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 static 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_first_entry(&our->our_req_list,
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_first_entry(&our->our_req_list,
390 struct osp_update_request_sub,
393 /* Let's check if it can be packed inline */
394 if (object_update_request_size(ours->ours_req) +
395 sizeof(struct out_update_header) <
396 OUT_UPDATE_MAX_INLINE_SIZE) {
397 rc = osp_prep_inline_update_req(env, req, our, repsize);
404 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_HEADER, RCL_CLIENT,
405 sizeof(struct out_update_header));
407 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_BUF, RCL_CLIENT,
408 buf_count * sizeof(*oub));
410 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, OUT_UPDATE);
414 ouh = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_HEADER);
415 ouh->ouh_magic = OUT_UPDATE_HEADER_MAGIC;
416 ouh->ouh_count = buf_count;
417 ouh->ouh_inline_length = 0;
418 ouh->ouh_reply_size = repsize;
419 oub = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE_BUF);
420 list_for_each_entry(ours, &our->our_req_list, ours_list) {
421 oub->oub_size = ours->ours_req_size;
423 /* First *and* last might be partial pages, hence +1 */
424 page_count += DIV_ROUND_UP(ours->ours_req_size, PAGE_SIZE) + 1;
427 req->rq_bulk_write = 1;
428 desc = ptlrpc_prep_bulk_imp(req, page_count,
429 MD_MAX_BRW_SIZE >> LNET_MTU_BITS,
430 PTLRPC_BULK_GET_SOURCE,
431 MDS_BULK_PORTAL, &ptlrpc_bulk_kiov_nopin_ops);
433 GOTO(out_req, rc = -ENOMEM);
435 /* NB req now owns desc and will free it when it gets freed */
436 list_for_each_entry(ours, &our->our_req_list, ours_list) {
437 desc->bd_frag_ops->add_iov_frag(desc, ours->ours_req,
438 ours->ours_req_size);
439 total += ours->ours_req_size;
441 CDEBUG(D_OTHER, "total %d in %u\n", total, our->our_update_nr);
443 req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_REPLY,
444 RCL_SERVER, repsize);
446 ptlrpc_request_set_replen(req);
447 req->rq_request_portal = OUT_PORTAL;
448 req->rq_reply_portal = OSC_REPLY_PORTAL;
453 ptlrpc_req_finished(req);
461 * Send update request to the remote MDT synchronously.
463 * \param[in] env execution environment
464 * \param[in] imp import on which ptlrpc request will be sent
465 * \param[in] our hold all of updates which will be packed into the req
466 * \param[in] reqp request to be created
468 * \retval 0 if RPC succeeds.
469 * \retval negative errno if RPC fails.
471 int osp_remote_sync(const struct lu_env *env, struct osp_device *osp,
472 struct osp_update_request *our,
473 struct ptlrpc_request **reqp)
475 struct obd_import *imp = osp->opd_obd->u.cli.cl_import;
476 struct ptlrpc_request *req = NULL;
480 rc = osp_prep_update_req(env, imp, our, &req);
484 osp_set_req_replay(osp, req);
485 req->rq_allow_intr = 1;
487 /* Note: some dt index api might return non-zero result here, like
488 * osd_index_ea_lookup, so we should only check rc < 0 here */
489 rc = ptlrpc_queue_wait(req);
491 if (rc < 0 || reqp == NULL)
492 ptlrpc_req_finished(req);
500 * Invalidate all objects in the osp thandle
502 * invalidate all of objects in the update request, which will be called
503 * when the transaction is aborted.
505 * \param[in] oth osp thandle.
507 static void osp_thandle_invalidate_object(const struct lu_env *env,
508 struct osp_thandle *oth,
511 struct osp_update_request *our = oth->ot_our;
512 struct osp_object *obj;
513 struct osp_object *next;
518 list_for_each_entry_safe(obj, next, &our->our_invalidate_cb_list,
519 opo_invalidate_cb_list) {
521 osp_invalidate(env, &obj->opo_obj);
523 spin_lock(&obj->opo_lock);
524 list_del_init(&obj->opo_invalidate_cb_list);
525 spin_unlock(&obj->opo_lock);
527 dt_object_put(env, &obj->opo_obj);
531 static void osp_trans_stop_cb(const struct lu_env *env,
532 struct osp_thandle *oth, int result)
534 struct dt_txn_commit_cb *dcb;
535 struct dt_txn_commit_cb *tmp;
537 /* call per-transaction stop callbacks if any */
538 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
540 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
541 "commit callback entry: magic=%x name='%s'\n",
542 dcb->dcb_magic, dcb->dcb_name);
543 list_del_init(&dcb->dcb_linkage);
544 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
547 osp_thandle_invalidate_object(env, oth, result);
551 * Allocate an osp request and initialize it with the given parameters.
553 * \param[in] obj pointer to the operation target
554 * \param[in] data pointer to the data used by the interpreter
555 * \param[in] interpreter pointer to the interpreter function
557 * \retval pointer to the asychronous request
558 * \retval NULL if the allocation failed
560 static struct osp_update_callback *
561 osp_update_callback_init(struct osp_object *obj, void *data,
562 osp_update_interpreter_t interpreter)
564 struct osp_update_callback *ouc;
570 lu_object_get(osp2lu_obj(obj));
571 INIT_LIST_HEAD(&ouc->ouc_list);
573 ouc->ouc_data = data;
574 ouc->ouc_interpreter = interpreter;
580 * Destroy the osp_update_callback.
582 * \param[in] env pointer to the thread context
583 * \param[in] ouc pointer to osp_update_callback
585 static void osp_update_callback_fini(const struct lu_env *env,
586 struct osp_update_callback *ouc)
588 LASSERT(list_empty(&ouc->ouc_list));
590 lu_object_put(env, osp2lu_obj(ouc->ouc_obj));
595 * Interpret the packaged OUT RPC results.
597 * For every packaged sub-request, call its registered interpreter function.
598 * Then destroy the sub-request.
600 * \param[in] env pointer to the thread context
601 * \param[in] req pointer to the RPC
602 * \param[in] arg pointer to data used by the interpreter
603 * \param[in] rc the RPC return value
605 * \retval 0 for success
606 * \retval negative error number on failure
608 static int osp_update_interpret(const struct lu_env *env,
609 struct ptlrpc_request *req, void *args, int rc)
611 struct object_update_reply *reply = NULL;
612 struct osp_update_args *oaua = args;
613 struct osp_update_request *our = oaua->oaua_update;
614 struct osp_thandle *oth;
615 struct osp_update_callback *ouc;
616 struct osp_update_callback *next;
626 /* Sigh env might be NULL in some cases, see
628 * osp_send_update_thread()
629 * ptlrpc_set_wait() ----> null env.
631 * osp_update_interpret()
632 * Let's use env in oaua for this case.
635 env = oaua->oaua_update_env;
637 oaua->oaua_update = NULL;
639 if (oaua->oaua_flow_control) {
640 struct osp_device *osp;
642 LASSERT(oth != NULL);
643 osp = dt2osp_dev(oth->ot_super.th_dev);
644 obd_put_request_slot(&osp->opd_obd->u.cli);
647 /* Unpack the results from the reply message. */
648 if (req->rq_repmsg != NULL && req->rq_replied) {
649 reply = req_capsule_server_sized_get(&req->rq_pill,
650 &RMF_OUT_UPDATE_REPLY,
651 OUT_UPDATE_REPLY_SIZE);
652 if (reply == NULL || reply->ourp_magic != UPDATE_REPLY_MAGIC) {
656 count = reply->ourp_count;
660 list_for_each_entry_safe(ouc, next, &our->our_cb_items, ouc_list) {
661 list_del_init(&ouc->ouc_list);
663 /* The peer may only have handled some requests (indicated
664 * by the 'count') in the packaged OUT RPC, we can only get
665 * results for the handled part. */
666 if (index < count && reply->ourp_lens[index] > 0 && rc >= 0) {
667 struct object_update_result *result;
669 result = object_update_result_get(reply, index, NULL);
673 rc1 = rc = result->our_rc;
674 } else if (rc1 >= 0) {
675 /* The peer did not handle these request, let's return
676 * -EINVAL to update interpret for now */
683 if (ouc->ouc_interpreter != NULL)
684 ouc->ouc_interpreter(env, reply, req, ouc->ouc_obj,
685 ouc->ouc_data, index, rc1);
687 osp_update_callback_fini(env, ouc);
691 if (oaua->oaua_count != NULL && atomic_dec_and_test(oaua->oaua_count))
692 wake_up(oaua->oaua_waitq);
695 /* oth and osp_update_requests will be destoryed in
697 osp_trans_stop_cb(env, oth, rc);
698 osp_thandle_put(env, oth);
700 osp_update_request_destroy(env, our);
707 * Pack all the requests in the shared asynchronous idempotent request queue
708 * into a single OUT RPC that will be given to the background ptlrpcd daemon.
710 * \param[in] env pointer to the thread context
711 * \param[in] osp pointer to the OSP device
712 * \param[in] our pointer to the shared queue
714 * \retval 0 for success
715 * \retval negative error number on failure
717 int osp_unplug_async_request(const struct lu_env *env,
718 struct osp_device *osp,
719 struct osp_update_request *our)
721 struct osp_update_args *args;
722 struct ptlrpc_request *req = NULL;
725 rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
728 struct osp_update_callback *ouc;
729 struct osp_update_callback *next;
731 list_for_each_entry_safe(ouc, next,
732 &our->our_cb_items, ouc_list) {
733 list_del_init(&ouc->ouc_list);
734 if (ouc->ouc_interpreter != NULL)
735 ouc->ouc_interpreter(env, NULL, NULL,
737 ouc->ouc_data, 0, rc);
738 osp_update_callback_fini(env, ouc);
740 osp_update_request_destroy(env, our);
742 args = ptlrpc_req_async_args(args, req);
743 args->oaua_update = our;
744 args->oaua_count = NULL;
745 args->oaua_waitq = NULL;
746 /* Note: this is asynchronous call for the request, so the
747 * interrupte cb and current function will be different
748 * thread, so we need use different env */
749 args->oaua_update_env = NULL;
750 args->oaua_flow_control = false;
751 req->rq_interpret_reply = osp_update_interpret;
752 ptlrpcd_add_req(req);
759 * Find or create (if NOT exist or purged) the shared asynchronous idempotent
760 * request queue - osp_device::opd_async_requests.
762 * If the osp_device::opd_async_requests is not NULL, then return it directly;
763 * otherwise create new osp_update_request and attach it to opd_async_requests.
765 * \param[in] osp pointer to the OSP device
767 * \retval pointer to the shared queue
768 * \retval negative error number on failure
770 static struct osp_update_request *
771 osp_find_or_create_async_update_request(struct osp_device *osp)
773 struct osp_update_request *our = osp->opd_async_requests;
778 our = osp_update_request_create(&osp->opd_dt_dev);
782 osp->opd_async_requests = our;
788 * Insert an osp_update_callback into the osp_update_request.
790 * Insert an osp_update_callback to the osp_update_request. Usually each update
791 * in the osp_update_request will have one correspondent callback, and these
792 * callbacks will be called in rq_interpret_reply.
794 * \param[in] env pointer to the thread context
795 * \param[in] obj pointer to the operation target object
796 * \param[in] data pointer to the data used by the interpreter
797 * \param[in] interpreter pointer to the interpreter function
799 * \retval 0 for success
800 * \retval negative error number on failure
802 int osp_insert_update_callback(const struct lu_env *env,
803 struct osp_update_request *our,
804 struct osp_object *obj, void *data,
805 osp_update_interpreter_t interpreter)
807 struct osp_update_callback *ouc;
809 ouc = osp_update_callback_init(obj, data, interpreter);
813 list_add_tail(&ouc->ouc_list, &our->our_cb_items);
819 * Insert an asynchronous idempotent request to the shared request queue that
820 * is attached to the osp_device.
822 * This function generates a new osp_async_request with the given parameters,
823 * then tries to insert the request into the osp_device-based shared request
824 * queue. If the queue is full, then triggers the packaged OUT RPC to purge
825 * the shared queue firstly, and then re-tries.
827 * NOTE: must hold the osp::opd_async_requests_mutex to serialize concurrent
828 * osp_insert_async_request call from others.
830 * \param[in] env pointer to the thread context
831 * \param[in] op operation type, see 'enum update_type'
832 * \param[in] obj pointer to the operation target
833 * \param[in] count array size of the subsequent \a lens and \a bufs
834 * \param[in] lens buffer length array for the subsequent \a bufs
835 * \param[in] bufs the buffers to compose the request
836 * \param[in] data pointer to the data used by the interpreter
837 * \param[in] repsize how many bytes the caller allocated for \a data
838 * \param[in] interpreter pointer to the interpreter function
840 * \retval 0 for success
841 * \retval negative error number on failure
843 int osp_insert_async_request(const struct lu_env *env, enum update_type op,
844 struct osp_object *obj, int count,
845 __u16 *lens, const void **bufs,
846 void *data, __u32 repsize,
847 osp_update_interpreter_t interpreter)
849 struct osp_device *osp;
850 struct osp_update_request *our;
851 struct object_update *object_update;
852 size_t max_update_size;
853 struct object_update_request *ureq;
854 struct osp_update_request_sub *ours;
858 osp = lu2osp_dev(osp2lu_obj(obj)->lo_dev);
859 our = osp_find_or_create_async_update_request(osp);
861 RETURN(PTR_ERR(our));
864 ours = osp_current_object_update_request(our);
866 ureq = ours->ours_req;
867 max_update_size = ours->ours_req_size -
868 object_update_request_size(ureq);
870 object_update = update_buffer_get_update(ureq, ureq->ourq_count);
871 rc = out_update_pack(env, object_update, &max_update_size, op,
872 lu_object_fid(osp2lu_obj(obj)), count, lens, bufs,
874 /* The queue is full. */
876 osp->opd_async_requests = NULL;
877 mutex_unlock(&osp->opd_async_requests_mutex);
879 rc = osp_unplug_async_request(env, osp, our);
880 mutex_lock(&osp->opd_async_requests_mutex);
884 our = osp_find_or_create_async_update_request(osp);
886 RETURN(PTR_ERR(our));
894 our->our_update_nr++;
897 rc = osp_insert_update_callback(env, our, obj, data, interpreter);
902 int osp_trans_update_request_create(struct thandle *th)
904 struct osp_thandle *oth = thandle_to_osp_thandle(th);
905 struct osp_update_request *our;
907 if (oth->ot_our != NULL)
910 our = osp_update_request_create(th->th_dev);
912 th->th_result = PTR_ERR(our);
922 void osp_thandle_destroy(const struct lu_env *env,
923 struct osp_thandle *oth)
925 LASSERT(oth->ot_magic == OSP_THANDLE_MAGIC);
926 LASSERT(list_empty(&oth->ot_commit_dcb_list));
927 LASSERT(list_empty(&oth->ot_stop_dcb_list));
928 if (oth->ot_our != NULL)
929 osp_update_request_destroy(env, oth->ot_our);
934 * The OSP layer dt_device_operations::dt_trans_create() interface
935 * to create a transaction.
937 * There are two kinds of transactions that will involve OSP:
939 * 1) If the transaction only contains the updates on remote server
940 * (MDT or OST), such as re-generating the lost OST-object for
941 * LFSCK, then it is a remote transaction. For remote transaction,
942 * the upper layer caller (such as the LFSCK engine) will call the
943 * dt_trans_create() (with the OSP dt_device as the parameter),
944 * then the call will be directed to the osp_trans_create() that
945 * creates the transaction handler and returns it to the caller.
947 * 2) If the transcation contains both local and remote updates,
948 * such as cross MDTs create under DNE mode, then the upper layer
949 * caller will not trigger osp_trans_create(). Instead, it will
950 * call dt_trans_create() on other dt_device, such as LOD that
951 * will generate the transaction handler. Such handler will be
952 * used by the whole transaction in subsequent sub-operations.
954 * \param[in] env pointer to the thread context
955 * \param[in] d pointer to the OSP dt_device
957 * \retval pointer to the transaction handler
958 * \retval negative error number on failure
960 struct thandle *osp_trans_create(const struct lu_env *env, struct dt_device *d)
962 struct osp_thandle *oth;
963 struct thandle *th = NULL;
967 if (unlikely(oth == NULL))
968 RETURN(ERR_PTR(-ENOMEM));
970 oth->ot_magic = OSP_THANDLE_MAGIC;
974 atomic_set(&oth->ot_refcount, 1);
975 INIT_LIST_HEAD(&oth->ot_commit_dcb_list);
976 INIT_LIST_HEAD(&oth->ot_stop_dcb_list);
982 * Add commit callback to transaction.
984 * Add commit callback to the osp thandle, which will be called
985 * when the thandle is committed remotely.
987 * \param[in] th the thandle
988 * \param[in] dcb commit callback structure
990 * \retval only return 0 for now.
992 int osp_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
994 struct osp_thandle *oth = thandle_to_osp_thandle(th);
996 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
997 LASSERT(&dcb->dcb_func != NULL);
998 if (dcb->dcb_flags & DCB_TRANS_STOP)
999 list_add(&dcb->dcb_linkage, &oth->ot_stop_dcb_list);
1001 list_add(&dcb->dcb_linkage, &oth->ot_commit_dcb_list);
1005 static void osp_trans_commit_cb(struct osp_thandle *oth, int result)
1007 struct dt_txn_commit_cb *dcb;
1008 struct dt_txn_commit_cb *tmp;
1010 LASSERT(atomic_read(&oth->ot_refcount) > 0);
1011 /* call per-transaction callbacks if any */
1012 list_for_each_entry_safe(dcb, tmp, &oth->ot_commit_dcb_list,
1014 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1015 "commit callback entry: magic=%x name='%s'\n",
1016 dcb->dcb_magic, dcb->dcb_name);
1017 list_del_init(&dcb->dcb_linkage);
1018 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1022 static void osp_request_commit_cb(struct ptlrpc_request *req)
1024 struct thandle *th = req->rq_cb_data;
1025 struct osp_thandle *oth;
1026 __u64 last_committed_transno = 0;
1027 int result = req->rq_status;
1033 oth = thandle_to_osp_thandle(th);
1034 if (req->rq_repmsg != NULL &&
1035 lustre_msg_get_last_committed(req->rq_repmsg))
1036 last_committed_transno =
1037 lustre_msg_get_last_committed(req->rq_repmsg);
1039 if (last_committed_transno <
1040 req->rq_import->imp_peer_committed_transno)
1041 last_committed_transno =
1042 req->rq_import->imp_peer_committed_transno;
1044 CDEBUG(D_HA, "trans no %llu committed transno %llu\n",
1045 req->rq_transno, last_committed_transno);
1047 /* If the transaction is not really committed, mark result = 1 */
1048 if (req->rq_transno != 0 &&
1049 (req->rq_transno > last_committed_transno) && result == 0)
1052 osp_trans_commit_cb(oth, result);
1053 req->rq_committed = 1;
1054 osp_thandle_put(NULL, oth);
1059 * callback of osp transaction
1061 * Call all of callbacks for this osp thandle. This will only be
1062 * called in error handler path. In the normal processing path,
1063 * these callback will be called in osp_request_commit_cb() and
1064 * osp_update_interpret().
1066 * \param [in] env execution environment
1067 * \param [in] oth osp thandle
1068 * \param [in] rc result of the osp thandle
1070 void osp_trans_callback(const struct lu_env *env,
1071 struct osp_thandle *oth, int rc)
1073 struct osp_update_callback *ouc;
1074 struct osp_update_callback *next;
1076 if (oth->ot_our != NULL) {
1077 list_for_each_entry_safe(ouc, next,
1078 &oth->ot_our->our_cb_items, ouc_list) {
1079 list_del_init(&ouc->ouc_list);
1080 if (ouc->ouc_interpreter != NULL)
1081 ouc->ouc_interpreter(env, NULL, NULL,
1083 ouc->ouc_data, 0, rc);
1084 osp_update_callback_fini(env, ouc);
1087 osp_trans_stop_cb(env, oth, rc);
1088 osp_trans_commit_cb(oth, rc);
1092 * Send the request for remote updates.
1094 * Send updates to the remote MDT. Prepare the request by osp_update_req
1095 * and send them to remote MDT, for sync request, it will wait
1096 * until the reply return, otherwise hand it to ptlrpcd.
1098 * Please refer to osp_trans_create() for transaction type.
1100 * \param[in] env pointer to the thread context
1101 * \param[in] osp pointer to the OSP device
1102 * \param[in] our pointer to the osp_update_request
1104 * \retval 0 for success
1105 * \retval negative error number on failure
1107 static int osp_send_update_req(const struct lu_env *env,
1108 struct osp_device *osp,
1109 struct osp_update_request *our)
1111 struct osp_update_args *args;
1112 struct ptlrpc_request *req;
1113 struct osp_thandle *oth = our->our_th;
1114 struct osp_updates *ou = osp->opd_update;
1118 LASSERT(oth != NULL);
1119 LASSERT(osp->opd_obd);
1121 if (ou && ou->ou_generation != our->our_generation) {
1122 const struct lnet_processid *peer;
1125 osp_trans_callback(env, oth, rc);
1126 peer = &osp->opd_obd->u.cli.cl_import->imp_connection->c_peer;
1127 CDEBUG(D_HA, "%s: stale tx to %s: gen %llu != %llu: rc = %d\n",
1128 osp->opd_obd->obd_name, libcfs_nidstr(&peer->nid),
1129 osp->opd_update->ou_generation, our->our_generation, rc);
1133 rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
1136 osp_trans_callback(env, oth, rc);
1140 args = ptlrpc_req_async_args(args, req);
1141 args->oaua_update = our;
1142 /* set env to NULL, in case the interrupt cb and current function
1143 * are in different thread */
1144 args->oaua_update_env = NULL;
1145 osp_thandle_get(oth); /* hold for update interpret */
1146 req->rq_interpret_reply = osp_update_interpret;
1147 if (!oth->ot_super.th_wait_submit && !oth->ot_super.th_sync) {
1148 if (!osp->opd_imp_active || !osp->opd_imp_connected) {
1149 osp_trans_callback(env, oth, rc);
1150 osp_thandle_put(env, oth);
1151 GOTO(out, rc = -ENOTCONN);
1154 rc = obd_get_request_slot(&osp->opd_obd->u.cli);
1156 osp_trans_callback(env, oth, rc);
1157 osp_thandle_put(env, oth);
1158 GOTO(out, rc = -ENOTCONN);
1160 args->oaua_flow_control = true;
1162 if (!osp->opd_connect_mdt) {
1163 down_read(&osp->opd_async_updates_rwsem);
1164 args->oaua_count = &osp->opd_async_updates_count;
1165 args->oaua_waitq = &osp->opd_sync_barrier_waitq;
1166 up_read(&osp->opd_async_updates_rwsem);
1167 atomic_inc(args->oaua_count);
1170 ptlrpcd_add_req(req);
1173 osp_thandle_get(oth); /* hold for commit callback */
1174 req->rq_commit_cb = osp_request_commit_cb;
1175 req->rq_cb_data = &oth->ot_super;
1176 args->oaua_flow_control = false;
1178 /* If the transaction is created during MDT recoverying
1179 * process, it means this is an recovery update, we need
1180 * to let OSP send it anyway without checking recoverying
1181 * status, in case the other target is being recoveried
1182 * at the same time, and if we wait here for the import
1183 * to be recoveryed, it might cause deadlock */
1184 osp_set_req_replay(osp, req);
1186 /* Because this req will be synchronus, i.e. it will be called
1187 * in the same thread, so it will be safe to use current
1189 args->oaua_update_env = env;
1190 if (osp->opd_connect_mdt)
1191 ptlrpc_get_mod_rpc_slot(req);
1192 rc = ptlrpc_queue_wait(req);
1193 if (osp->opd_connect_mdt)
1194 ptlrpc_put_mod_rpc_slot(req);
1196 /* We use rq_queued_time to distinguish between local
1197 * and remote -ENOMEM. */
1198 if ((rc == -ENOMEM && req->rq_queued_time == 0) ||
1199 (req->rq_transno == 0 && !req->rq_committed)) {
1200 if (args->oaua_update != NULL) {
1201 /* If osp_update_interpret is not being called,
1202 * release the osp_thandle */
1203 args->oaua_update = NULL;
1204 osp_thandle_put(env, oth);
1207 req->rq_cb_data = NULL;
1208 rc = rc == 0 ? req->rq_status : rc;
1209 osp_trans_callback(env, oth, rc);
1210 osp_thandle_put(env, oth);
1216 ptlrpc_req_finished(req);
1222 * Get local thandle for osp_thandle
1224 * Get the local OSD thandle from the OSP thandle. Currently, there
1225 * are a few OSP API (osp_create() and osp_sync_add()) needs
1226 * to update the object on local OSD device.
1228 * If the osp_thandle comes from normal stack (MDD->LOD->OSP), then
1229 * we will get local thandle by thandle_get_sub_by_dt.
1231 * If the osp_thandle is remote thandle (th_top == NULL, only used
1232 * by LFSCK), then it will create a local thandle, and stop it in
1233 * osp_trans_stop(). And this only happens on OSP for OST.
1235 * These are temporary solution, once OSP accessing OSD object is
1236 * being fixed properly, this function should be removed. XXX
1238 * \param[in] env pointer to the thread context
1239 * \param[in] th pointer to the transaction handler
1240 * \param[in] dt pointer to the OSP device
1242 * \retval pointer to the local thandle
1243 * \retval ERR_PTR(errno) if it fails.
1245 struct thandle *osp_get_storage_thandle(const struct lu_env *env,
1247 struct osp_device *osp)
1249 struct osp_thandle *oth;
1250 struct thandle *local_th;
1252 if (th->th_top != NULL)
1253 return thandle_get_sub_by_dt(env, th->th_top,
1256 LASSERT(!osp->opd_connect_mdt);
1257 oth = thandle_to_osp_thandle(th);
1258 if (oth->ot_storage_th != NULL)
1259 return oth->ot_storage_th;
1261 local_th = dt_trans_create(env, osp->opd_storage);
1262 if (IS_ERR(local_th))
1265 oth->ot_storage_th = local_th;
1271 * Set version for the transaction
1273 * Set the version for the transaction and add the request to
1274 * the sending list, then after transaction stop, the request
1275 * will be sent in the order of version by the sending thread.
1277 * \param [in] oth osp thandle to be set version.
1279 * \retval 0 if set version succeeds
1280 * negative errno if set version fails.
1282 int osp_check_and_set_rpc_version(struct osp_thandle *oth,
1283 struct osp_object *obj)
1285 struct osp_device *osp = dt2osp_dev(oth->ot_super.th_dev);
1286 struct osp_updates *ou = osp->opd_update;
1291 if (oth->ot_our->our_version != 0)
1294 spin_lock(&ou->ou_lock);
1295 spin_lock(&oth->ot_our->our_list_lock);
1296 if (obj->opo_stale) {
1297 spin_unlock(&oth->ot_our->our_list_lock);
1298 spin_unlock(&ou->ou_lock);
1302 /* Assign the version and add it to the sending list */
1303 osp_thandle_get(oth);
1304 oth->ot_our->our_version = ou->ou_version++;
1305 oth->ot_our->our_generation = ou->ou_generation;
1306 list_add_tail(&oth->ot_our->our_list,
1307 &osp->opd_update->ou_list);
1308 oth->ot_our->our_req_ready = 0;
1309 spin_unlock(&oth->ot_our->our_list_lock);
1310 spin_unlock(&ou->ou_lock);
1312 LASSERT(oth->ot_super.th_wait_submit == 1);
1313 CDEBUG(D_INFO, "%s: version %llu gen %llu oth:version %p:%llu\n",
1314 osp->opd_obd->obd_name, ou->ou_version, ou->ou_generation, oth,
1315 oth->ot_our->our_version);
1321 * Get next OSP update request in the sending list
1322 * Get next OSP update request in the sending list by version number, next
1324 * 1. transaction which does not have a version number.
1325 * 2. transaction whose version == opd_rpc_version.
1327 * \param [in] ou osp update structure.
1328 * \param [out] ourp the pointer holding the next update request.
1330 * \retval true if getting the next transaction.
1331 * \retval false if not getting the next transaction.
1334 osp_get_next_request(struct osp_updates *ou, struct osp_update_request **ourp)
1336 struct osp_update_request *our;
1337 struct osp_update_request *tmp;
1338 bool got_req = false;
1340 spin_lock(&ou->ou_lock);
1341 list_for_each_entry_safe(our, tmp, &ou->ou_list, our_list) {
1342 LASSERT(our->our_th != NULL);
1343 CDEBUG(D_HA, "ou %p version %llu rpc_version %llu\n",
1344 ou, our->our_version, ou->ou_rpc_version);
1345 spin_lock(&our->our_list_lock);
1346 /* Find next osp_update_request in the list */
1347 if (our->our_version == ou->ou_rpc_version &&
1348 our->our_req_ready) {
1349 list_del_init(&our->our_list);
1350 spin_unlock(&our->our_list_lock);
1355 spin_unlock(&our->our_list_lock);
1357 spin_unlock(&ou->ou_lock);
1363 * Invalidate update request
1365 * Invalidate update request in the OSP sending list, so all of
1366 * requests in the sending list will return error, which happens
1367 * when it finds one update (with writing llog) requests fails or
1368 * the OSP is evicted by remote target. see osp_send_update_thread().
1370 * \param[in] osp OSP device whose update requests will be
1373 void osp_invalidate_request(struct osp_device *osp)
1376 struct osp_updates *ou = osp->opd_update;
1377 struct osp_update_request *our;
1378 struct osp_update_request *tmp;
1386 rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1388 CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
1391 spin_lock(&ou->ou_lock);
1392 ou->ou_generation++;
1393 spin_unlock(&ou->ou_lock);
1398 spin_lock(&ou->ou_lock);
1399 /* invalidate all of request in the sending list */
1400 list_for_each_entry_safe(our, tmp, &ou->ou_list, our_list) {
1401 spin_lock(&our->our_list_lock);
1402 if (our->our_req_ready) {
1403 list_move(&our->our_list, &list);
1405 /* this thandle won't be forwarded to
1406 * the dedicated thread, so drop the
1408 osp_thandle_put(&env, our->our_th);
1409 list_del_init(&our->our_list);
1412 if (our->our_th->ot_super.th_result == 0)
1413 our->our_th->ot_super.th_result = -EIO;
1415 if (our->our_version >= ou->ou_rpc_version)
1416 ou->ou_rpc_version = our->our_version + 1;
1417 spin_unlock(&our->our_list_lock);
1419 CDEBUG(D_HA, "%s invalidate our %p\n", osp->opd_obd->obd_name,
1423 /* Increase the generation, then the update request with old generation
1424 * will fail with -EIO. */
1425 ou->ou_generation++;
1426 spin_unlock(&ou->ou_lock);
1428 /* invalidate all of request in the sending list */
1429 list_for_each_entry_safe(our, tmp, &list, our_list) {
1430 spin_lock(&our->our_list_lock);
1431 list_del_init(&our->our_list);
1432 spin_unlock(&our->our_list_lock);
1433 osp_trans_callback(&env, our->our_th,
1434 our->our_th->ot_super.th_result);
1435 osp_thandle_put(&env, our->our_th);
1441 * Sending update thread
1443 * Create thread to send update request to other MDTs, this thread will pull
1444 * out update request from the list in OSP by version number, i.e. it will
1445 * make sure the update request with lower version number will be sent first.
1447 * \param[in] arg hold the OSP device.
1449 * \retval 0 if the thread is created successfully.
1450 * \retal negative error if the thread is not created
1453 int osp_send_update_thread(void *arg)
1456 struct osp_device *osp = arg;
1457 struct osp_updates *ou = osp->opd_update;
1458 struct osp_update_request *our = NULL;
1462 LASSERT(ou != NULL);
1467 wait_event_idle(ou->ou_waitq,
1468 kthread_should_stop() ||
1469 osp_get_next_request(ou, &our));
1471 if (kthread_should_stop()) {
1473 osp_trans_callback(env, our->our_th, -EINTR);
1474 osp_thandle_put(env, our->our_th);
1479 LASSERT(our->our_th != NULL);
1480 if (our->our_th->ot_super.th_result != 0) {
1481 osp_trans_callback(env, our->our_th,
1482 our->our_th->ot_super.th_result);
1483 rc = our->our_th->ot_super.th_result;
1484 } else if (CFS_FAIL_CHECK(OBD_FAIL_INVALIDATE_UPDATE)) {
1486 osp_trans_callback(env, our->our_th, rc);
1488 rc = osp_send_update_req(env, osp, our);
1491 /* Update the rpc version */
1492 spin_lock(&ou->ou_lock);
1493 if (our->our_version == ou->ou_rpc_version)
1494 ou->ou_rpc_version++;
1495 spin_unlock(&ou->ou_lock);
1497 /* If one update request fails, let's fail all of the requests
1498 * in the sending list, because the request in the sending
1499 * list are dependent on either other, continue sending these
1500 * request might cause llog or filesystem corruption */
1502 osp_invalidate_request(osp);
1504 /* Balanced for thandle_get in osp_check_and_set_rpc_version */
1505 osp_thandle_put(env, our->our_th);
1512 * The OSP layer dt_device_operations::dt_trans_start() interface
1513 * to start the transaction.
1515 * If the transaction is a remote transaction, then related remote
1516 * updates will be triggered in the osp_trans_stop().
1517 * Please refer to osp_trans_create() for transaction type.
1519 * \param[in] env pointer to the thread context
1520 * \param[in] dt pointer to the OSP dt_device
1521 * \param[in] th pointer to the transaction handler
1523 * \retval 0 for success
1524 * \retval negative error number on failure
1526 int osp_trans_start(const struct lu_env *env, struct dt_device *dt,
1529 struct osp_thandle *oth = thandle_to_osp_thandle(th);
1530 struct osp_device *osp = dt2osp_dev(dt);
1531 struct osp_updates *ou = osp->opd_update;
1534 LASSERT(oth->ot_our);
1535 oth->ot_our->our_generation = ou->ou_generation;
1537 if (oth->ot_super.th_sync)
1538 oth->ot_our->our_flags |= UPDATE_FL_SYNC;
1539 /* For remote thandle, if there are local thandle, start it here*/
1540 if (is_only_remote_trans(th) && oth->ot_storage_th != NULL)
1541 return dt_trans_start(env, oth->ot_storage_th->th_dev,
1542 oth->ot_storage_th);
1547 * The OSP layer dt_device_operations::dt_trans_stop() interface
1548 * to stop the transaction.
1550 * If the transaction is a remote transaction, related remote
1551 * updates will be triggered at the end of this function.
1553 * For synchronous mode update or any failed update, the request
1554 * will be destroyed explicitly when the osp_trans_stop().
1556 * Please refer to osp_trans_create() for transaction type.
1558 * \param[in] env pointer to the thread context
1559 * \param[in] dt pointer to the OSP dt_device
1560 * \param[in] th pointer to the transaction handler
1562 * \retval 0 for success
1563 * \retval negative error number on failure
1565 int osp_trans_stop(const struct lu_env *env, struct dt_device *dt,
1568 struct osp_thandle *oth = thandle_to_osp_thandle(th);
1569 struct osp_update_request *our = oth->ot_our;
1570 struct osp_device *osp = dt2osp_dev(dt);
1574 /* For remote transaction, if there is local storage thandle,
1576 if (oth->ot_storage_th != NULL && th->th_top == NULL) {
1577 dt_trans_stop(env, oth->ot_storage_th->th_dev,
1578 oth->ot_storage_th);
1579 oth->ot_storage_th = NULL;
1582 if (our == NULL || list_empty(&our->our_req_list)) {
1583 osp_trans_callback(env, oth, th->th_result);
1584 GOTO(out, rc = th->th_result);
1587 if (!osp->opd_connect_mdt) {
1588 osp_trans_callback(env, oth, th->th_result);
1589 rc = osp_send_update_req(env, osp, oth->ot_our);
1593 if (osp->opd_update == NULL) {
1594 osp_trans_callback(env, oth, -EIO);
1595 GOTO(out, rc = -EIO);
1598 CDEBUG(D_HA, "%s: add oth %p with version %llu\n",
1599 osp->opd_obd->obd_name, oth, our->our_version);
1601 LASSERT(our->our_req_ready == 0);
1602 spin_lock(&our->our_list_lock);
1603 if (likely(!list_empty(&our->our_list))) {
1604 /* notify sending thread */
1605 our->our_req_ready = 1;
1606 wake_up(&osp->opd_update->ou_waitq);
1607 spin_unlock(&our->our_list_lock);
1608 } else if (th->th_result == 0) {
1609 /* if the request does not needs to be serialized,
1610 * read-only request etc, let's send it right away */
1611 spin_unlock(&our->our_list_lock);
1612 rc = osp_send_update_req(env, osp, our);
1614 spin_unlock(&our->our_list_lock);
1615 osp_trans_callback(env, oth, th->th_result);
1618 osp_thandle_put(env, oth);