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, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2014, 2015, Intel Corporation.
26 * lustre/osp/osp_trans.c
29 * 1. OSP (Object Storage Proxy) transaction methods
31 * Implement OSP layer transaction related interfaces for the dt_device API
32 * dt_device_operations.
35 * 2. Handle asynchronous idempotent operations
37 * The OSP uses OUT (Object Unified Target) RPC to talk with other server
38 * (MDT or OST) for kinds of operations, such as create, unlink, insert,
39 * delete, lookup, set_(x)attr, get_(x)attr, and etc. To reduce the number
40 * of RPCs, we allow multiple operations to be packaged together in single
43 * For the asynchronous idempotent operations, such as get_(x)attr, related
44 * RPCs will be inserted into an osp_device based shared asynchronous request
45 * queue - osp_device::opd_async_requests. When the queue is full, all the
46 * requests in the queue will be packaged into a single OUT RPC and given to
47 * the ptlrpcd daemon (for sending), then the queue is purged and other new
48 * requests can be inserted into it.
50 * When the asynchronous idempotent operation inserts the request into the
51 * shared queue, it will register an interpreter. When the packaged OUT RPC
52 * is replied (or failed to be sent out), all the registered interpreters
53 * will be called one by one to handle each own result.
56 * There are three kinds of transactions
58 * 1. Local transaction, all of updates of the transaction are in the local MDT.
59 * 2. Remote transaction, all of updates of the transaction are in one remote
60 * MDT, which only happens in LFSCK now.
61 * 3. Distribute transaction, updates for the transaction are in mulitple MDTs.
63 * Author: Di Wang <di.wang@intel.com>
64 * Author: Fan, Yong <fan.yong@intel.com>
67 #define DEBUG_SUBSYSTEM S_MDS
69 #include <lustre_net.h>
70 #include "osp_internal.h"
73 * The argument for the interpreter callback of osp request.
75 struct osp_update_args {
76 struct osp_update_request *oaua_update;
78 wait_queue_head_t *oaua_waitq;
79 bool oaua_flow_control;
83 * Call back for each update request.
85 struct osp_update_callback {
86 /* list in the osp_update_request::our_cb_items */
87 struct list_head ouc_list;
89 /* The target of the async update request. */
90 struct osp_object *ouc_obj;
92 /* The data used by or_interpreter. */
95 /* The interpreter function called after the async request handled. */
96 osp_update_interpreter_t ouc_interpreter;
99 static struct object_update_request *object_update_request_alloc(size_t size)
101 struct object_update_request *ourq;
103 OBD_ALLOC_LARGE(ourq, size);
105 return ERR_PTR(-ENOMEM);
107 ourq->ourq_magic = UPDATE_REQUEST_MAGIC;
108 ourq->ourq_count = 0;
114 * Allocate new update request
116 * Allocate new update request and insert it to the req_update_list.
118 * \param [in] our osp_udate_request where to create a new
121 * \retval 0 if creation succeeds.
122 * \retval negative errno if creation fails.
124 int osp_object_update_request_create(struct osp_update_request *our,
127 struct osp_update_request_sub *ours;
133 if (size < OUT_UPDATE_INIT_BUFFER_SIZE)
134 size = OUT_UPDATE_INIT_BUFFER_SIZE;
136 ours->ours_req = object_update_request_alloc(size);
138 if (IS_ERR(ours->ours_req)) {
143 ours->ours_req_size = size;
144 INIT_LIST_HEAD(&ours->ours_list);
145 list_add_tail(&ours->ours_list, &our->our_req_list);
152 * Get current update request
154 * Get current object update request from our_req_list in
155 * osp_update_request, because we always insert the new update
156 * request in the last position, so the last update request
157 * in the list will be the current update req.
159 * \param[in] our osp update request where to get the
160 * current object update.
162 * \retval the current object update.
164 struct osp_update_request_sub *
165 osp_current_object_update_request(struct osp_update_request *our)
167 if (list_empty(&our->our_req_list))
170 return list_entry(our->our_req_list.prev, struct osp_update_request_sub,
175 * Allocate and initialize osp_update_request
177 * osp_update_request is being used to track updates being executed on
178 * this dt_device(OSD or OSP). The update buffer will be 4k initially,
179 * and increased if needed.
181 * \param [in] dt dt device
183 * \retval osp_update_request being allocated if succeed
184 * \retval ERR_PTR(errno) if failed
186 struct osp_update_request *osp_update_request_create(struct dt_device *dt)
188 struct osp_update_request *our;
192 return ERR_PTR(-ENOMEM);
194 INIT_LIST_HEAD(&our->our_req_list);
195 INIT_LIST_HEAD(&our->our_cb_items);
196 INIT_LIST_HEAD(&our->our_list);
198 osp_object_update_request_create(our, OUT_UPDATE_INIT_BUFFER_SIZE);
202 void osp_update_request_destroy(struct osp_update_request *our)
204 struct osp_update_request_sub *ours;
205 struct osp_update_request_sub *tmp;
210 list_for_each_entry_safe(ours, tmp, &our->our_req_list, ours_list) {
211 list_del(&ours->ours_list);
212 if (ours->ours_req != NULL)
213 OBD_FREE(ours->ours_req, ours->ours_req_size);
220 object_update_request_dump(const struct object_update_request *ourq,
224 size_t total_size = 0;
226 for (i = 0; i < ourq->ourq_count; i++) {
227 struct object_update *update;
230 update = object_update_request_get(ourq, i, &size);
231 LASSERT(update != NULL);
232 CDEBUG(mask, "i = %u fid = "DFID" op = %s "
233 "params = %d batchid = "LPU64" size = %zu repsize %u\n",
234 i, PFID(&update->ou_fid),
235 update_op_str(update->ou_type),
236 update->ou_params_count,
237 update->ou_batchid, size,
238 (unsigned)update->ou_result_size);
243 CDEBUG(mask, "updates = %p magic = %x count = %d size = %zu\n", ourq,
244 ourq->ourq_magic, ourq->ourq_count, total_size);
248 * Prepare inline update request
250 * Prepare OUT update ptlrpc inline request, and the request usually includes
251 * one update buffer, which does not need bulk transfer.
253 * \param[in] env execution environment
254 * \param[in] req ptlrpc request
255 * \param[in] ours sub osp_update_request to be packed
257 * \retval 0 if packing succeeds
258 * \retval negative errno if packing fails
260 int osp_prep_inline_update_req(const struct lu_env *env,
261 struct ptlrpc_request *req,
262 struct osp_update_request *our,
265 struct osp_update_request_sub *ours;
266 struct out_update_header *ouh;
267 __u32 update_req_size;
270 ours = list_entry(our->our_req_list.next,
271 struct osp_update_request_sub, ours_list);
272 update_req_size = object_update_request_size(ours->ours_req);
273 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_HEADER, RCL_CLIENT,
274 update_req_size + sizeof(*ouh));
276 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, OUT_UPDATE);
280 ouh = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_HEADER);
281 ouh->ouh_magic = OUT_UPDATE_HEADER_MAGIC;
283 ouh->ouh_inline_length = update_req_size;
284 ouh->ouh_reply_size = repsize;
286 memcpy(ouh->ouh_inline_data, ours->ours_req, update_req_size);
288 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_REPLY,
289 RCL_SERVER, repsize);
291 ptlrpc_request_set_replen(req);
292 req->rq_request_portal = OUT_PORTAL;
293 req->rq_reply_portal = OSC_REPLY_PORTAL;
299 * Prepare update request.
301 * Prepare OUT update ptlrpc request, and the request usually includes
302 * all of updates (stored in \param ureq) from one operation.
304 * \param[in] env execution environment
305 * \param[in] imp import on which ptlrpc request will be sent
306 * \param[in] ureq hold all of updates which will be packed into the req
307 * \param[in] reqp request to be created
309 * \retval 0 if preparation succeeds.
310 * \retval negative errno if preparation fails.
312 int osp_prep_update_req(const struct lu_env *env, struct obd_import *imp,
313 struct osp_update_request *our,
314 struct ptlrpc_request **reqp)
316 struct ptlrpc_request *req;
317 struct ptlrpc_bulk_desc *desc;
318 struct osp_update_request_sub *ours;
319 const struct object_update_request *ourq;
320 struct out_update_header *ouh;
321 struct out_update_buffer *oub;
324 struct object_update_reply *reply;
329 list_for_each_entry(ours, &our->our_req_list, ours_list) {
330 object_update_request_dump(ours->ours_req, D_INFO);
332 ourq = ours->ours_req;
333 for (i = 0; i < ourq->ourq_count; i++) {
334 struct object_update *update;
338 /* XXX: it's very inefficient to lookup update
339 * this way, iterating from the beginning
341 update = object_update_request_get(ourq, i, &size);
342 LASSERT(update != NULL);
344 repsize += sizeof(reply->ourp_lens[0]);
345 repsize += sizeof(struct object_update_result);
346 repsize += update->ou_result_size;
351 repsize += sizeof(*reply);
352 repsize = (repsize + OUT_UPDATE_REPLY_SIZE - 1) &
353 ~(OUT_UPDATE_REPLY_SIZE - 1);
354 LASSERT(buf_count > 0);
356 req = ptlrpc_request_alloc(imp, &RQF_OUT_UPDATE);
360 if (buf_count == 1) {
361 ours = list_entry(our->our_req_list.next,
362 struct osp_update_request_sub, ours_list);
364 /* Let's check if it can be packed inline */
365 if (object_update_request_size(ours->ours_req) +
366 sizeof(struct out_update_header) <
367 OUT_UPDATE_MAX_INLINE_SIZE) {
368 rc = osp_prep_inline_update_req(env, req, our, repsize);
375 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_HEADER, RCL_CLIENT,
376 sizeof(struct osp_update_request));
378 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_BUF, RCL_CLIENT,
379 buf_count * sizeof(*oub));
381 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, OUT_UPDATE);
385 ouh = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_HEADER);
386 ouh->ouh_magic = OUT_UPDATE_HEADER_MAGIC;
387 ouh->ouh_count = buf_count;
388 ouh->ouh_inline_length = 0;
389 ouh->ouh_reply_size = repsize;
390 oub = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_BUF);
391 list_for_each_entry(ours, &our->our_req_list, ours_list) {
392 oub->oub_size = ours->ours_req_size;
396 req->rq_bulk_write = 1;
397 desc = ptlrpc_prep_bulk_imp(req, buf_count,
398 MD_MAX_BRW_SIZE >> LNET_MTU_BITS,
399 PTLRPC_BULK_GET_SOURCE | PTLRPC_BULK_BUF_KVEC,
400 MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
402 GOTO(out_req, rc = -ENOMEM);
404 /* NB req now owns desc and will free it when it gets freed */
405 list_for_each_entry(ours, &our->our_req_list, ours_list) {
406 desc->bd_frag_ops->add_iov_frag(desc, ours->ours_req,
407 ours->ours_req_size);
408 total += ours->ours_req_size;
410 CDEBUG(D_OTHER, "total %d in %u\n", total, our->our_update_nr);
412 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_REPLY,
413 RCL_SERVER, repsize);
415 ptlrpc_request_set_replen(req);
416 req->rq_request_portal = OUT_PORTAL;
417 req->rq_reply_portal = OSC_REPLY_PORTAL;
422 ptlrpc_req_finished(req);
430 * Send update request to the remote MDT synchronously.
432 * \param[in] env execution environment
433 * \param[in] imp import on which ptlrpc request will be sent
434 * \param[in] our hold all of updates which will be packed into the req
435 * \param[in] reqp request to be created
437 * \retval 0 if RPC succeeds.
438 * \retval negative errno if RPC fails.
440 int osp_remote_sync(const struct lu_env *env, struct osp_device *osp,
441 struct osp_update_request *our,
442 struct ptlrpc_request **reqp)
444 struct obd_import *imp = osp->opd_obd->u.cli.cl_import;
445 struct ptlrpc_request *req = NULL;
449 rc = osp_prep_update_req(env, imp, our, &req);
453 /* This will only be called with read-only update, and these updates
454 * might be used to retrieve update log during recovery process, so
455 * it will be allowed to send during recovery process */
456 req->rq_allow_replay = 1;
458 /* Note: some dt index api might return non-zero result here, like
459 * osd_index_ea_lookup, so we should only check rc < 0 here */
460 rc = ptlrpc_queue_wait(req);
462 if (rc < 0 || reqp == NULL)
463 ptlrpc_req_finished(req);
470 static void osp_trans_stop_cb(struct osp_thandle *oth, int result)
472 struct dt_txn_commit_cb *dcb;
473 struct dt_txn_commit_cb *tmp;
475 /* call per-transaction stop callbacks if any */
476 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
478 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
479 "commit callback entry: magic=%x name='%s'\n",
480 dcb->dcb_magic, dcb->dcb_name);
481 list_del_init(&dcb->dcb_linkage);
482 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
487 * Allocate an osp request and initialize it with the given parameters.
489 * \param[in] obj pointer to the operation target
490 * \param[in] data pointer to the data used by the interpreter
491 * \param[in] interpreter pointer to the interpreter function
493 * \retval pointer to the asychronous request
494 * \retval NULL if the allocation failed
496 static struct osp_update_callback *
497 osp_update_callback_init(struct osp_object *obj, void *data,
498 osp_update_interpreter_t interpreter)
500 struct osp_update_callback *ouc;
506 lu_object_get(osp2lu_obj(obj));
507 INIT_LIST_HEAD(&ouc->ouc_list);
509 ouc->ouc_data = data;
510 ouc->ouc_interpreter = interpreter;
516 * Destroy the osp_update_callback.
518 * \param[in] env pointer to the thread context
519 * \param[in] ouc pointer to osp_update_callback
521 static void osp_update_callback_fini(const struct lu_env *env,
522 struct osp_update_callback *ouc)
524 LASSERT(list_empty(&ouc->ouc_list));
526 lu_object_put(env, osp2lu_obj(ouc->ouc_obj));
531 * Interpret the packaged OUT RPC results.
533 * For every packaged sub-request, call its registered interpreter function.
534 * Then destroy the sub-request.
536 * \param[in] env pointer to the thread context
537 * \param[in] req pointer to the RPC
538 * \param[in] arg pointer to data used by the interpreter
539 * \param[in] rc the RPC return value
541 * \retval 0 for success
542 * \retval negative error number on failure
544 static int osp_update_interpret(const struct lu_env *env,
545 struct ptlrpc_request *req, void *arg, int rc)
547 struct object_update_reply *reply = NULL;
548 struct osp_update_args *oaua = arg;
549 struct osp_update_request *our = oaua->oaua_update;
550 struct osp_thandle *oth;
551 struct osp_update_callback *ouc;
552 struct osp_update_callback *next;
562 oaua->oaua_update = NULL;
564 if (oaua->oaua_flow_control) {
565 struct osp_device *osp;
567 LASSERT(oth != NULL);
568 osp = dt2osp_dev(oth->ot_super.th_dev);
569 obd_put_request_slot(&osp->opd_obd->u.cli);
572 /* Unpack the results from the reply message. */
573 if (req->rq_repmsg != NULL) {
574 reply = req_capsule_server_sized_get(&req->rq_pill,
575 &RMF_OUT_UPDATE_REPLY,
576 OUT_UPDATE_REPLY_SIZE);
577 if (reply == NULL || reply->ourp_magic != UPDATE_REPLY_MAGIC)
580 count = reply->ourp_count;
585 list_for_each_entry_safe(ouc, next, &our->our_cb_items, ouc_list) {
586 list_del_init(&ouc->ouc_list);
588 /* The peer may only have handled some requests (indicated
589 * by the 'count') in the packaged OUT RPC, we can only get
590 * results for the handled part. */
591 if (index < count && reply->ourp_lens[index] > 0) {
592 struct object_update_result *result;
594 result = object_update_result_get(reply, index, NULL);
598 rc1 = result->our_rc;
601 if (unlikely(rc1 == 0))
605 if (ouc->ouc_interpreter != NULL)
606 ouc->ouc_interpreter(env, reply, req, ouc->ouc_obj,
607 ouc->ouc_data, index, rc1);
609 osp_update_callback_fini(env, ouc);
613 if (oaua->oaua_count != NULL && atomic_dec_and_test(oaua->oaua_count))
614 wake_up_all(oaua->oaua_waitq);
617 /* oth and osp_update_requests will be destoryed in
619 osp_trans_stop_cb(oth, rc);
620 osp_thandle_put(oth);
622 osp_update_request_destroy(our);
629 * Pack all the requests in the shared asynchronous idempotent request queue
630 * into a single OUT RPC that will be given to the background ptlrpcd daemon.
632 * \param[in] env pointer to the thread context
633 * \param[in] osp pointer to the OSP device
634 * \param[in] our pointer to the shared queue
636 * \retval 0 for success
637 * \retval negative error number on failure
639 int osp_unplug_async_request(const struct lu_env *env,
640 struct osp_device *osp,
641 struct osp_update_request *our)
643 struct osp_update_args *args;
644 struct ptlrpc_request *req = NULL;
647 rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
650 struct osp_update_callback *ouc;
651 struct osp_update_callback *next;
653 list_for_each_entry_safe(ouc, next,
654 &our->our_cb_items, ouc_list) {
655 list_del_init(&ouc->ouc_list);
656 if (ouc->ouc_interpreter != NULL)
657 ouc->ouc_interpreter(env, NULL, NULL,
659 ouc->ouc_data, 0, rc);
660 osp_update_callback_fini(env, ouc);
662 osp_update_request_destroy(our);
664 args = ptlrpc_req_async_args(req);
665 args->oaua_update = our;
666 args->oaua_count = NULL;
667 args->oaua_waitq = NULL;
668 args->oaua_flow_control = false;
669 req->rq_interpret_reply = osp_update_interpret;
670 ptlrpcd_add_req(req);
677 * Find or create (if NOT exist or purged) the shared asynchronous idempotent
678 * request queue - osp_device::opd_async_requests.
680 * If the osp_device::opd_async_requests is not NULL, then return it directly;
681 * otherwise create new osp_update_request and attach it to opd_async_requests.
683 * \param[in] osp pointer to the OSP device
685 * \retval pointer to the shared queue
686 * \retval negative error number on failure
688 static struct osp_update_request *
689 osp_find_or_create_async_update_request(struct osp_device *osp)
691 struct osp_update_request *our = osp->opd_async_requests;
696 our = osp_update_request_create(&osp->opd_dt_dev);
700 osp->opd_async_requests = our;
706 * Insert an osp_update_callback into the osp_update_request.
708 * Insert an osp_update_callback to the osp_update_request. Usually each update
709 * in the osp_update_request will have one correspondent callback, and these
710 * callbacks will be called in rq_interpret_reply.
712 * \param[in] env pointer to the thread context
713 * \param[in] obj pointer to the operation target object
714 * \param[in] data pointer to the data used by the interpreter
715 * \param[in] interpreter pointer to the interpreter function
717 * \retval 0 for success
718 * \retval negative error number on failure
720 int osp_insert_update_callback(const struct lu_env *env,
721 struct osp_update_request *our,
722 struct osp_object *obj, void *data,
723 osp_update_interpreter_t interpreter)
725 struct osp_update_callback *ouc;
727 ouc = osp_update_callback_init(obj, data, interpreter);
731 list_add_tail(&ouc->ouc_list, &our->our_cb_items);
737 * Insert an asynchronous idempotent request to the shared request queue that
738 * is attached to the osp_device.
740 * This function generates a new osp_async_request with the given parameters,
741 * then tries to insert the request into the osp_device-based shared request
742 * queue. If the queue is full, then triggers the packaged OUT RPC to purge
743 * the shared queue firstly, and then re-tries.
745 * NOTE: must hold the osp::opd_async_requests_mutex to serialize concurrent
746 * osp_insert_async_request call from others.
748 * \param[in] env pointer to the thread context
749 * \param[in] op operation type, see 'enum update_type'
750 * \param[in] obj pointer to the operation target
751 * \param[in] count array size of the subsequent \a lens and \a bufs
752 * \param[in] lens buffer length array for the subsequent \a bufs
753 * \param[in] bufs the buffers to compose the request
754 * \param[in] data pointer to the data used by the interpreter
755 * \param[in] repsize how many bytes the caller allocated for \a data
756 * \param[in] interpreter pointer to the interpreter function
758 * \retval 0 for success
759 * \retval negative error number on failure
761 int osp_insert_async_request(const struct lu_env *env, enum update_type op,
762 struct osp_object *obj, int count,
763 __u16 *lens, const void **bufs,
764 void *data, __u32 repsize,
765 osp_update_interpreter_t interpreter)
767 struct osp_device *osp;
768 struct osp_update_request *our;
769 struct object_update *object_update;
770 size_t max_update_size;
771 struct object_update_request *ureq;
772 struct osp_update_request_sub *ours;
776 osp = lu2osp_dev(osp2lu_obj(obj)->lo_dev);
777 our = osp_find_or_create_async_update_request(osp);
779 RETURN(PTR_ERR(our));
782 ours = osp_current_object_update_request(our);
784 ureq = ours->ours_req;
785 max_update_size = ours->ours_req_size -
786 object_update_request_size(ureq);
788 object_update = update_buffer_get_update(ureq, ureq->ourq_count);
789 rc = out_update_pack(env, object_update, &max_update_size, op,
790 lu_object_fid(osp2lu_obj(obj)), count, lens, bufs,
792 /* The queue is full. */
794 osp->opd_async_requests = NULL;
795 mutex_unlock(&osp->opd_async_requests_mutex);
797 rc = osp_unplug_async_request(env, osp, our);
798 mutex_lock(&osp->opd_async_requests_mutex);
802 our = osp_find_or_create_async_update_request(osp);
804 RETURN(PTR_ERR(our));
812 our->our_update_nr++;
815 rc = osp_insert_update_callback(env, our, obj, data, interpreter);
820 int osp_trans_update_request_create(struct thandle *th)
822 struct osp_thandle *oth = thandle_to_osp_thandle(th);
823 struct osp_update_request *our;
825 if (oth->ot_our != NULL)
828 our = osp_update_request_create(th->th_dev);
830 th->th_result = PTR_ERR(our);
840 void osp_thandle_destroy(struct osp_thandle *oth)
842 LASSERT(oth->ot_magic == OSP_THANDLE_MAGIC);
843 LASSERT(list_empty(&oth->ot_commit_dcb_list));
844 LASSERT(list_empty(&oth->ot_stop_dcb_list));
845 if (oth->ot_our != NULL)
846 osp_update_request_destroy(oth->ot_our);
851 * The OSP layer dt_device_operations::dt_trans_create() interface
852 * to create a transaction.
854 * There are two kinds of transactions that will involve OSP:
856 * 1) If the transaction only contains the updates on remote server
857 * (MDT or OST), such as re-generating the lost OST-object for
858 * LFSCK, then it is a remote transaction. For remote transaction,
859 * the upper layer caller (such as the LFSCK engine) will call the
860 * dt_trans_create() (with the OSP dt_device as the parameter),
861 * then the call will be directed to the osp_trans_create() that
862 * creates the transaction handler and returns it to the caller.
864 * 2) If the transcation contains both local and remote updates,
865 * such as cross MDTs create under DNE mode, then the upper layer
866 * caller will not trigger osp_trans_create(). Instead, it will
867 * call dt_trans_create() on other dt_device, such as LOD that
868 * will generate the transaction handler. Such handler will be
869 * used by the whole transaction in subsequent sub-operations.
871 * \param[in] env pointer to the thread context
872 * \param[in] d pointer to the OSP dt_device
874 * \retval pointer to the transaction handler
875 * \retval negative error number on failure
877 struct thandle *osp_trans_create(const struct lu_env *env, struct dt_device *d)
879 struct osp_thandle *oth;
880 struct thandle *th = NULL;
884 if (unlikely(oth == NULL))
885 RETURN(ERR_PTR(-ENOMEM));
887 oth->ot_magic = OSP_THANDLE_MAGIC;
890 th->th_tags = LCT_TX_HANDLE;
892 atomic_set(&oth->ot_refcount, 1);
893 INIT_LIST_HEAD(&oth->ot_commit_dcb_list);
894 INIT_LIST_HEAD(&oth->ot_stop_dcb_list);
900 * Add commit callback to transaction.
902 * Add commit callback to the osp thandle, which will be called
903 * when the thandle is committed remotely.
905 * \param[in] th the thandle
906 * \param[in] dcb commit callback structure
908 * \retval only return 0 for now.
910 int osp_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
912 struct osp_thandle *oth = thandle_to_osp_thandle(th);
914 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
915 LASSERT(&dcb->dcb_func != NULL);
916 if (dcb->dcb_flags & DCB_TRANS_STOP)
917 list_add(&dcb->dcb_linkage, &oth->ot_stop_dcb_list);
919 list_add(&dcb->dcb_linkage, &oth->ot_commit_dcb_list);
923 static void osp_trans_commit_cb(struct osp_thandle *oth, int result)
925 struct dt_txn_commit_cb *dcb;
926 struct dt_txn_commit_cb *tmp;
928 LASSERT(atomic_read(&oth->ot_refcount) > 0);
929 /* call per-transaction callbacks if any */
930 list_for_each_entry_safe(dcb, tmp, &oth->ot_commit_dcb_list,
932 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
933 "commit callback entry: magic=%x name='%s'\n",
934 dcb->dcb_magic, dcb->dcb_name);
935 list_del_init(&dcb->dcb_linkage);
936 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
940 static void osp_request_commit_cb(struct ptlrpc_request *req)
942 struct thandle *th = req->rq_cb_data;
943 struct osp_thandle *oth;
944 __u64 last_committed_transno = 0;
945 int result = req->rq_status;
951 oth = thandle_to_osp_thandle(th);
952 if (req->rq_repmsg != NULL &&
953 lustre_msg_get_last_committed(req->rq_repmsg))
954 last_committed_transno =
955 lustre_msg_get_last_committed(req->rq_repmsg);
957 if (last_committed_transno <
958 req->rq_import->imp_peer_committed_transno)
959 last_committed_transno =
960 req->rq_import->imp_peer_committed_transno;
962 CDEBUG(D_HA, "trans no "LPU64" committed transno "LPU64"\n",
963 req->rq_transno, last_committed_transno);
965 /* If the transaction is not really committed, mark result = 1 */
966 if (req->rq_transno != 0 &&
967 (req->rq_transno > last_committed_transno) && result == 0)
970 osp_trans_commit_cb(oth, result);
971 req->rq_committed = 1;
972 osp_thandle_put(oth);
977 * callback of osp transaction
979 * Call all of callbacks for this osp thandle. This will only be
980 * called in error handler path. In the normal processing path,
981 * these callback will be called in osp_request_commit_cb() and
982 * osp_update_interpret().
984 * \param [in] env execution environment
985 * \param [in] oth osp thandle
986 * \param [in] rc result of the osp thandle
988 void osp_trans_callback(const struct lu_env *env,
989 struct osp_thandle *oth, int rc)
991 struct osp_update_callback *ouc;
992 struct osp_update_callback *next;
994 if (oth->ot_our != NULL) {
995 list_for_each_entry_safe(ouc, next,
996 &oth->ot_our->our_cb_items, ouc_list) {
997 list_del_init(&ouc->ouc_list);
998 if (ouc->ouc_interpreter != NULL)
999 ouc->ouc_interpreter(env, NULL, NULL,
1001 ouc->ouc_data, 0, rc);
1002 osp_update_callback_fini(env, ouc);
1005 osp_trans_stop_cb(oth, rc);
1006 osp_trans_commit_cb(oth, rc);
1010 * Send the request for remote updates.
1012 * Send updates to the remote MDT. Prepare the request by osp_update_req
1013 * and send them to remote MDT, for sync request, it will wait
1014 * until the reply return, otherwise hand it to ptlrpcd.
1016 * Please refer to osp_trans_create() for transaction type.
1018 * \param[in] env pointer to the thread context
1019 * \param[in] osp pointer to the OSP device
1020 * \param[in] our pointer to the osp_update_request
1022 * \retval 0 for success
1023 * \retval negative error number on failure
1025 static int osp_send_update_req(const struct lu_env *env,
1026 struct osp_device *osp,
1027 struct osp_update_request *our)
1029 struct osp_update_args *args;
1030 struct ptlrpc_request *req;
1031 struct lu_device *top_device;
1032 struct osp_thandle *oth = our->our_th;
1036 LASSERT(oth != NULL);
1037 LASSERT(our->our_req_sent == 0);
1038 rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
1041 osp_trans_callback(env, oth, rc);
1045 args = ptlrpc_req_async_args(req);
1046 args->oaua_update = our;
1047 osp_thandle_get(oth); /* hold for update interpret */
1048 req->rq_interpret_reply = osp_update_interpret;
1049 if (!oth->ot_super.th_wait_submit && !oth->ot_super.th_sync) {
1050 if (!osp->opd_imp_active || !osp->opd_imp_connected) {
1051 osp_trans_callback(env, oth, rc);
1052 osp_thandle_put(oth);
1053 GOTO(out, rc = -ENOTCONN);
1056 rc = obd_get_request_slot(&osp->opd_obd->u.cli);
1058 osp_trans_callback(env, oth, rc);
1059 osp_thandle_put(oth);
1060 GOTO(out, rc = -ENOTCONN);
1062 args->oaua_flow_control = true;
1064 if (!osp->opd_connect_mdt) {
1065 down_read(&osp->opd_async_updates_rwsem);
1066 args->oaua_count = &osp->opd_async_updates_count;
1067 args->oaua_waitq = &osp->opd_syn_barrier_waitq;
1068 up_read(&osp->opd_async_updates_rwsem);
1069 atomic_inc(args->oaua_count);
1072 ptlrpcd_add_req(req);
1075 osp_thandle_get(oth); /* hold for commit callback */
1076 req->rq_commit_cb = osp_request_commit_cb;
1077 req->rq_cb_data = &oth->ot_super;
1078 args->oaua_flow_control = false;
1080 /* If the transaction is created during MDT recoverying
1081 * process, it means this is an recovery update, we need
1082 * to let OSP send it anyway without checking recoverying
1083 * status, in case the other target is being recoveried
1084 * at the same time, and if we wait here for the import
1085 * to be recoveryed, it might cause deadlock */
1086 top_device = osp->opd_dt_dev.dd_lu_dev.ld_site->ls_top_dev;
1087 if (top_device->ld_obd->obd_recovering)
1088 req->rq_allow_replay = 1;
1090 if (osp->opd_connect_mdt)
1091 osp_get_rpc_lock(osp);
1092 rc = ptlrpc_queue_wait(req);
1093 if (osp->opd_connect_mdt)
1094 osp_put_rpc_lock(osp);
1095 if ((rc == -ENOMEM && req->rq_set == NULL) ||
1096 (req->rq_transno == 0 && !req->rq_committed)) {
1097 if (args->oaua_update != NULL) {
1098 /* If osp_update_interpret is not being called,
1099 * release the osp_thandle */
1100 args->oaua_update = NULL;
1101 osp_thandle_put(oth);
1104 req->rq_cb_data = NULL;
1105 rc = rc == 0 ? req->rq_status : rc;
1106 osp_trans_callback(env, oth, rc);
1107 osp_thandle_put(oth);
1113 ptlrpc_req_finished(req);
1119 * Get local thandle for osp_thandle
1121 * Get the local OSD thandle from the OSP thandle. Currently, there
1122 * are a few OSP API (osp_object_create() and osp_sync_add()) needs
1123 * to update the object on local OSD device.
1125 * If the osp_thandle comes from normal stack (MDD->LOD->OSP), then
1126 * we will get local thandle by thandle_get_sub_by_dt.
1128 * If the osp_thandle is remote thandle (th_top == NULL, only used
1129 * by LFSCK), then it will create a local thandle, and stop it in
1130 * osp_trans_stop(). And this only happens on OSP for OST.
1132 * These are temporary solution, once OSP accessing OSD object is
1133 * being fixed properly, this function should be removed. XXX
1135 * \param[in] env pointer to the thread context
1136 * \param[in] th pointer to the transaction handler
1137 * \param[in] dt pointer to the OSP device
1139 * \retval pointer to the local thandle
1140 * \retval ERR_PTR(errno) if it fails.
1142 struct thandle *osp_get_storage_thandle(const struct lu_env *env,
1144 struct osp_device *osp)
1146 struct osp_thandle *oth;
1147 struct thandle *local_th;
1149 if (th->th_top != NULL)
1150 return thandle_get_sub_by_dt(env, th->th_top,
1153 LASSERT(!osp->opd_connect_mdt);
1154 oth = thandle_to_osp_thandle(th);
1155 if (oth->ot_storage_th != NULL)
1156 return oth->ot_storage_th;
1158 local_th = dt_trans_create(env, osp->opd_storage);
1159 if (IS_ERR(local_th))
1162 oth->ot_storage_th = local_th;
1168 * Set version for the transaction
1170 * Set the version for the transaction, then the osp RPC will be
1171 * sent in the order of version, i.e. the transaction with lower
1172 * version will be sent first.
1174 * \param [in] oth osp thandle to be set version.
1176 * \retval 0 if set version succeeds
1177 * negative errno if set version fails.
1179 int osp_check_and_set_rpc_version(struct osp_thandle *oth)
1181 struct osp_device *osp = dt2osp_dev(oth->ot_super.th_dev);
1182 struct osp_updates *ou = osp->opd_update;
1187 if (oth->ot_version != 0)
1190 spin_lock(&ou->ou_lock);
1191 oth->ot_version = ou->ou_version++;
1192 spin_unlock(&ou->ou_lock);
1194 CDEBUG(D_INFO, "%s: version "LPU64" oth:version %p:"LPU64"\n",
1195 osp->opd_obd->obd_name, ou->ou_version, oth, oth->ot_version);
1201 * Get next OSP update request in the sending list
1202 * Get next OSP update request in the sending list by version number, next
1204 * 1. transaction which does not have a version number.
1205 * 2. transaction whose version == opd_rpc_version.
1207 * \param [in] ou osp update structure.
1208 * \param [out] ourp the pointer holding the next update request.
1210 * \retval true if getting the next transaction.
1211 * \retval false if not getting the next transaction.
1214 osp_get_next_request(struct osp_updates *ou, struct osp_update_request **ourp)
1216 struct osp_update_request *our;
1217 struct osp_update_request *tmp;
1218 bool got_req = false;
1220 spin_lock(&ou->ou_lock);
1221 list_for_each_entry_safe(our, tmp, &ou->ou_list, our_list) {
1222 LASSERT(our->our_th != NULL);
1223 CDEBUG(D_INFO, "our %p version "LPU64" rpc_version "LPU64"\n",
1224 our, our->our_th->ot_version, ou->ou_rpc_version);
1225 if (our->our_th->ot_version == 0) {
1226 list_del_init(&our->our_list);
1232 /* Find next osp_update_request in the list */
1233 if (our->our_th->ot_version == ou->ou_rpc_version) {
1234 list_del_init(&our->our_list);
1240 spin_unlock(&ou->ou_lock);
1245 static void osp_update_rpc_version(struct osp_updates *ou,
1246 struct osp_thandle *oth)
1248 if (oth->ot_version == 0)
1251 LASSERT(oth->ot_version == ou->ou_rpc_version);
1252 spin_lock(&ou->ou_lock);
1253 ou->ou_rpc_version++;
1254 spin_unlock(&ou->ou_lock);
1258 * Sending update thread
1260 * Create thread to send update request to other MDTs, this thread will pull
1261 * out update request from the list in OSP by version number, i.e. it will
1262 * make sure the update request with lower version number will be sent first.
1264 * \param[in] arg hold the OSP device.
1266 * \retval 0 if the thread is created successfully.
1267 * \retal negative error if the thread is not created
1270 int osp_send_update_thread(void *arg)
1273 struct osp_device *osp = arg;
1274 struct l_wait_info lwi = { 0 };
1275 struct osp_updates *ou = osp->opd_update;
1276 struct ptlrpc_thread *thread = &osp->opd_update_thread;
1277 struct osp_update_request *our = NULL;
1281 LASSERT(ou != NULL);
1282 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1284 CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
1289 thread->t_flags = SVC_RUNNING;
1290 wake_up(&thread->t_ctl_waitq);
1293 l_wait_event(ou->ou_waitq,
1294 !osp_send_update_thread_running(osp) ||
1295 osp_get_next_request(ou, &our),
1298 if (!osp_send_update_thread_running(osp)) {
1299 if (our != NULL && our->our_th != NULL) {
1300 osp_trans_callback(&env, our->our_th, -EINTR);
1301 osp_thandle_put(our->our_th);
1306 if (our->our_req_sent == 0) {
1307 if (our->our_th != NULL &&
1308 our->our_th->ot_super.th_result != 0)
1309 osp_trans_callback(&env, our->our_th,
1310 our->our_th->ot_super.th_result);
1312 rc = osp_send_update_req(&env, osp, our);
1315 if (our->our_th != NULL) {
1316 /* Update the rpc version */
1317 osp_update_rpc_version(ou, our->our_th);
1318 /* Balanced for thandle_get in osp_trans_trigger() */
1319 osp_thandle_put(our->our_th);
1323 thread->t_flags = SVC_STOPPED;
1325 wake_up(&thread->t_ctl_waitq);
1331 * Trigger the request for remote updates.
1333 * Add the request to the sending list, and wake up osp update
1336 * \param[in] env pointer to the thread context
1337 * \param[in] osp pointer to the OSP device
1338 * \param[in] oth pointer to the transaction handler
1341 static void osp_trans_trigger(const struct lu_env *env,
1342 struct osp_device *osp,
1343 struct osp_thandle *oth)
1346 CDEBUG(D_INFO, "%s: add oth %p with version "LPU64"\n",
1347 osp->opd_obd->obd_name, oth, oth->ot_version);
1349 LASSERT(oth->ot_magic == OSP_THANDLE_MAGIC);
1350 osp_thandle_get(oth);
1351 LASSERT(oth->ot_our != NULL);
1352 spin_lock(&osp->opd_update->ou_lock);
1353 list_add_tail(&oth->ot_our->our_list,
1354 &osp->opd_update->ou_list);
1355 spin_unlock(&osp->opd_update->ou_lock);
1357 wake_up(&osp->opd_update->ou_waitq);
1361 * The OSP layer dt_device_operations::dt_trans_start() interface
1362 * to start the transaction.
1364 * If the transaction is a remote transaction, then related remote
1365 * updates will be triggered in the osp_trans_stop().
1366 * Please refer to osp_trans_create() for transaction type.
1368 * \param[in] env pointer to the thread context
1369 * \param[in] dt pointer to the OSP dt_device
1370 * \param[in] th pointer to the transaction handler
1372 * \retval 0 for success
1373 * \retval negative error number on failure
1375 int osp_trans_start(const struct lu_env *env, struct dt_device *dt,
1378 struct osp_thandle *oth = thandle_to_osp_thandle(th);
1380 if (oth->ot_super.th_sync)
1381 oth->ot_our->our_flags |= UPDATE_FL_SYNC;
1382 /* For remote thandle, if there are local thandle, start it here*/
1383 if (is_only_remote_trans(th) && oth->ot_storage_th != NULL)
1384 return dt_trans_start(env, oth->ot_storage_th->th_dev,
1385 oth->ot_storage_th);
1390 * The OSP layer dt_device_operations::dt_trans_stop() interface
1391 * to stop the transaction.
1393 * If the transaction is a remote transaction, related remote
1394 * updates will be triggered here via osp_trans_trigger().
1396 * For synchronous mode update or any failed update, the request
1397 * will be destroyed explicitly when the osp_trans_stop().
1399 * Please refer to osp_trans_create() for transaction type.
1401 * \param[in] env pointer to the thread context
1402 * \param[in] dt pointer to the OSP dt_device
1403 * \param[in] th pointer to the transaction handler
1405 * \retval 0 for success
1406 * \retval negative error number on failure
1408 int osp_trans_stop(const struct lu_env *env, struct dt_device *dt,
1411 struct osp_thandle *oth = thandle_to_osp_thandle(th);
1412 struct osp_update_request *our = oth->ot_our;
1413 struct osp_device *osp = dt2osp_dev(dt);
1417 /* For remote transaction, if there is local storage thandle,
1419 if (oth->ot_storage_th != NULL && th->th_top == NULL) {
1420 dt_trans_stop(env, oth->ot_storage_th->th_dev,
1421 oth->ot_storage_th);
1422 oth->ot_storage_th = NULL;
1425 if (our == NULL || list_empty(&our->our_req_list)) {
1426 osp_trans_callback(env, oth, th->th_result);
1427 GOTO(out, rc = th->th_result);
1430 if (!osp->opd_connect_mdt) {
1431 osp_trans_callback(env, oth, th->th_result);
1432 rc = osp_send_update_req(env, osp, oth->ot_our);
1436 if (osp->opd_update == NULL ||
1437 !osp_send_update_thread_running(osp)) {
1438 osp_trans_callback(env, oth, -EIO);
1439 GOTO(out, rc = -EIO);
1443 /* if th_sync is set, then it needs to be sent
1444 * right away. Note: even thought the RPC has been
1445 * sent, it still needs to be added to the sending
1446 * list (see osp_trans_trigger()), so ou_rpc_version
1447 * can be updated correctly. */
1448 rc = osp_send_update_req(env, osp, our);
1449 our->our_req_sent = 1;
1452 osp_trans_trigger(env, osp, oth);
1454 osp_thandle_put(oth);