hlist_entry(hnode, struct nrs_orr_object, oo_hnode);
415 static void nrs_orr_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
417 struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
418 struct nrs_orr_object,
424 * Removes an nrs_orr_object the hash and frees its memory, if the object has
427 static void nrs_orr_hop_put_free(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
429 struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
430 struct nrs_orr_object,
432 struct nrs_orr_data *orrd = container_of(orro->oo_res.res_parent,
433 struct nrs_orr_data, od_res);
436 cfs_hash_bd_get_and_lock(hs, &orro->oo_key, &bd, 1);
438 if (--orro->oo_ref > 1) {
439 cfs_hash_bd_unlock(hs, &bd, 1);
443 LASSERT(orro->oo_ref == 1);
445 cfs_hash_bd_del_locked(hs, &bd, hnode);
446 cfs_hash_bd_unlock(hs, &bd, 1);
448 OBD_SLAB_FREE_PTR(orro, orrd->od_cache);
451 static void nrs_orr_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
453 struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
454 struct nrs_orr_object,
459 static int nrs_trr_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
461 struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
462 struct nrs_orr_object,
465 return orro->oo_key.ok_idx == ((struct nrs_orr_key *)key)->ok_idx;
468 static void nrs_trr_hop_exit(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
470 struct nrs_orr_object *orro = cfs_hlist_entry(hnode,
471 struct nrs_orr_object,
473 struct nrs_orr_data *orrd = container_of(orro->oo_res.res_parent,
474 struct nrs_orr_data, od_res);
476 LASSERTF(orro->oo_ref == 0,
477 "Busy NRS TRR policy object for OST with index %u, with %ld "
478 "refs\n", orro->oo_key.ok_idx, orro->oo_ref);
480 OBD_SLAB_FREE_PTR(orro, orrd->od_cache);
483 static cfs_hash_ops_t nrs_orr_hash_ops = {
484 .hs_hash = nrs_orr_hop_hash,
485 .hs_key = nrs_orr_hop_key,
486 .hs_keycmp = nrs_orr_hop_keycmp,
487 .hs_object = nrs_orr_hop_object,
488 .hs_get = nrs_orr_hop_get,
489 .hs_put = nrs_orr_hop_put_free,
490 .hs_put_locked = nrs_orr_hop_put,
493 static cfs_hash_ops_t nrs_trr_hash_ops = {
494 .hs_hash = nrs_orr_hop_hash,
495 .hs_key = nrs_orr_hop_key,
496 .hs_keycmp = nrs_trr_hop_keycmp,
497 .hs_object = nrs_orr_hop_object,
498 .hs_get = nrs_orr_hop_get,
499 .hs_put = nrs_orr_hop_put,
500 .hs_put_locked = nrs_orr_hop_put,
501 .hs_exit = nrs_trr_hop_exit,
504 #define NRS_ORR_QUANTUM_DFLT 256
507 * Binary heap predicate.
510 * ptlrpc_nrs_request::nr_u::orr::or_round,
511 * ptlrpc_nrs_request::nr_u::orr::or_sequence, and
512 * ptlrpc_nrs_request::nr_u::orr::or_range to compare two binheap nodes and
513 * produce a binary predicate that indicates their relative priority, so that
514 * the binary heap can perform the necessary sorting operations.
516 * \param[in] e1 the first binheap node to compare
517 * \param[in] e2 the second binheap node to compare
522 static int orr_req_compare(cfs_binheap_node_t *e1, cfs_binheap_node_t *e2)
524 struct ptlrpc_nrs_request *nrq1;
525 struct ptlrpc_nrs_request *nrq2;
527 nrq1 = container_of(e1, struct ptlrpc_nrs_request, nr_node);
528 nrq2 = container_of(e2, struct ptlrpc_nrs_request, nr_node);
531 * Requests have been scheduled against a different scheduling round.
533 if (nrq1->nr_u.orr.or_round < nrq2->nr_u.orr.or_round)
535 else if (nrq1->nr_u.orr.or_round > nrq2->nr_u.orr.or_round)
539 * Requests have been scheduled against the same scheduling round, but
540 * belong to a different batch, i.e. they pertain to a different
541 * backend-fs object (for ORR policy instances) or OST (for TRR policy
544 if (nrq1->nr_u.orr.or_sequence < nrq2->nr_u.orr.or_sequence)
546 else if (nrq1->nr_u.orr.or_sequence > nrq2->nr_u.orr.or_sequence)
550 * If round numbers and sequence numbers are equal, the two requests
551 * have been scheduled on the same round, and belong to the same batch,
552 * which means they pertain to the same backend-fs object (if this is an
553 * ORR policy instance), or to the same OST (if this is a TRR policy
554 * instance), so these requests should be sorted by ascending offset
557 if (nrq1->nr_u.orr.or_range.or_start <
558 nrq2->nr_u.orr.or_range.or_start) {
560 } else if (nrq1->nr_u.orr.or_range.or_start >
561 nrq2->nr_u.orr.or_range.or_start) {
565 * Requests start from the same offset; Dispatch the shorter one
566 * first; perhaps slightly more chances of hitting caches like
569 return nrq1->nr_u.orr.or_range.or_end <
570 nrq2->nr_u.orr.or_range.or_end;
575 * ORR binary heap operations
577 static cfs_binheap_ops_t nrs_orr_heap_ops = {
580 .hop_compare = orr_req_compare,
584 * Prints a warning message if an ORR/TRR policy is started on a service with
585 * more than one CPT. Not printed on the console for now, since we don't
586 * have any performance metrics in the first place, and it is annoying.
588 * \param[in] policy the policy instance
592 static int nrs_orr_init(struct ptlrpc_nrs_policy *policy)
594 if (policy->pol_nrs->nrs_svcpt->scp_service->srv_ncpts > 1)
595 CDEBUG(D_CONFIG, "%s: The %s NRS policy was registered on a "
596 "service with multiple service partitions. This policy "
597 "may perform better with a single partition.\n",
598 policy->pol_nrs->nrs_svcpt->scp_service->srv_name,
599 policy->pol_desc->pd_name);
605 * Called when an ORR policy instance is started.
607 * \param[in] policy the policy
609 * \retval -ENOMEM OOM error
612 static int nrs_orr_start(struct ptlrpc_nrs_policy *policy)
614 struct nrs_orr_data *orrd;
623 OBD_CPT_ALLOC_PTR(orrd, nrs_pol2cptab(policy), nrs_pol2cptid(policy));
628 * Binary heap instance for sorted incoming requests.
630 orrd->od_binheap = cfs_binheap_create(&nrs_orr_heap_ops,
631 CBH_FLAG_ATOMIC_GROW, 4096, NULL,
632 nrs_pol2cptab(policy),
633 nrs_pol2cptid(policy));
634 if (orrd->od_binheap == NULL)
635 GOTO(failed, rc = -ENOMEM);
637 nrs_orr_genobjname(policy, orrd->od_objname);
640 * Slab cache for NRS ORR/TRR objects.
642 orrd->od_cache = kmem_cache_create(orrd->od_objname,
643 sizeof(struct nrs_orr_object),
645 if (orrd->od_cache == NULL)
646 GOTO(failed, rc = -ENOMEM);
648 if (strncmp(policy->pol_desc->pd_name, NRS_POL_NAME_ORR,
649 NRS_POL_NAME_MAX) == 0) {
650 ops = &nrs_orr_hash_ops;
651 cur_bits = NRS_ORR_BITS;
652 max_bits = NRS_ORR_BITS;
653 bkt_bits = NRS_ORR_BKT_BITS;
654 flags = NRS_ORR_HASH_FLAGS;
656 ops = &nrs_trr_hash_ops;
657 cur_bits = NRS_TRR_BITS;
658 max_bits = NRS_TRR_BITS;
659 bkt_bits = NRS_TRR_BKT_BITS;
660 flags = NRS_TRR_HASH_FLAGS;
664 * Hash for finding objects by struct nrs_orr_key.
665 * XXX: For TRR, it might be better to avoid using libcfs_hash?
666 * All that needs to be resolved are OST indices, and they
667 * will stay relatively stable during an OSS node's lifetime.
669 orrd->od_obj_hash = cfs_hash_create(orrd->od_objname, cur_bits,
670 max_bits, bkt_bits, 0,
672 CFS_HASH_MAX_THETA, ops, flags);
673 if (orrd->od_obj_hash == NULL)
674 GOTO(failed, rc = -ENOMEM);
676 /* XXX: Fields accessed unlocked */
677 orrd->od_quantum = NRS_ORR_QUANTUM_DFLT;
678 orrd->od_supp = NOS_DFLT;
679 orrd->od_physical = true;
681 * Set to 1 so that the test inside nrs_orr_req_add() can evaluate to
684 orrd->od_sequence = 1;
686 policy->pol_private = orrd;
691 if (orrd->od_cache) {
692 kmem_cache_destroy(orrd->od_cache);
693 LASSERTF(rc == 0, "Could not destroy od_cache slab\n");
695 if (orrd->od_binheap != NULL)
696 cfs_binheap_destroy(orrd->od_binheap);
704 * Called when an ORR/TRR policy instance is stopped.
706 * Called when the policy has been instructed to transition to the
707 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state and has no more
708 * pending requests to serve.
710 * \param[in] policy the policy
712 static void nrs_orr_stop(struct ptlrpc_nrs_policy *policy)
714 struct nrs_orr_data *orrd = policy->pol_private;
717 LASSERT(orrd != NULL);
718 LASSERT(orrd->od_binheap != NULL);
719 LASSERT(orrd->od_obj_hash != NULL);
720 LASSERT(orrd->od_cache != NULL);
721 LASSERT(cfs_binheap_is_empty(orrd->od_binheap));
723 cfs_binheap_destroy(orrd->od_binheap);
724 cfs_hash_putref(orrd->od_obj_hash);
725 kmem_cache_destroy(orrd->od_cache);
731 * Performs a policy-specific ctl function on ORR/TRR policy instances; similar
734 * \param[in] policy the policy instance
735 * \param[in] opc the opcode
736 * \param[in,out] arg used for passing parameters and information
738 * \pre spin_is_locked(&policy->pol_nrs->->nrs_lock)
739 * \post spin_is_locked(&policy->pol_nrs->->nrs_lock)
741 * \retval 0 operation carried successfully
744 int nrs_orr_ctl(struct ptlrpc_nrs_policy *policy, enum ptlrpc_nrs_ctl opc,
747 LASSERT(spin_is_locked(&policy->pol_nrs->nrs_lock));
749 switch((enum nrs_ctl_orr)opc) {
753 case NRS_CTL_ORR_RD_QUANTUM: {
754 struct nrs_orr_data *orrd = policy->pol_private;
756 *(__u16 *)arg = orrd->od_quantum;
760 case NRS_CTL_ORR_WR_QUANTUM: {
761 struct nrs_orr_data *orrd = policy->pol_private;
763 orrd->od_quantum = *(__u16 *)arg;
764 LASSERT(orrd->od_quantum != 0);
768 case NRS_CTL_ORR_RD_OFF_TYPE: {
769 struct nrs_orr_data *orrd = policy->pol_private;
771 *(bool *)arg = orrd->od_physical;
775 case NRS_CTL_ORR_WR_OFF_TYPE: {
776 struct nrs_orr_data *orrd = policy->pol_private;
778 orrd->od_physical = *(bool *)arg;
782 case NRS_CTL_ORR_RD_SUPP_REQ: {
783 struct nrs_orr_data *orrd = policy->pol_private;
785 *(enum nrs_orr_supp *)arg = orrd->od_supp;
789 case NRS_CTL_ORR_WR_SUPP_REQ: {
790 struct nrs_orr_data *orrd = policy->pol_private;
792 orrd->od_supp = *(enum nrs_orr_supp *)arg;
793 LASSERT((orrd->od_supp & NOS_OST_RW) != 0);
801 * Obtains resources for ORR/TRR policy instances. The top-level resource lives
802 * inside \e nrs_orr_data and the second-level resource inside
803 * \e nrs_orr_object instances.
805 * \param[in] policy the policy for which resources are being taken for
807 * \param[in] nrq the request for which resources are being taken
808 * \param[in] parent parent resource, embedded in nrs_orr_data for the
810 * \param[out] resp used to return resource references
811 * \param[in] moving_req signifies limited caller context; used to perform
812 * memory allocations in an atomic context in this
815 * \retval 0 we are returning a top-level, parent resource, one that is
816 * embedded in an nrs_orr_data object
817 * \retval 1 we are returning a bottom-level resource, one that is embedded
818 * in an nrs_orr_object object
820 * \see nrs_resource_get_safe()
822 int nrs_orr_res_get(struct ptlrpc_nrs_policy *policy,
823 struct ptlrpc_nrs_request *nrq,
824 const struct ptlrpc_nrs_resource *parent,
825 struct ptlrpc_nrs_resource **resp, bool moving_req)
827 struct nrs_orr_data *orrd;
828 struct nrs_orr_object *orro;
829 struct nrs_orr_object *tmp;
830 struct nrs_orr_key key = { { { 0 } } };
835 * struct nrs_orr_data is requested.
837 if (parent == NULL) {
838 *resp = &((struct nrs_orr_data *)policy->pol_private)->od_res;
842 orrd = container_of(parent, struct nrs_orr_data, od_res);
845 * If the request type is not supported, fail the enqueuing; the RPC
846 * will be handled by the fallback NRS policy.
848 if (!nrs_orr_req_supported(orrd, nrq, &opc))
852 * Fill in the key for the request; OST FID for ORR policy instances,
853 * and OST index for TRR policy instances.
855 rc = nrs_orr_key_fill(orrd, nrq, opc, policy->pol_desc->pd_name, &key);
860 * Set the offset range the request covers
862 rc = nrs_orr_range_fill(nrq, orrd, opc, moving_req);
866 orro = cfs_hash_lookup(orrd->od_obj_hash, &key);
870 OBD_SLAB_CPT_ALLOC_PTR_GFP(orro, orrd->od_cache,
871 nrs_pol2cptab(policy), nrs_pol2cptid(policy),
872 moving_req ? GFP_ATOMIC : __GFP_IO);
879 tmp = cfs_hash_findadd_unique(orrd->od_obj_hash, &orro->oo_key,
882 OBD_SLAB_FREE_PTR(orro, orrd->od_cache);
887 * For debugging purposes
889 nrq->nr_u.orr.or_key = orro->oo_key;
891 *resp = &orro->oo_res;
897 * Called when releasing references to the resource hierachy obtained for a
898 * request for scheduling using ORR/TRR policy instances
900 * \param[in] policy the policy the resource belongs to
901 * \param[in] res the resource to be released
903 static void nrs_orr_res_put(struct ptlrpc_nrs_policy *policy,
904 const struct ptlrpc_nrs_resource *res)
906 struct nrs_orr_data *orrd;
907 struct nrs_orr_object *orro;
910 * Do nothing for freeing parent, nrs_orr_data resources.
912 if (res->res_parent == NULL)
915 orro = container_of(res, struct nrs_orr_object, oo_res);
916 orrd = container_of(res->res_parent, struct nrs_orr_data, od_res);
918 cfs_hash_put(orrd->od_obj_hash, &orro->oo_hnode);
922 * Called when polling an ORR/TRR policy instance for a request so that it can
923 * be served. Returns the request that is at the root of the binary heap, as
924 * that is the lowest priority one (i.e. libcfs_heap is an implementation of a
927 * \param[in] policy the policy instance being polled
928 * \param[in] peek when set, signifies that we just want to examine the
929 * request, and not handle it, so the request is not removed
931 * \param[in] force force the policy to return a request; unused in this policy
933 * \retval the request to be handled
934 * \retval NULL no request available
936 * \see ptlrpc_nrs_req_get_nolock()
937 * \see nrs_request_get()
940 struct ptlrpc_nrs_request *nrs_orr_req_get(struct ptlrpc_nrs_policy *policy,
941 bool peek, bool force)
943 struct nrs_orr_data *orrd = policy->pol_private;
944 cfs_binheap_node_t *node = cfs_binheap_root(orrd->od_binheap);
945 struct ptlrpc_nrs_request *nrq;
947 nrq = unlikely(node == NULL) ? NULL :
948 container_of(node, struct ptlrpc_nrs_request, nr_node);
950 if (likely(!peek && nrq != NULL)) {
951 struct nrs_orr_object *orro;
953 orro = container_of(nrs_request_resource(nrq),
954 struct nrs_orr_object, oo_res);
956 LASSERT(nrq->nr_u.orr.or_round <= orro->oo_round);
958 cfs_binheap_remove(orrd->od_binheap, &nrq->nr_node);
961 if (strncmp(policy->pol_desc->pd_name, NRS_POL_NAME_ORR,
962 NRS_POL_NAME_MAX) == 0)
964 "NRS: starting to handle %s request for object "
965 "with FID "DFID", from OST with index %u, with "
966 "round "LPU64"\n", NRS_POL_NAME_ORR,
967 PFID(&orro->oo_key.ok_fid),
968 nrq->nr_u.orr.or_key.ok_idx,
969 nrq->nr_u.orr.or_round);
972 "NRS: starting to handle %s request from OST "
973 "with index %u, with round "LPU64"\n",
974 NRS_POL_NAME_TRR, nrq->nr_u.orr.or_key.ok_idx,
975 nrq->nr_u.orr.or_round);
977 /** Peek at the next request to be served */
978 node = cfs_binheap_root(orrd->od_binheap);
980 /** No more requests */
981 if (unlikely(node == NULL)) {
984 struct ptlrpc_nrs_request *next;
986 next = container_of(node, struct ptlrpc_nrs_request,
989 if (orrd->od_round < next->nr_u.orr.or_round)
990 orrd->od_round = next->nr_u.orr.or_round;
998 * Sort-adds request \a nrq to an ORR/TRR \a policy instance's set of queued
999 * requests in the policy's binary heap.
1001 * A scheduling round is a stream of requests that have been sorted in batches
1002 * according to the backend-fs object (for ORR policy instances) or OST (for TRR
1003 * policy instances) that they pertain to (as identified by its IDIF FID or OST
1004 * index respectively); there can be only one batch for each object or OST in
1005 * each round. The batches are of maximum size nrs_orr_data:od_quantum. When a
1006 * new request arrives for scheduling for an object or OST that has exhausted
1007 * its quantum in its current round, the request will be scheduled on the next
1008 * scheduling round. Requests are allowed to be scheduled against a round until
1009 * all requests for the round are serviced, so an object or OST might miss a
1010 * round if requests are not scheduled for it for a long enough period of time.
1011 * Objects or OSTs that miss a round will continue with having their next
1012 * request scheduled, starting at the round that requests are being dispatched
1013 * for, at the time of arrival of this request.
1015 * Requests are tagged with the round number and a sequence number; the sequence
1016 * number indicates the relative ordering amongst the batches of requests in a
1017 * round, and is identical for all requests in a batch, as is the round number.
1018 * The round and sequence numbers are used by orr_req_compare() in order to use
1019 * nrs_orr_data::od_binheap in order to maintain an ordered set of rounds, with
1020 * each round consisting of an ordered set of batches of requests, and each
1021 * batch consisting of an ordered set of requests according to their logical
1022 * file or physical disk offsets.
1024 * \param[in] policy the policy
1025 * \param[in] nrq the request to add
1027 * \retval 0 request successfully added
1028 * \retval != 0 error
1030 static int nrs_orr_req_add(struct ptlrpc_nrs_policy *policy,
1031 struct ptlrpc_nrs_request *nrq)
1033 struct nrs_orr_data *orrd;
1034 struct nrs_orr_object *orro;
1037 orro = container_of(nrs_request_resource(nrq),
1038 struct nrs_orr_object, oo_res);
1039 orrd = container_of(nrs_request_resource(nrq)->res_parent,
1040 struct nrs_orr_data, od_res);
1042 if (orro->oo_quantum == 0 || orro->oo_round < orrd->od_round ||
1043 (orro->oo_active == 0 && orro->oo_quantum > 0)) {
1046 * If there are no pending requests for the object/OST, but some
1047 * of its quantum still remains unused, which implies we did not
1048 * get a chance to schedule up to its maximum allowed batch size
1049 * of requests in the previous round this object/OST
1050 * participated in, schedule this next request on a new round;
1051 * this avoids fragmentation of request batches caused by
1052 * intermittent inactivity on the object/OST, at the expense of
1053 * potentially slightly increased service time for the request
1054 * batch this request will be a part of.
1056 if (orro->oo_active == 0 && orro->oo_quantum > 0)
1059 /** A new scheduling round has commenced */
1060 if (orro->oo_round < orrd->od_round)
1061 orro->oo_round = orrd->od_round;
1063 /** I was not the last object/OST that scheduled a request */
1064 if (orro->oo_sequence < orrd->od_sequence)
1065 orro->oo_sequence = ++orrd->od_sequence;
1067 * Reset the quantum if we have reached the maximum quantum
1068 * size for this batch, or even if we have not managed to
1069 * complete a batch size up to its maximum allowed size.
1070 * XXX: Accessed unlocked
1072 orro->oo_quantum = orrd->od_quantum;
1075 nrq->nr_u.orr.or_round = orro->oo_round;
1076 nrq->nr_u.orr.or_sequence = orro->oo_sequence;
1078 rc = cfs_binheap_insert(orrd->od_binheap, &nrq->nr_node);
1081 if (--orro->oo_quantum == 0)
1088 * Removes request \a nrq from an ORR/TRR \a policy instance's set of queued
1091 * \param[in] policy the policy
1092 * \param[in] nrq the request to remove
1094 static void nrs_orr_req_del(struct ptlrpc_nrs_policy *policy,
1095 struct ptlrpc_nrs_request *nrq)
1097 struct nrs_orr_data *orrd;
1098 struct nrs_orr_object *orro;
1101 orro = container_of(nrs_request_resource(nrq),
1102 struct nrs_orr_object, oo_res);
1103 orrd = container_of(nrs_request_resource(nrq)->res_parent,
1104 struct nrs_orr_data, od_res);
1106 LASSERT(nrq->nr_u.orr.or_round <= orro->oo_round);
1108 is_root = &nrq->nr_node == cfs_binheap_root(orrd->od_binheap);
1110 cfs_binheap_remove(orrd->od_binheap, &nrq->nr_node);
1114 * If we just deleted the node at the root of the binheap, we may have
1115 * to adjust round numbers.
1117 if (unlikely(is_root)) {
1118 /** Peek at the next request to be served */
1119 cfs_binheap_node_t *node = cfs_binheap_root(orrd->od_binheap);
1121 /** No more requests */
1122 if (unlikely(node == NULL)) {
1125 nrq = container_of(node, struct ptlrpc_nrs_request,
1128 if (orrd->od_round < nrq->nr_u.orr.or_round)
1129 orrd->od_round = nrq->nr_u.orr.or_round;
1135 * Called right after the request \a nrq finishes being handled by ORR policy
1136 * instance \a policy.
1138 * \param[in] policy the policy that handled the request
1139 * \param[in] nrq the request that was handled
1141 static void nrs_orr_req_stop(struct ptlrpc_nrs_policy *policy,
1142 struct ptlrpc_nrs_request *nrq)
1144 /** NB: resource control, credits etc can be added here */
1145 if (strncmp(policy->pol_desc->pd_name, NRS_POL_NAME_ORR,
1146 NRS_POL_NAME_MAX) == 0)
1148 "NRS: finished handling %s request for object with FID "
1149 DFID", from OST with index %u, with round "LPU64"\n",
1150 NRS_POL_NAME_ORR, PFID(&nrq->nr_u.orr.or_key.ok_fid),
1151 nrq->nr_u.orr.or_key.ok_idx, nrq->nr_u.orr.or_round);
1154 "NRS: finished handling %s request from OST with index %u,"
1155 " with round "LPU64"\n",
1156 NRS_POL_NAME_TRR, nrq->nr_u.orr.or_key.ok_idx,
1157 nrq->nr_u.orr.or_round);
1167 * This allows to bundle the policy name into the lprocfs_vars::data pointer
1168 * so that lprocfs read/write functions can be used by both the ORR and TRR
1171 struct nrs_lprocfs_orr_data {
1172 struct ptlrpc_service *svc;
1174 } lprocfs_orr_data = {
1175 .name = NRS_POL_NAME_ORR
1176 }, lprocfs_trr_data = {
1177 .name = NRS_POL_NAME_TRR
1181 * Retrieves the value of the Round Robin quantum (i.e. the maximum batch size)
1182 * for ORR/TRR policy instances on both the regular and high-priority NRS head
1183 * of a service, as long as a policy instance is not in the
1184 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state; policy instances in this
1185 * state are skipped later by nrs_orr_ctl().
1187 * Quantum values are in # of RPCs, and the output is in YAML format.
1194 * XXX: the CRR-N version of this, ptlrpc_lprocfs_rd_nrs_crrn_quantum() is
1195 * almost identical; it can be reworked and then reused for ORR/TRR.
1197 static int ptlrpc_lprocfs_rd_nrs_orr_quantum(char *page, char **start,
1198 off_t off, int count, int *eof,
1201 struct nrs_lprocfs_orr_data *orr_data = data;
1202 struct ptlrpc_service *svc = orr_data->svc;
1208 * Perform two separate calls to this as only one of the NRS heads'
1209 * policies may be in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
1210 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING state.
1212 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1214 NRS_CTL_ORR_RD_QUANTUM,
1218 rc2 = snprintf(page, count, NRS_LPROCFS_QUANTUM_NAME_REG
1221 * Ignore -ENODEV as the regular NRS head's policy may be in the
1222 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1224 } else if (rc != -ENODEV) {
1229 * We know the ost_io service which is the only one ORR/TRR policies are
1230 * compatible with, do have an HP NRS head, but it may be best to guard
1231 * against a possible change of this in the future.
1233 if (!nrs_svc_has_hp(svc))
1236 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1237 orr_data->name, NRS_CTL_ORR_RD_QUANTUM,
1241 rc2 += snprintf(page + rc2, count - rc2,
1242 NRS_LPROCFS_QUANTUM_NAME_HP"%-5d\n", quantum);
1244 * Ignore -ENODEV as the high priority NRS head's policy may be
1245 * in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1247 } else if (rc != -ENODEV) {
1257 * Sets the value of the Round Robin quantum (i.e. the maximum batch size)
1258 * for ORR/TRR policy instances of a service. The user can set the quantum size
1259 * for the regular and high priority NRS head separately by specifying each
1260 * value, or both together in a single invocation.
1264 * lctl set_param ost.OSS.ost_io.nrs_orr_quantum=req_quantum:64, to set the
1265 * request quantum size of the ORR policy instance on the regular NRS head of
1266 * the ost_io service to 64
1268 * lctl set_param ost.OSS.ost_io.nrs_trr_quantum=hp_quantum:8 to set the request
1269 * quantum size of the TRR policy instance on the high priority NRS head of the
1270 * ost_io service to 8
1272 * lctl set_param ost.OSS.ost_io.nrs_orr_quantum=32, to set both the request
1273 * quantum size of the ORR policy instance on both the regular and the high
1274 * priority NRS head of the ost_io service to 32
1276 * policy instances in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state
1277 * are skipped later by nrs_orr_ctl().
1279 * XXX: the CRR-N version of this, ptlrpc_lprocfs_wr_nrs_crrn_quantum() is
1280 * almost identical; it can be reworked and then reused for ORR/TRR.
1282 static int ptlrpc_lprocfs_wr_nrs_orr_quantum(struct file *file,
1284 unsigned long count, void *data)
1286 struct nrs_lprocfs_orr_data *orr_data = data;
1287 struct ptlrpc_service *svc = orr_data->svc;
1288 enum ptlrpc_nrs_queue_type queue = 0;
1289 char kernbuf[LPROCFS_NRS_WR_QUANTUM_MAX_CMD];
1293 /** lprocfs_find_named_value() modifies its argument, so keep a copy */
1294 unsigned long count_copy;
1298 if (count > (sizeof(kernbuf) - 1))
1301 if (copy_from_user(kernbuf, buffer, count))
1304 kernbuf[count] = '\0';
1309 * Check if the regular quantum value has been specified
1311 val = lprocfs_find_named_value(kernbuf, NRS_LPROCFS_QUANTUM_NAME_REG,
1313 if (val != kernbuf) {
1314 quantum_reg = simple_strtol(val, NULL, 10);
1316 queue |= PTLRPC_NRS_QUEUE_REG;
1322 * Check if the high priority quantum value has been specified
1324 val = lprocfs_find_named_value(kernbuf, NRS_LPROCFS_QUANTUM_NAME_HP,
1326 if (val != kernbuf) {
1327 if (!nrs_svc_has_hp(svc))
1330 quantum_hp = simple_strtol(val, NULL, 10);
1332 queue |= PTLRPC_NRS_QUEUE_HP;
1336 * If none of the queues has been specified, look for a valid numerical
1340 if (!isdigit(kernbuf[0]))
1343 quantum_reg = simple_strtol(kernbuf, NULL, 10);
1345 queue = PTLRPC_NRS_QUEUE_REG;
1347 if (nrs_svc_has_hp(svc)) {
1348 queue |= PTLRPC_NRS_QUEUE_HP;
1349 quantum_hp = quantum_reg;
1353 if ((((queue & PTLRPC_NRS_QUEUE_REG) != 0) &&
1354 ((quantum_reg > LPROCFS_NRS_QUANTUM_MAX || quantum_reg <= 0))) ||
1355 (((queue & PTLRPC_NRS_QUEUE_HP) != 0) &&
1356 ((quantum_hp > LPROCFS_NRS_QUANTUM_MAX || quantum_hp <= 0))))
1360 * We change the values on regular and HP NRS heads separately, so that
1361 * we do not exit early from ptlrpc_nrs_policy_control() with an error
1362 * returned by nrs_policy_ctl_locked(), in cases where the user has not
1363 * started the policy on either the regular or HP NRS head; i.e. we are
1364 * ignoring -ENODEV within nrs_policy_ctl_locked(). -ENODEV is returned
1365 * only if the operation fails with -ENODEV on all heads that have been
1366 * specified by the command; if at least one operation succeeds,
1367 * success is returned.
1369 if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1370 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1372 NRS_CTL_ORR_WR_QUANTUM, false,
1374 if ((rc < 0 && rc != -ENODEV) ||
1375 (rc == -ENODEV && queue == PTLRPC_NRS_QUEUE_REG))
1379 if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1380 rc2 = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1382 NRS_CTL_ORR_WR_QUANTUM, false,
1384 if ((rc2 < 0 && rc2 != -ENODEV) ||
1385 (rc2 == -ENODEV && queue == PTLRPC_NRS_QUEUE_HP))
1389 return rc == -ENODEV && rc2 == -ENODEV ? -ENODEV : count;
1392 #define LPROCFS_NRS_OFF_NAME_REG "reg_offset_type:"
1393 #define LPROCFS_NRS_OFF_NAME_HP "hp_offset_type:"
1395 #define LPROCFS_NRS_OFF_NAME_PHYSICAL "physical"
1396 #define LPROCFS_NRS_OFF_NAME_LOGICAL "logical"
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().
1404 * Offset type information is a (physical|logical) string, and output is
1409 * reg_offset_type:physical
1410 * hp_offset_type:logical
1412 static int ptlrpc_lprocfs_rd_nrs_orr_offset_type(char *page, char **start,
1413 off_t off, int count, int *eof,
1416 struct nrs_lprocfs_orr_data *orr_data = data;
1417 struct ptlrpc_service *svc = orr_data->svc;
1423 * Perform two separate calls to this as only one of the NRS heads'
1424 * policies may be in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED
1425 * or ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING state.
1427 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1428 orr_data->name, NRS_CTL_ORR_RD_OFF_TYPE,
1432 rc2 = snprintf(page, count,
1433 LPROCFS_NRS_OFF_NAME_REG"%s\n",
1434 physical ? LPROCFS_NRS_OFF_NAME_PHYSICAL :
1435 LPROCFS_NRS_OFF_NAME_LOGICAL);
1437 * Ignore -ENODEV as the regular NRS head's policy may be in the
1438 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1440 } else if (rc != -ENODEV) {
1445 * We know the ost_io service which is the only one ORR/TRR policies are
1446 * compatible with, do have an HP NRS head, but it may be best to guard
1447 * against a possible change of this in the future.
1449 if (!nrs_svc_has_hp(svc))
1452 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1453 orr_data->name, NRS_CTL_ORR_RD_OFF_TYPE,
1457 rc2 += snprintf(page + rc2, count - rc2,
1458 LPROCFS_NRS_OFF_NAME_HP"%s\n",
1459 physical ? LPROCFS_NRS_OFF_NAME_PHYSICAL :
1460 LPROCFS_NRS_OFF_NAME_LOGICAL);
1462 * Ignore -ENODEV as the high priority NRS head's policy may be
1463 * in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1465 } else if (rc != -ENODEV) {
1475 * Max valid command string is the size of the labels, plus "physical" twice.
1476 * plus a separating ' '
1478 #define LPROCFS_NRS_WR_OFF_TYPE_MAX_CMD \
1479 sizeof(LPROCFS_NRS_OFF_NAME_REG LPROCFS_NRS_OFF_NAME_PHYSICAL " " \
1480 LPROCFS_NRS_OFF_NAME_HP LPROCFS_NRS_OFF_NAME_PHYSICAL)
1483 * Sets the type of offsets used to order RPCs in ORR/TRR policy instances. The
1484 * user can set offset type for the regular or high priority NRS head
1485 * separately by specifying each value, or both together in a single invocation.
1489 * lctl set_param ost.OSS.ost_io.nrs_orr_offset_type=
1490 * reg_offset_type:physical, to enable the ORR policy instance on the regular
1491 * NRS head of the ost_io service to use physical disk offset ordering.
1493 * lctl set_param ost.OSS.ost_io.nrs_trr_offset_type=logical, to enable the TRR
1494 * policy instances on both the regular ang high priority NRS heads of the
1495 * ost_io service to use logical file offset ordering.
1497 * policy instances in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state are
1498 * are skipped later by nrs_orr_ctl().
1500 static int ptlrpc_lprocfs_wr_nrs_orr_offset_type(struct file *file,
1502 unsigned long count,
1505 struct nrs_lprocfs_orr_data *orr_data = data;
1506 struct ptlrpc_service *svc = orr_data->svc;
1507 enum ptlrpc_nrs_queue_type queue = 0;
1508 char kernbuf[LPROCFS_NRS_WR_OFF_TYPE_MAX_CMD];
1513 unsigned long count_copy;
1517 if (count > (sizeof(kernbuf) - 1))
1520 if (copy_from_user(kernbuf, buffer, count))
1523 kernbuf[count] = '\0';
1528 * Check if the regular offset type has been specified
1530 val_reg = lprocfs_find_named_value(kernbuf,
1531 LPROCFS_NRS_OFF_NAME_REG,
1533 if (val_reg != kernbuf)
1534 queue |= PTLRPC_NRS_QUEUE_REG;
1539 * Check if the high priority offset type has been specified
1541 val_hp = lprocfs_find_named_value(kernbuf, LPROCFS_NRS_OFF_NAME_HP,
1543 if (val_hp != kernbuf) {
1544 if (!nrs_svc_has_hp(svc))
1547 queue |= PTLRPC_NRS_QUEUE_HP;
1551 * If none of the queues has been specified, there may be a valid
1552 * command string at the start of the buffer.
1555 queue = PTLRPC_NRS_QUEUE_REG;
1557 if (nrs_svc_has_hp(svc))
1558 queue |= PTLRPC_NRS_QUEUE_HP;
1561 if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1562 if (strncmp(val_reg, LPROCFS_NRS_OFF_NAME_PHYSICAL,
1563 sizeof(LPROCFS_NRS_OFF_NAME_PHYSICAL) - 1) == 0)
1564 physical_reg = true;
1565 else if (strncmp(val_reg, LPROCFS_NRS_OFF_NAME_LOGICAL,
1566 sizeof(LPROCFS_NRS_OFF_NAME_LOGICAL) - 1) == 0)
1567 physical_reg = false;
1572 if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1573 if (strncmp(val_hp, LPROCFS_NRS_OFF_NAME_PHYSICAL,
1574 sizeof(LPROCFS_NRS_OFF_NAME_PHYSICAL) - 1) == 0)
1576 else if (strncmp(val_hp, LPROCFS_NRS_OFF_NAME_LOGICAL,
1577 sizeof(LPROCFS_NRS_OFF_NAME_LOGICAL) - 1) == 0)
1578 physical_hp = false;
1584 * We change the values on regular and HP NRS heads separately, so that
1585 * we do not exit early from ptlrpc_nrs_policy_control() with an error
1586 * returned by nrs_policy_ctl_locked(), in cases where the user has not
1587 * started the policy on either the regular or HP NRS head; i.e. we are
1588 * ignoring -ENODEV within nrs_policy_ctl_locked(). -ENODEV is returned
1589 * only if the operation fails with -ENODEV on all heads that have been
1590 * specified by the command; if at least one operation succeeds,
1591 * success is returned.
1593 if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1594 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1596 NRS_CTL_ORR_WR_OFF_TYPE, false,
1598 if ((rc < 0 && rc != -ENODEV) ||
1599 (rc == -ENODEV && queue == PTLRPC_NRS_QUEUE_REG))
1603 if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1604 rc2 = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1606 NRS_CTL_ORR_WR_OFF_TYPE, false,
1608 if ((rc2 < 0 && rc2 != -ENODEV) ||
1609 (rc2 == -ENODEV && queue == PTLRPC_NRS_QUEUE_HP))
1613 return rc == -ENODEV && rc2 == -ENODEV ? -ENODEV : count;
1616 #define NRS_LPROCFS_REQ_SUPP_NAME_REG "reg_supported:"
1617 #define NRS_LPROCFS_REQ_SUPP_NAME_HP "hp_supported:"
1619 #define LPROCFS_NRS_SUPP_NAME_READS "reads"
1620 #define LPROCFS_NRS_SUPP_NAME_WRITES "writes"
1621 #define LPROCFS_NRS_SUPP_NAME_READWRITES "reads_and_writes"
1624 * Translates enum nrs_orr_supp values to a corresponding string.
1626 static const char *nrs_orr_supp2str(enum nrs_orr_supp supp)
1632 return LPROCFS_NRS_SUPP_NAME_READS;
1634 return LPROCFS_NRS_SUPP_NAME_WRITES;
1636 return LPROCFS_NRS_SUPP_NAME_READWRITES;
1641 * Translates strings to the corresponding enum nrs_orr_supp value
1643 static enum nrs_orr_supp nrs_orr_str2supp(const char *val)
1645 if (strncmp(val, LPROCFS_NRS_SUPP_NAME_READWRITES,
1646 sizeof(LPROCFS_NRS_SUPP_NAME_READWRITES) - 1) == 0)
1648 else if (strncmp(val, LPROCFS_NRS_SUPP_NAME_READS,
1649 sizeof(LPROCFS_NRS_SUPP_NAME_READS) - 1) == 0)
1650 return NOS_OST_READ;
1651 else if (strncmp(val, LPROCFS_NRS_SUPP_NAME_WRITES,
1652 sizeof(LPROCFS_NRS_SUPP_NAME_WRITES) - 1) == 0)
1653 return NOS_OST_WRITE;
1659 * Retrieves the type of RPCs handled at the point of invocation by ORR/TRR
1660 * policy instances on both the regular and high-priority NRS head of a service,
1661 * as long as a policy instance is not in the
1662 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state; policy instances in this
1663 * state are skipped later by nrs_orr_ctl().
1665 * Supported RPC type information is a (reads|writes|reads_and_writes) string,
1666 * and output is in YAML format.
1670 * reg_supported:reads
1671 * hp_supported:reads_and_writes
1673 static int ptlrpc_lprocfs_rd_nrs_orr_supported(char *page, char **start,
1674 off_t off, int count, int *eof,
1677 struct nrs_lprocfs_orr_data *orr_data = data;
1678 struct ptlrpc_service *svc = orr_data->svc;
1679 enum nrs_orr_supp supported;
1684 * Perform two separate calls to this as only one of the NRS heads'
1685 * policies may be in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED
1686 * or ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING state.
1688 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1690 NRS_CTL_ORR_RD_SUPP_REQ, true,
1695 rc2 = snprintf(page, count,
1696 NRS_LPROCFS_REQ_SUPP_NAME_REG"%s\n",
1697 nrs_orr_supp2str(supported));
1699 * Ignore -ENODEV as the regular NRS head's policy may be in the
1700 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1702 } else if (rc != -ENODEV) {
1707 * We know the ost_io service which is the only one ORR/TRR policies are
1708 * compatible with, do have an HP NRS head, but it may be best to guard
1709 * against a possible change of this in the future.
1711 if (!nrs_svc_has_hp(svc))
1714 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1716 NRS_CTL_ORR_RD_SUPP_REQ, true,
1720 rc2 += snprintf(page + rc2, count - rc2,
1721 NRS_LPROCFS_REQ_SUPP_NAME_HP"%s\n",
1722 nrs_orr_supp2str(supported));
1724 * Ignore -ENODEV as the high priority NRS head's policy may be
1725 * in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1727 } else if (rc != -ENODEV) {
1737 * Max valid command string is the size of the labels, plus "reads_and_writes"
1738 * twice, plus a separating ' '
1740 #define LPROCFS_NRS_WR_REQ_SUPP_MAX_CMD \
1741 sizeof(NRS_LPROCFS_REQ_SUPP_NAME_REG LPROCFS_NRS_SUPP_NAME_READWRITES \
1742 NRS_LPROCFS_REQ_SUPP_NAME_HP LPROCFS_NRS_SUPP_NAME_READWRITES \
1746 * Sets the type of RPCs handled by ORR/TRR policy instances. The user can
1747 * modify this setting for the regular or high priority NRS heads separately, or
1748 * both together in a single invocation.
1752 * lctl set_param ost.OSS.ost_io.nrs_orr_supported=
1753 * "reg_supported:reads", to enable the ORR policy instance on the regular NRS
1754 * head of the ost_io service to handle OST_READ RPCs.
1756 * lctl set_param ost.OSS.ost_io.nrs_trr_supported=reads_and_writes, to enable
1757 * the TRR policy instances on both the regular ang high priority NRS heads of
1758 * the ost_io service to use handle OST_READ and OST_WRITE RPCs.
1760 * policy instances in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state are
1761 * are skipped later by nrs_orr_ctl().
1763 static int ptlrpc_lprocfs_wr_nrs_orr_supported(struct file *file,
1765 unsigned long count, void *data)
1767 struct nrs_lprocfs_orr_data *orr_data = data;
1768 struct ptlrpc_service *svc = orr_data->svc;
1769 enum ptlrpc_nrs_queue_type queue = 0;
1770 char kernbuf[LPROCFS_NRS_WR_REQ_SUPP_MAX_CMD];
1773 enum nrs_orr_supp supp_reg;
1774 enum nrs_orr_supp supp_hp;
1775 unsigned long count_copy;
1779 if (count > (sizeof(kernbuf) - 1))
1782 if (copy_from_user(kernbuf, buffer, count))
1785 kernbuf[count] = '\0';
1790 * Check if the regular supported requests setting has been specified
1792 val_reg = lprocfs_find_named_value(kernbuf,
1793 NRS_LPROCFS_REQ_SUPP_NAME_REG,
1795 if (val_reg != kernbuf)
1796 queue |= PTLRPC_NRS_QUEUE_REG;
1801 * Check if the high priority supported requests setting has been
1804 val_hp = lprocfs_find_named_value(kernbuf, NRS_LPROCFS_REQ_SUPP_NAME_HP,
1806 if (val_hp != kernbuf) {
1807 if (!nrs_svc_has_hp(svc))
1810 queue |= PTLRPC_NRS_QUEUE_HP;
1814 * If none of the queues has been specified, there may be a valid
1815 * command string at the start of the buffer.
1818 queue = PTLRPC_NRS_QUEUE_REG;
1820 if (nrs_svc_has_hp(svc))
1821 queue |= PTLRPC_NRS_QUEUE_HP;
1824 if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1825 supp_reg = nrs_orr_str2supp(val_reg);
1826 if (supp_reg == -EINVAL)
1830 if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1831 supp_hp = nrs_orr_str2supp(val_hp);
1832 if (supp_hp == -EINVAL)
1837 * We change the values on regular and HP NRS heads separately, so that
1838 * we do not exit early from ptlrpc_nrs_policy_control() with an error
1839 * returned by nrs_policy_ctl_locked(), in cases where the user has not
1840 * started the policy on either the regular or HP NRS head; i.e. we are
1841 * ignoring -ENODEV within nrs_policy_ctl_locked(). -ENODEV is returned
1842 * only if the operation fails with -ENODEV on all heads that have been
1843 * specified by the command; if at least one operation succeeds,
1844 * success is returned.
1846 if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1847 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1849 NRS_CTL_ORR_WR_SUPP_REQ, false,
1851 if ((rc < 0 && rc != -ENODEV) ||
1852 (rc == -ENODEV && queue == PTLRPC_NRS_QUEUE_REG))
1856 if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1857 rc2 = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1859 NRS_CTL_ORR_WR_SUPP_REQ, false,
1861 if ((rc2 < 0 && rc2 != -ENODEV) ||
1862 (rc2 == -ENODEV && queue == PTLRPC_NRS_QUEUE_HP))
1866 return rc == -ENODEV && rc2 == -ENODEV ? -ENODEV : count;
1869 int nrs_orr_lprocfs_init(struct ptlrpc_service *svc)
1874 struct lprocfs_vars nrs_orr_lprocfs_vars[] = {
1875 { .name = "nrs_orr_quantum",
1876 .read_fptr = ptlrpc_lprocfs_rd_nrs_orr_quantum,
1877 .write_fptr = ptlrpc_lprocfs_wr_nrs_orr_quantum },
1878 { .name = "nrs_orr_offset_type",
1879 .read_fptr = ptlrpc_lprocfs_rd_nrs_orr_offset_type,
1880 .write_fptr = ptlrpc_lprocfs_wr_nrs_orr_offset_type },
1881 { .name = "nrs_orr_supported",
1882 .read_fptr = ptlrpc_lprocfs_rd_nrs_orr_supported,
1883 .write_fptr = ptlrpc_lprocfs_wr_nrs_orr_supported },
1887 if (svc->srv_procroot == NULL)
1890 lprocfs_orr_data.svc = svc;
1892 for (i = 0; i < ARRAY_SIZE(nrs_orr_lprocfs_vars); i++)
1893 nrs_orr_lprocfs_vars[i].data = &lprocfs_orr_data;
1895 rc = lprocfs_add_vars(svc->srv_procroot, nrs_orr_lprocfs_vars, NULL);
1900 void nrs_orr_lprocfs_fini(struct ptlrpc_service *svc)
1902 if (svc->srv_procroot == NULL)
1905 lprocfs_remove_proc_entry("nrs_orr_quantum", svc->srv_procroot);
1906 lprocfs_remove_proc_entry("nrs_orr_offset_type", svc->srv_procroot);
1907 lprocfs_remove_proc_entry("nrs_orr_supported", svc->srv_procroot);
1910 #endif /* LPROCFS */
1912 static const struct ptlrpc_nrs_pol_ops nrs_orr_ops = {
1913 .op_policy_init = nrs_orr_init,
1914 .op_policy_start = nrs_orr_start,
1915 .op_policy_stop = nrs_orr_stop,
1916 .op_policy_ctl = nrs_orr_ctl,
1917 .op_res_get = nrs_orr_res_get,
1918 .op_res_put = nrs_orr_res_put,
1919 .op_req_get = nrs_orr_req_get,
1920 .op_req_enqueue = nrs_orr_req_add,
1921 .op_req_dequeue = nrs_orr_req_del,
1922 .op_req_stop = nrs_orr_req_stop,
1924 .op_lprocfs_init = nrs_orr_lprocfs_init,
1925 .op_lprocfs_fini = nrs_orr_lprocfs_fini,
1929 struct ptlrpc_nrs_pol_conf nrs_conf_orr = {
1930 .nc_name = NRS_POL_NAME_ORR,
1931 .nc_ops = &nrs_orr_ops,
1932 .nc_compat = nrs_policy_compat_one,
1933 .nc_compat_svc_name = "ost_io",
1937 * TRR, Target-based Round Robin policy
1939 * TRR reuses much of the functions and data structures of ORR
1944 int nrs_trr_lprocfs_init(struct ptlrpc_service *svc)
1949 struct lprocfs_vars nrs_trr_lprocfs_vars[] = {
1950 { .name = "nrs_trr_quantum",
1951 .read_fptr = ptlrpc_lprocfs_rd_nrs_orr_quantum,
1952 .write_fptr = ptlrpc_lprocfs_wr_nrs_orr_quantum },
1953 { .name = "nrs_trr_offset_type",
1954 .read_fptr = ptlrpc_lprocfs_rd_nrs_orr_offset_type,
1955 .write_fptr = ptlrpc_lprocfs_wr_nrs_orr_offset_type },
1956 { .name = "nrs_trr_supported",
1957 .read_fptr = ptlrpc_lprocfs_rd_nrs_orr_supported,
1958 .write_fptr = ptlrpc_lprocfs_wr_nrs_orr_supported },
1962 if (svc->srv_procroot == NULL)
1965 lprocfs_trr_data.svc = svc;
1967 for (i = 0; i < ARRAY_SIZE(nrs_trr_lprocfs_vars); i++)
1968 nrs_trr_lprocfs_vars[i].data = &lprocfs_trr_data;
1970 rc = lprocfs_add_vars(svc->srv_procroot, nrs_trr_lprocfs_vars, NULL);
1975 void nrs_trr_lprocfs_fini(struct ptlrpc_service *svc)
1977 if (svc->srv_procroot == NULL)
1980 lprocfs_remove_proc_entry("nrs_trr_quantum", svc->srv_procroot);
1981 lprocfs_remove_proc_entry("nrs_trr_offset_type", svc->srv_procroot);
1982 lprocfs_remove_proc_entry("nrs_trr_supported", svc->srv_procroot);
1985 #endif /* LPROCFS */
1988 * Reuse much of the ORR functionality for TRR.
1990 static const struct ptlrpc_nrs_pol_ops nrs_trr_ops = {
1991 .op_policy_init = nrs_orr_init,
1992 .op_policy_start = nrs_orr_start,
1993 .op_policy_stop = nrs_orr_stop,
1994 .op_policy_ctl = nrs_orr_ctl,
1995 .op_res_get = nrs_orr_res_get,
1996 .op_res_put = nrs_orr_res_put,
1997 .op_req_get = nrs_orr_req_get,
1998 .op_req_enqueue = nrs_orr_req_add,
1999 .op_req_dequeue = nrs_orr_req_del,
2000 .op_req_stop = nrs_orr_req_stop,
2002 .op_lprocfs_init = nrs_trr_lprocfs_init,
2003 .op_lprocfs_fini = nrs_trr_lprocfs_fini,
2007 struct ptlrpc_nrs_pol_conf nrs_conf_trr = {
2008 .nc_name = NRS_POL_NAME_TRR,
2009 .nc_ops = &nrs_trr_ops,
2010 .nc_compat = nrs_policy_compat_one,
2011 .nc_compat_svc_name = "ost_io",
2014 /** @} ORR/TRR policy */
2018 #endif /* HAVE_SERVER_SUPPORT */