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, 2017, 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;
80 const struct lu_env *oaua_update_env;
84 * Call back for each update request.
86 struct osp_update_callback {
87 /* list in the osp_update_request::our_cb_items */
88 struct list_head ouc_list;
90 /* The target of the async update request. */
91 struct osp_object *ouc_obj;
93 /* The data used by or_interpreter. */
96 /* The interpreter function called after the async request handled. */
97 osp_update_interpreter_t ouc_interpreter;
101 * Allocate new update request
103 * Allocate new update request and insert it to the req_update_list.
105 * \param [in] our osp_udate_request where to create a new
108 * \retval 0 if creation succeeds.
109 * \retval negative errno if creation fails.
111 int osp_object_update_request_create(struct osp_update_request *our,
114 struct osp_update_request_sub *ours;
115 struct object_update_request *ourq;
121 /* The object update request will be added to an SG list for
122 * bulk transfer. Some IB HW cannot handle partial pages in SG
123 * lists (since they create gaps in memory regions) so we
124 * round the size up to the next multiple of PAGE_SIZE. See
127 size = round_up(size, PAGE_SIZE);
128 OBD_ALLOC_LARGE(ourq, size);
134 ourq->ourq_magic = UPDATE_REQUEST_MAGIC;
135 ourq->ourq_count = 0;
136 ours->ours_req = ourq;
137 ours->ours_req_size = size;
138 INIT_LIST_HEAD(&ours->ours_list);
139 list_add_tail(&ours->ours_list, &our->our_req_list);
146 * Get current update request
148 * Get current object update request from our_req_list in
149 * osp_update_request, because we always insert the new update
150 * request in the last position, so the last update request
151 * in the list will be the current update req.
153 * \param[in] our osp update request where to get the
154 * current object update.
156 * \retval the current object update.
158 struct osp_update_request_sub *
159 osp_current_object_update_request(struct osp_update_request *our)
161 if (list_empty(&our->our_req_list))
164 return list_entry(our->our_req_list.prev, struct osp_update_request_sub,
169 * Allocate and initialize osp_update_request
171 * osp_update_request is being used to track updates being executed on
172 * this dt_device(OSD or OSP). The update buffer will be 4k initially,
173 * and increased if needed.
175 * \param [in] dt dt device
177 * \retval osp_update_request being allocated if succeed
178 * \retval ERR_PTR(errno) if failed
180 struct osp_update_request *osp_update_request_create(struct dt_device *dt)
182 struct osp_update_request *our;
187 return ERR_PTR(-ENOMEM);
189 INIT_LIST_HEAD(&our->our_req_list);
190 INIT_LIST_HEAD(&our->our_cb_items);
191 INIT_LIST_HEAD(&our->our_list);
192 INIT_LIST_HEAD(&our->our_invalidate_cb_list);
193 spin_lock_init(&our->our_list_lock);
195 rc = osp_object_update_request_create(our, PAGE_SIZE);
203 void osp_update_request_destroy(const struct lu_env *env,
204 struct osp_update_request *our)
206 struct osp_update_request_sub *ours;
207 struct osp_update_request_sub *tmp;
212 list_for_each_entry_safe(ours, tmp, &our->our_req_list, ours_list) {
213 list_del(&ours->ours_list);
214 if (ours->ours_req != NULL)
215 OBD_FREE_LARGE(ours->ours_req, ours->ours_req_size);
219 if (!list_empty(&our->our_invalidate_cb_list)) {
221 struct osp_object *obj;
222 struct osp_object *next;
225 lu_env_init(&lenv, LCT_MD_THREAD | LCT_DT_THREAD);
229 list_for_each_entry_safe(obj, next,
230 &our->our_invalidate_cb_list,
231 opo_invalidate_cb_list) {
232 spin_lock(&obj->opo_lock);
233 list_del_init(&obj->opo_invalidate_cb_list);
234 spin_unlock(&obj->opo_lock);
236 dt_object_put(env, &obj->opo_obj);
247 object_update_request_dump(const struct object_update_request *ourq,
251 size_t total_size = 0;
253 for (i = 0; i < ourq->ourq_count; i++) {
254 struct object_update *update;
257 update = object_update_request_get(ourq, i, &size);
258 LASSERT(update != NULL);
259 CDEBUG(mask, "i = %u fid = "DFID" op = %s "
260 "params = %d batchid = %llu size = %zu repsize %u\n",
261 i, PFID(&update->ou_fid),
262 update_op_str(update->ou_type),
263 update->ou_params_count,
264 update->ou_batchid, size,
265 (unsigned)update->ou_result_size);
270 CDEBUG(mask, "updates = %p magic = %x count = %d size = %zu\n", ourq,
271 ourq->ourq_magic, ourq->ourq_count, total_size);
275 * Prepare inline update request
277 * Prepare OUT update ptlrpc inline request, and the request usually includes
278 * one update buffer, which does not need bulk transfer.
280 * \param[in] env execution environment
281 * \param[in] req ptlrpc request
282 * \param[in] ours sub osp_update_request to be packed
284 * \retval 0 if packing succeeds
285 * \retval negative errno if packing fails
287 int osp_prep_inline_update_req(const struct lu_env *env,
288 struct ptlrpc_request *req,
289 struct osp_update_request *our,
292 struct osp_update_request_sub *ours;
293 struct out_update_header *ouh;
294 __u32 update_req_size;
297 ours = list_entry(our->our_req_list.next,
298 struct osp_update_request_sub, ours_list);
299 update_req_size = object_update_request_size(ours->ours_req);
300 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_HEADER, RCL_CLIENT,
301 update_req_size + sizeof(*ouh));
303 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, OUT_UPDATE);
307 ouh = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_HEADER);
308 ouh->ouh_magic = OUT_UPDATE_HEADER_MAGIC;
310 ouh->ouh_inline_length = update_req_size;
311 ouh->ouh_reply_size = repsize;
313 memcpy(ouh->ouh_inline_data, ours->ours_req, update_req_size);
315 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_REPLY,
316 RCL_SERVER, repsize);
318 ptlrpc_request_set_replen(req);
319 req->rq_request_portal = OUT_PORTAL;
320 req->rq_reply_portal = OSC_REPLY_PORTAL;
326 * Prepare update request.
328 * Prepare OUT update ptlrpc request, and the request usually includes
329 * all of updates (stored in \param ureq) from one operation.
331 * \param[in] env execution environment
332 * \param[in] imp import on which ptlrpc request will be sent
333 * \param[in] ureq hold all of updates which will be packed into the req
334 * \param[in] reqp request to be created
336 * \retval 0 if preparation succeeds.
337 * \retval negative errno if preparation fails.
339 int osp_prep_update_req(const struct lu_env *env, struct obd_import *imp,
340 struct osp_update_request *our,
341 struct ptlrpc_request **reqp)
343 struct ptlrpc_request *req;
344 struct ptlrpc_bulk_desc *desc;
345 struct osp_update_request_sub *ours;
346 const struct object_update_request *ourq;
347 struct out_update_header *ouh;
348 struct out_update_buffer *oub;
352 struct object_update_reply *reply;
357 list_for_each_entry(ours, &our->our_req_list, ours_list) {
358 object_update_request_dump(ours->ours_req, D_INFO);
360 ourq = ours->ours_req;
361 for (i = 0; i < ourq->ourq_count; i++) {
362 struct object_update *update;
366 /* XXX: it's very inefficient to lookup update
367 * this way, iterating from the beginning
369 update = object_update_request_get(ourq, i, &size);
370 LASSERT(update != NULL);
372 repsize += sizeof(reply->ourp_lens[0]);
373 repsize += sizeof(struct object_update_result);
374 repsize += update->ou_result_size;
379 repsize += sizeof(*reply);
380 if (repsize < OUT_UPDATE_REPLY_SIZE)
381 repsize = OUT_UPDATE_REPLY_SIZE;
382 LASSERT(buf_count > 0);
384 req = ptlrpc_request_alloc(imp, &RQF_OUT_UPDATE);
388 if (buf_count == 1) {
389 ours = list_entry(our->our_req_list.next,
390 struct osp_update_request_sub, ours_list);
392 /* Let's check if it can be packed inline */
393 if (object_update_request_size(ours->ours_req) +
394 sizeof(struct out_update_header) <
395 OUT_UPDATE_MAX_INLINE_SIZE) {
396 rc = osp_prep_inline_update_req(env, req, our, repsize);
403 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_HEADER, RCL_CLIENT,
404 sizeof(struct osp_update_request));
406 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_BUF, RCL_CLIENT,
407 buf_count * sizeof(*oub));
409 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, OUT_UPDATE);
413 ouh = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_HEADER);
414 ouh->ouh_magic = OUT_UPDATE_HEADER_MAGIC;
415 ouh->ouh_count = buf_count;
416 ouh->ouh_inline_length = 0;
417 ouh->ouh_reply_size = repsize;
418 oub = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_BUF);
419 list_for_each_entry(ours, &our->our_req_list, ours_list) {
420 oub->oub_size = ours->ours_req_size;
422 page_count += round_up(ours->ours_req_size, PAGE_SIZE) + 1;
425 req->rq_bulk_write = 1;
426 desc = ptlrpc_prep_bulk_imp(req, page_count,
427 MD_MAX_BRW_SIZE >> LNET_MTU_BITS,
428 PTLRPC_BULK_GET_SOURCE | PTLRPC_BULK_BUF_KIOV,
429 MDS_BULK_PORTAL, &ptlrpc_bulk_kiov_nopin_ops);
431 GOTO(out_req, rc = -ENOMEM);
433 /* NB req now owns desc and will free it when it gets freed */
434 list_for_each_entry(ours, &our->our_req_list, ours_list) {
435 desc->bd_frag_ops->add_iov_frag(desc, ours->ours_req,
436 ours->ours_req_size);
437 total += ours->ours_req_size;
439 CDEBUG(D_OTHER, "total %d in %u\n", total, our->our_update_nr);
441 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_REPLY,
442 RCL_SERVER, repsize);
444 ptlrpc_request_set_replen(req);
445 req->rq_request_portal = OUT_PORTAL;
446 req->rq_reply_portal = OSC_REPLY_PORTAL;
451 ptlrpc_req_finished(req);
459 * Send update request to the remote MDT synchronously.
461 * \param[in] env execution environment
462 * \param[in] imp import on which ptlrpc request will be sent
463 * \param[in] our hold all of updates which will be packed into the req
464 * \param[in] reqp request to be created
466 * \retval 0 if RPC succeeds.
467 * \retval negative errno if RPC fails.
469 int osp_remote_sync(const struct lu_env *env, struct osp_device *osp,
470 struct osp_update_request *our,
471 struct ptlrpc_request **reqp)
473 struct obd_import *imp = osp->opd_obd->u.cli.cl_import;
474 struct ptlrpc_request *req = NULL;
478 rc = osp_prep_update_req(env, imp, our, &req);
482 osp_set_req_replay(osp, req);
483 req->rq_allow_intr = 1;
485 /* Note: some dt index api might return non-zero result here, like
486 * osd_index_ea_lookup, so we should only check rc < 0 here */
487 rc = ptlrpc_queue_wait(req);
489 if (rc < 0 || reqp == NULL)
490 ptlrpc_req_finished(req);
498 * Invalidate all objects in the osp thandle
500 * invalidate all of objects in the update request, which will be called
501 * when the transaction is aborted.
503 * \param[in] oth osp thandle.
505 static void osp_thandle_invalidate_object(const struct lu_env *env,
506 struct osp_thandle *oth,
509 struct osp_update_request *our = oth->ot_our;
510 struct osp_object *obj;
511 struct osp_object *next;
516 list_for_each_entry_safe(obj, next, &our->our_invalidate_cb_list,
517 opo_invalidate_cb_list) {
519 osp_invalidate(env, &obj->opo_obj);
521 spin_lock(&obj->opo_lock);
522 list_del_init(&obj->opo_invalidate_cb_list);
523 spin_unlock(&obj->opo_lock);
525 dt_object_put(env, &obj->opo_obj);
529 static void osp_trans_stop_cb(const struct lu_env *env,
530 struct osp_thandle *oth, int result)
532 struct dt_txn_commit_cb *dcb;
533 struct dt_txn_commit_cb *tmp;
535 /* call per-transaction stop callbacks if any */
536 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
538 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
539 "commit callback entry: magic=%x name='%s'\n",
540 dcb->dcb_magic, dcb->dcb_name);
541 list_del_init(&dcb->dcb_linkage);
542 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
545 osp_thandle_invalidate_object(env, oth, result);
549 * Allocate an osp request and initialize it with the given parameters.
551 * \param[in] obj pointer to the operation target
552 * \param[in] data pointer to the data used by the interpreter
553 * \param[in] interpreter pointer to the interpreter function
555 * \retval pointer to the asychronous request
556 * \retval NULL if the allocation failed
558 static struct osp_update_callback *
559 osp_update_callback_init(struct osp_object *obj, void *data,
560 osp_update_interpreter_t interpreter)
562 struct osp_update_callback *ouc;
568 lu_object_get(osp2lu_obj(obj));
569 INIT_LIST_HEAD(&ouc->ouc_list);
571 ouc->ouc_data = data;
572 ouc->ouc_interpreter = interpreter;
578 * Destroy the osp_update_callback.
580 * \param[in] env pointer to the thread context
581 * \param[in] ouc pointer to osp_update_callback
583 static void osp_update_callback_fini(const struct lu_env *env,
584 struct osp_update_callback *ouc)
586 LASSERT(list_empty(&ouc->ouc_list));
588 lu_object_put(env, osp2lu_obj(ouc->ouc_obj));
593 * Interpret the packaged OUT RPC results.
595 * For every packaged sub-request, call its registered interpreter function.
596 * Then destroy the sub-request.
598 * \param[in] env pointer to the thread context
599 * \param[in] req pointer to the RPC
600 * \param[in] arg pointer to data used by the interpreter
601 * \param[in] rc the RPC return value
603 * \retval 0 for success
604 * \retval negative error number on failure
606 static int osp_update_interpret(const struct lu_env *env,
607 struct ptlrpc_request *req, void *args, int rc)
609 struct object_update_reply *reply = NULL;
610 struct osp_update_args *oaua = args;
611 struct osp_update_request *our = oaua->oaua_update;
612 struct osp_thandle *oth;
613 struct osp_update_callback *ouc;
614 struct osp_update_callback *next;
624 /* Sigh env might be NULL in some cases, see
626 * osp_send_update_thread()
627 * ptlrpc_set_wait() ----> null env.
629 * osp_update_interpret()
630 * Let's use env in oaua for this case.
633 env = oaua->oaua_update_env;
635 oaua->oaua_update = NULL;
637 if (oaua->oaua_flow_control) {
638 struct osp_device *osp;
640 LASSERT(oth != NULL);
641 osp = dt2osp_dev(oth->ot_super.th_dev);
642 obd_put_request_slot(&osp->opd_obd->u.cli);
645 /* Unpack the results from the reply message. */
646 if (req->rq_repmsg != NULL && req->rq_replied) {
647 reply = req_capsule_server_sized_get(&req->rq_pill,
648 &RMF_OUT_UPDATE_REPLY,
649 OUT_UPDATE_REPLY_SIZE);
650 if (reply == NULL || reply->ourp_magic != UPDATE_REPLY_MAGIC) {
654 count = reply->ourp_count;
658 list_for_each_entry_safe(ouc, next, &our->our_cb_items, ouc_list) {
659 list_del_init(&ouc->ouc_list);
661 /* The peer may only have handled some requests (indicated
662 * by the 'count') in the packaged OUT RPC, we can only get
663 * results for the handled part. */
664 if (index < count && reply->ourp_lens[index] > 0 && rc >= 0) {
665 struct object_update_result *result;
667 result = object_update_result_get(reply, index, NULL);
671 rc1 = rc = result->our_rc;
672 } else if (rc1 >= 0) {
673 /* The peer did not handle these request, let's return
674 * -EINVAL to update interpret for now */
681 if (ouc->ouc_interpreter != NULL)
682 ouc->ouc_interpreter(env, reply, req, ouc->ouc_obj,
683 ouc->ouc_data, index, rc1);
685 osp_update_callback_fini(env, ouc);
689 if (oaua->oaua_count != NULL && atomic_dec_and_test(oaua->oaua_count))
690 wake_up_all(oaua->oaua_waitq);
693 /* oth and osp_update_requests will be destoryed in
695 osp_trans_stop_cb(env, oth, rc);
696 osp_thandle_put(env, oth);
698 osp_update_request_destroy(env, our);
705 * Pack all the requests in the shared asynchronous idempotent request queue
706 * into a single OUT RPC that will be given to the background ptlrpcd daemon.
708 * \param[in] env pointer to the thread context
709 * \param[in] osp pointer to the OSP device
710 * \param[in] our pointer to the shared queue
712 * \retval 0 for success
713 * \retval negative error number on failure
715 int osp_unplug_async_request(const struct lu_env *env,
716 struct osp_device *osp,
717 struct osp_update_request *our)
719 struct osp_update_args *args;
720 struct ptlrpc_request *req = NULL;
723 rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
726 struct osp_update_callback *ouc;
727 struct osp_update_callback *next;
729 list_for_each_entry_safe(ouc, next,
730 &our->our_cb_items, ouc_list) {
731 list_del_init(&ouc->ouc_list);
732 if (ouc->ouc_interpreter != NULL)
733 ouc->ouc_interpreter(env, NULL, NULL,
735 ouc->ouc_data, 0, rc);
736 osp_update_callback_fini(env, ouc);
738 osp_update_request_destroy(env, our);
740 args = ptlrpc_req_async_args(args, req);
741 args->oaua_update = our;
742 args->oaua_count = NULL;
743 args->oaua_waitq = NULL;
744 /* Note: this is asynchronous call for the request, so the
745 * interrupte cb and current function will be different
746 * thread, so we need use different env */
747 args->oaua_update_env = NULL;
748 args->oaua_flow_control = false;
749 req->rq_interpret_reply = osp_update_interpret;
750 ptlrpcd_add_req(req);
757 * Find or create (if NOT exist or purged) the shared asynchronous idempotent
758 * request queue - osp_device::opd_async_requests.
760 * If the osp_device::opd_async_requests is not NULL, then return it directly;
761 * otherwise create new osp_update_request and attach it to opd_async_requests.
763 * \param[in] osp pointer to the OSP device
765 * \retval pointer to the shared queue
766 * \retval negative error number on failure
768 static struct osp_update_request *
769 osp_find_or_create_async_update_request(struct osp_device *osp)
771 struct osp_update_request *our = osp->opd_async_requests;
776 our = osp_update_request_create(&osp->opd_dt_dev);
780 osp->opd_async_requests = our;
786 * Insert an osp_update_callback into the osp_update_request.
788 * Insert an osp_update_callback to the osp_update_request. Usually each update
789 * in the osp_update_request will have one correspondent callback, and these
790 * callbacks will be called in rq_interpret_reply.
792 * \param[in] env pointer to the thread context
793 * \param[in] obj pointer to the operation target object
794 * \param[in] data pointer to the data used by the interpreter
795 * \param[in] interpreter pointer to the interpreter function
797 * \retval 0 for success
798 * \retval negative error number on failure
800 int osp_insert_update_callback(const struct lu_env *env,
801 struct osp_update_request *our,
802 struct osp_object *obj, void *data,
803 osp_update_interpreter_t interpreter)
805 struct osp_update_callback *ouc;
807 ouc = osp_update_callback_init(obj, data, interpreter);
811 list_add_tail(&ouc->ouc_list, &our->our_cb_items);
817 * Insert an asynchronous idempotent request to the shared request queue that
818 * is attached to the osp_device.
820 * This function generates a new osp_async_request with the given parameters,
821 * then tries to insert the request into the osp_device-based shared request
822 * queue. If the queue is full, then triggers the packaged OUT RPC to purge
823 * the shared queue firstly, and then re-tries.
825 * NOTE: must hold the osp::opd_async_requests_mutex to serialize concurrent
826 * osp_insert_async_request call from others.
828 * \param[in] env pointer to the thread context
829 * \param[in] op operation type, see 'enum update_type'
830 * \param[in] obj pointer to the operation target
831 * \param[in] count array size of the subsequent \a lens and \a bufs
832 * \param[in] lens buffer length array for the subsequent \a bufs
833 * \param[in] bufs the buffers to compose the request
834 * \param[in] data pointer to the data used by the interpreter
835 * \param[in] repsize how many bytes the caller allocated for \a data
836 * \param[in] interpreter pointer to the interpreter function
838 * \retval 0 for success
839 * \retval negative error number on failure
841 int osp_insert_async_request(const struct lu_env *env, enum update_type op,
842 struct osp_object *obj, int count,
843 __u16 *lens, const void **bufs,
844 void *data, __u32 repsize,
845 osp_update_interpreter_t interpreter)
847 struct osp_device *osp;
848 struct osp_update_request *our;
849 struct object_update *object_update;
850 size_t max_update_size;
851 struct object_update_request *ureq;
852 struct osp_update_request_sub *ours;
856 osp = lu2osp_dev(osp2lu_obj(obj)->lo_dev);
857 our = osp_find_or_create_async_update_request(osp);
859 RETURN(PTR_ERR(our));
862 ours = osp_current_object_update_request(our);
864 ureq = ours->ours_req;
865 max_update_size = ours->ours_req_size -
866 object_update_request_size(ureq);
868 object_update = update_buffer_get_update(ureq, ureq->ourq_count);
869 rc = out_update_pack(env, object_update, &max_update_size, op,
870 lu_object_fid(osp2lu_obj(obj)), count, lens, bufs,
872 /* The queue is full. */
874 osp->opd_async_requests = NULL;
875 mutex_unlock(&osp->opd_async_requests_mutex);
877 rc = osp_unplug_async_request(env, osp, our);
878 mutex_lock(&osp->opd_async_requests_mutex);
882 our = osp_find_or_create_async_update_request(osp);
884 RETURN(PTR_ERR(our));
892 our->our_update_nr++;
895 rc = osp_insert_update_callback(env, our, obj, data, interpreter);
900 int osp_trans_update_request_create(struct thandle *th)
902 struct osp_thandle *oth = thandle_to_osp_thandle(th);
903 struct osp_update_request *our;
905 if (oth->ot_our != NULL)
908 our = osp_update_request_create(th->th_dev);
910 th->th_result = PTR_ERR(our);
920 void osp_thandle_destroy(const struct lu_env *env,
921 struct osp_thandle *oth)
923 LASSERT(oth->ot_magic == OSP_THANDLE_MAGIC);
924 LASSERT(list_empty(&oth->ot_commit_dcb_list));
925 LASSERT(list_empty(&oth->ot_stop_dcb_list));
926 if (oth->ot_our != NULL)
927 osp_update_request_destroy(env, oth->ot_our);
932 * The OSP layer dt_device_operations::dt_trans_create() interface
933 * to create a transaction.
935 * There are two kinds of transactions that will involve OSP:
937 * 1) If the transaction only contains the updates on remote server
938 * (MDT or OST), such as re-generating the lost OST-object for
939 * LFSCK, then it is a remote transaction. For remote transaction,
940 * the upper layer caller (such as the LFSCK engine) will call the
941 * dt_trans_create() (with the OSP dt_device as the parameter),
942 * then the call will be directed to the osp_trans_create() that
943 * creates the transaction handler and returns it to the caller.
945 * 2) If the transcation contains both local and remote updates,
946 * such as cross MDTs create under DNE mode, then the upper layer
947 * caller will not trigger osp_trans_create(). Instead, it will
948 * call dt_trans_create() on other dt_device, such as LOD that
949 * will generate the transaction handler. Such handler will be
950 * used by the whole transaction in subsequent sub-operations.
952 * \param[in] env pointer to the thread context
953 * \param[in] d pointer to the OSP dt_device
955 * \retval pointer to the transaction handler
956 * \retval negative error number on failure
958 struct thandle *osp_trans_create(const struct lu_env *env, struct dt_device *d)
960 struct osp_thandle *oth;
961 struct thandle *th = NULL;
965 if (unlikely(oth == NULL))
966 RETURN(ERR_PTR(-ENOMEM));
968 oth->ot_magic = OSP_THANDLE_MAGIC;
972 atomic_set(&oth->ot_refcount, 1);
973 INIT_LIST_HEAD(&oth->ot_commit_dcb_list);
974 INIT_LIST_HEAD(&oth->ot_stop_dcb_list);
980 * Add commit callback to transaction.
982 * Add commit callback to the osp thandle, which will be called
983 * when the thandle is committed remotely.
985 * \param[in] th the thandle
986 * \param[in] dcb commit callback structure
988 * \retval only return 0 for now.
990 int osp_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
992 struct osp_thandle *oth = thandle_to_osp_thandle(th);
994 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
995 LASSERT(&dcb->dcb_func != NULL);
996 if (dcb->dcb_flags & DCB_TRANS_STOP)
997 list_add(&dcb->dcb_linkage, &oth->ot_stop_dcb_list);
999 list_add(&dcb->dcb_linkage, &oth->ot_commit_dcb_list);
1003 static void osp_trans_commit_cb(struct osp_thandle *oth, int result)
1005 struct dt_txn_commit_cb *dcb;
1006 struct dt_txn_commit_cb *tmp;
1008 LASSERT(atomic_read(&oth->ot_refcount) > 0);
1009 /* call per-transaction callbacks if any */
1010 list_for_each_entry_safe(dcb, tmp, &oth->ot_commit_dcb_list,
1012 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1013 "commit callback entry: magic=%x name='%s'\n",
1014 dcb->dcb_magic, dcb->dcb_name);
1015 list_del_init(&dcb->dcb_linkage);
1016 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1020 static void osp_request_commit_cb(struct ptlrpc_request *req)
1022 struct thandle *th = req->rq_cb_data;
1023 struct osp_thandle *oth;
1024 __u64 last_committed_transno = 0;
1025 int result = req->rq_status;
1031 oth = thandle_to_osp_thandle(th);
1032 if (req->rq_repmsg != NULL &&
1033 lustre_msg_get_last_committed(req->rq_repmsg))
1034 last_committed_transno =
1035 lustre_msg_get_last_committed(req->rq_repmsg);
1037 if (last_committed_transno <
1038 req->rq_import->imp_peer_committed_transno)
1039 last_committed_transno =
1040 req->rq_import->imp_peer_committed_transno;
1042 CDEBUG(D_HA, "trans no %llu committed transno %llu\n",
1043 req->rq_transno, last_committed_transno);
1045 /* If the transaction is not really committed, mark result = 1 */
1046 if (req->rq_transno != 0 &&
1047 (req->rq_transno > last_committed_transno) && result == 0)
1050 osp_trans_commit_cb(oth, result);
1051 req->rq_committed = 1;
1052 osp_thandle_put(NULL, oth);
1057 * callback of osp transaction
1059 * Call all of callbacks for this osp thandle. This will only be
1060 * called in error handler path. In the normal processing path,
1061 * these callback will be called in osp_request_commit_cb() and
1062 * osp_update_interpret().
1064 * \param [in] env execution environment
1065 * \param [in] oth osp thandle
1066 * \param [in] rc result of the osp thandle
1068 void osp_trans_callback(const struct lu_env *env,
1069 struct osp_thandle *oth, int rc)
1071 struct osp_update_callback *ouc;
1072 struct osp_update_callback *next;
1074 if (oth->ot_our != NULL) {
1075 list_for_each_entry_safe(ouc, next,
1076 &oth->ot_our->our_cb_items, ouc_list) {
1077 list_del_init(&ouc->ouc_list);
1078 if (ouc->ouc_interpreter != NULL)
1079 ouc->ouc_interpreter(env, NULL, NULL,
1081 ouc->ouc_data, 0, rc);
1082 osp_update_callback_fini(env, ouc);
1085 osp_trans_stop_cb(env, oth, rc);
1086 osp_trans_commit_cb(oth, rc);
1090 * Send the request for remote updates.
1092 * Send updates to the remote MDT. Prepare the request by osp_update_req
1093 * and send them to remote MDT, for sync request, it will wait
1094 * until the reply return, otherwise hand it to ptlrpcd.
1096 * Please refer to osp_trans_create() for transaction type.
1098 * \param[in] env pointer to the thread context
1099 * \param[in] osp pointer to the OSP device
1100 * \param[in] our pointer to the osp_update_request
1102 * \retval 0 for success
1103 * \retval negative error number on failure
1105 static int osp_send_update_req(const struct lu_env *env,
1106 struct osp_device *osp,
1107 struct osp_update_request *our)
1109 struct osp_update_args *args;
1110 struct ptlrpc_request *req;
1111 struct osp_thandle *oth = our->our_th;
1115 LASSERT(oth != NULL);
1116 rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
1119 osp_trans_callback(env, oth, rc);
1123 args = ptlrpc_req_async_args(args, req);
1124 args->oaua_update = our;
1125 /* set env to NULL, in case the interrupt cb and current function
1126 * are in different thread */
1127 args->oaua_update_env = NULL;
1128 osp_thandle_get(oth); /* hold for update interpret */
1129 req->rq_interpret_reply = osp_update_interpret;
1130 if (!oth->ot_super.th_wait_submit && !oth->ot_super.th_sync) {
1131 if (!osp->opd_imp_active || !osp->opd_imp_connected) {
1132 osp_trans_callback(env, oth, rc);
1133 osp_thandle_put(env, oth);
1134 GOTO(out, rc = -ENOTCONN);
1137 rc = obd_get_request_slot(&osp->opd_obd->u.cli);
1139 osp_trans_callback(env, oth, rc);
1140 osp_thandle_put(env, oth);
1141 GOTO(out, rc = -ENOTCONN);
1143 args->oaua_flow_control = true;
1145 if (!osp->opd_connect_mdt) {
1146 down_read(&osp->opd_async_updates_rwsem);
1147 args->oaua_count = &osp->opd_async_updates_count;
1148 args->oaua_waitq = &osp->opd_sync_barrier_waitq;
1149 up_read(&osp->opd_async_updates_rwsem);
1150 atomic_inc(args->oaua_count);
1153 ptlrpcd_add_req(req);
1156 osp_thandle_get(oth); /* hold for commit callback */
1157 req->rq_commit_cb = osp_request_commit_cb;
1158 req->rq_cb_data = &oth->ot_super;
1159 args->oaua_flow_control = false;
1161 /* If the transaction is created during MDT recoverying
1162 * process, it means this is an recovery update, we need
1163 * to let OSP send it anyway without checking recoverying
1164 * status, in case the other target is being recoveried
1165 * at the same time, and if we wait here for the import
1166 * to be recoveryed, it might cause deadlock */
1167 osp_set_req_replay(osp, req);
1169 /* Because this req will be synchronus, i.e. it will be called
1170 * in the same thread, so it will be safe to use current
1172 args->oaua_update_env = env;
1173 if (osp->opd_connect_mdt)
1174 osp_get_rpc_lock(osp);
1175 rc = ptlrpc_queue_wait(req);
1176 if (osp->opd_connect_mdt)
1177 osp_put_rpc_lock(osp);
1179 /* We use rq_queued_time to distinguish between local
1180 * and remote -ENOMEM. */
1181 if ((rc == -ENOMEM && req->rq_queued_time == 0) ||
1182 (req->rq_transno == 0 && !req->rq_committed)) {
1183 if (args->oaua_update != NULL) {
1184 /* If osp_update_interpret is not being called,
1185 * release the osp_thandle */
1186 args->oaua_update = NULL;
1187 osp_thandle_put(env, oth);
1190 req->rq_cb_data = NULL;
1191 rc = rc == 0 ? req->rq_status : rc;
1192 osp_trans_callback(env, oth, rc);
1193 osp_thandle_put(env, oth);
1199 ptlrpc_req_finished(req);
1205 * Get local thandle for osp_thandle
1207 * Get the local OSD thandle from the OSP thandle. Currently, there
1208 * are a few OSP API (osp_create() and osp_sync_add()) needs
1209 * to update the object on local OSD device.
1211 * If the osp_thandle comes from normal stack (MDD->LOD->OSP), then
1212 * we will get local thandle by thandle_get_sub_by_dt.
1214 * If the osp_thandle is remote thandle (th_top == NULL, only used
1215 * by LFSCK), then it will create a local thandle, and stop it in
1216 * osp_trans_stop(). And this only happens on OSP for OST.
1218 * These are temporary solution, once OSP accessing OSD object is
1219 * being fixed properly, this function should be removed. XXX
1221 * \param[in] env pointer to the thread context
1222 * \param[in] th pointer to the transaction handler
1223 * \param[in] dt pointer to the OSP device
1225 * \retval pointer to the local thandle
1226 * \retval ERR_PTR(errno) if it fails.
1228 struct thandle *osp_get_storage_thandle(const struct lu_env *env,
1230 struct osp_device *osp)
1232 struct osp_thandle *oth;
1233 struct thandle *local_th;
1235 if (th->th_top != NULL)
1236 return thandle_get_sub_by_dt(env, th->th_top,
1239 LASSERT(!osp->opd_connect_mdt);
1240 oth = thandle_to_osp_thandle(th);
1241 if (oth->ot_storage_th != NULL)
1242 return oth->ot_storage_th;
1244 local_th = dt_trans_create(env, osp->opd_storage);
1245 if (IS_ERR(local_th))
1248 oth->ot_storage_th = local_th;
1254 * Set version for the transaction
1256 * Set the version for the transaction and add the request to
1257 * the sending list, then after transaction stop, the request
1258 * will be sent in the order of version by the sending thread.
1260 * \param [in] oth osp thandle to be set version.
1262 * \retval 0 if set version succeeds
1263 * negative errno if set version fails.
1265 int osp_check_and_set_rpc_version(struct osp_thandle *oth,
1266 struct osp_object *obj)
1268 struct osp_device *osp = dt2osp_dev(oth->ot_super.th_dev);
1269 struct osp_updates *ou = osp->opd_update;
1274 if (oth->ot_our->our_version != 0)
1277 spin_lock(&ou->ou_lock);
1278 spin_lock(&oth->ot_our->our_list_lock);
1279 if (obj->opo_stale) {
1280 spin_unlock(&oth->ot_our->our_list_lock);
1281 spin_unlock(&ou->ou_lock);
1285 /* Assign the version and add it to the sending list */
1286 osp_thandle_get(oth);
1287 oth->ot_our->our_version = ou->ou_version++;
1288 oth->ot_our->our_generation = ou->ou_generation;
1289 list_add_tail(&oth->ot_our->our_list,
1290 &osp->opd_update->ou_list);
1291 oth->ot_our->our_req_ready = 0;
1292 spin_unlock(&oth->ot_our->our_list_lock);
1293 spin_unlock(&ou->ou_lock);
1295 LASSERT(oth->ot_super.th_wait_submit == 1);
1296 CDEBUG(D_INFO, "%s: version %llu gen %llu oth:version %p:%llu\n",
1297 osp->opd_obd->obd_name, ou->ou_version, ou->ou_generation, oth,
1298 oth->ot_our->our_version);
1304 * Get next OSP update request in the sending list
1305 * Get next OSP update request in the sending list by version number, next
1307 * 1. transaction which does not have a version number.
1308 * 2. transaction whose version == opd_rpc_version.
1310 * \param [in] ou osp update structure.
1311 * \param [out] ourp the pointer holding the next update request.
1313 * \retval true if getting the next transaction.
1314 * \retval false if not getting the next transaction.
1317 osp_get_next_request(struct osp_updates *ou, struct osp_update_request **ourp)
1319 struct osp_update_request *our;
1320 struct osp_update_request *tmp;
1321 bool got_req = false;
1323 spin_lock(&ou->ou_lock);
1324 list_for_each_entry_safe(our, tmp, &ou->ou_list, our_list) {
1325 LASSERT(our->our_th != NULL);
1326 CDEBUG(D_HA, "ou %p version %llu rpc_version %llu\n",
1327 ou, our->our_version, ou->ou_rpc_version);
1328 spin_lock(&our->our_list_lock);
1329 /* Find next osp_update_request in the list */
1330 if (our->our_version == ou->ou_rpc_version &&
1331 our->our_req_ready) {
1332 list_del_init(&our->our_list);
1333 spin_unlock(&our->our_list_lock);
1338 spin_unlock(&our->our_list_lock);
1340 spin_unlock(&ou->ou_lock);
1346 * Invalidate update request
1348 * Invalidate update request in the OSP sending list, so all of
1349 * requests in the sending list will return error, which happens
1350 * when it finds one update (with writing llog) requests fails or
1351 * the OSP is evicted by remote target. see osp_send_update_thread().
1353 * \param[in] osp OSP device whose update requests will be
1356 void osp_invalidate_request(struct osp_device *osp)
1359 struct osp_updates *ou = osp->opd_update;
1360 struct osp_update_request *our;
1361 struct osp_update_request *tmp;
1369 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1371 CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
1374 spin_lock(&ou->ou_lock);
1375 ou->ou_generation++;
1376 spin_unlock(&ou->ou_lock);
1381 spin_lock(&ou->ou_lock);
1382 /* invalidate all of request in the sending list */
1383 list_for_each_entry_safe(our, tmp, &ou->ou_list, our_list) {
1384 spin_lock(&our->our_list_lock);
1385 if (our->our_req_ready)
1386 list_move(&our->our_list, &list);
1388 list_del_init(&our->our_list);
1390 if (our->our_th->ot_super.th_result == 0)
1391 our->our_th->ot_super.th_result = -EIO;
1393 if (our->our_version >= ou->ou_rpc_version)
1394 ou->ou_rpc_version = our->our_version + 1;
1395 spin_unlock(&our->our_list_lock);
1397 CDEBUG(D_HA, "%s invalidate our %p\n", osp->opd_obd->obd_name,
1401 /* Increase the generation, then the update request with old generation
1402 * will fail with -EIO. */
1403 ou->ou_generation++;
1404 spin_unlock(&ou->ou_lock);
1406 /* invalidate all of request in the sending list */
1407 list_for_each_entry_safe(our, tmp, &list, our_list) {
1408 spin_lock(&our->our_list_lock);
1409 list_del_init(&our->our_list);
1410 spin_unlock(&our->our_list_lock);
1411 osp_trans_callback(&env, our->our_th,
1412 our->our_th->ot_super.th_result);
1413 osp_thandle_put(&env, our->our_th);
1419 * Sending update thread
1421 * Create thread to send update request to other MDTs, this thread will pull
1422 * out update request from the list in OSP by version number, i.e. it will
1423 * make sure the update request with lower version number will be sent first.
1425 * \param[in] arg hold the OSP device.
1427 * \retval 0 if the thread is created successfully.
1428 * \retal negative error if the thread is not created
1431 int osp_send_update_thread(void *arg)
1434 struct osp_device *osp = arg;
1435 struct osp_updates *ou = osp->opd_update;
1436 struct ptlrpc_thread *thread = &osp->opd_update_thread;
1437 struct osp_update_request *our = NULL;
1441 LASSERT(ou != NULL);
1442 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1444 CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
1449 thread->t_flags = SVC_RUNNING;
1450 wake_up(&thread->t_ctl_waitq);
1453 wait_event_idle(ou->ou_waitq,
1454 !osp_send_update_thread_running(osp) ||
1455 osp_get_next_request(ou, &our));
1457 if (!osp_send_update_thread_running(osp)) {
1459 osp_trans_callback(&env, our->our_th, -EINTR);
1460 osp_thandle_put(&env, our->our_th);
1465 LASSERT(our->our_th != NULL);
1466 if (our->our_th->ot_super.th_result != 0) {
1467 osp_trans_callback(&env, our->our_th,
1468 our->our_th->ot_super.th_result);
1469 rc = our->our_th->ot_super.th_result;
1470 } else if (ou->ou_generation != our->our_generation ||
1471 OBD_FAIL_CHECK(OBD_FAIL_INVALIDATE_UPDATE)) {
1473 osp_trans_callback(&env, our->our_th, rc);
1475 rc = osp_send_update_req(&env, osp, our);
1478 /* Update the rpc version */
1479 spin_lock(&ou->ou_lock);
1480 if (our->our_version == ou->ou_rpc_version)
1481 ou->ou_rpc_version++;
1482 spin_unlock(&ou->ou_lock);
1484 /* If one update request fails, let's fail all of the requests
1485 * in the sending list, because the request in the sending
1486 * list are dependent on either other, continue sending these
1487 * request might cause llog or filesystem corruption */
1489 osp_invalidate_request(osp);
1491 /* Balanced for thandle_get in osp_check_and_set_rpc_version */
1492 osp_thandle_put(&env, our->our_th);
1495 thread->t_flags = SVC_STOPPED;
1497 wake_up(&thread->t_ctl_waitq);
1503 * The OSP layer dt_device_operations::dt_trans_start() interface
1504 * to start the transaction.
1506 * If the transaction is a remote transaction, then related remote
1507 * updates will be triggered in the osp_trans_stop().
1508 * Please refer to osp_trans_create() for transaction type.
1510 * \param[in] env pointer to the thread context
1511 * \param[in] dt pointer to the OSP dt_device
1512 * \param[in] th pointer to the transaction handler
1514 * \retval 0 for success
1515 * \retval negative error number on failure
1517 int osp_trans_start(const struct lu_env *env, struct dt_device *dt,
1520 struct osp_thandle *oth = thandle_to_osp_thandle(th);
1522 if (oth->ot_super.th_sync)
1523 oth->ot_our->our_flags |= UPDATE_FL_SYNC;
1524 /* For remote thandle, if there are local thandle, start it here*/
1525 if (is_only_remote_trans(th) && oth->ot_storage_th != NULL)
1526 return dt_trans_start(env, oth->ot_storage_th->th_dev,
1527 oth->ot_storage_th);
1532 * The OSP layer dt_device_operations::dt_trans_stop() interface
1533 * to stop the transaction.
1535 * If the transaction is a remote transaction, related remote
1536 * updates will be triggered at the end of this function.
1538 * For synchronous mode update or any failed update, the request
1539 * will be destroyed explicitly when the osp_trans_stop().
1541 * Please refer to osp_trans_create() for transaction type.
1543 * \param[in] env pointer to the thread context
1544 * \param[in] dt pointer to the OSP dt_device
1545 * \param[in] th pointer to the transaction handler
1547 * \retval 0 for success
1548 * \retval negative error number on failure
1550 int osp_trans_stop(const struct lu_env *env, struct dt_device *dt,
1553 struct osp_thandle *oth = thandle_to_osp_thandle(th);
1554 struct osp_update_request *our = oth->ot_our;
1555 struct osp_device *osp = dt2osp_dev(dt);
1559 /* For remote transaction, if there is local storage thandle,
1561 if (oth->ot_storage_th != NULL && th->th_top == NULL) {
1562 dt_trans_stop(env, oth->ot_storage_th->th_dev,
1563 oth->ot_storage_th);
1564 oth->ot_storage_th = NULL;
1567 if (our == NULL || list_empty(&our->our_req_list)) {
1568 osp_trans_callback(env, oth, th->th_result);
1569 GOTO(out, rc = th->th_result);
1572 if (!osp->opd_connect_mdt) {
1573 osp_trans_callback(env, oth, th->th_result);
1574 rc = osp_send_update_req(env, osp, oth->ot_our);
1578 if (osp->opd_update == NULL ||
1579 !osp_send_update_thread_running(osp)) {
1580 osp_trans_callback(env, oth, -EIO);
1581 GOTO(out, rc = -EIO);
1584 CDEBUG(D_HA, "%s: add oth %p with version %llu\n",
1585 osp->opd_obd->obd_name, oth, our->our_version);
1587 LASSERT(our->our_req_ready == 0);
1588 spin_lock(&our->our_list_lock);
1589 if (likely(!list_empty(&our->our_list))) {
1590 /* notify sending thread */
1591 our->our_req_ready = 1;
1592 wake_up(&osp->opd_update->ou_waitq);
1593 spin_unlock(&our->our_list_lock);
1594 } else if (th->th_result == 0) {
1595 /* if the request does not needs to be serialized,
1596 * read-only request etc, let's send it right away */
1597 spin_unlock(&our->our_list_lock);
1598 rc = osp_send_update_req(env, osp, our);
1600 spin_unlock(&our->our_list_lock);
1601 osp_trans_callback(env, oth, th->th_result);
1604 osp_thandle_put(env, oth);