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;
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;
100 static struct object_update_request *object_update_request_alloc(size_t size)
102 struct object_update_request *ourq;
104 OBD_ALLOC_LARGE(ourq, size);
106 return ERR_PTR(-ENOMEM);
108 ourq->ourq_magic = UPDATE_REQUEST_MAGIC;
109 ourq->ourq_count = 0;
115 * Allocate new update request
117 * Allocate new update request and insert it to the req_update_list.
119 * \param [in] our osp_udate_request where to create a new
122 * \retval 0 if creation succeeds.
123 * \retval negative errno if creation fails.
125 int osp_object_update_request_create(struct osp_update_request *our,
128 struct osp_update_request_sub *ours;
134 if (size < OUT_UPDATE_INIT_BUFFER_SIZE)
135 size = OUT_UPDATE_INIT_BUFFER_SIZE;
137 ours->ours_req = object_update_request_alloc(size);
139 if (IS_ERR(ours->ours_req)) {
144 ours->ours_req_size = size;
145 INIT_LIST_HEAD(&ours->ours_list);
146 list_add_tail(&ours->ours_list, &our->our_req_list);
153 * Get current update request
155 * Get current object update request from our_req_list in
156 * osp_update_request, because we always insert the new update
157 * request in the last position, so the last update request
158 * in the list will be the current update req.
160 * \param[in] our osp update request where to get the
161 * current object update.
163 * \retval the current object update.
165 struct osp_update_request_sub *
166 osp_current_object_update_request(struct osp_update_request *our)
168 if (list_empty(&our->our_req_list))
171 return list_entry(our->our_req_list.prev, struct osp_update_request_sub,
176 * Allocate and initialize osp_update_request
178 * osp_update_request is being used to track updates being executed on
179 * this dt_device(OSD or OSP). The update buffer will be 4k initially,
180 * and increased if needed.
182 * \param [in] dt dt device
184 * \retval osp_update_request being allocated if succeed
185 * \retval ERR_PTR(errno) if failed
187 struct osp_update_request *osp_update_request_create(struct dt_device *dt)
189 struct osp_update_request *our;
194 return ERR_PTR(-ENOMEM);
196 INIT_LIST_HEAD(&our->our_req_list);
197 INIT_LIST_HEAD(&our->our_cb_items);
198 INIT_LIST_HEAD(&our->our_list);
199 spin_lock_init(&our->our_list_lock);
201 rc = osp_object_update_request_create(our, OUT_UPDATE_INIT_BUFFER_SIZE);
209 void osp_update_request_destroy(struct osp_update_request *our)
211 struct osp_update_request_sub *ours;
212 struct osp_update_request_sub *tmp;
217 list_for_each_entry_safe(ours, tmp, &our->our_req_list, ours_list) {
218 list_del(&ours->ours_list);
219 if (ours->ours_req != NULL)
220 OBD_FREE_LARGE(ours->ours_req, ours->ours_req_size);
227 object_update_request_dump(const struct object_update_request *ourq,
231 size_t total_size = 0;
233 for (i = 0; i < ourq->ourq_count; i++) {
234 struct object_update *update;
237 update = object_update_request_get(ourq, i, &size);
238 LASSERT(update != NULL);
239 CDEBUG(mask, "i = %u fid = "DFID" op = %s "
240 "params = %d batchid = "LPU64" size = %zu repsize %u\n",
241 i, PFID(&update->ou_fid),
242 update_op_str(update->ou_type),
243 update->ou_params_count,
244 update->ou_batchid, size,
245 (unsigned)update->ou_result_size);
250 CDEBUG(mask, "updates = %p magic = %x count = %d size = %zu\n", ourq,
251 ourq->ourq_magic, ourq->ourq_count, total_size);
255 * Prepare inline update request
257 * Prepare OUT update ptlrpc inline request, and the request usually includes
258 * one update buffer, which does not need bulk transfer.
260 * \param[in] env execution environment
261 * \param[in] req ptlrpc request
262 * \param[in] ours sub osp_update_request to be packed
264 * \retval 0 if packing succeeds
265 * \retval negative errno if packing fails
267 int osp_prep_inline_update_req(const struct lu_env *env,
268 struct ptlrpc_request *req,
269 struct osp_update_request *our,
272 struct osp_update_request_sub *ours;
273 struct out_update_header *ouh;
274 __u32 update_req_size;
277 ours = list_entry(our->our_req_list.next,
278 struct osp_update_request_sub, ours_list);
279 update_req_size = object_update_request_size(ours->ours_req);
280 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_HEADER, RCL_CLIENT,
281 update_req_size + sizeof(*ouh));
283 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, OUT_UPDATE);
287 ouh = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_HEADER);
288 ouh->ouh_magic = OUT_UPDATE_HEADER_MAGIC;
290 ouh->ouh_inline_length = update_req_size;
291 ouh->ouh_reply_size = repsize;
293 memcpy(ouh->ouh_inline_data, ours->ours_req, update_req_size);
295 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_REPLY,
296 RCL_SERVER, repsize);
298 ptlrpc_request_set_replen(req);
299 req->rq_request_portal = OUT_PORTAL;
300 req->rq_reply_portal = OSC_REPLY_PORTAL;
306 * Prepare update request.
308 * Prepare OUT update ptlrpc request, and the request usually includes
309 * all of updates (stored in \param ureq) from one operation.
311 * \param[in] env execution environment
312 * \param[in] imp import on which ptlrpc request will be sent
313 * \param[in] ureq hold all of updates which will be packed into the req
314 * \param[in] reqp request to be created
316 * \retval 0 if preparation succeeds.
317 * \retval negative errno if preparation fails.
319 int osp_prep_update_req(const struct lu_env *env, struct obd_import *imp,
320 struct osp_update_request *our,
321 struct ptlrpc_request **reqp)
323 struct ptlrpc_request *req;
324 struct ptlrpc_bulk_desc *desc;
325 struct osp_update_request_sub *ours;
326 const struct object_update_request *ourq;
327 struct out_update_header *ouh;
328 struct out_update_buffer *oub;
331 struct object_update_reply *reply;
336 list_for_each_entry(ours, &our->our_req_list, ours_list) {
337 object_update_request_dump(ours->ours_req, D_INFO);
339 ourq = ours->ours_req;
340 for (i = 0; i < ourq->ourq_count; i++) {
341 struct object_update *update;
345 /* XXX: it's very inefficient to lookup update
346 * this way, iterating from the beginning
348 update = object_update_request_get(ourq, i, &size);
349 LASSERT(update != NULL);
351 repsize += sizeof(reply->ourp_lens[0]);
352 repsize += sizeof(struct object_update_result);
353 repsize += update->ou_result_size;
358 repsize += sizeof(*reply);
359 repsize = (repsize + OUT_UPDATE_REPLY_SIZE - 1) &
360 ~(OUT_UPDATE_REPLY_SIZE - 1);
361 LASSERT(buf_count > 0);
363 req = ptlrpc_request_alloc(imp, &RQF_OUT_UPDATE);
367 if (buf_count == 1) {
368 ours = list_entry(our->our_req_list.next,
369 struct osp_update_request_sub, ours_list);
371 /* Let's check if it can be packed inline */
372 if (object_update_request_size(ours->ours_req) +
373 sizeof(struct out_update_header) <
374 OUT_UPDATE_MAX_INLINE_SIZE) {
375 rc = osp_prep_inline_update_req(env, req, our, repsize);
382 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_HEADER, RCL_CLIENT,
383 sizeof(struct osp_update_request));
385 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_BUF, RCL_CLIENT,
386 buf_count * sizeof(*oub));
388 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, OUT_UPDATE);
392 ouh = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_HEADER);
393 ouh->ouh_magic = OUT_UPDATE_HEADER_MAGIC;
394 ouh->ouh_count = buf_count;
395 ouh->ouh_inline_length = 0;
396 ouh->ouh_reply_size = repsize;
397 oub = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_BUF);
398 list_for_each_entry(ours, &our->our_req_list, ours_list) {
399 oub->oub_size = ours->ours_req_size;
403 req->rq_bulk_write = 1;
404 desc = ptlrpc_prep_bulk_imp(req, buf_count,
405 MD_MAX_BRW_SIZE >> LNET_MTU_BITS,
406 PTLRPC_BULK_GET_SOURCE | PTLRPC_BULK_BUF_KVEC,
407 MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
409 GOTO(out_req, rc = -ENOMEM);
411 /* NB req now owns desc and will free it when it gets freed */
412 list_for_each_entry(ours, &our->our_req_list, ours_list) {
413 desc->bd_frag_ops->add_iov_frag(desc, ours->ours_req,
414 ours->ours_req_size);
415 total += ours->ours_req_size;
417 CDEBUG(D_OTHER, "total %d in %u\n", total, our->our_update_nr);
419 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_REPLY,
420 RCL_SERVER, repsize);
422 ptlrpc_request_set_replen(req);
423 req->rq_request_portal = OUT_PORTAL;
424 req->rq_reply_portal = OSC_REPLY_PORTAL;
429 ptlrpc_req_finished(req);
437 * Send update request to the remote MDT synchronously.
439 * \param[in] env execution environment
440 * \param[in] imp import on which ptlrpc request will be sent
441 * \param[in] our hold all of updates which will be packed into the req
442 * \param[in] reqp request to be created
444 * \retval 0 if RPC succeeds.
445 * \retval negative errno if RPC fails.
447 int osp_remote_sync(const struct lu_env *env, struct osp_device *osp,
448 struct osp_update_request *our,
449 struct ptlrpc_request **reqp)
451 struct obd_import *imp = osp->opd_obd->u.cli.cl_import;
452 struct ptlrpc_request *req = NULL;
456 rc = osp_prep_update_req(env, imp, our, &req);
460 /* This will only be called with read-only update, and these updates
461 * might be used to retrieve update log during recovery process, so
462 * it will be allowed to send during recovery process */
463 req->rq_allow_replay = 1;
464 req->rq_allow_intr = 1;
466 /* Note: some dt index api might return non-zero result here, like
467 * osd_index_ea_lookup, so we should only check rc < 0 here */
468 rc = ptlrpc_queue_wait(req);
470 if (rc < 0 || reqp == NULL)
471 ptlrpc_req_finished(req);
479 * Invalidate all objects in the osp thandle
481 * invalidate all of objects in the update request, which will be called
482 * when the transaction is aborted.
484 * \param[in] oth osp thandle.
486 static void osp_thandle_invalidate_object(const struct lu_env *env,
487 struct osp_thandle *oth)
489 struct osp_update_request *our = oth->ot_our;
490 struct osp_update_request_sub *ours;
495 list_for_each_entry(ours, &our->our_req_list, ours_list) {
496 struct object_update_request *our_req = ours->ours_req;
498 struct lu_object *obj;
500 for (i = 0; i < our_req->ourq_count; i++) {
501 struct object_update *update;
503 update = object_update_request_get(our_req, i, NULL);
507 if (update->ou_type != OUT_WRITE)
510 if (!fid_is_sane(&update->ou_fid))
513 obj = lu_object_find_slice(env,
514 &oth->ot_super.th_dev->dd_lu_dev,
515 &update->ou_fid, NULL);
519 osp_invalidate(env, lu2dt_obj(obj));
520 lu_object_put(env, obj);
525 static void osp_trans_stop_cb(const struct lu_env *env,
526 struct osp_thandle *oth, int result)
528 struct dt_txn_commit_cb *dcb;
529 struct dt_txn_commit_cb *tmp;
531 /* call per-transaction stop callbacks if any */
532 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
534 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
535 "commit callback entry: magic=%x name='%s'\n",
536 dcb->dcb_magic, dcb->dcb_name);
537 list_del_init(&dcb->dcb_linkage);
538 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
542 osp_thandle_invalidate_object(env, oth);
546 * Allocate an osp request and initialize it with the given parameters.
548 * \param[in] obj pointer to the operation target
549 * \param[in] data pointer to the data used by the interpreter
550 * \param[in] interpreter pointer to the interpreter function
552 * \retval pointer to the asychronous request
553 * \retval NULL if the allocation failed
555 static struct osp_update_callback *
556 osp_update_callback_init(struct osp_object *obj, void *data,
557 osp_update_interpreter_t interpreter)
559 struct osp_update_callback *ouc;
565 lu_object_get(osp2lu_obj(obj));
566 INIT_LIST_HEAD(&ouc->ouc_list);
568 ouc->ouc_data = data;
569 ouc->ouc_interpreter = interpreter;
575 * Destroy the osp_update_callback.
577 * \param[in] env pointer to the thread context
578 * \param[in] ouc pointer to osp_update_callback
580 static void osp_update_callback_fini(const struct lu_env *env,
581 struct osp_update_callback *ouc)
583 LASSERT(list_empty(&ouc->ouc_list));
585 lu_object_put(env, osp2lu_obj(ouc->ouc_obj));
590 * Interpret the packaged OUT RPC results.
592 * For every packaged sub-request, call its registered interpreter function.
593 * Then destroy the sub-request.
595 * \param[in] env pointer to the thread context
596 * \param[in] req pointer to the RPC
597 * \param[in] arg pointer to data used by the interpreter
598 * \param[in] rc the RPC return value
600 * \retval 0 for success
601 * \retval negative error number on failure
603 static int osp_update_interpret(const struct lu_env *env,
604 struct ptlrpc_request *req, void *arg, int rc)
606 struct object_update_reply *reply = NULL;
607 struct osp_update_args *oaua = arg;
608 struct osp_update_request *our = oaua->oaua_update;
609 struct osp_thandle *oth;
610 struct osp_update_callback *ouc;
611 struct osp_update_callback *next;
621 /* Sigh env might be NULL in some cases, see
623 * osp_send_update_thread()
624 * ptlrpc_set_wait() ----> null env.
626 * osp_update_interpret()
627 * Let's use env in oaua for this case.
630 env = oaua->oaua_update_env;
632 oaua->oaua_update = NULL;
634 if (oaua->oaua_flow_control) {
635 struct osp_device *osp;
637 LASSERT(oth != NULL);
638 osp = dt2osp_dev(oth->ot_super.th_dev);
639 obd_put_request_slot(&osp->opd_obd->u.cli);
642 /* Unpack the results from the reply message. */
643 if (req->rq_repmsg != NULL && req->rq_replied) {
644 reply = req_capsule_server_sized_get(&req->rq_pill,
645 &RMF_OUT_UPDATE_REPLY,
646 OUT_UPDATE_REPLY_SIZE);
647 if (reply == NULL || reply->ourp_magic != UPDATE_REPLY_MAGIC) {
651 count = reply->ourp_count;
655 list_for_each_entry_safe(ouc, next, &our->our_cb_items, ouc_list) {
656 list_del_init(&ouc->ouc_list);
658 /* The peer may only have handled some requests (indicated
659 * by the 'count') in the packaged OUT RPC, we can only get
660 * results for the handled part. */
661 if (index < count && reply->ourp_lens[index] > 0 && rc >= 0) {
662 struct object_update_result *result;
664 result = object_update_result_get(reply, index, NULL);
668 rc1 = rc = result->our_rc;
669 } else if (rc1 >= 0) {
670 /* The peer did not handle these request, let's return
671 * -EINVAL to update interpret for now */
678 if (ouc->ouc_interpreter != NULL)
679 ouc->ouc_interpreter(env, reply, req, ouc->ouc_obj,
680 ouc->ouc_data, index, rc1);
682 osp_update_callback_fini(env, ouc);
686 if (oaua->oaua_count != NULL && atomic_dec_and_test(oaua->oaua_count))
687 wake_up_all(oaua->oaua_waitq);
690 /* oth and osp_update_requests will be destoryed in
692 osp_trans_stop_cb(env, oth, rc);
693 osp_thandle_put(oth);
695 osp_update_request_destroy(our);
702 * Pack all the requests in the shared asynchronous idempotent request queue
703 * into a single OUT RPC that will be given to the background ptlrpcd daemon.
705 * \param[in] env pointer to the thread context
706 * \param[in] osp pointer to the OSP device
707 * \param[in] our pointer to the shared queue
709 * \retval 0 for success
710 * \retval negative error number on failure
712 int osp_unplug_async_request(const struct lu_env *env,
713 struct osp_device *osp,
714 struct osp_update_request *our)
716 struct osp_update_args *args;
717 struct ptlrpc_request *req = NULL;
720 rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
723 struct osp_update_callback *ouc;
724 struct osp_update_callback *next;
726 list_for_each_entry_safe(ouc, next,
727 &our->our_cb_items, ouc_list) {
728 list_del_init(&ouc->ouc_list);
729 if (ouc->ouc_interpreter != NULL)
730 ouc->ouc_interpreter(env, NULL, NULL,
732 ouc->ouc_data, 0, rc);
733 osp_update_callback_fini(env, ouc);
735 osp_update_request_destroy(our);
737 args = ptlrpc_req_async_args(req);
738 args->oaua_update = our;
739 args->oaua_count = NULL;
740 args->oaua_waitq = NULL;
741 /* Note: this is asynchronous call for the request, so the
742 * interrupte cb and current function will be different
743 * thread, so we need use different env */
744 args->oaua_update_env = NULL;
745 args->oaua_flow_control = false;
746 req->rq_interpret_reply = osp_update_interpret;
747 ptlrpcd_add_req(req);
754 * Find or create (if NOT exist or purged) the shared asynchronous idempotent
755 * request queue - osp_device::opd_async_requests.
757 * If the osp_device::opd_async_requests is not NULL, then return it directly;
758 * otherwise create new osp_update_request and attach it to opd_async_requests.
760 * \param[in] osp pointer to the OSP device
762 * \retval pointer to the shared queue
763 * \retval negative error number on failure
765 static struct osp_update_request *
766 osp_find_or_create_async_update_request(struct osp_device *osp)
768 struct osp_update_request *our = osp->opd_async_requests;
773 our = osp_update_request_create(&osp->opd_dt_dev);
777 osp->opd_async_requests = our;
783 * Insert an osp_update_callback into the osp_update_request.
785 * Insert an osp_update_callback to the osp_update_request. Usually each update
786 * in the osp_update_request will have one correspondent callback, and these
787 * callbacks will be called in rq_interpret_reply.
789 * \param[in] env pointer to the thread context
790 * \param[in] obj pointer to the operation target object
791 * \param[in] data pointer to the data used by the interpreter
792 * \param[in] interpreter pointer to the interpreter function
794 * \retval 0 for success
795 * \retval negative error number on failure
797 int osp_insert_update_callback(const struct lu_env *env,
798 struct osp_update_request *our,
799 struct osp_object *obj, void *data,
800 osp_update_interpreter_t interpreter)
802 struct osp_update_callback *ouc;
804 ouc = osp_update_callback_init(obj, data, interpreter);
808 list_add_tail(&ouc->ouc_list, &our->our_cb_items);
814 * Insert an asynchronous idempotent request to the shared request queue that
815 * is attached to the osp_device.
817 * This function generates a new osp_async_request with the given parameters,
818 * then tries to insert the request into the osp_device-based shared request
819 * queue. If the queue is full, then triggers the packaged OUT RPC to purge
820 * the shared queue firstly, and then re-tries.
822 * NOTE: must hold the osp::opd_async_requests_mutex to serialize concurrent
823 * osp_insert_async_request call from others.
825 * \param[in] env pointer to the thread context
826 * \param[in] op operation type, see 'enum update_type'
827 * \param[in] obj pointer to the operation target
828 * \param[in] count array size of the subsequent \a lens and \a bufs
829 * \param[in] lens buffer length array for the subsequent \a bufs
830 * \param[in] bufs the buffers to compose the request
831 * \param[in] data pointer to the data used by the interpreter
832 * \param[in] repsize how many bytes the caller allocated for \a data
833 * \param[in] interpreter pointer to the interpreter function
835 * \retval 0 for success
836 * \retval negative error number on failure
838 int osp_insert_async_request(const struct lu_env *env, enum update_type op,
839 struct osp_object *obj, int count,
840 __u16 *lens, const void **bufs,
841 void *data, __u32 repsize,
842 osp_update_interpreter_t interpreter)
844 struct osp_device *osp;
845 struct osp_update_request *our;
846 struct object_update *object_update;
847 size_t max_update_size;
848 struct object_update_request *ureq;
849 struct osp_update_request_sub *ours;
853 osp = lu2osp_dev(osp2lu_obj(obj)->lo_dev);
854 our = osp_find_or_create_async_update_request(osp);
856 RETURN(PTR_ERR(our));
859 ours = osp_current_object_update_request(our);
861 ureq = ours->ours_req;
862 max_update_size = ours->ours_req_size -
863 object_update_request_size(ureq);
865 object_update = update_buffer_get_update(ureq, ureq->ourq_count);
866 rc = out_update_pack(env, object_update, &max_update_size, op,
867 lu_object_fid(osp2lu_obj(obj)), count, lens, bufs,
869 /* The queue is full. */
871 osp->opd_async_requests = NULL;
872 mutex_unlock(&osp->opd_async_requests_mutex);
874 rc = osp_unplug_async_request(env, osp, our);
875 mutex_lock(&osp->opd_async_requests_mutex);
879 our = osp_find_or_create_async_update_request(osp);
881 RETURN(PTR_ERR(our));
889 our->our_update_nr++;
892 rc = osp_insert_update_callback(env, our, obj, data, interpreter);
897 int osp_trans_update_request_create(struct thandle *th)
899 struct osp_thandle *oth = thandle_to_osp_thandle(th);
900 struct osp_update_request *our;
902 if (oth->ot_our != NULL)
905 our = osp_update_request_create(th->th_dev);
907 th->th_result = PTR_ERR(our);
917 void osp_thandle_destroy(struct osp_thandle *oth)
919 LASSERT(oth->ot_magic == OSP_THANDLE_MAGIC);
920 LASSERT(list_empty(&oth->ot_commit_dcb_list));
921 LASSERT(list_empty(&oth->ot_stop_dcb_list));
922 if (oth->ot_our != NULL)
923 osp_update_request_destroy(oth->ot_our);
928 * The OSP layer dt_device_operations::dt_trans_create() interface
929 * to create a transaction.
931 * There are two kinds of transactions that will involve OSP:
933 * 1) If the transaction only contains the updates on remote server
934 * (MDT or OST), such as re-generating the lost OST-object for
935 * LFSCK, then it is a remote transaction. For remote transaction,
936 * the upper layer caller (such as the LFSCK engine) will call the
937 * dt_trans_create() (with the OSP dt_device as the parameter),
938 * then the call will be directed to the osp_trans_create() that
939 * creates the transaction handler and returns it to the caller.
941 * 2) If the transcation contains both local and remote updates,
942 * such as cross MDTs create under DNE mode, then the upper layer
943 * caller will not trigger osp_trans_create(). Instead, it will
944 * call dt_trans_create() on other dt_device, such as LOD that
945 * will generate the transaction handler. Such handler will be
946 * used by the whole transaction in subsequent sub-operations.
948 * \param[in] env pointer to the thread context
949 * \param[in] d pointer to the OSP dt_device
951 * \retval pointer to the transaction handler
952 * \retval negative error number on failure
954 struct thandle *osp_trans_create(const struct lu_env *env, struct dt_device *d)
956 struct osp_thandle *oth;
957 struct thandle *th = NULL;
961 if (unlikely(oth == NULL))
962 RETURN(ERR_PTR(-ENOMEM));
964 oth->ot_magic = OSP_THANDLE_MAGIC;
967 th->th_tags = LCT_TX_HANDLE;
969 atomic_set(&oth->ot_refcount, 1);
970 INIT_LIST_HEAD(&oth->ot_commit_dcb_list);
971 INIT_LIST_HEAD(&oth->ot_stop_dcb_list);
977 * Add commit callback to transaction.
979 * Add commit callback to the osp thandle, which will be called
980 * when the thandle is committed remotely.
982 * \param[in] th the thandle
983 * \param[in] dcb commit callback structure
985 * \retval only return 0 for now.
987 int osp_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
989 struct osp_thandle *oth = thandle_to_osp_thandle(th);
991 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
992 LASSERT(&dcb->dcb_func != NULL);
993 if (dcb->dcb_flags & DCB_TRANS_STOP)
994 list_add(&dcb->dcb_linkage, &oth->ot_stop_dcb_list);
996 list_add(&dcb->dcb_linkage, &oth->ot_commit_dcb_list);
1000 static void osp_trans_commit_cb(struct osp_thandle *oth, int result)
1002 struct dt_txn_commit_cb *dcb;
1003 struct dt_txn_commit_cb *tmp;
1005 LASSERT(atomic_read(&oth->ot_refcount) > 0);
1006 /* call per-transaction callbacks if any */
1007 list_for_each_entry_safe(dcb, tmp, &oth->ot_commit_dcb_list,
1009 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1010 "commit callback entry: magic=%x name='%s'\n",
1011 dcb->dcb_magic, dcb->dcb_name);
1012 list_del_init(&dcb->dcb_linkage);
1013 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1017 static void osp_request_commit_cb(struct ptlrpc_request *req)
1019 struct thandle *th = req->rq_cb_data;
1020 struct osp_thandle *oth;
1021 __u64 last_committed_transno = 0;
1022 int result = req->rq_status;
1028 oth = thandle_to_osp_thandle(th);
1029 if (req->rq_repmsg != NULL &&
1030 lustre_msg_get_last_committed(req->rq_repmsg))
1031 last_committed_transno =
1032 lustre_msg_get_last_committed(req->rq_repmsg);
1034 if (last_committed_transno <
1035 req->rq_import->imp_peer_committed_transno)
1036 last_committed_transno =
1037 req->rq_import->imp_peer_committed_transno;
1039 CDEBUG(D_HA, "trans no "LPU64" committed transno "LPU64"\n",
1040 req->rq_transno, last_committed_transno);
1042 /* If the transaction is not really committed, mark result = 1 */
1043 if (req->rq_transno != 0 &&
1044 (req->rq_transno > last_committed_transno) && result == 0)
1047 osp_trans_commit_cb(oth, result);
1048 req->rq_committed = 1;
1049 osp_thandle_put(oth);
1054 * callback of osp transaction
1056 * Call all of callbacks for this osp thandle. This will only be
1057 * called in error handler path. In the normal processing path,
1058 * these callback will be called in osp_request_commit_cb() and
1059 * osp_update_interpret().
1061 * \param [in] env execution environment
1062 * \param [in] oth osp thandle
1063 * \param [in] rc result of the osp thandle
1065 void osp_trans_callback(const struct lu_env *env,
1066 struct osp_thandle *oth, int rc)
1068 struct osp_update_callback *ouc;
1069 struct osp_update_callback *next;
1071 if (oth->ot_our != NULL) {
1072 list_for_each_entry_safe(ouc, next,
1073 &oth->ot_our->our_cb_items, ouc_list) {
1074 list_del_init(&ouc->ouc_list);
1075 if (ouc->ouc_interpreter != NULL)
1076 ouc->ouc_interpreter(env, NULL, NULL,
1078 ouc->ouc_data, 0, rc);
1079 osp_update_callback_fini(env, ouc);
1082 osp_trans_stop_cb(env, oth, rc);
1083 osp_trans_commit_cb(oth, rc);
1087 * Send the request for remote updates.
1089 * Send updates to the remote MDT. Prepare the request by osp_update_req
1090 * and send them to remote MDT, for sync request, it will wait
1091 * until the reply return, otherwise hand it to ptlrpcd.
1093 * Please refer to osp_trans_create() for transaction type.
1095 * \param[in] env pointer to the thread context
1096 * \param[in] osp pointer to the OSP device
1097 * \param[in] our pointer to the osp_update_request
1099 * \retval 0 for success
1100 * \retval negative error number on failure
1102 static int osp_send_update_req(const struct lu_env *env,
1103 struct osp_device *osp,
1104 struct osp_update_request *our)
1106 struct osp_update_args *args;
1107 struct ptlrpc_request *req;
1108 struct lu_device *top_device;
1109 struct osp_thandle *oth = our->our_th;
1113 LASSERT(oth != NULL);
1114 rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
1117 osp_trans_callback(env, oth, rc);
1121 args = ptlrpc_req_async_args(req);
1122 args->oaua_update = our;
1123 /* set env to NULL, in case the interrupt cb and current function
1124 * are in different thread */
1125 args->oaua_update_env = NULL;
1126 osp_thandle_get(oth); /* hold for update interpret */
1127 req->rq_interpret_reply = osp_update_interpret;
1128 if (!oth->ot_super.th_wait_submit && !oth->ot_super.th_sync) {
1129 if (!osp->opd_imp_active || !osp->opd_imp_connected) {
1130 osp_trans_callback(env, oth, rc);
1131 osp_thandle_put(oth);
1132 GOTO(out, rc = -ENOTCONN);
1135 rc = obd_get_request_slot(&osp->opd_obd->u.cli);
1137 osp_trans_callback(env, oth, rc);
1138 osp_thandle_put(oth);
1139 GOTO(out, rc = -ENOTCONN);
1141 args->oaua_flow_control = true;
1143 if (!osp->opd_connect_mdt) {
1144 down_read(&osp->opd_async_updates_rwsem);
1145 args->oaua_count = &osp->opd_async_updates_count;
1146 args->oaua_waitq = &osp->opd_syn_barrier_waitq;
1147 up_read(&osp->opd_async_updates_rwsem);
1148 atomic_inc(args->oaua_count);
1151 ptlrpcd_add_req(req);
1154 osp_thandle_get(oth); /* hold for commit callback */
1155 req->rq_commit_cb = osp_request_commit_cb;
1156 req->rq_cb_data = &oth->ot_super;
1157 args->oaua_flow_control = false;
1159 /* If the transaction is created during MDT recoverying
1160 * process, it means this is an recovery update, we need
1161 * to let OSP send it anyway without checking recoverying
1162 * status, in case the other target is being recoveried
1163 * at the same time, and if we wait here for the import
1164 * to be recoveryed, it might cause deadlock */
1165 top_device = osp->opd_dt_dev.dd_lu_dev.ld_site->ls_top_dev;
1166 if (top_device->ld_obd->obd_recovering)
1167 req->rq_allow_replay = 1;
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);
1178 if ((rc == -ENOMEM && req->rq_set == NULL) ||
1179 (req->rq_transno == 0 && !req->rq_committed)) {
1180 if (args->oaua_update != NULL) {
1181 /* If osp_update_interpret is not being called,
1182 * release the osp_thandle */
1183 args->oaua_update = NULL;
1184 osp_thandle_put(oth);
1187 req->rq_cb_data = NULL;
1188 rc = rc == 0 ? req->rq_status : rc;
1189 osp_trans_callback(env, oth, rc);
1190 osp_thandle_put(oth);
1196 ptlrpc_req_finished(req);
1202 * Get local thandle for osp_thandle
1204 * Get the local OSD thandle from the OSP thandle. Currently, there
1205 * are a few OSP API (osp_object_create() and osp_sync_add()) needs
1206 * to update the object on local OSD device.
1208 * If the osp_thandle comes from normal stack (MDD->LOD->OSP), then
1209 * we will get local thandle by thandle_get_sub_by_dt.
1211 * If the osp_thandle is remote thandle (th_top == NULL, only used
1212 * by LFSCK), then it will create a local thandle, and stop it in
1213 * osp_trans_stop(). And this only happens on OSP for OST.
1215 * These are temporary solution, once OSP accessing OSD object is
1216 * being fixed properly, this function should be removed. XXX
1218 * \param[in] env pointer to the thread context
1219 * \param[in] th pointer to the transaction handler
1220 * \param[in] dt pointer to the OSP device
1222 * \retval pointer to the local thandle
1223 * \retval ERR_PTR(errno) if it fails.
1225 struct thandle *osp_get_storage_thandle(const struct lu_env *env,
1227 struct osp_device *osp)
1229 struct osp_thandle *oth;
1230 struct thandle *local_th;
1232 if (th->th_top != NULL)
1233 return thandle_get_sub_by_dt(env, th->th_top,
1236 LASSERT(!osp->opd_connect_mdt);
1237 oth = thandle_to_osp_thandle(th);
1238 if (oth->ot_storage_th != NULL)
1239 return oth->ot_storage_th;
1241 local_th = dt_trans_create(env, osp->opd_storage);
1242 if (IS_ERR(local_th))
1245 oth->ot_storage_th = local_th;
1251 * Set version for the transaction
1253 * Set the version for the transaction and add the request to
1254 * the sending list, then after transaction stop, the request
1255 * will be picked in the order of version, by sending thread.
1257 * \param [in] oth osp thandle to be set version.
1259 * \retval 0 if set version succeeds
1260 * negative errno if set version fails.
1262 int osp_check_and_set_rpc_version(struct osp_thandle *oth,
1263 struct osp_object *obj)
1265 struct osp_device *osp = dt2osp_dev(oth->ot_super.th_dev);
1266 struct osp_updates *ou = osp->opd_update;
1271 if (oth->ot_our->our_version != 0)
1274 spin_lock(&ou->ou_lock);
1275 spin_lock(&oth->ot_our->our_list_lock);
1276 if (obj->opo_stale) {
1277 spin_unlock(&oth->ot_our->our_list_lock);
1278 spin_unlock(&ou->ou_lock);
1282 /* Assign the version and add it to the sending list */
1283 osp_thandle_get(oth);
1284 oth->ot_our->our_version = ou->ou_version++;
1285 list_add_tail(&oth->ot_our->our_list,
1286 &osp->opd_update->ou_list);
1287 oth->ot_our->our_req_ready = 0;
1288 spin_unlock(&oth->ot_our->our_list_lock);
1289 spin_unlock(&ou->ou_lock);
1291 LASSERT(oth->ot_super.th_wait_submit == 1);
1292 CDEBUG(D_INFO, "%s: version "LPU64" oth:version %p:"LPU64"\n",
1293 osp->opd_obd->obd_name, ou->ou_version, oth,
1294 oth->ot_our->our_version);
1300 * Get next OSP update request in the sending list
1301 * Get next OSP update request in the sending list by version number, next
1303 * 1. transaction which does not have a version number.
1304 * 2. transaction whose version == opd_rpc_version.
1306 * \param [in] ou osp update structure.
1307 * \param [out] ourp the pointer holding the next update request.
1309 * \retval true if getting the next transaction.
1310 * \retval false if not getting the next transaction.
1313 osp_get_next_request(struct osp_updates *ou, struct osp_update_request **ourp)
1315 struct osp_update_request *our;
1316 struct osp_update_request *tmp;
1317 bool got_req = false;
1319 spin_lock(&ou->ou_lock);
1320 list_for_each_entry_safe(our, tmp, &ou->ou_list, our_list) {
1321 LASSERT(our->our_th != NULL);
1322 CDEBUG(D_HA, "ou %p version "LPU64" rpc_version "LPU64"\n",
1323 ou, our->our_version, ou->ou_rpc_version);
1324 spin_lock(&our->our_list_lock);
1325 /* Find next osp_update_request in the list */
1326 if (our->our_version == ou->ou_rpc_version &&
1327 our->our_req_ready) {
1328 list_del_init(&our->our_list);
1329 spin_unlock(&our->our_list_lock);
1334 spin_unlock(&our->our_list_lock);
1336 spin_unlock(&ou->ou_lock);
1342 * Invalidate update request
1344 * Invalidate update request in the OSP sending list, so all of
1345 * requests in the sending list will return error, which happens
1346 * when it finds one update (with writing llog) requests fails or
1347 * the OSP is evicted by remote target. see osp_send_update_thread().
1349 * \param[in] osp OSP device whose update requests will be
1352 void osp_invalidate_request(struct osp_device *osp)
1355 struct osp_updates *ou = osp->opd_update;
1356 struct osp_update_request *our;
1357 struct osp_update_request *tmp;
1365 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1367 CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
1372 INIT_LIST_HEAD(&list);
1374 spin_lock(&ou->ou_lock);
1375 /* invalidate all of request in the sending list */
1376 list_for_each_entry_safe(our, tmp, &ou->ou_list, our_list) {
1377 spin_lock(&our->our_list_lock);
1378 if (our->our_req_ready)
1379 list_move(&our->our_list, &list);
1381 list_del_init(&our->our_list);
1383 if (our->our_th->ot_super.th_result == 0)
1384 our->our_th->ot_super.th_result = -EIO;
1386 if (our->our_version >= ou->ou_rpc_version)
1387 ou->ou_rpc_version = our->our_version + 1;
1388 spin_unlock(&our->our_list_lock);
1390 CDEBUG(D_HA, "%s invalidate our %p\n", osp->opd_obd->obd_name,
1394 spin_unlock(&ou->ou_lock);
1396 /* invalidate all of request in the sending list */
1397 list_for_each_entry_safe(our, tmp, &list, our_list) {
1398 spin_lock(&our->our_list_lock);
1399 list_del_init(&our->our_list);
1400 spin_unlock(&our->our_list_lock);
1401 osp_trans_callback(&env, our->our_th,
1402 our->our_th->ot_super.th_result);
1403 osp_thandle_put(our->our_th);
1409 * Sending update thread
1411 * Create thread to send update request to other MDTs, this thread will pull
1412 * out update request from the list in OSP by version number, i.e. it will
1413 * make sure the update request with lower version number will be sent first.
1415 * \param[in] arg hold the OSP device.
1417 * \retval 0 if the thread is created successfully.
1418 * \retal negative error if the thread is not created
1421 int osp_send_update_thread(void *arg)
1424 struct osp_device *osp = arg;
1425 struct l_wait_info lwi = { 0 };
1426 struct osp_updates *ou = osp->opd_update;
1427 struct ptlrpc_thread *thread = &osp->opd_update_thread;
1428 struct osp_update_request *our = NULL;
1432 LASSERT(ou != NULL);
1433 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1435 CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
1440 thread->t_flags = SVC_RUNNING;
1441 wake_up(&thread->t_ctl_waitq);
1444 l_wait_event(ou->ou_waitq,
1445 !osp_send_update_thread_running(osp) ||
1446 osp_get_next_request(ou, &our), &lwi);
1448 if (!osp_send_update_thread_running(osp)) {
1450 osp_trans_callback(&env, our->our_th, -EINTR);
1451 osp_thandle_put(our->our_th);
1456 LASSERT(our->our_th != NULL);
1457 if (our->our_th->ot_super.th_result != 0) {
1458 osp_trans_callback(&env, our->our_th,
1459 our->our_th->ot_super.th_result);
1460 rc = our->our_th->ot_super.th_result;
1461 } else if (OBD_FAIL_CHECK(OBD_FAIL_INVALIDATE_UPDATE)) {
1463 osp_trans_callback(&env, our->our_th, rc);
1465 rc = osp_send_update_req(&env, osp, our);
1468 /* Update the rpc version */
1469 spin_lock(&ou->ou_lock);
1470 if (our->our_version == ou->ou_rpc_version)
1471 ou->ou_rpc_version++;
1472 spin_unlock(&ou->ou_lock);
1474 /* If one update request fails, let's fail all of the requests
1475 * in the sending list, because the request in the sending
1476 * list are dependent on either other, continue sending these
1477 * request might cause llog or filesystem corruption */
1479 osp_invalidate_request(osp);
1481 /* Balanced for thandle_get in osp_check_and_set_rpc_version */
1482 osp_thandle_put(our->our_th);
1485 thread->t_flags = SVC_STOPPED;
1487 wake_up(&thread->t_ctl_waitq);
1493 * The OSP layer dt_device_operations::dt_trans_start() interface
1494 * to start the transaction.
1496 * If the transaction is a remote transaction, then related remote
1497 * updates will be triggered in the osp_trans_stop().
1498 * Please refer to osp_trans_create() for transaction type.
1500 * \param[in] env pointer to the thread context
1501 * \param[in] dt pointer to the OSP dt_device
1502 * \param[in] th pointer to the transaction handler
1504 * \retval 0 for success
1505 * \retval negative error number on failure
1507 int osp_trans_start(const struct lu_env *env, struct dt_device *dt,
1510 struct osp_thandle *oth = thandle_to_osp_thandle(th);
1512 if (oth->ot_super.th_sync)
1513 oth->ot_our->our_flags |= UPDATE_FL_SYNC;
1514 /* For remote thandle, if there are local thandle, start it here*/
1515 if (is_only_remote_trans(th) && oth->ot_storage_th != NULL)
1516 return dt_trans_start(env, oth->ot_storage_th->th_dev,
1517 oth->ot_storage_th);
1522 * The OSP layer dt_device_operations::dt_trans_stop() interface
1523 * to stop the transaction.
1525 * If the transaction is a remote transaction, related remote
1526 * updates will be triggered at the end of this function.
1528 * For synchronous mode update or any failed update, the request
1529 * will be destroyed explicitly when the osp_trans_stop().
1531 * Please refer to osp_trans_create() for transaction type.
1533 * \param[in] env pointer to the thread context
1534 * \param[in] dt pointer to the OSP dt_device
1535 * \param[in] th pointer to the transaction handler
1537 * \retval 0 for success
1538 * \retval negative error number on failure
1540 int osp_trans_stop(const struct lu_env *env, struct dt_device *dt,
1543 struct osp_thandle *oth = thandle_to_osp_thandle(th);
1544 struct osp_update_request *our = oth->ot_our;
1545 struct osp_device *osp = dt2osp_dev(dt);
1549 /* For remote transaction, if there is local storage thandle,
1551 if (oth->ot_storage_th != NULL && th->th_top == NULL) {
1552 dt_trans_stop(env, oth->ot_storage_th->th_dev,
1553 oth->ot_storage_th);
1554 oth->ot_storage_th = NULL;
1557 if (our == NULL || list_empty(&our->our_req_list)) {
1558 osp_trans_callback(env, oth, th->th_result);
1559 GOTO(out, rc = th->th_result);
1562 if (!osp->opd_connect_mdt) {
1563 osp_trans_callback(env, oth, th->th_result);
1564 rc = osp_send_update_req(env, osp, oth->ot_our);
1568 if (osp->opd_update == NULL ||
1569 !osp_send_update_thread_running(osp)) {
1570 osp_trans_callback(env, oth, -EIO);
1571 GOTO(out, rc = -EIO);
1574 CDEBUG(D_HA, "%s: add oth %p with version "LPU64"\n",
1575 osp->opd_obd->obd_name, oth, our->our_version);
1577 LASSERT(our->our_req_ready == 0);
1578 spin_lock(&our->our_list_lock);
1579 if (likely(!list_empty(&our->our_list))) {
1580 /* notify sending thread */
1581 our->our_req_ready = 1;
1582 wake_up(&osp->opd_update->ou_waitq);
1583 spin_unlock(&our->our_list_lock);
1584 } else if (th->th_result == 0) {
1585 /* if the request does not needs to be serialized,
1586 * read-only request etc, let's send it right away */
1587 spin_unlock(&our->our_list_lock);
1588 rc = osp_send_update_req(env, osp, our);
1590 spin_unlock(&our->our_list_lock);
1591 osp_trans_callback(env, oth, th->th_result);
1594 osp_thandle_put(oth);