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, 2016, 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 INIT_LIST_HEAD(&our->our_invalidate_cb_list);
200 spin_lock_init(&our->our_list_lock);
202 rc = osp_object_update_request_create(our, OUT_UPDATE_INIT_BUFFER_SIZE);
210 void osp_update_request_destroy(const struct lu_env *env,
211 struct osp_update_request *our)
213 struct osp_update_request_sub *ours;
214 struct osp_update_request_sub *tmp;
219 list_for_each_entry_safe(ours, tmp, &our->our_req_list, ours_list) {
220 list_del(&ours->ours_list);
221 if (ours->ours_req != NULL)
222 OBD_FREE_LARGE(ours->ours_req, ours->ours_req_size);
226 if (!list_empty(&our->our_invalidate_cb_list)) {
228 struct osp_object *obj;
229 struct osp_object *next;
232 lu_env_init(&lenv, LCT_MD_THREAD | LCT_DT_THREAD);
236 list_for_each_entry_safe(obj, next,
237 &our->our_invalidate_cb_list,
238 opo_invalidate_cb_list) {
239 spin_lock(&obj->opo_lock);
240 list_del_init(&obj->opo_invalidate_cb_list);
241 spin_unlock(&obj->opo_lock);
243 dt_object_put(env, &obj->opo_obj);
254 object_update_request_dump(const struct object_update_request *ourq,
258 size_t total_size = 0;
260 for (i = 0; i < ourq->ourq_count; i++) {
261 struct object_update *update;
264 update = object_update_request_get(ourq, i, &size);
265 LASSERT(update != NULL);
266 CDEBUG(mask, "i = %u fid = "DFID" op = %s "
267 "params = %d batchid = %llu size = %zu repsize %u\n",
268 i, PFID(&update->ou_fid),
269 update_op_str(update->ou_type),
270 update->ou_params_count,
271 update->ou_batchid, size,
272 (unsigned)update->ou_result_size);
277 CDEBUG(mask, "updates = %p magic = %x count = %d size = %zu\n", ourq,
278 ourq->ourq_magic, ourq->ourq_count, total_size);
282 * Prepare inline update request
284 * Prepare OUT update ptlrpc inline request, and the request usually includes
285 * one update buffer, which does not need bulk transfer.
287 * \param[in] env execution environment
288 * \param[in] req ptlrpc request
289 * \param[in] ours sub osp_update_request to be packed
291 * \retval 0 if packing succeeds
292 * \retval negative errno if packing fails
294 int osp_prep_inline_update_req(const struct lu_env *env,
295 struct ptlrpc_request *req,
296 struct osp_update_request *our,
299 struct osp_update_request_sub *ours;
300 struct out_update_header *ouh;
301 __u32 update_req_size;
304 ours = list_entry(our->our_req_list.next,
305 struct osp_update_request_sub, ours_list);
306 update_req_size = object_update_request_size(ours->ours_req);
307 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_HEADER, RCL_CLIENT,
308 update_req_size + sizeof(*ouh));
310 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, OUT_UPDATE);
314 ouh = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_HEADER);
315 ouh->ouh_magic = OUT_UPDATE_HEADER_MAGIC;
317 ouh->ouh_inline_length = update_req_size;
318 ouh->ouh_reply_size = repsize;
320 memcpy(ouh->ouh_inline_data, ours->ours_req, update_req_size);
322 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_REPLY,
323 RCL_SERVER, repsize);
325 ptlrpc_request_set_replen(req);
326 req->rq_request_portal = OUT_PORTAL;
327 req->rq_reply_portal = OSC_REPLY_PORTAL;
333 * Prepare update request.
335 * Prepare OUT update ptlrpc request, and the request usually includes
336 * all of updates (stored in \param ureq) from one operation.
338 * \param[in] env execution environment
339 * \param[in] imp import on which ptlrpc request will be sent
340 * \param[in] ureq hold all of updates which will be packed into the req
341 * \param[in] reqp request to be created
343 * \retval 0 if preparation succeeds.
344 * \retval negative errno if preparation fails.
346 int osp_prep_update_req(const struct lu_env *env, struct obd_import *imp,
347 struct osp_update_request *our,
348 struct ptlrpc_request **reqp)
350 struct ptlrpc_request *req;
351 struct ptlrpc_bulk_desc *desc;
352 struct osp_update_request_sub *ours;
353 const struct object_update_request *ourq;
354 struct out_update_header *ouh;
355 struct out_update_buffer *oub;
358 struct object_update_reply *reply;
363 list_for_each_entry(ours, &our->our_req_list, ours_list) {
364 object_update_request_dump(ours->ours_req, D_INFO);
366 ourq = ours->ours_req;
367 for (i = 0; i < ourq->ourq_count; i++) {
368 struct object_update *update;
372 /* XXX: it's very inefficient to lookup update
373 * this way, iterating from the beginning
375 update = object_update_request_get(ourq, i, &size);
376 LASSERT(update != NULL);
378 repsize += sizeof(reply->ourp_lens[0]);
379 repsize += sizeof(struct object_update_result);
380 repsize += update->ou_result_size;
385 repsize += sizeof(*reply);
386 repsize = (repsize + OUT_UPDATE_REPLY_SIZE - 1) &
387 ~(OUT_UPDATE_REPLY_SIZE - 1);
388 LASSERT(buf_count > 0);
390 req = ptlrpc_request_alloc(imp, &RQF_OUT_UPDATE);
394 if (buf_count == 1) {
395 ours = list_entry(our->our_req_list.next,
396 struct osp_update_request_sub, ours_list);
398 /* Let's check if it can be packed inline */
399 if (object_update_request_size(ours->ours_req) +
400 sizeof(struct out_update_header) <
401 OUT_UPDATE_MAX_INLINE_SIZE) {
402 rc = osp_prep_inline_update_req(env, req, our, repsize);
409 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_HEADER, RCL_CLIENT,
410 sizeof(struct osp_update_request));
412 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_BUF, RCL_CLIENT,
413 buf_count * sizeof(*oub));
415 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, OUT_UPDATE);
419 ouh = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_HEADER);
420 ouh->ouh_magic = OUT_UPDATE_HEADER_MAGIC;
421 ouh->ouh_count = buf_count;
422 ouh->ouh_inline_length = 0;
423 ouh->ouh_reply_size = repsize;
424 oub = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_BUF);
425 list_for_each_entry(ours, &our->our_req_list, ours_list) {
426 oub->oub_size = ours->ours_req_size;
430 req->rq_bulk_write = 1;
431 desc = ptlrpc_prep_bulk_imp(req, buf_count,
432 MD_MAX_BRW_SIZE >> LNET_MTU_BITS,
433 PTLRPC_BULK_GET_SOURCE | PTLRPC_BULK_BUF_KVEC,
434 MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
436 GOTO(out_req, rc = -ENOMEM);
438 /* NB req now owns desc and will free it when it gets freed */
439 list_for_each_entry(ours, &our->our_req_list, ours_list) {
440 desc->bd_frag_ops->add_iov_frag(desc, ours->ours_req,
441 ours->ours_req_size);
442 total += ours->ours_req_size;
444 CDEBUG(D_OTHER, "total %d in %u\n", total, our->our_update_nr);
446 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_REPLY,
447 RCL_SERVER, repsize);
449 ptlrpc_request_set_replen(req);
450 req->rq_request_portal = OUT_PORTAL;
451 req->rq_reply_portal = OSC_REPLY_PORTAL;
456 ptlrpc_req_finished(req);
464 * Send update request to the remote MDT synchronously.
466 * \param[in] env execution environment
467 * \param[in] imp import on which ptlrpc request will be sent
468 * \param[in] our hold all of updates which will be packed into the req
469 * \param[in] reqp request to be created
471 * \retval 0 if RPC succeeds.
472 * \retval negative errno if RPC fails.
474 int osp_remote_sync(const struct lu_env *env, struct osp_device *osp,
475 struct osp_update_request *our,
476 struct ptlrpc_request **reqp)
478 struct obd_import *imp = osp->opd_obd->u.cli.cl_import;
479 struct ptlrpc_request *req = NULL;
483 rc = osp_prep_update_req(env, imp, our, &req);
487 osp_set_req_replay(osp, req);
488 req->rq_allow_intr = 1;
490 /* Note: some dt index api might return non-zero result here, like
491 * osd_index_ea_lookup, so we should only check rc < 0 here */
492 rc = ptlrpc_queue_wait(req);
494 if (rc < 0 || reqp == NULL)
495 ptlrpc_req_finished(req);
503 * Invalidate all objects in the osp thandle
505 * invalidate all of objects in the update request, which will be called
506 * when the transaction is aborted.
508 * \param[in] oth osp thandle.
510 static void osp_thandle_invalidate_object(const struct lu_env *env,
511 struct osp_thandle *oth,
514 struct osp_update_request *our = oth->ot_our;
515 struct osp_object *obj;
516 struct osp_object *next;
521 list_for_each_entry_safe(obj, next, &our->our_invalidate_cb_list,
522 opo_invalidate_cb_list) {
524 osp_invalidate(env, &obj->opo_obj);
526 spin_lock(&obj->opo_lock);
527 list_del_init(&obj->opo_invalidate_cb_list);
528 spin_unlock(&obj->opo_lock);
530 dt_object_put(env, &obj->opo_obj);
534 static void osp_trans_stop_cb(const struct lu_env *env,
535 struct osp_thandle *oth, int result)
537 struct dt_txn_commit_cb *dcb;
538 struct dt_txn_commit_cb *tmp;
540 /* call per-transaction stop callbacks if any */
541 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
543 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
544 "commit callback entry: magic=%x name='%s'\n",
545 dcb->dcb_magic, dcb->dcb_name);
546 list_del_init(&dcb->dcb_linkage);
547 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
550 osp_thandle_invalidate_object(env, oth, result);
554 * Allocate an osp request and initialize it with the given parameters.
556 * \param[in] obj pointer to the operation target
557 * \param[in] data pointer to the data used by the interpreter
558 * \param[in] interpreter pointer to the interpreter function
560 * \retval pointer to the asychronous request
561 * \retval NULL if the allocation failed
563 static struct osp_update_callback *
564 osp_update_callback_init(struct osp_object *obj, void *data,
565 osp_update_interpreter_t interpreter)
567 struct osp_update_callback *ouc;
573 lu_object_get(osp2lu_obj(obj));
574 INIT_LIST_HEAD(&ouc->ouc_list);
576 ouc->ouc_data = data;
577 ouc->ouc_interpreter = interpreter;
583 * Destroy the osp_update_callback.
585 * \param[in] env pointer to the thread context
586 * \param[in] ouc pointer to osp_update_callback
588 static void osp_update_callback_fini(const struct lu_env *env,
589 struct osp_update_callback *ouc)
591 LASSERT(list_empty(&ouc->ouc_list));
593 lu_object_put(env, osp2lu_obj(ouc->ouc_obj));
598 * Interpret the packaged OUT RPC results.
600 * For every packaged sub-request, call its registered interpreter function.
601 * Then destroy the sub-request.
603 * \param[in] env pointer to the thread context
604 * \param[in] req pointer to the RPC
605 * \param[in] arg pointer to data used by the interpreter
606 * \param[in] rc the RPC return value
608 * \retval 0 for success
609 * \retval negative error number on failure
611 static int osp_update_interpret(const struct lu_env *env,
612 struct ptlrpc_request *req, void *arg, int rc)
614 struct object_update_reply *reply = NULL;
615 struct osp_update_args *oaua = arg;
616 struct osp_update_request *our = oaua->oaua_update;
617 struct osp_thandle *oth;
618 struct osp_update_callback *ouc;
619 struct osp_update_callback *next;
629 /* Sigh env might be NULL in some cases, see
631 * osp_send_update_thread()
632 * ptlrpc_set_wait() ----> null env.
634 * osp_update_interpret()
635 * Let's use env in oaua for this case.
638 env = oaua->oaua_update_env;
640 oaua->oaua_update = NULL;
642 if (oaua->oaua_flow_control) {
643 struct osp_device *osp;
645 LASSERT(oth != NULL);
646 osp = dt2osp_dev(oth->ot_super.th_dev);
647 obd_put_request_slot(&osp->opd_obd->u.cli);
650 /* Unpack the results from the reply message. */
651 if (req->rq_repmsg != NULL && req->rq_replied) {
652 reply = req_capsule_server_sized_get(&req->rq_pill,
653 &RMF_OUT_UPDATE_REPLY,
654 OUT_UPDATE_REPLY_SIZE);
655 if (reply == NULL || reply->ourp_magic != UPDATE_REPLY_MAGIC) {
659 count = reply->ourp_count;
663 list_for_each_entry_safe(ouc, next, &our->our_cb_items, ouc_list) {
664 list_del_init(&ouc->ouc_list);
666 /* The peer may only have handled some requests (indicated
667 * by the 'count') in the packaged OUT RPC, we can only get
668 * results for the handled part. */
669 if (index < count && reply->ourp_lens[index] > 0 && rc >= 0) {
670 struct object_update_result *result;
672 result = object_update_result_get(reply, index, NULL);
676 rc1 = rc = result->our_rc;
677 } else if (rc1 >= 0) {
678 /* The peer did not handle these request, let's return
679 * -EINVAL to update interpret for now */
686 if (ouc->ouc_interpreter != NULL)
687 ouc->ouc_interpreter(env, reply, req, ouc->ouc_obj,
688 ouc->ouc_data, index, rc1);
690 osp_update_callback_fini(env, ouc);
694 if (oaua->oaua_count != NULL && atomic_dec_and_test(oaua->oaua_count))
695 wake_up_all(oaua->oaua_waitq);
698 /* oth and osp_update_requests will be destoryed in
700 osp_trans_stop_cb(env, oth, rc);
701 osp_thandle_put(env, oth);
703 osp_update_request_destroy(env, our);
710 * Pack all the requests in the shared asynchronous idempotent request queue
711 * into a single OUT RPC that will be given to the background ptlrpcd daemon.
713 * \param[in] env pointer to the thread context
714 * \param[in] osp pointer to the OSP device
715 * \param[in] our pointer to the shared queue
717 * \retval 0 for success
718 * \retval negative error number on failure
720 int osp_unplug_async_request(const struct lu_env *env,
721 struct osp_device *osp,
722 struct osp_update_request *our)
724 struct osp_update_args *args;
725 struct ptlrpc_request *req = NULL;
728 rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
731 struct osp_update_callback *ouc;
732 struct osp_update_callback *next;
734 list_for_each_entry_safe(ouc, next,
735 &our->our_cb_items, ouc_list) {
736 list_del_init(&ouc->ouc_list);
737 if (ouc->ouc_interpreter != NULL)
738 ouc->ouc_interpreter(env, NULL, NULL,
740 ouc->ouc_data, 0, rc);
741 osp_update_callback_fini(env, ouc);
743 osp_update_request_destroy(env, our);
745 args = ptlrpc_req_async_args(req);
746 args->oaua_update = our;
747 args->oaua_count = NULL;
748 args->oaua_waitq = NULL;
749 /* Note: this is asynchronous call for the request, so the
750 * interrupte cb and current function will be different
751 * thread, so we need use different env */
752 args->oaua_update_env = NULL;
753 args->oaua_flow_control = false;
754 req->rq_interpret_reply = osp_update_interpret;
755 ptlrpcd_add_req(req);
762 * Find or create (if NOT exist or purged) the shared asynchronous idempotent
763 * request queue - osp_device::opd_async_requests.
765 * If the osp_device::opd_async_requests is not NULL, then return it directly;
766 * otherwise create new osp_update_request and attach it to opd_async_requests.
768 * \param[in] osp pointer to the OSP device
770 * \retval pointer to the shared queue
771 * \retval negative error number on failure
773 static struct osp_update_request *
774 osp_find_or_create_async_update_request(struct osp_device *osp)
776 struct osp_update_request *our = osp->opd_async_requests;
781 our = osp_update_request_create(&osp->opd_dt_dev);
785 osp->opd_async_requests = our;
791 * Insert an osp_update_callback into the osp_update_request.
793 * Insert an osp_update_callback to the osp_update_request. Usually each update
794 * in the osp_update_request will have one correspondent callback, and these
795 * callbacks will be called in rq_interpret_reply.
797 * \param[in] env pointer to the thread context
798 * \param[in] obj pointer to the operation target object
799 * \param[in] data pointer to the data used by the interpreter
800 * \param[in] interpreter pointer to the interpreter function
802 * \retval 0 for success
803 * \retval negative error number on failure
805 int osp_insert_update_callback(const struct lu_env *env,
806 struct osp_update_request *our,
807 struct osp_object *obj, void *data,
808 osp_update_interpreter_t interpreter)
810 struct osp_update_callback *ouc;
812 ouc = osp_update_callback_init(obj, data, interpreter);
816 list_add_tail(&ouc->ouc_list, &our->our_cb_items);
822 * Insert an asynchronous idempotent request to the shared request queue that
823 * is attached to the osp_device.
825 * This function generates a new osp_async_request with the given parameters,
826 * then tries to insert the request into the osp_device-based shared request
827 * queue. If the queue is full, then triggers the packaged OUT RPC to purge
828 * the shared queue firstly, and then re-tries.
830 * NOTE: must hold the osp::opd_async_requests_mutex to serialize concurrent
831 * osp_insert_async_request call from others.
833 * \param[in] env pointer to the thread context
834 * \param[in] op operation type, see 'enum update_type'
835 * \param[in] obj pointer to the operation target
836 * \param[in] count array size of the subsequent \a lens and \a bufs
837 * \param[in] lens buffer length array for the subsequent \a bufs
838 * \param[in] bufs the buffers to compose the request
839 * \param[in] data pointer to the data used by the interpreter
840 * \param[in] repsize how many bytes the caller allocated for \a data
841 * \param[in] interpreter pointer to the interpreter function
843 * \retval 0 for success
844 * \retval negative error number on failure
846 int osp_insert_async_request(const struct lu_env *env, enum update_type op,
847 struct osp_object *obj, int count,
848 __u16 *lens, const void **bufs,
849 void *data, __u32 repsize,
850 osp_update_interpreter_t interpreter)
852 struct osp_device *osp;
853 struct osp_update_request *our;
854 struct object_update *object_update;
855 size_t max_update_size;
856 struct object_update_request *ureq;
857 struct osp_update_request_sub *ours;
861 osp = lu2osp_dev(osp2lu_obj(obj)->lo_dev);
862 our = osp_find_or_create_async_update_request(osp);
864 RETURN(PTR_ERR(our));
867 ours = osp_current_object_update_request(our);
869 ureq = ours->ours_req;
870 max_update_size = ours->ours_req_size -
871 object_update_request_size(ureq);
873 object_update = update_buffer_get_update(ureq, ureq->ourq_count);
874 rc = out_update_pack(env, object_update, &max_update_size, op,
875 lu_object_fid(osp2lu_obj(obj)), count, lens, bufs,
877 /* The queue is full. */
879 osp->opd_async_requests = NULL;
880 mutex_unlock(&osp->opd_async_requests_mutex);
882 rc = osp_unplug_async_request(env, osp, our);
883 mutex_lock(&osp->opd_async_requests_mutex);
887 our = osp_find_or_create_async_update_request(osp);
889 RETURN(PTR_ERR(our));
897 our->our_update_nr++;
900 rc = osp_insert_update_callback(env, our, obj, data, interpreter);
905 int osp_trans_update_request_create(struct thandle *th)
907 struct osp_thandle *oth = thandle_to_osp_thandle(th);
908 struct osp_update_request *our;
910 if (oth->ot_our != NULL)
913 our = osp_update_request_create(th->th_dev);
915 th->th_result = PTR_ERR(our);
925 void osp_thandle_destroy(const struct lu_env *env,
926 struct osp_thandle *oth)
928 LASSERT(oth->ot_magic == OSP_THANDLE_MAGIC);
929 LASSERT(list_empty(&oth->ot_commit_dcb_list));
930 LASSERT(list_empty(&oth->ot_stop_dcb_list));
931 if (oth->ot_our != NULL)
932 osp_update_request_destroy(env, oth->ot_our);
937 * The OSP layer dt_device_operations::dt_trans_create() interface
938 * to create a transaction.
940 * There are two kinds of transactions that will involve OSP:
942 * 1) If the transaction only contains the updates on remote server
943 * (MDT or OST), such as re-generating the lost OST-object for
944 * LFSCK, then it is a remote transaction. For remote transaction,
945 * the upper layer caller (such as the LFSCK engine) will call the
946 * dt_trans_create() (with the OSP dt_device as the parameter),
947 * then the call will be directed to the osp_trans_create() that
948 * creates the transaction handler and returns it to the caller.
950 * 2) If the transcation contains both local and remote updates,
951 * such as cross MDTs create under DNE mode, then the upper layer
952 * caller will not trigger osp_trans_create(). Instead, it will
953 * call dt_trans_create() on other dt_device, such as LOD that
954 * will generate the transaction handler. Such handler will be
955 * used by the whole transaction in subsequent sub-operations.
957 * \param[in] env pointer to the thread context
958 * \param[in] d pointer to the OSP dt_device
960 * \retval pointer to the transaction handler
961 * \retval negative error number on failure
963 struct thandle *osp_trans_create(const struct lu_env *env, struct dt_device *d)
965 struct osp_thandle *oth;
966 struct thandle *th = NULL;
970 if (unlikely(oth == NULL))
971 RETURN(ERR_PTR(-ENOMEM));
973 oth->ot_magic = OSP_THANDLE_MAGIC;
976 th->th_tags = LCT_TX_HANDLE;
978 atomic_set(&oth->ot_refcount, 1);
979 INIT_LIST_HEAD(&oth->ot_commit_dcb_list);
980 INIT_LIST_HEAD(&oth->ot_stop_dcb_list);
986 * Add commit callback to transaction.
988 * Add commit callback to the osp thandle, which will be called
989 * when the thandle is committed remotely.
991 * \param[in] th the thandle
992 * \param[in] dcb commit callback structure
994 * \retval only return 0 for now.
996 int osp_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
998 struct osp_thandle *oth = thandle_to_osp_thandle(th);
1000 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1001 LASSERT(&dcb->dcb_func != NULL);
1002 if (dcb->dcb_flags & DCB_TRANS_STOP)
1003 list_add(&dcb->dcb_linkage, &oth->ot_stop_dcb_list);
1005 list_add(&dcb->dcb_linkage, &oth->ot_commit_dcb_list);
1009 static void osp_trans_commit_cb(struct osp_thandle *oth, int result)
1011 struct dt_txn_commit_cb *dcb;
1012 struct dt_txn_commit_cb *tmp;
1014 LASSERT(atomic_read(&oth->ot_refcount) > 0);
1015 /* call per-transaction callbacks if any */
1016 list_for_each_entry_safe(dcb, tmp, &oth->ot_commit_dcb_list,
1018 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1019 "commit callback entry: magic=%x name='%s'\n",
1020 dcb->dcb_magic, dcb->dcb_name);
1021 list_del_init(&dcb->dcb_linkage);
1022 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1026 static void osp_request_commit_cb(struct ptlrpc_request *req)
1028 struct thandle *th = req->rq_cb_data;
1029 struct osp_thandle *oth;
1030 __u64 last_committed_transno = 0;
1031 int result = req->rq_status;
1037 oth = thandle_to_osp_thandle(th);
1038 if (req->rq_repmsg != NULL &&
1039 lustre_msg_get_last_committed(req->rq_repmsg))
1040 last_committed_transno =
1041 lustre_msg_get_last_committed(req->rq_repmsg);
1043 if (last_committed_transno <
1044 req->rq_import->imp_peer_committed_transno)
1045 last_committed_transno =
1046 req->rq_import->imp_peer_committed_transno;
1048 CDEBUG(D_HA, "trans no %llu committed transno %llu\n",
1049 req->rq_transno, last_committed_transno);
1051 /* If the transaction is not really committed, mark result = 1 */
1052 if (req->rq_transno != 0 &&
1053 (req->rq_transno > last_committed_transno) && result == 0)
1056 osp_trans_commit_cb(oth, result);
1057 req->rq_committed = 1;
1058 osp_thandle_put(NULL, oth);
1063 * callback of osp transaction
1065 * Call all of callbacks for this osp thandle. This will only be
1066 * called in error handler path. In the normal processing path,
1067 * these callback will be called in osp_request_commit_cb() and
1068 * osp_update_interpret().
1070 * \param [in] env execution environment
1071 * \param [in] oth osp thandle
1072 * \param [in] rc result of the osp thandle
1074 void osp_trans_callback(const struct lu_env *env,
1075 struct osp_thandle *oth, int rc)
1077 struct osp_update_callback *ouc;
1078 struct osp_update_callback *next;
1080 if (oth->ot_our != NULL) {
1081 list_for_each_entry_safe(ouc, next,
1082 &oth->ot_our->our_cb_items, ouc_list) {
1083 list_del_init(&ouc->ouc_list);
1084 if (ouc->ouc_interpreter != NULL)
1085 ouc->ouc_interpreter(env, NULL, NULL,
1087 ouc->ouc_data, 0, rc);
1088 osp_update_callback_fini(env, ouc);
1091 osp_trans_stop_cb(env, oth, rc);
1092 osp_trans_commit_cb(oth, rc);
1096 * Send the request for remote updates.
1098 * Send updates to the remote MDT. Prepare the request by osp_update_req
1099 * and send them to remote MDT, for sync request, it will wait
1100 * until the reply return, otherwise hand it to ptlrpcd.
1102 * Please refer to osp_trans_create() for transaction type.
1104 * \param[in] env pointer to the thread context
1105 * \param[in] osp pointer to the OSP device
1106 * \param[in] our pointer to the osp_update_request
1108 * \retval 0 for success
1109 * \retval negative error number on failure
1111 static int osp_send_update_req(const struct lu_env *env,
1112 struct osp_device *osp,
1113 struct osp_update_request *our)
1115 struct osp_update_args *args;
1116 struct ptlrpc_request *req;
1117 struct osp_thandle *oth = our->our_th;
1121 LASSERT(oth != NULL);
1122 rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
1125 osp_trans_callback(env, oth, rc);
1129 args = ptlrpc_req_async_args(req);
1130 args->oaua_update = our;
1131 /* set env to NULL, in case the interrupt cb and current function
1132 * are in different thread */
1133 args->oaua_update_env = NULL;
1134 osp_thandle_get(oth); /* hold for update interpret */
1135 req->rq_interpret_reply = osp_update_interpret;
1136 if (!oth->ot_super.th_wait_submit && !oth->ot_super.th_sync) {
1137 if (!osp->opd_imp_active || !osp->opd_imp_connected) {
1138 osp_trans_callback(env, oth, rc);
1139 osp_thandle_put(env, oth);
1140 GOTO(out, rc = -ENOTCONN);
1143 rc = obd_get_request_slot(&osp->opd_obd->u.cli);
1145 osp_trans_callback(env, oth, rc);
1146 osp_thandle_put(env, oth);
1147 GOTO(out, rc = -ENOTCONN);
1149 args->oaua_flow_control = true;
1151 if (!osp->opd_connect_mdt) {
1152 down_read(&osp->opd_async_updates_rwsem);
1153 args->oaua_count = &osp->opd_async_updates_count;
1154 args->oaua_waitq = &osp->opd_sync_barrier_waitq;
1155 up_read(&osp->opd_async_updates_rwsem);
1156 atomic_inc(args->oaua_count);
1159 ptlrpcd_add_req(req);
1162 osp_thandle_get(oth); /* hold for commit callback */
1163 req->rq_commit_cb = osp_request_commit_cb;
1164 req->rq_cb_data = &oth->ot_super;
1165 args->oaua_flow_control = false;
1167 /* If the transaction is created during MDT recoverying
1168 * process, it means this is an recovery update, we need
1169 * to let OSP send it anyway without checking recoverying
1170 * status, in case the other target is being recoveried
1171 * at the same time, and if we wait here for the import
1172 * to be recoveryed, it might cause deadlock */
1173 osp_set_req_replay(osp, req);
1175 /* Because this req will be synchronus, i.e. it will be called
1176 * in the same thread, so it will be safe to use current
1178 args->oaua_update_env = env;
1179 if (osp->opd_connect_mdt)
1180 osp_get_rpc_lock(osp);
1181 rc = ptlrpc_queue_wait(req);
1182 if (osp->opd_connect_mdt)
1183 osp_put_rpc_lock(osp);
1184 if ((rc == -ENOMEM && req->rq_set == NULL) ||
1185 (req->rq_transno == 0 && !req->rq_committed)) {
1186 if (args->oaua_update != NULL) {
1187 /* If osp_update_interpret is not being called,
1188 * release the osp_thandle */
1189 args->oaua_update = NULL;
1190 osp_thandle_put(env, oth);
1193 req->rq_cb_data = NULL;
1194 rc = rc == 0 ? req->rq_status : rc;
1195 osp_trans_callback(env, oth, rc);
1196 osp_thandle_put(env, oth);
1202 ptlrpc_req_finished(req);
1208 * Get local thandle for osp_thandle
1210 * Get the local OSD thandle from the OSP thandle. Currently, there
1211 * are a few OSP API (osp_create() and osp_sync_add()) needs
1212 * to update the object on local OSD device.
1214 * If the osp_thandle comes from normal stack (MDD->LOD->OSP), then
1215 * we will get local thandle by thandle_get_sub_by_dt.
1217 * If the osp_thandle is remote thandle (th_top == NULL, only used
1218 * by LFSCK), then it will create a local thandle, and stop it in
1219 * osp_trans_stop(). And this only happens on OSP for OST.
1221 * These are temporary solution, once OSP accessing OSD object is
1222 * being fixed properly, this function should be removed. XXX
1224 * \param[in] env pointer to the thread context
1225 * \param[in] th pointer to the transaction handler
1226 * \param[in] dt pointer to the OSP device
1228 * \retval pointer to the local thandle
1229 * \retval ERR_PTR(errno) if it fails.
1231 struct thandle *osp_get_storage_thandle(const struct lu_env *env,
1233 struct osp_device *osp)
1235 struct osp_thandle *oth;
1236 struct thandle *local_th;
1238 if (th->th_top != NULL)
1239 return thandle_get_sub_by_dt(env, th->th_top,
1242 LASSERT(!osp->opd_connect_mdt);
1243 oth = thandle_to_osp_thandle(th);
1244 if (oth->ot_storage_th != NULL)
1245 return oth->ot_storage_th;
1247 local_th = dt_trans_create(env, osp->opd_storage);
1248 if (IS_ERR(local_th))
1251 oth->ot_storage_th = local_th;
1257 * Set version for the transaction
1259 * Set the version for the transaction and add the request to
1260 * the sending list, then after transaction stop, the request
1261 * will be sent in the order of version by the sending thread.
1263 * \param [in] oth osp thandle to be set version.
1265 * \retval 0 if set version succeeds
1266 * negative errno if set version fails.
1268 int osp_check_and_set_rpc_version(struct osp_thandle *oth,
1269 struct osp_object *obj)
1271 struct osp_device *osp = dt2osp_dev(oth->ot_super.th_dev);
1272 struct osp_updates *ou = osp->opd_update;
1277 if (oth->ot_our->our_version != 0)
1280 spin_lock(&ou->ou_lock);
1281 spin_lock(&oth->ot_our->our_list_lock);
1282 if (obj->opo_stale) {
1283 spin_unlock(&oth->ot_our->our_list_lock);
1284 spin_unlock(&ou->ou_lock);
1288 /* Assign the version and add it to the sending list */
1289 osp_thandle_get(oth);
1290 oth->ot_our->our_version = ou->ou_version++;
1291 oth->ot_our->our_generation = ou->ou_generation;
1292 list_add_tail(&oth->ot_our->our_list,
1293 &osp->opd_update->ou_list);
1294 oth->ot_our->our_req_ready = 0;
1295 spin_unlock(&oth->ot_our->our_list_lock);
1296 spin_unlock(&ou->ou_lock);
1298 LASSERT(oth->ot_super.th_wait_submit == 1);
1299 CDEBUG(D_INFO, "%s: version %llu gen %llu oth:version %p:%llu\n",
1300 osp->opd_obd->obd_name, ou->ou_version, ou->ou_generation, oth,
1301 oth->ot_our->our_version);
1307 * Get next OSP update request in the sending list
1308 * Get next OSP update request in the sending list by version number, next
1310 * 1. transaction which does not have a version number.
1311 * 2. transaction whose version == opd_rpc_version.
1313 * \param [in] ou osp update structure.
1314 * \param [out] ourp the pointer holding the next update request.
1316 * \retval true if getting the next transaction.
1317 * \retval false if not getting the next transaction.
1320 osp_get_next_request(struct osp_updates *ou, struct osp_update_request **ourp)
1322 struct osp_update_request *our;
1323 struct osp_update_request *tmp;
1324 bool got_req = false;
1326 spin_lock(&ou->ou_lock);
1327 list_for_each_entry_safe(our, tmp, &ou->ou_list, our_list) {
1328 LASSERT(our->our_th != NULL);
1329 CDEBUG(D_HA, "ou %p version %llu rpc_version %llu\n",
1330 ou, our->our_version, ou->ou_rpc_version);
1331 spin_lock(&our->our_list_lock);
1332 /* Find next osp_update_request in the list */
1333 if (our->our_version == ou->ou_rpc_version &&
1334 our->our_req_ready) {
1335 list_del_init(&our->our_list);
1336 spin_unlock(&our->our_list_lock);
1341 spin_unlock(&our->our_list_lock);
1343 spin_unlock(&ou->ou_lock);
1349 * Invalidate update request
1351 * Invalidate update request in the OSP sending list, so all of
1352 * requests in the sending list will return error, which happens
1353 * when it finds one update (with writing llog) requests fails or
1354 * the OSP is evicted by remote target. see osp_send_update_thread().
1356 * \param[in] osp OSP device whose update requests will be
1359 void osp_invalidate_request(struct osp_device *osp)
1362 struct osp_updates *ou = osp->opd_update;
1363 struct osp_update_request *our;
1364 struct osp_update_request *tmp;
1372 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1374 CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
1377 spin_lock(&ou->ou_lock);
1378 ou->ou_generation++;
1379 spin_unlock(&ou->ou_lock);
1384 INIT_LIST_HEAD(&list);
1386 spin_lock(&ou->ou_lock);
1387 /* invalidate all of request in the sending list */
1388 list_for_each_entry_safe(our, tmp, &ou->ou_list, our_list) {
1389 spin_lock(&our->our_list_lock);
1390 if (our->our_req_ready)
1391 list_move(&our->our_list, &list);
1393 list_del_init(&our->our_list);
1395 if (our->our_th->ot_super.th_result == 0)
1396 our->our_th->ot_super.th_result = -EIO;
1398 if (our->our_version >= ou->ou_rpc_version)
1399 ou->ou_rpc_version = our->our_version + 1;
1400 spin_unlock(&our->our_list_lock);
1402 CDEBUG(D_HA, "%s invalidate our %p\n", osp->opd_obd->obd_name,
1406 /* Increase the generation, then the update request with old generation
1407 * will fail with -EIO. */
1408 ou->ou_generation++;
1409 spin_unlock(&ou->ou_lock);
1411 /* invalidate all of request in the sending list */
1412 list_for_each_entry_safe(our, tmp, &list, our_list) {
1413 spin_lock(&our->our_list_lock);
1414 list_del_init(&our->our_list);
1415 spin_unlock(&our->our_list_lock);
1416 osp_trans_callback(&env, our->our_th,
1417 our->our_th->ot_super.th_result);
1418 osp_thandle_put(&env, our->our_th);
1424 * Sending update thread
1426 * Create thread to send update request to other MDTs, this thread will pull
1427 * out update request from the list in OSP by version number, i.e. it will
1428 * make sure the update request with lower version number will be sent first.
1430 * \param[in] arg hold the OSP device.
1432 * \retval 0 if the thread is created successfully.
1433 * \retal negative error if the thread is not created
1436 int osp_send_update_thread(void *arg)
1439 struct osp_device *osp = arg;
1440 struct l_wait_info lwi = { 0 };
1441 struct osp_updates *ou = osp->opd_update;
1442 struct ptlrpc_thread *thread = &osp->opd_update_thread;
1443 struct osp_update_request *our = NULL;
1447 LASSERT(ou != NULL);
1448 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1450 CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
1455 thread->t_flags = SVC_RUNNING;
1456 wake_up(&thread->t_ctl_waitq);
1459 l_wait_event(ou->ou_waitq,
1460 !osp_send_update_thread_running(osp) ||
1461 osp_get_next_request(ou, &our), &lwi);
1463 if (!osp_send_update_thread_running(osp)) {
1465 osp_trans_callback(&env, our->our_th, -EINTR);
1466 osp_thandle_put(&env, our->our_th);
1471 LASSERT(our->our_th != NULL);
1472 if (our->our_th->ot_super.th_result != 0) {
1473 osp_trans_callback(&env, our->our_th,
1474 our->our_th->ot_super.th_result);
1475 rc = our->our_th->ot_super.th_result;
1476 } else if (ou->ou_generation != our->our_generation ||
1477 OBD_FAIL_CHECK(OBD_FAIL_INVALIDATE_UPDATE)) {
1479 osp_trans_callback(&env, our->our_th, rc);
1481 rc = osp_send_update_req(&env, osp, our);
1484 /* Update the rpc version */
1485 spin_lock(&ou->ou_lock);
1486 if (our->our_version == ou->ou_rpc_version)
1487 ou->ou_rpc_version++;
1488 spin_unlock(&ou->ou_lock);
1490 /* If one update request fails, let's fail all of the requests
1491 * in the sending list, because the request in the sending
1492 * list are dependent on either other, continue sending these
1493 * request might cause llog or filesystem corruption */
1495 osp_invalidate_request(osp);
1497 /* Balanced for thandle_get in osp_check_and_set_rpc_version */
1498 osp_thandle_put(&env, our->our_th);
1501 thread->t_flags = SVC_STOPPED;
1503 wake_up(&thread->t_ctl_waitq);
1509 * The OSP layer dt_device_operations::dt_trans_start() interface
1510 * to start the transaction.
1512 * If the transaction is a remote transaction, then related remote
1513 * updates will be triggered in the osp_trans_stop().
1514 * Please refer to osp_trans_create() for transaction type.
1516 * \param[in] env pointer to the thread context
1517 * \param[in] dt pointer to the OSP dt_device
1518 * \param[in] th pointer to the transaction handler
1520 * \retval 0 for success
1521 * \retval negative error number on failure
1523 int osp_trans_start(const struct lu_env *env, struct dt_device *dt,
1526 struct osp_thandle *oth = thandle_to_osp_thandle(th);
1528 if (oth->ot_super.th_sync)
1529 oth->ot_our->our_flags |= UPDATE_FL_SYNC;
1530 /* For remote thandle, if there are local thandle, start it here*/
1531 if (is_only_remote_trans(th) && oth->ot_storage_th != NULL)
1532 return dt_trans_start(env, oth->ot_storage_th->th_dev,
1533 oth->ot_storage_th);
1538 * The OSP layer dt_device_operations::dt_trans_stop() interface
1539 * to stop the transaction.
1541 * If the transaction is a remote transaction, related remote
1542 * updates will be triggered at the end of this function.
1544 * For synchronous mode update or any failed update, the request
1545 * will be destroyed explicitly when the osp_trans_stop().
1547 * Please refer to osp_trans_create() for transaction type.
1549 * \param[in] env pointer to the thread context
1550 * \param[in] dt pointer to the OSP dt_device
1551 * \param[in] th pointer to the transaction handler
1553 * \retval 0 for success
1554 * \retval negative error number on failure
1556 int osp_trans_stop(const struct lu_env *env, struct dt_device *dt,
1559 struct osp_thandle *oth = thandle_to_osp_thandle(th);
1560 struct osp_update_request *our = oth->ot_our;
1561 struct osp_device *osp = dt2osp_dev(dt);
1565 /* For remote transaction, if there is local storage thandle,
1567 if (oth->ot_storage_th != NULL && th->th_top == NULL) {
1568 dt_trans_stop(env, oth->ot_storage_th->th_dev,
1569 oth->ot_storage_th);
1570 oth->ot_storage_th = NULL;
1573 if (our == NULL || list_empty(&our->our_req_list)) {
1574 osp_trans_callback(env, oth, th->th_result);
1575 GOTO(out, rc = th->th_result);
1578 if (!osp->opd_connect_mdt) {
1579 osp_trans_callback(env, oth, th->th_result);
1580 rc = osp_send_update_req(env, osp, oth->ot_our);
1584 if (osp->opd_update == NULL ||
1585 !osp_send_update_thread_running(osp)) {
1586 osp_trans_callback(env, oth, -EIO);
1587 GOTO(out, rc = -EIO);
1590 CDEBUG(D_HA, "%s: add oth %p with version %llu\n",
1591 osp->opd_obd->obd_name, oth, our->our_version);
1593 LASSERT(our->our_req_ready == 0);
1594 spin_lock(&our->our_list_lock);
1595 if (likely(!list_empty(&our->our_list))) {
1596 /* notify sending thread */
1597 our->our_req_ready = 1;
1598 wake_up(&osp->opd_update->ou_waitq);
1599 spin_unlock(&our->our_list_lock);
1600 } else if (th->th_result == 0) {
1601 /* if the request does not needs to be serialized,
1602 * read-only request etc, let's send it right away */
1603 spin_unlock(&our->our_list_lock);
1604 rc = osp_send_update_req(env, osp, our);
1606 spin_unlock(&our->our_list_lock);
1607 osp_trans_callback(env, oth, th->th_result);
1610 osp_thandle_put(env, oth);