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, 2015, 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 /* Bounce unconnected requests to the default policy. */
158 if (req->rq_export == NULL)
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));
164 ost_idx = class_server_data(req->rq_export->exp_obd)->lsd_osd_index;
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.
174 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
178 rc = ostid_to_fid(&key->ok_fid, &body->oa.o_oi, ost_idx);
182 nrq->nr_u.orr.or_orr_set = 1;
184 key->ok_idx = ost_idx;
185 nrq->nr_u.orr.or_trr_set = 1;
192 * Populates the range values in \a range with logical offsets obtained via
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
199 static void nrs_orr_range_fill_logical(struct niobuf_remote *nb, int niocount,
200 struct nrs_orr_req_range *range)
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;
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.
212 #define ORR_NUM_EXTENTS 1
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.
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
225 * \retval 0 physical offsets obtained successfully
228 static int nrs_orr_range_fill_physical(struct ptlrpc_nrs_request *nrq,
230 struct nrs_orr_req_range *range)
232 struct ptlrpc_request *req = container_of(nrq,
233 struct ptlrpc_request,
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;
243 key = (typeof(key)) {
244 .lfik_name = KEY_FIEMAP,
247 .fm_start = range->or_start,
248 .fm_length = range->or_end - range->or_start,
249 .fm_extent_count = ORR_NUM_EXTENTS
253 rc = obd_get_info(req->rq_svc_thread->t_env, req->rq_export,
254 sizeof(key), &key, NULL, fiemap);
258 if (fiemap->fm_mapped_extents == 0 ||
259 fiemap->fm_mapped_extents > ORR_NUM_EXTENTS)
260 GOTO(out, rc = -EFAULT);
263 * Calculate the physical offset ranges for the request from the extent
264 * information and the logical request offsets.
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;
270 range->or_start = start;
273 nrq->nr_u.orr.or_physical_set = 1;
279 * Sets the offset range the request covers; either in logical file
280 * offsets or in physical disk offsets.
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?
288 * \retval 0 range filled successfully
291 static int nrs_orr_range_fill(struct ptlrpc_nrs_request *nrq,
292 struct nrs_orr_data *orrd, __u32 opc,
295 struct ptlrpc_request *req = container_of(nrq,
296 struct ptlrpc_request,
298 struct obd_ioobj *ioo;
299 struct niobuf_remote *nb;
300 struct ost_body *body;
301 struct nrs_orr_req_range range;
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.
316 if (orrd->od_physical && nrq->nr_u.orr.or_physical_set)
319 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
321 GOTO(out, rc = -EFAULT);
323 niocount = ioo->ioo_bufcnt;
325 nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
327 GOTO(out, rc = -EFAULT);
330 * Use logical information from niobuf_remote structures.
332 nrs_orr_range_fill_logical(nb, niocount, &range);
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.
342 if (orrd->od_physical && opc == OST_READ && !moving_req) {
343 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
345 GOTO(out, rc = -EFAULT);
348 * Translate to physical block offsets from backend filesystem
350 * Ignore return values; if obtaining the physical offsets
351 * fails, use the logical offsets.
353 nrs_orr_range_fill_physical(nrq, &body->oa, &range);
356 nrq->nr_u.orr.or_range = range;
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
369 * \param[in] policy the policy instance
370 * \param[out] name the character array that will hold the generated name
372 static void nrs_orr_genobjname(struct ptlrpc_nrs_policy *policy, char *name)
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));
381 * ORR/TRR hash operations
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)
387 #define NRS_TRR_BITS 4
388 #define NRS_TRR_BKT_BITS 2
389 #define NRS_TRR_HASH_FLAGS CFS_HASH_SPIN_BKTLOCK
392 nrs_orr_hop_hash(struct cfs_hash *hs, const void *key, unsigned mask)
394 return cfs_hash_djb2_hash(key, sizeof(struct nrs_orr_key), mask);
397 static void *nrs_orr_hop_key(struct hlist_node *hnode)
399 struct nrs_orr_object *orro = hlist_entry(hnode,
400 struct nrs_orr_object,
402 return &orro->oo_key;
405 static int nrs_orr_hop_keycmp(const void *key, struct hlist_node *hnode)
407 struct nrs_orr_object *orro = hlist_entry(hnode,
408 struct nrs_orr_object,
411 return lu_fid_eq(&orro->oo_key.ok_fid,
412 &((struct nrs_orr_key *)key)->ok_fid);
415 static void *nrs_orr_hop_object(struct hlist_node *hnode)
417 return hlist_entry(hnode, struct nrs_orr_object, oo_hnode);
420 static void nrs_orr_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
422 struct nrs_orr_object *orro = hlist_entry(hnode,
423 struct nrs_orr_object,
429 * Removes an nrs_orr_object the hash and frees its memory, if the object has
432 static void nrs_orr_hop_put_free(struct cfs_hash *hs, struct hlist_node *hnode)
434 struct nrs_orr_object *orro = hlist_entry(hnode,
435 struct nrs_orr_object,
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;
441 cfs_hash_bd_get_and_lock(hs, &orro->oo_key, &bd, 1);
443 if (--orro->oo_ref > 1) {
444 cfs_hash_bd_unlock(hs, &bd, 1);
448 LASSERT(orro->oo_ref == 1);
450 cfs_hash_bd_del_locked(hs, &bd, hnode);
451 cfs_hash_bd_unlock(hs, &bd, 1);
453 OBD_SLAB_FREE_PTR(orro, orrd->od_cache);
456 static void nrs_orr_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
458 struct nrs_orr_object *orro = hlist_entry(hnode,
459 struct nrs_orr_object,
464 static int nrs_trr_hop_keycmp(const void *key, struct hlist_node *hnode)
466 struct nrs_orr_object *orro = hlist_entry(hnode,
467 struct nrs_orr_object,
470 return orro->oo_key.ok_idx == ((struct nrs_orr_key *)key)->ok_idx;
473 static void nrs_trr_hop_exit(struct cfs_hash *hs, struct hlist_node *hnode)
475 struct nrs_orr_object *orro = hlist_entry(hnode,
476 struct nrs_orr_object,
478 struct nrs_orr_data *orrd = container_of(orro->oo_res.res_parent,
479 struct nrs_orr_data, od_res);
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);
485 OBD_SLAB_FREE_PTR(orro, orrd->od_cache);
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,
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,
509 #define NRS_ORR_QUANTUM_DFLT 256
512 * Binary heap predicate.
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.
521 * \param[in] e1 the first binheap node to compare
522 * \param[in] e2 the second binheap node to compare
528 orr_req_compare(struct cfs_binheap_node *e1, struct cfs_binheap_node *e2)
530 struct ptlrpc_nrs_request *nrq1;
531 struct ptlrpc_nrs_request *nrq2;
533 nrq1 = container_of(e1, struct ptlrpc_nrs_request, nr_node);
534 nrq2 = container_of(e2, struct ptlrpc_nrs_request, nr_node);
537 * Requests have been scheduled against a different scheduling round.
539 if (nrq1->nr_u.orr.or_round < nrq2->nr_u.orr.or_round)
541 else if (nrq1->nr_u.orr.or_round > nrq2->nr_u.orr.or_round)
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
550 if (nrq1->nr_u.orr.or_sequence < nrq2->nr_u.orr.or_sequence)
552 else if (nrq1->nr_u.orr.or_sequence > nrq2->nr_u.orr.or_sequence)
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
563 if (nrq1->nr_u.orr.or_range.or_start <
564 nrq2->nr_u.orr.or_range.or_start) {
566 } else if (nrq1->nr_u.orr.or_range.or_start >
567 nrq2->nr_u.orr.or_range.or_start) {
571 * Requests start from the same offset; Dispatch the shorter one
572 * first; perhaps slightly more chances of hitting caches like
575 return nrq1->nr_u.orr.or_range.or_end <
576 nrq2->nr_u.orr.or_range.or_end;
581 * ORR binary heap operations
583 static struct cfs_binheap_ops nrs_orr_heap_ops = {
586 .hop_compare = orr_req_compare,
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.
594 * \param[in] policy the policy instance
598 static int nrs_orr_init(struct ptlrpc_nrs_policy *policy)
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);
611 * Called when an ORR policy instance is started.
613 * \param[in] policy the policy
615 * \retval -ENOMEM OOM error
618 static int nrs_orr_start(struct ptlrpc_nrs_policy *policy, char *arg)
620 struct nrs_orr_data *orrd;
621 struct cfs_hash_ops *ops;
629 OBD_CPT_ALLOC_PTR(orrd, nrs_pol2cptab(policy), nrs_pol2cptid(policy));
634 * Binary heap instance for sorted incoming requests.
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(failed, rc = -ENOMEM);
643 nrs_orr_genobjname(policy, orrd->od_objname);
646 * Slab cache for NRS ORR/TRR objects.
648 orrd->od_cache = kmem_cache_create(orrd->od_objname,
649 sizeof(struct nrs_orr_object),
651 if (orrd->od_cache == NULL)
652 GOTO(failed, rc = -ENOMEM);
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;
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;
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.
675 orrd->od_obj_hash = cfs_hash_create(orrd->od_objname, cur_bits,
676 max_bits, bkt_bits, 0,
678 CFS_HASH_MAX_THETA, ops, flags);
679 if (orrd->od_obj_hash == NULL)
680 GOTO(failed, rc = -ENOMEM);
682 /* XXX: Fields accessed unlocked */
683 orrd->od_quantum = NRS_ORR_QUANTUM_DFLT;
684 orrd->od_supp = NOS_DFLT;
685 orrd->od_physical = true;
687 * Set to 1 so that the test inside nrs_orr_req_add() can evaluate to
690 orrd->od_sequence = 1;
692 policy->pol_private = orrd;
698 kmem_cache_destroy(orrd->od_cache);
699 if (orrd->od_binheap != NULL)
700 cfs_binheap_destroy(orrd->od_binheap);
708 * Called when an ORR/TRR policy instance is stopped.
710 * Called when the policy has been instructed to transition to the
711 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state and has no more
712 * pending requests to serve.
714 * \param[in] policy the policy
716 static void nrs_orr_stop(struct ptlrpc_nrs_policy *policy)
718 struct nrs_orr_data *orrd = policy->pol_private;
721 LASSERT(orrd != NULL);
722 LASSERT(orrd->od_binheap != NULL);
723 LASSERT(orrd->od_obj_hash != NULL);
724 LASSERT(orrd->od_cache != NULL);
725 LASSERT(cfs_binheap_is_empty(orrd->od_binheap));
727 cfs_binheap_destroy(orrd->od_binheap);
728 cfs_hash_putref(orrd->od_obj_hash);
729 kmem_cache_destroy(orrd->od_cache);
735 * Performs a policy-specific ctl function on ORR/TRR policy instances; similar
738 * \param[in] policy the policy instance
739 * \param[in] opc the opcode
740 * \param[in,out] arg used for passing parameters and information
742 * \pre assert_spin_locked(&policy->pol_nrs->->nrs_lock)
743 * \post assert_spin_locked(&policy->pol_nrs->->nrs_lock)
745 * \retval 0 operation carried successfully
748 static int nrs_orr_ctl(struct ptlrpc_nrs_policy *policy,
749 enum ptlrpc_nrs_ctl opc, void *arg)
751 assert_spin_locked(&policy->pol_nrs->nrs_lock);
753 switch((enum nrs_ctl_orr)opc) {
757 case NRS_CTL_ORR_RD_QUANTUM: {
758 struct nrs_orr_data *orrd = policy->pol_private;
760 *(__u16 *)arg = orrd->od_quantum;
764 case NRS_CTL_ORR_WR_QUANTUM: {
765 struct nrs_orr_data *orrd = policy->pol_private;
767 orrd->od_quantum = *(__u16 *)arg;
768 LASSERT(orrd->od_quantum != 0);
772 case NRS_CTL_ORR_RD_OFF_TYPE: {
773 struct nrs_orr_data *orrd = policy->pol_private;
775 *(bool *)arg = orrd->od_physical;
779 case NRS_CTL_ORR_WR_OFF_TYPE: {
780 struct nrs_orr_data *orrd = policy->pol_private;
782 orrd->od_physical = *(bool *)arg;
786 case NRS_CTL_ORR_RD_SUPP_REQ: {
787 struct nrs_orr_data *orrd = policy->pol_private;
789 *(enum nrs_orr_supp *)arg = orrd->od_supp;
793 case NRS_CTL_ORR_WR_SUPP_REQ: {
794 struct nrs_orr_data *orrd = policy->pol_private;
796 orrd->od_supp = *(enum nrs_orr_supp *)arg;
797 LASSERT((orrd->od_supp & NOS_OST_RW) != 0);
805 * Obtains resources for ORR/TRR policy instances. The top-level resource lives
806 * inside \e nrs_orr_data and the second-level resource inside
807 * \e nrs_orr_object instances.
809 * \param[in] policy the policy for which resources are being taken for
811 * \param[in] nrq the request for which resources are being taken
812 * \param[in] parent parent resource, embedded in nrs_orr_data for the
814 * \param[out] resp used to return resource references
815 * \param[in] moving_req signifies limited caller context; used to perform
816 * memory allocations in an atomic context in this
819 * \retval 0 we are returning a top-level, parent resource, one that is
820 * embedded in an nrs_orr_data object
821 * \retval 1 we are returning a bottom-level resource, one that is embedded
822 * in an nrs_orr_object object
824 * \see nrs_resource_get_safe()
826 static int nrs_orr_res_get(struct ptlrpc_nrs_policy *policy,
827 struct ptlrpc_nrs_request *nrq,
828 const struct ptlrpc_nrs_resource *parent,
829 struct ptlrpc_nrs_resource **resp, bool moving_req)
831 struct nrs_orr_data *orrd;
832 struct nrs_orr_object *orro;
833 struct nrs_orr_object *tmp;
834 struct nrs_orr_key key = { { { 0 } } };
839 * struct nrs_orr_data is requested.
841 if (parent == NULL) {
842 *resp = &((struct nrs_orr_data *)policy->pol_private)->od_res;
846 orrd = container_of(parent, struct nrs_orr_data, od_res);
849 * If the request type is not supported, fail the enqueuing; the RPC
850 * will be handled by the fallback NRS policy.
852 if (!nrs_orr_req_supported(orrd, nrq, &opc))
856 * Fill in the key for the request; OST FID for ORR policy instances,
857 * and OST index for TRR policy instances.
859 rc = nrs_orr_key_fill(orrd, nrq, opc, policy->pol_desc->pd_name, &key);
864 * Set the offset range the request covers
866 rc = nrs_orr_range_fill(nrq, orrd, opc, moving_req);
870 orro = cfs_hash_lookup(orrd->od_obj_hash, &key);
874 OBD_SLAB_CPT_ALLOC_PTR_GFP(orro, orrd->od_cache,
875 nrs_pol2cptab(policy), nrs_pol2cptid(policy),
876 moving_req ? GFP_ATOMIC : GFP_NOFS);
883 tmp = cfs_hash_findadd_unique(orrd->od_obj_hash, &orro->oo_key,
886 OBD_SLAB_FREE_PTR(orro, orrd->od_cache);
891 * For debugging purposes
893 nrq->nr_u.orr.or_key = orro->oo_key;
895 *resp = &orro->oo_res;
901 * Called when releasing references to the resource hierachy obtained for a
902 * request for scheduling using ORR/TRR policy instances
904 * \param[in] policy the policy the resource belongs to
905 * \param[in] res the resource to be released
907 static void nrs_orr_res_put(struct ptlrpc_nrs_policy *policy,
908 const struct ptlrpc_nrs_resource *res)
910 struct nrs_orr_data *orrd;
911 struct nrs_orr_object *orro;
914 * Do nothing for freeing parent, nrs_orr_data resources.
916 if (res->res_parent == NULL)
919 orro = container_of(res, struct nrs_orr_object, oo_res);
920 orrd = container_of(res->res_parent, struct nrs_orr_data, od_res);
922 cfs_hash_put(orrd->od_obj_hash, &orro->oo_hnode);
926 * Called when polling an ORR/TRR policy instance for a request so that it can
927 * be served. Returns the request that is at the root of the binary heap, as
928 * that is the lowest priority one (i.e. libcfs_heap is an implementation of a
931 * \param[in] policy the policy instance being polled
932 * \param[in] peek when set, signifies that we just want to examine the
933 * request, and not handle it, so the request is not removed
935 * \param[in] force force the policy to return a request; unused in this policy
937 * \retval the request to be handled
938 * \retval NULL no request available
940 * \see ptlrpc_nrs_req_get_nolock()
941 * \see nrs_request_get()
944 struct ptlrpc_nrs_request *nrs_orr_req_get(struct ptlrpc_nrs_policy *policy,
945 bool peek, bool force)
947 struct nrs_orr_data *orrd = policy->pol_private;
948 struct cfs_binheap_node *node = cfs_binheap_root(orrd->od_binheap);
949 struct ptlrpc_nrs_request *nrq;
951 nrq = unlikely(node == NULL) ? NULL :
952 container_of(node, struct ptlrpc_nrs_request, nr_node);
954 if (likely(!peek && nrq != NULL)) {
955 struct nrs_orr_object *orro;
957 orro = container_of(nrs_request_resource(nrq),
958 struct nrs_orr_object, oo_res);
960 LASSERT(nrq->nr_u.orr.or_round <= orro->oo_round);
962 cfs_binheap_remove(orrd->od_binheap, &nrq->nr_node);
965 if (strncmp(policy->pol_desc->pd_name, NRS_POL_NAME_ORR,
966 NRS_POL_NAME_MAX) == 0)
968 "NRS: starting to handle %s request for object "
969 "with FID "DFID", from OST with index %u, with "
970 "round %llu\n", NRS_POL_NAME_ORR,
971 PFID(&orro->oo_key.ok_fid),
972 nrq->nr_u.orr.or_key.ok_idx,
973 nrq->nr_u.orr.or_round);
976 "NRS: starting to handle %s request from OST "
977 "with index %u, with round %llu\n",
978 NRS_POL_NAME_TRR, nrq->nr_u.orr.or_key.ok_idx,
979 nrq->nr_u.orr.or_round);
981 /** Peek at the next request to be served */
982 node = cfs_binheap_root(orrd->od_binheap);
984 /** No more requests */
985 if (unlikely(node == NULL)) {
988 struct ptlrpc_nrs_request *next;
990 next = container_of(node, struct ptlrpc_nrs_request,
993 if (orrd->od_round < next->nr_u.orr.or_round)
994 orrd->od_round = next->nr_u.orr.or_round;
1002 * Sort-adds request \a nrq to an ORR/TRR \a policy instance's set of queued
1003 * requests in the policy's binary heap.
1005 * A scheduling round is a stream of requests that have been sorted in batches
1006 * according to the backend-fs object (for ORR policy instances) or OST (for TRR
1007 * policy instances) that they pertain to (as identified by its IDIF FID or OST
1008 * index respectively); there can be only one batch for each object or OST in
1009 * each round. The batches are of maximum size nrs_orr_data:od_quantum. When a
1010 * new request arrives for scheduling for an object or OST that has exhausted
1011 * its quantum in its current round, the request will be scheduled on the next
1012 * scheduling round. Requests are allowed to be scheduled against a round until
1013 * all requests for the round are serviced, so an object or OST might miss a
1014 * round if requests are not scheduled for it for a long enough period of time.
1015 * Objects or OSTs that miss a round will continue with having their next
1016 * request scheduled, starting at the round that requests are being dispatched
1017 * for, at the time of arrival of this request.
1019 * Requests are tagged with the round number and a sequence number; the sequence
1020 * number indicates the relative ordering amongst the batches of requests in a
1021 * round, and is identical for all requests in a batch, as is the round number.
1022 * The round and sequence numbers are used by orr_req_compare() in order to use
1023 * nrs_orr_data::od_binheap in order to maintain an ordered set of rounds, with
1024 * each round consisting of an ordered set of batches of requests, and each
1025 * batch consisting of an ordered set of requests according to their logical
1026 * file or physical disk offsets.
1028 * \param[in] policy the policy
1029 * \param[in] nrq the request to add
1031 * \retval 0 request successfully added
1032 * \retval != 0 error
1034 static int nrs_orr_req_add(struct ptlrpc_nrs_policy *policy,
1035 struct ptlrpc_nrs_request *nrq)
1037 struct nrs_orr_data *orrd;
1038 struct nrs_orr_object *orro;
1041 orro = container_of(nrs_request_resource(nrq),
1042 struct nrs_orr_object, oo_res);
1043 orrd = container_of(nrs_request_resource(nrq)->res_parent,
1044 struct nrs_orr_data, od_res);
1046 if (orro->oo_quantum == 0 || orro->oo_round < orrd->od_round ||
1047 (orro->oo_active == 0 && orro->oo_quantum > 0)) {
1050 * If there are no pending requests for the object/OST, but some
1051 * of its quantum still remains unused, which implies we did not
1052 * get a chance to schedule up to its maximum allowed batch size
1053 * of requests in the previous round this object/OST
1054 * participated in, schedule this next request on a new round;
1055 * this avoids fragmentation of request batches caused by
1056 * intermittent inactivity on the object/OST, at the expense of
1057 * potentially slightly increased service time for the request
1058 * batch this request will be a part of.
1060 if (orro->oo_active == 0 && orro->oo_quantum > 0)
1063 /** A new scheduling round has commenced */
1064 if (orro->oo_round < orrd->od_round)
1065 orro->oo_round = orrd->od_round;
1067 /** I was not the last object/OST that scheduled a request */
1068 if (orro->oo_sequence < orrd->od_sequence)
1069 orro->oo_sequence = ++orrd->od_sequence;
1071 * Reset the quantum if we have reached the maximum quantum
1072 * size for this batch, or even if we have not managed to
1073 * complete a batch size up to its maximum allowed size.
1074 * XXX: Accessed unlocked
1076 orro->oo_quantum = orrd->od_quantum;
1079 nrq->nr_u.orr.or_round = orro->oo_round;
1080 nrq->nr_u.orr.or_sequence = orro->oo_sequence;
1082 rc = cfs_binheap_insert(orrd->od_binheap, &nrq->nr_node);
1085 if (--orro->oo_quantum == 0)
1092 * Removes request \a nrq from an ORR/TRR \a policy instance's set of queued
1095 * \param[in] policy the policy
1096 * \param[in] nrq the request to remove
1098 static void nrs_orr_req_del(struct ptlrpc_nrs_policy *policy,
1099 struct ptlrpc_nrs_request *nrq)
1101 struct nrs_orr_data *orrd;
1102 struct nrs_orr_object *orro;
1105 orro = container_of(nrs_request_resource(nrq),
1106 struct nrs_orr_object, oo_res);
1107 orrd = container_of(nrs_request_resource(nrq)->res_parent,
1108 struct nrs_orr_data, od_res);
1110 LASSERT(nrq->nr_u.orr.or_round <= orro->oo_round);
1112 is_root = &nrq->nr_node == cfs_binheap_root(orrd->od_binheap);
1114 cfs_binheap_remove(orrd->od_binheap, &nrq->nr_node);
1118 * If we just deleted the node at the root of the binheap, we may have
1119 * to adjust round numbers.
1121 if (unlikely(is_root)) {
1122 /** Peek at the next request to be served */
1123 struct cfs_binheap_node *node = cfs_binheap_root(orrd->od_binheap);
1125 /** No more requests */
1126 if (unlikely(node == NULL)) {
1129 nrq = container_of(node, struct ptlrpc_nrs_request,
1132 if (orrd->od_round < nrq->nr_u.orr.or_round)
1133 orrd->od_round = nrq->nr_u.orr.or_round;
1139 * Called right after the request \a nrq finishes being handled by ORR policy
1140 * instance \a policy.
1142 * \param[in] policy the policy that handled the request
1143 * \param[in] nrq the request that was handled
1145 static void nrs_orr_req_stop(struct ptlrpc_nrs_policy *policy,
1146 struct ptlrpc_nrs_request *nrq)
1148 /** NB: resource control, credits etc can be added here */
1149 if (strncmp(policy->pol_desc->pd_name, NRS_POL_NAME_ORR,
1150 NRS_POL_NAME_MAX) == 0)
1152 "NRS: finished handling %s request for object with FID "
1153 DFID", from OST with index %u, with round %llu\n",
1154 NRS_POL_NAME_ORR, PFID(&nrq->nr_u.orr.or_key.ok_fid),
1155 nrq->nr_u.orr.or_key.ok_idx, nrq->nr_u.orr.or_round);
1158 "NRS: finished handling %s request from OST with index %u,"
1159 " with round %llu\n",
1160 NRS_POL_NAME_TRR, nrq->nr_u.orr.or_key.ok_idx,
1161 nrq->nr_u.orr.or_round);
1168 #ifdef CONFIG_PROC_FS
1171 * This allows to bundle the policy name into the lprocfs_vars::data pointer
1172 * so that lprocfs read/write functions can be used by both the ORR and TRR
1175 static struct nrs_lprocfs_orr_data {
1176 struct ptlrpc_service *svc;
1178 } lprocfs_orr_data = {
1179 .name = NRS_POL_NAME_ORR
1180 }, lprocfs_trr_data = {
1181 .name = NRS_POL_NAME_TRR
1185 * Retrieves the value of the Round Robin quantum (i.e. the maximum batch size)
1186 * for ORR/TRR policy instances on both the regular and high-priority NRS head
1187 * of a service, as long as a policy instance is not in the
1188 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state; policy instances in this
1189 * state are skipped later by nrs_orr_ctl().
1191 * Quantum values are in # of RPCs, and the output is in YAML format.
1198 * XXX: the CRR-N version of this, ptlrpc_lprocfs_rd_nrs_crrn_quantum() is
1199 * almost identical; it can be reworked and then reused for ORR/TRR.
1202 ptlrpc_lprocfs_nrs_orr_quantum_seq_show(struct seq_file *m, void *data)
1204 struct nrs_lprocfs_orr_data *orr_data = m->private;
1205 struct ptlrpc_service *svc = orr_data->svc;
1210 * Perform two separate calls to this as only one of the NRS heads'
1211 * policies may be in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
1212 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING state.
1214 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1216 NRS_CTL_ORR_RD_QUANTUM,
1219 seq_printf(m, NRS_LPROCFS_QUANTUM_NAME_REG "%-5d\n", quantum);
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,
1240 seq_printf(m, NRS_LPROCFS_QUANTUM_NAME_HP"%-5d\n", quantum);
1242 * Ignore -ENODEV as the high priority NRS head's policy may be
1243 * in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1245 } else if (rc != -ENODEV) {
1255 * Sets the value of the Round Robin quantum (i.e. the maximum batch size)
1256 * for ORR/TRR policy instances of a service. The user can set the quantum size
1257 * for the regular and high priority NRS head separately by specifying each
1258 * value, or both together in a single invocation.
1262 * lctl set_param ost.OSS.ost_io.nrs_orr_quantum=req_quantum:64, to set the
1263 * request quantum size of the ORR policy instance on the regular NRS head of
1264 * the ost_io service to 64
1266 * lctl set_param ost.OSS.ost_io.nrs_trr_quantum=hp_quantum:8 to set the request
1267 * quantum size of the TRR policy instance on the high priority NRS head of the
1268 * ost_io service to 8
1270 * lctl set_param ost.OSS.ost_io.nrs_orr_quantum=32, to set both the request
1271 * quantum size of the ORR policy instance on both the regular and the high
1272 * priority NRS head of the ost_io service to 32
1274 * policy instances in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state
1275 * are skipped later by nrs_orr_ctl().
1277 * XXX: the CRR-N version of this, ptlrpc_lprocfs_wr_nrs_crrn_quantum() is
1278 * almost identical; it can be reworked and then reused for ORR/TRR.
1281 ptlrpc_lprocfs_nrs_orr_quantum_seq_write(struct file *file,
1282 const char __user *buffer,
1283 size_t count, loff_t *off)
1285 struct seq_file *m = file->private_data;
1286 struct nrs_lprocfs_orr_data *orr_data = m->private;
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 */
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;
1391 LPROC_SEQ_FOPS(ptlrpc_lprocfs_nrs_orr_quantum);
1393 #define LPROCFS_NRS_OFF_NAME_REG "reg_offset_type:"
1394 #define LPROCFS_NRS_OFF_NAME_HP "hp_offset_type:"
1396 #define LPROCFS_NRS_OFF_NAME_PHYSICAL "physical"
1397 #define LPROCFS_NRS_OFF_NAME_LOGICAL "logical"
1400 * Retrieves the offset type used by ORR/TRR policy instances on both the
1401 * regular and high-priority NRS head of a service, as long as a policy
1402 * instance is not in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state;
1403 * policy instances in this state are skipped later by nrs_orr_ctl().
1405 * Offset type information is a (physical|logical) string, and output is
1410 * reg_offset_type:physical
1411 * hp_offset_type:logical
1414 ptlrpc_lprocfs_nrs_orr_offset_type_seq_show(struct seq_file *m, void *data)
1416 struct nrs_lprocfs_orr_data *orr_data = m->private;
1417 struct ptlrpc_service *svc = orr_data->svc;
1422 * Perform two separate calls to this as only one of the NRS heads'
1423 * policies may be in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED
1424 * or ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING state.
1426 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1427 orr_data->name, NRS_CTL_ORR_RD_OFF_TYPE,
1430 seq_printf(m, LPROCFS_NRS_OFF_NAME_REG"%s\n",
1431 physical ? LPROCFS_NRS_OFF_NAME_PHYSICAL :
1432 LPROCFS_NRS_OFF_NAME_LOGICAL);
1434 * Ignore -ENODEV as the regular NRS head's policy may be in the
1435 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1437 } else if (rc != -ENODEV) {
1442 * We know the ost_io service which is the only one ORR/TRR policies are
1443 * compatible with, do have an HP NRS head, but it may be best to guard
1444 * against a possible change of this in the future.
1446 if (!nrs_svc_has_hp(svc))
1449 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1450 orr_data->name, NRS_CTL_ORR_RD_OFF_TYPE,
1453 seq_printf(m, LPROCFS_NRS_OFF_NAME_HP"%s\n",
1454 physical ? LPROCFS_NRS_OFF_NAME_PHYSICAL :
1455 LPROCFS_NRS_OFF_NAME_LOGICAL);
1457 * Ignore -ENODEV as the high priority NRS head's policy may be
1458 * in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1460 } else if (rc != -ENODEV) {
1469 * Max valid command string is the size of the labels, plus "physical" twice.
1470 * plus a separating ' '
1472 #define LPROCFS_NRS_WR_OFF_TYPE_MAX_CMD \
1473 sizeof(LPROCFS_NRS_OFF_NAME_REG LPROCFS_NRS_OFF_NAME_PHYSICAL " " \
1474 LPROCFS_NRS_OFF_NAME_HP LPROCFS_NRS_OFF_NAME_PHYSICAL)
1477 * Sets the type of offsets used to order RPCs in ORR/TRR policy instances. The
1478 * user can set offset type for the regular or high priority NRS head
1479 * separately by specifying each value, or both together in a single invocation.
1483 * lctl set_param ost.OSS.ost_io.nrs_orr_offset_type=
1484 * reg_offset_type:physical, to enable the ORR policy instance on the regular
1485 * NRS head of the ost_io service to use physical disk offset ordering.
1487 * lctl set_param ost.OSS.ost_io.nrs_trr_offset_type=logical, to enable the TRR
1488 * policy instances on both the regular ang high priority NRS heads of the
1489 * ost_io service to use logical file offset ordering.
1491 * policy instances in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state are
1492 * are skipped later by nrs_orr_ctl().
1495 ptlrpc_lprocfs_nrs_orr_offset_type_seq_write(struct file *file,
1496 const char __user *buffer,
1500 struct seq_file *m = file->private_data;
1501 struct nrs_lprocfs_orr_data *orr_data = m->private;
1502 struct ptlrpc_service *svc = orr_data->svc;
1503 enum ptlrpc_nrs_queue_type queue = 0;
1504 char kernbuf[LPROCFS_NRS_WR_OFF_TYPE_MAX_CMD];
1513 if (count > (sizeof(kernbuf) - 1))
1516 if (copy_from_user(kernbuf, buffer, count))
1519 kernbuf[count] = '\0';
1524 * Check if the regular offset type has been specified
1526 val_reg = lprocfs_find_named_value(kernbuf,
1527 LPROCFS_NRS_OFF_NAME_REG,
1529 if (val_reg != kernbuf)
1530 queue |= PTLRPC_NRS_QUEUE_REG;
1535 * Check if the high priority offset type has been specified
1537 val_hp = lprocfs_find_named_value(kernbuf, LPROCFS_NRS_OFF_NAME_HP,
1539 if (val_hp != kernbuf) {
1540 if (!nrs_svc_has_hp(svc))
1543 queue |= PTLRPC_NRS_QUEUE_HP;
1547 * If none of the queues has been specified, there may be a valid
1548 * command string at the start of the buffer.
1551 queue = PTLRPC_NRS_QUEUE_REG;
1553 if (nrs_svc_has_hp(svc))
1554 queue |= PTLRPC_NRS_QUEUE_HP;
1557 if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1558 if (strncmp(val_reg, LPROCFS_NRS_OFF_NAME_PHYSICAL,
1559 sizeof(LPROCFS_NRS_OFF_NAME_PHYSICAL) - 1) == 0)
1560 physical_reg = true;
1561 else if (strncmp(val_reg, LPROCFS_NRS_OFF_NAME_LOGICAL,
1562 sizeof(LPROCFS_NRS_OFF_NAME_LOGICAL) - 1) == 0)
1563 physical_reg = false;
1568 if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1569 if (strncmp(val_hp, LPROCFS_NRS_OFF_NAME_PHYSICAL,
1570 sizeof(LPROCFS_NRS_OFF_NAME_PHYSICAL) - 1) == 0)
1572 else if (strncmp(val_hp, LPROCFS_NRS_OFF_NAME_LOGICAL,
1573 sizeof(LPROCFS_NRS_OFF_NAME_LOGICAL) - 1) == 0)
1574 physical_hp = false;
1580 * We change the values on regular and HP NRS heads separately, so that
1581 * we do not exit early from ptlrpc_nrs_policy_control() with an error
1582 * returned by nrs_policy_ctl_locked(), in cases where the user has not
1583 * started the policy on either the regular or HP NRS head; i.e. we are
1584 * ignoring -ENODEV within nrs_policy_ctl_locked(). -ENODEV is returned
1585 * only if the operation fails with -ENODEV on all heads that have been
1586 * specified by the command; if at least one operation succeeds,
1587 * success is returned.
1589 if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1590 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1592 NRS_CTL_ORR_WR_OFF_TYPE, false,
1594 if ((rc < 0 && rc != -ENODEV) ||
1595 (rc == -ENODEV && queue == PTLRPC_NRS_QUEUE_REG))
1599 if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1600 rc2 = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1602 NRS_CTL_ORR_WR_OFF_TYPE, false,
1604 if ((rc2 < 0 && rc2 != -ENODEV) ||
1605 (rc2 == -ENODEV && queue == PTLRPC_NRS_QUEUE_HP))
1609 return rc == -ENODEV && rc2 == -ENODEV ? -ENODEV : count;
1611 LPROC_SEQ_FOPS(ptlrpc_lprocfs_nrs_orr_offset_type);
1613 #define NRS_LPROCFS_REQ_SUPP_NAME_REG "reg_supported:"
1614 #define NRS_LPROCFS_REQ_SUPP_NAME_HP "hp_supported:"
1616 #define LPROCFS_NRS_SUPP_NAME_READS "reads"
1617 #define LPROCFS_NRS_SUPP_NAME_WRITES "writes"
1618 #define LPROCFS_NRS_SUPP_NAME_READWRITES "reads_and_writes"
1621 * Translates enum nrs_orr_supp values to a corresponding string.
1623 static const char *nrs_orr_supp2str(enum nrs_orr_supp supp)
1629 return LPROCFS_NRS_SUPP_NAME_READS;
1631 return LPROCFS_NRS_SUPP_NAME_WRITES;
1633 return LPROCFS_NRS_SUPP_NAME_READWRITES;
1638 * Translates strings to the corresponding enum nrs_orr_supp value
1640 static enum nrs_orr_supp nrs_orr_str2supp(const char *val)
1642 if (strncmp(val, LPROCFS_NRS_SUPP_NAME_READWRITES,
1643 sizeof(LPROCFS_NRS_SUPP_NAME_READWRITES) - 1) == 0)
1645 else if (strncmp(val, LPROCFS_NRS_SUPP_NAME_READS,
1646 sizeof(LPROCFS_NRS_SUPP_NAME_READS) - 1) == 0)
1647 return NOS_OST_READ;
1648 else if (strncmp(val, LPROCFS_NRS_SUPP_NAME_WRITES,
1649 sizeof(LPROCFS_NRS_SUPP_NAME_WRITES) - 1) == 0)
1650 return NOS_OST_WRITE;
1656 * Retrieves the type of RPCs handled at the point of invocation by ORR/TRR
1657 * policy instances on both the regular and high-priority NRS head of a service,
1658 * as long as a policy instance is not in the
1659 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state; policy instances in this
1660 * state are skipped later by nrs_orr_ctl().
1662 * Supported RPC type information is a (reads|writes|reads_and_writes) string,
1663 * and output is in YAML format.
1667 * reg_supported:reads
1668 * hp_supported:reads_and_writes
1671 ptlrpc_lprocfs_nrs_orr_supported_seq_show(struct seq_file *m, void *data)
1673 struct nrs_lprocfs_orr_data *orr_data = m->private;
1674 struct ptlrpc_service *svc = orr_data->svc;
1675 enum nrs_orr_supp supported;
1679 * Perform two separate calls to this as only one of the NRS heads'
1680 * policies may be in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED
1681 * or ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING state.
1683 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1685 NRS_CTL_ORR_RD_SUPP_REQ, true,
1689 seq_printf(m, NRS_LPROCFS_REQ_SUPP_NAME_REG"%s\n",
1690 nrs_orr_supp2str(supported));
1692 * Ignore -ENODEV as the regular NRS head's policy may be in the
1693 * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1695 } else if (rc != -ENODEV) {
1700 * We know the ost_io service which is the only one ORR/TRR policies are
1701 * compatible with, do have an HP NRS head, but it may be best to guard
1702 * against a possible change of this in the future.
1704 if (!nrs_svc_has_hp(svc))
1707 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1709 NRS_CTL_ORR_RD_SUPP_REQ, true,
1712 seq_printf(m, NRS_LPROCFS_REQ_SUPP_NAME_HP"%s\n",
1713 nrs_orr_supp2str(supported));
1715 * Ignore -ENODEV as the high priority NRS head's policy may be
1716 * in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1718 } else if (rc != -ENODEV) {
1728 * Max valid command string is the size of the labels, plus "reads_and_writes"
1729 * twice, plus a separating ' '
1731 #define LPROCFS_NRS_WR_REQ_SUPP_MAX_CMD \
1732 sizeof(NRS_LPROCFS_REQ_SUPP_NAME_REG LPROCFS_NRS_SUPP_NAME_READWRITES \
1733 NRS_LPROCFS_REQ_SUPP_NAME_HP LPROCFS_NRS_SUPP_NAME_READWRITES \
1737 * Sets the type of RPCs handled by ORR/TRR policy instances. The user can
1738 * modify this setting for the regular or high priority NRS heads separately, or
1739 * both together in a single invocation.
1743 * lctl set_param ost.OSS.ost_io.nrs_orr_supported=
1744 * "reg_supported:reads", to enable the ORR policy instance on the regular NRS
1745 * head of the ost_io service to handle OST_READ RPCs.
1747 * lctl set_param ost.OSS.ost_io.nrs_trr_supported=reads_and_writes, to enable
1748 * the TRR policy instances on both the regular ang high priority NRS heads of
1749 * the ost_io service to use handle OST_READ and OST_WRITE RPCs.
1751 * policy instances in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state are
1752 * are skipped later by nrs_orr_ctl().
1755 ptlrpc_lprocfs_nrs_orr_supported_seq_write(struct file *file,
1756 const char __user *buffer,
1760 struct seq_file *m = file->private_data;
1761 struct nrs_lprocfs_orr_data *orr_data = m->private;
1762 struct ptlrpc_service *svc = orr_data->svc;
1763 enum ptlrpc_nrs_queue_type queue = 0;
1764 char kernbuf[LPROCFS_NRS_WR_REQ_SUPP_MAX_CMD];
1767 enum nrs_orr_supp supp_reg;
1768 enum nrs_orr_supp supp_hp;
1773 if (count > (sizeof(kernbuf) - 1))
1776 if (copy_from_user(kernbuf, buffer, count))
1779 kernbuf[count] = '\0';
1784 * Check if the regular supported requests setting has been specified
1786 val_reg = lprocfs_find_named_value(kernbuf,
1787 NRS_LPROCFS_REQ_SUPP_NAME_REG,
1789 if (val_reg != kernbuf)
1790 queue |= PTLRPC_NRS_QUEUE_REG;
1795 * Check if the high priority supported requests setting has been
1798 val_hp = lprocfs_find_named_value(kernbuf, NRS_LPROCFS_REQ_SUPP_NAME_HP,
1800 if (val_hp != kernbuf) {
1801 if (!nrs_svc_has_hp(svc))
1804 queue |= PTLRPC_NRS_QUEUE_HP;
1808 * If none of the queues has been specified, there may be a valid
1809 * command string at the start of the buffer.
1812 queue = PTLRPC_NRS_QUEUE_REG;
1814 if (nrs_svc_has_hp(svc))
1815 queue |= PTLRPC_NRS_QUEUE_HP;
1818 if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1819 supp_reg = nrs_orr_str2supp(val_reg);
1820 if (supp_reg == -EINVAL)
1824 if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1825 supp_hp = nrs_orr_str2supp(val_hp);
1826 if (supp_hp == -EINVAL)
1831 * We change the values on regular and HP NRS heads separately, so that
1832 * we do not exit early from ptlrpc_nrs_policy_control() with an error
1833 * returned by nrs_policy_ctl_locked(), in cases where the user has not
1834 * started the policy on either the regular or HP NRS head; i.e. we are
1835 * ignoring -ENODEV within nrs_policy_ctl_locked(). -ENODEV is returned
1836 * only if the operation fails with -ENODEV on all heads that have been
1837 * specified by the command; if at least one operation succeeds,
1838 * success is returned.
1840 if ((queue & PTLRPC_NRS_QUEUE_REG) != 0) {
1841 rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1843 NRS_CTL_ORR_WR_SUPP_REQ, false,
1845 if ((rc < 0 && rc != -ENODEV) ||
1846 (rc == -ENODEV && queue == PTLRPC_NRS_QUEUE_REG))
1850 if ((queue & PTLRPC_NRS_QUEUE_HP) != 0) {
1851 rc2 = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1853 NRS_CTL_ORR_WR_SUPP_REQ, false,
1855 if ((rc2 < 0 && rc2 != -ENODEV) ||
1856 (rc2 == -ENODEV && queue == PTLRPC_NRS_QUEUE_HP))
1860 return rc == -ENODEV && rc2 == -ENODEV ? -ENODEV : count;
1862 LPROC_SEQ_FOPS(ptlrpc_lprocfs_nrs_orr_supported);
1864 static int nrs_orr_lprocfs_init(struct ptlrpc_service *svc)
1868 struct lprocfs_vars nrs_orr_lprocfs_vars[] = {
1869 { .name = "nrs_orr_quantum",
1870 .fops = &ptlrpc_lprocfs_nrs_orr_quantum_fops },
1871 { .name = "nrs_orr_offset_type",
1872 .fops = &ptlrpc_lprocfs_nrs_orr_offset_type_fops },
1873 { .name = "nrs_orr_supported",
1874 .fops = &ptlrpc_lprocfs_nrs_orr_supported_fops },
1878 if (svc->srv_procroot == NULL)
1881 lprocfs_orr_data.svc = svc;
1883 for (i = 0; i < ARRAY_SIZE(nrs_orr_lprocfs_vars); i++)
1884 nrs_orr_lprocfs_vars[i].data = &lprocfs_orr_data;
1886 return lprocfs_add_vars(svc->srv_procroot, nrs_orr_lprocfs_vars, NULL);
1889 static void nrs_orr_lprocfs_fini(struct ptlrpc_service *svc)
1891 if (svc->srv_procroot == NULL)
1894 lprocfs_remove_proc_entry("nrs_orr_quantum", svc->srv_procroot);
1895 lprocfs_remove_proc_entry("nrs_orr_offset_type", svc->srv_procroot);
1896 lprocfs_remove_proc_entry("nrs_orr_supported", svc->srv_procroot);
1899 #endif /* CONFIG_PROC_FS */
1901 static const struct ptlrpc_nrs_pol_ops nrs_orr_ops = {
1902 .op_policy_init = nrs_orr_init,
1903 .op_policy_start = nrs_orr_start,
1904 .op_policy_stop = nrs_orr_stop,
1905 .op_policy_ctl = nrs_orr_ctl,
1906 .op_res_get = nrs_orr_res_get,
1907 .op_res_put = nrs_orr_res_put,
1908 .op_req_get = nrs_orr_req_get,
1909 .op_req_enqueue = nrs_orr_req_add,
1910 .op_req_dequeue = nrs_orr_req_del,
1911 .op_req_stop = nrs_orr_req_stop,
1912 #ifdef CONFIG_PROC_FS
1913 .op_lprocfs_init = nrs_orr_lprocfs_init,
1914 .op_lprocfs_fini = nrs_orr_lprocfs_fini,
1918 struct ptlrpc_nrs_pol_conf nrs_conf_orr = {
1919 .nc_name = NRS_POL_NAME_ORR,
1920 .nc_ops = &nrs_orr_ops,
1921 .nc_compat = nrs_policy_compat_one,
1922 .nc_compat_svc_name = "ost_io",
1926 * TRR, Target-based Round Robin policy
1928 * TRR reuses much of the functions and data structures of ORR
1931 #ifdef CONFIG_PROC_FS
1933 static int nrs_trr_lprocfs_init(struct ptlrpc_service *svc)
1937 struct lprocfs_vars nrs_trr_lprocfs_vars[] = {
1938 { .name = "nrs_trr_quantum",
1939 .fops = &ptlrpc_lprocfs_nrs_orr_quantum_fops },
1940 { .name = "nrs_trr_offset_type",
1941 .fops = &ptlrpc_lprocfs_nrs_orr_offset_type_fops },
1942 { .name = "nrs_trr_supported",
1943 .fops = &ptlrpc_lprocfs_nrs_orr_supported_fops },
1947 if (svc->srv_procroot == NULL)
1950 lprocfs_trr_data.svc = svc;
1952 for (i = 0; i < ARRAY_SIZE(nrs_trr_lprocfs_vars); i++)
1953 nrs_trr_lprocfs_vars[i].data = &lprocfs_trr_data;
1955 return lprocfs_add_vars(svc->srv_procroot, nrs_trr_lprocfs_vars, NULL);
1958 static void nrs_trr_lprocfs_fini(struct ptlrpc_service *svc)
1960 if (svc->srv_procroot == NULL)
1963 lprocfs_remove_proc_entry("nrs_trr_quantum", svc->srv_procroot);
1964 lprocfs_remove_proc_entry("nrs_trr_offset_type", svc->srv_procroot);
1965 lprocfs_remove_proc_entry("nrs_trr_supported", svc->srv_procroot);
1968 #endif /* CONFIG_PROC_FS */
1971 * Reuse much of the ORR functionality for TRR.
1973 static const struct ptlrpc_nrs_pol_ops nrs_trr_ops = {
1974 .op_policy_init = nrs_orr_init,
1975 .op_policy_start = nrs_orr_start,
1976 .op_policy_stop = nrs_orr_stop,
1977 .op_policy_ctl = nrs_orr_ctl,
1978 .op_res_get = nrs_orr_res_get,
1979 .op_res_put = nrs_orr_res_put,
1980 .op_req_get = nrs_orr_req_get,
1981 .op_req_enqueue = nrs_orr_req_add,
1982 .op_req_dequeue = nrs_orr_req_del,
1983 .op_req_stop = nrs_orr_req_stop,
1984 #ifdef CONFIG_PROC_FS
1985 .op_lprocfs_init = nrs_trr_lprocfs_init,
1986 .op_lprocfs_fini = nrs_trr_lprocfs_fini,
1990 struct ptlrpc_nrs_pol_conf nrs_conf_trr = {
1991 .nc_name = NRS_POL_NAME_TRR,
1992 .nc_ops = &nrs_trr_ops,
1993 .nc_compat = nrs_policy_compat_one,
1994 .nc_compat_svc_name = "ost_io",
1997 /** @} ORR/TRR policy */
2001 #endif /* HAVE_SERVER_SUPPORT */