Viewing: update_trans.c
// SPDX-License-Identifier: GPL-2.0
/*
* Copyright (c) 2015, 2017, Intel Corporation.
*/
/*
* This file implements the update distribute transaction API.
*
* To manage the cross-MDT operation (distribute operation) transaction,
* the transaction will also be separated two layers on MD stack, top
* transaction and sub transaction.
*
* During the distribute operation, top transaction is created in the LOD
* layer, and represent the operation. Sub transaction is created by
* each OSD or OSP. Top transaction start/stop will trigger all of its sub
* transaction start/stop. Top transaction (the whole operation) is committed
* only all of its sub transaction are committed.
*
* there are three kinds of transactions
* 1. local transaction: All updates are in a single local OSD.
* 2. Remote transaction: All Updates are only in the remote OSD,
* i.e. locally all updates are in OSP.
* 3. Mixed transaction: Updates are both in local OSD and remote
* OSD.
*
* Author: Di Wang <di.wang@intel.com>
*/
#define DEBUG_SUBSYSTEM S_CLASS
#include <linux/kthread.h>
#include <lu_target.h>
#include <lustre_log.h>
#include <lustre_update.h>
#include <obd.h>
#include <obd_class.h>
#include "tgt_internal.h"
/**
* top_multiple_thandle_dump() - Dump top mulitple thandle
* @tmt: top_thandle to be dumped
* @mask: debug mask
*
* Dump top multiple thandle and all of its sub thandle to the debug log.
*/
static void top_multiple_thandle_dump(struct top_multiple_thandle *tmt,
__u32 mask)
{
struct sub_thandle *st;
LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
CDEBUG(mask, "%s tmt %p refcount %d committed %d result %d batchid %llu\n",
tmt->tmt_master_sub_dt ?
tmt->tmt_master_sub_dt->dd_lu_dev.ld_obd->obd_name :
"NULL",
tmt, kref_read(&tmt->tmt_refcount), tmt->tmt_committed,
tmt->tmt_result, tmt->tmt_batchid);
list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
struct sub_thandle_cookie *stc;
CDEBUG(mask, "st %p obd %s committed %d started %d stopped %d result %d sub_th %p\n",
st, st->st_dt->dd_lu_dev.ld_obd->obd_name,
st->st_committed, st->st_started, st->st_stopped,
st->st_result, st->st_sub_th);
list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
CDEBUG(mask, " cookie "DFID".%u\n",
PLOGID(&stc->stc_cookie.lgc_lgl),
stc->stc_cookie.lgc_index);
}
}
}
/**
* sub_declare_updates_write() - Declare write update to sub device
* @env: execution environment
* @record: update records being written
* @sub_th: sub transaction handle
* @record_size: total update record size
*
* Declare Write updates llog records to the sub device during distribute
* transaction.
*
* Return:
* * %0 if writing succeeds
* * %negative errno if writing fails
*/
static int sub_declare_updates_write(const struct lu_env *env,
struct llog_update_record *record,
struct thandle *sub_th, size_t record_size)
{
struct llog_ctxt *ctxt;
struct dt_device *dt = sub_th->th_dev;
int left = record_size;
int rc;
/* If ctxt is NULL, it means not need to write update,
* for example if the OSP is used to connect to OST */
ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
LLOG_UPDATELOG_ORIG_CTXT);
/* Not ready to record updates yet. */
if (ctxt == NULL || ctxt->loc_handle == NULL) {
llog_ctxt_put(ctxt);
return 0;
}
rc = llog_declare_add(env, ctxt->loc_handle,
&record->lur_hdr, sub_th);
if (rc < 0)
GOTO(out_put, rc);
while (left > ctxt->loc_chunk_size) {
rc = llog_declare_add(env, ctxt->loc_handle,
&record->lur_hdr, sub_th);
if (rc < 0)
GOTO(out_put, rc);
left -= ctxt->loc_chunk_size;
}
out_put:
llog_ctxt_put(ctxt);
return rc;
}
/**
* sub_updates_write() - write update to sub device
* @env: execution environment
* @record: update records being written
* @sub_th: sub transaction handle
*
* Write llog update record to the sub device during distribute
* transaction. If it succeeds, llog cookie of the record will be
* returned by @cookie.
*
* Return:
* * %1 if writing succeeds
* * %negative errno if writing fails
*/
static int sub_updates_write(const struct lu_env *env,
struct llog_update_record *record,
struct sub_thandle *sub_th)
{
struct dt_device *dt = sub_th->st_dt;
struct llog_ctxt *ctxt;
struct llog_update_record *lur = NULL;
__u32 update_count = 0;
__u32 param_count = 0;
__u32 last_update_count = 0;
__u32 last_param_count = 0;
char *start;
char *cur;
char *next;
struct sub_thandle_cookie *stc;
size_t reclen;
bool eof = false;
int rc;
ENTRY;
ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
LLOG_UPDATELOG_ORIG_CTXT);
/* If ctxt == NULL, then it means updates on OST (only happens
* during migration), and we do not track those updates for now */
/* If ctxt->loc_handle == NULL, then it does not need to record
* update, usually happens in error handler path */
if (ctxt == NULL || ctxt->loc_handle == NULL) {
llog_ctxt_put(ctxt);
RETURN(0);
}
/* Since the cross-MDT updates will includes both local
* and remote updates, the update ops count must > 1 */
LASSERT(record->lur_update_rec.ur_update_count > 1);
LASSERTF(record->lur_hdr.lrh_len == llog_update_record_size(record),
"lrh_len %u record_size %zu\n", record->lur_hdr.lrh_len,
llog_update_record_size(record));
/*
* If its size > llog chunk_size, then write current chunk to the update
* llog, NB the padding should >= LLOG_MIN_REC_SIZE.
*
* So check padding length is either >= LLOG_MIN_REC_SIZE or is 0
* (record length just matches the chunk size).
*/
reclen = record->lur_hdr.lrh_len;
if (reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
reclen == ctxt->loc_chunk_size) {
OBD_ALLOC_PTR(stc);
if (stc == NULL)
GOTO(llog_put, rc = -ENOMEM);
INIT_LIST_HEAD(&stc->stc_list);
rc = llog_add(env, ctxt->loc_handle, &record->lur_hdr,
&stc->stc_cookie, sub_th->st_sub_th);
CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
dt->dd_lu_dev.ld_obd->obd_name,
PLOGID(&stc->stc_cookie.lgc_lgl),
stc->stc_cookie.lgc_index, rc);
if (rc > 0) {
list_add(&stc->stc_list, &sub_th->st_cookie_list);
rc = 0;
} else {
OBD_FREE_PTR(stc);
}
GOTO(llog_put, rc);
}
/* Split the records into chunk_size update record */
OBD_ALLOC_LARGE(lur, ctxt->loc_chunk_size);
if (lur == NULL)
GOTO(llog_put, rc = -ENOMEM);
memcpy(lur, &record->lur_hdr, sizeof(record->lur_hdr));
lur->lur_update_rec.ur_update_count = 0;
lur->lur_update_rec.ur_param_count = 0;
start = (char *)&record->lur_update_rec.ur_ops;
cur = next = start;
do {
if (update_count < record->lur_update_rec.ur_update_count)
next = (char *)update_op_next_op(
(struct update_op *)cur);
else if (param_count < record->lur_update_rec.ur_param_count)
next = (char *)update_param_next_param(
(struct update_param *)cur);
else
eof = true;
reclen = __llog_update_record_size(
__update_records_size(next - start));
if ((reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
reclen == ctxt->loc_chunk_size) &&
!eof) {
cur = next;
if (update_count <
record->lur_update_rec.ur_update_count)
update_count++;
else if (param_count <
record->lur_update_rec.ur_param_count)
param_count++;
continue;
}
lur->lur_update_rec.ur_update_count = update_count -
last_update_count;
lur->lur_update_rec.ur_param_count = param_count -
last_param_count;
memcpy(&lur->lur_update_rec.ur_ops, start, cur - start);
lur->lur_hdr.lrh_len = llog_update_record_size(lur);
LASSERT(lur->lur_hdr.lrh_len ==
__llog_update_record_size(
__update_records_size(cur - start)));
LASSERT(lur->lur_hdr.lrh_len <= ctxt->loc_chunk_size);
update_records_dump(&lur->lur_update_rec, D_INFO, true);
OBD_ALLOC_PTR(stc);
if (stc == NULL)
GOTO(llog_put, rc = -ENOMEM);
INIT_LIST_HEAD(&stc->stc_list);
rc = llog_add(env, ctxt->loc_handle, &lur->lur_hdr,
&stc->stc_cookie, sub_th->st_sub_th);
CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
dt->dd_lu_dev.ld_obd->obd_name,
PLOGID(&stc->stc_cookie.lgc_lgl),
stc->stc_cookie.lgc_index, rc);
if (rc > 0) {
list_add(&stc->stc_list, &sub_th->st_cookie_list);
rc = 0;
} else {
OBD_FREE_PTR(stc);
GOTO(llog_put, rc);
}
last_update_count = update_count;
last_param_count = param_count;
start = cur;
lur->lur_update_rec.ur_update_count = 0;
lur->lur_update_rec.ur_param_count = 0;
lur->lur_update_rec.ur_flags |= UPDATE_RECORD_CONTINUE;
} while (!eof);
llog_put:
if (lur != NULL)
OBD_FREE_LARGE(lur, ctxt->loc_chunk_size);
llog_ctxt_put(ctxt);
RETURN(rc);
}
/**
* prepare_writing_updates() - Prepare the update records.
* @env: execution environment
* @tmt: top_multiple_thandle for distribute txn
*
* Merge params and ops into the update records, then initializing
* the update buffer.
*
* During transaction execution phase, parameters and update ops
* are collected in two different buffers (see lod_updates_pack()),
* during transaction stop, it needs to be merged in one buffer,
* so it will be written in the update log.
*
* Return:
* * %0 if merging succeeds.
* * %negaitive errno if merging fails.
*/
static int prepare_writing_updates(const struct lu_env *env,
struct top_multiple_thandle *tmt)
{
struct thandle_update_records *tur = tmt->tmt_update_records;
struct llog_update_record *lur;
struct update_params *params;
size_t params_size;
size_t update_size;
if (tur == NULL || tur->tur_update_records == NULL ||
tur->tur_update_params == NULL)
return 0;
lur = tur->tur_update_records;
/* Extends the update records buffer if needed */
params_size = update_params_size(tur->tur_update_params,
tur->tur_update_param_count);
LASSERT(lur->lur_update_rec.ur_param_count == 0);
update_size = round_up(llog_update_record_size(lur) + params_size, 8);
if (update_size > tur->tur_update_records_buf_size) {
int rc;
rc = tur_update_records_extend(tur, update_size);
if (rc < 0)
return rc;
lur = tur->tur_update_records;
}
params = update_records_get_params(&lur->lur_update_rec);
memcpy(params, tur->tur_update_params, params_size);
lur->lur_update_rec.ur_param_count = tur->tur_update_param_count;
lur->lur_update_rec.ur_batchid = tmt->tmt_batchid;
/* Init update record header */
lur->lur_hdr.lrh_len = llog_update_record_size(lur);
lur->lur_hdr.lrh_type = UPDATE_REC;
/* Dump updates for debugging purpose */
update_records_dump(&lur->lur_update_rec, D_INFO, true);
return 0;
}
/**
* top_trans_committed_cb() - Top thandle commit callback
* @tmt: top thandle to be committed.
*
* This callback will be called when all of sub transactions are committed.
*/
static void top_trans_committed_cb(struct top_multiple_thandle *tmt)
{
struct lu_target *lut;
ENTRY;
LASSERT(kref_read(&tmt->tmt_refcount) > 0);
top_multiple_thandle_dump(tmt, D_HA);
tmt->tmt_committed = 1;
lut = dt2lu_dev(tmt->tmt_master_sub_dt)->ld_site->ls_tgt;
if (lut->lut_tdtd && lut->lut_tdtd->tdtd_commit_task)
wake_up_process(lut->lut_tdtd->tdtd_commit_task);
RETURN_EXIT;
}
/**
* lookup_sub_thandle() - Search for sub_handle within top handle
* @tmt: top_multiple_thandle for lookup
* @dt_dev: dt_device involved in the transaction
*
* Returns sub_handle on success or %NULL on failure
*/
struct sub_thandle *lookup_sub_thandle(struct top_multiple_thandle *tmt,
struct dt_device *dt_dev)
{
struct sub_thandle *st;
list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
if (st->st_dt == dt_dev)
return st;
}
return NULL;
}
EXPORT_SYMBOL(lookup_sub_thandle);
/**
* create_sub_thandle() - create sub_handle
* @tmt: top_multiple_thandle where sub_handle should be created
* @dt_dev: dt_device involved in the transaction
*
* Returns sub_handle on success or %NULL on failure
*/
struct sub_thandle *create_sub_thandle(struct top_multiple_thandle *tmt,
struct dt_device *dt_dev)
{
struct sub_thandle *st;
OBD_ALLOC_PTR(st);
if (st == NULL)
RETURN(ERR_PTR(-ENOMEM));
INIT_LIST_HEAD(&st->st_sub_list);
INIT_LIST_HEAD(&st->st_cookie_list);
st->st_dt = dt_dev;
list_add(&st->st_sub_list, &tmt->tmt_sub_thandle_list);
return st;
}
static void sub_trans_commit_cb_internal(struct top_multiple_thandle *tmt,
struct thandle *sub_th, int err)
{
struct sub_thandle *st;
bool all_committed = true;
/* Check if all sub thandles are committed */
spin_lock(&tmt->tmt_sub_lock);
list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
if (st->st_sub_th == sub_th) {
st->st_committed = 1;
st->st_result = err;
}
if (!st->st_committed)
all_committed = false;
}
spin_unlock(&tmt->tmt_sub_lock);
if (tmt->tmt_result == 0)
tmt->tmt_result = err;
if (all_committed)
top_trans_committed_cb(tmt);
top_multiple_thandle_dump(tmt, D_INFO);
top_multiple_thandle_put(tmt);
RETURN_EXIT;
}
/**
* sub_trans_commit_cb() - sub thandle commit callback
* @env: execution environment
* @sub_th: sub thandle being committed
* @cb: commit callback
* @err: trans result
*
* Mark the sub thandle to be committed and if all sub thandle are committed
* notify the top thandle.
*/
static void sub_trans_commit_cb(struct lu_env *env,
struct thandle *sub_th,
struct dt_txn_commit_cb *cb, int err)
{
struct top_multiple_thandle *tmt = cb->dcb_data;
sub_trans_commit_cb_internal(tmt, sub_th, err);
}
static void sub_thandle_register_commit_cb(struct sub_thandle *st,
struct top_multiple_thandle *tmt)
{
LASSERT(st->st_sub_th != NULL);
top_multiple_thandle_get(tmt);
st->st_commit_dcb.dcb_func = sub_trans_commit_cb;
st->st_commit_dcb.dcb_data = tmt;
INIT_LIST_HEAD(&st->st_commit_dcb.dcb_linkage);
dt_trans_cb_add(st->st_sub_th, &st->st_commit_dcb);
}
/**
* sub_trans_stop_cb() - Sub thandle stop call back
* @env: execution environment
* @sub_th: sub thandle to be stopped
* @cb: per-transaction callback
* @err: result of sub trans (@cb)
*
* After sub thandle is stopped, it will call this callback to notify
* the top thandle.
*/
static void sub_trans_stop_cb(struct lu_env *env,
struct thandle *sub_th,
struct dt_txn_commit_cb *cb, int err)
{
struct sub_thandle *st;
struct top_multiple_thandle *tmt = cb->dcb_data;
ENTRY;
spin_lock(&tmt->tmt_sub_lock);
list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
if (st->st_stopped)
continue;
if (st->st_dt == sub_th->th_dev) {
st->st_stopped = 1;
st->st_result = err;
break;
}
}
spin_unlock(&tmt->tmt_sub_lock);
wake_up(&tmt->tmt_stop_waitq);
RETURN_EXIT;
}
static void sub_thandle_register_stop_cb(struct sub_thandle *st,
struct top_multiple_thandle *tmt)
{
st->st_stop_dcb.dcb_func = sub_trans_stop_cb;
st->st_stop_dcb.dcb_data = tmt;
st->st_stop_dcb.dcb_flags = DCB_TRANS_STOP;
INIT_LIST_HEAD(&st->st_stop_dcb.dcb_linkage);
dt_trans_cb_add(st->st_sub_th, &st->st_stop_dcb);
}
/**
* sub_thandle_trans_create() - Create sub thandle
* @env: execution environment
* @top_th: top thandle
* @st: sub_thandle
*
* Create transaction handle for sub_thandle
*
* Return:
* * %0 if creation succeeds.
* * %negative errno if creation fails.
*/
int sub_thandle_trans_create(const struct lu_env *env,
struct top_thandle *top_th,
struct sub_thandle *st)
{
struct thandle *sub_th;
sub_th = dt_trans_create(env, st->st_dt);
if (IS_ERR(sub_th))
return PTR_ERR(sub_th);
sub_th->th_top = &top_th->tt_super;
st->st_sub_th = sub_th;
sub_th->th_wait_submit = 1;
sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
return 0;
}
/**
* top_trans_create() - Create the top transaction.
* @env: execution environment
* @master_dev: master_dev the top thandle will be created
*
* Create the top transaction on the master device. It will create a top
* thandle and a sub thandle on the master device.
*
* Return:
* * %pointer to the created thandle.
* * %ERR_PTR(errno) if creation failed.
*/
struct thandle *
top_trans_create(const struct lu_env *env, struct dt_device *master_dev)
{
struct top_thandle *top_th;
struct thandle *child_th;
OBD_ALLOC_GFP(top_th, sizeof(*top_th), GFP_NOFS);
if (top_th == NULL)
return ERR_PTR(-ENOMEM);
top_th->tt_super.th_top = &top_th->tt_super;
if (master_dev != NULL) {
child_th = dt_trans_create(env, master_dev);
if (IS_ERR(child_th)) {
OBD_FREE_PTR(top_th);
return child_th;
}
child_th->th_top = &top_th->tt_super;
child_th->th_wait_submit = 1;
top_th->tt_master_sub_thandle = child_th;
}
return &top_th->tt_super;
}
EXPORT_SYMBOL(top_trans_create);
/**
* declare_updates_write() - Declare write update transaction
* @env: execution environment
* @tmt: top multiple transaction handle
*
* Check if there are updates being recorded in this transaction,
* it will write the record into the disk.
*
* Return:
* * %0 if writing succeeds
* * %negative errno if writing fails
*/
static int declare_updates_write(const struct lu_env *env,
struct top_multiple_thandle *tmt)
{
struct llog_update_record *record;
struct sub_thandle *st;
int rc = 0;
record = tmt->tmt_update_records->tur_update_records;
/* Declare update write for all other target */
list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
if (st->st_sub_th == NULL)
continue;
rc = sub_declare_updates_write(env, record, st->st_sub_th,
tmt->tmt_record_size);
if (rc < 0)
break;
}
return rc;
}
/**
* distribute_txn_assign_batchid() - Assign batchid to the distribute
* transaction.
* @new: the distribute txn for assignment
*
* Assign batchid to the distribute transaction
*
*/
static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
{
struct target_distribute_txn_data *tdtd;
struct dt_device *dt = new->tmt_master_sub_dt;
struct sub_thandle *st;
LASSERT(dt != NULL);
tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
spin_lock(&tdtd->tdtd_batchid_lock);
new->tmt_batchid = tdtd->tdtd_batchid++;
list_add_tail(&new->tmt_commit_list, &tdtd->tdtd_list);
spin_unlock(&tdtd->tdtd_batchid_lock);
list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
if (st->st_sub_th != NULL)
sub_thandle_register_commit_cb(st, new);
}
top_multiple_thandle_get(new);
top_multiple_thandle_dump(new, D_INFO);
}
/**
* distribute_txn_insert_by_batchid() - Insert distribute transaction to the
* distribute txn list.
* @new: the distribute txn to be inserted.
*
* Insert distribute transaction to the distribute txn list.
*/
void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
{
struct dt_device *dt = new->tmt_master_sub_dt;
struct top_multiple_thandle *tmt;
struct target_distribute_txn_data *tdtd;
struct sub_thandle *st;
bool at_head = false;
LASSERT(dt != NULL);
tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
/* have a reference so the competing commit thread can't release it*/
top_multiple_thandle_get(new);
spin_lock(&tdtd->tdtd_batchid_lock);
list_for_each_entry_reverse(tmt, &tdtd->tdtd_list, tmt_commit_list) {
if (new->tmt_batchid > tmt->tmt_batchid) {
list_add(&new->tmt_commit_list, &tmt->tmt_commit_list);
break;
}
}
if (list_empty(&new->tmt_commit_list)) {
at_head = true;
list_add(&new->tmt_commit_list, &tdtd->tdtd_list);
}
spin_unlock(&tdtd->tdtd_batchid_lock);
list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
if (st->st_sub_th != NULL)
sub_thandle_register_commit_cb(st, new);
}
top_multiple_thandle_dump(new, D_INFO);
if (new->tmt_committed && at_head && tdtd->tdtd_commit_task)
wake_up_process(tdtd->tdtd_commit_task);
}
/**
* prepare_multiple_node_trans() - Prepare cross-MDT operation.
* @env: execution environment
* @tmt: top transaction handle
*
* Create the update record buffer to record updates for cross-MDT operation,
* add master sub transaction to tt_sub_trans_list, and declare the update
* writes.
*
* During updates packing, all of parameters will be packed in
* tur_update_params, and updates will be packed in tur_update_records.
* Then in transaction stop, parameters and updates will be merged
* into one updates buffer.
*
* And also master thandle will be added to the sub_th list, so it will be
* easy to track the commit status.
*
* Return:
* * %0 if preparation succeeds.
* * %negative errno if preparation fails.
*/
static int prepare_multiple_node_trans(const struct lu_env *env,
struct top_multiple_thandle *tmt)
{
struct thandle_update_records *tur;
int rc;
ENTRY;
if (tmt->tmt_update_records == NULL) {
tur = &update_env_info(env)->uti_tur;
rc = check_and_prepare_update_record(env, tur);
if (rc < 0)
RETURN(rc);
tmt->tmt_update_records = tur;
distribute_txn_assign_batchid(tmt);
}
rc = declare_updates_write(env, tmt);
RETURN(rc);
}
/**
* top_trans_start() - start the top transaction.
* @env: execution environment
* @master_dev: master_dev the top thandle will be start
* @th: top thandle
*
* Start all of its sub transactions, then start master sub transaction.
*
* Return:
* * %0 if transaction start succeeds.
* * %negative errno if start fails.
*/
int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
struct thandle *th)
{
struct top_thandle *top_th = container_of(th, struct top_thandle,
tt_super);
struct sub_thandle *st;
struct top_multiple_thandle *tmt = top_th->tt_multiple_thandle;
int rc = 0;
ENTRY;
if (tmt == NULL) {
if (th->th_sync)
top_th->tt_master_sub_thandle->th_sync = th->th_sync;
if (th->th_local)
top_th->tt_master_sub_thandle->th_local = th->th_local;
rc = dt_trans_start(env, top_th->tt_master_sub_thandle->th_dev,
top_th->tt_master_sub_thandle);
RETURN(rc);
}
tmt = top_th->tt_multiple_thandle;
rc = prepare_multiple_node_trans(env, tmt);
if (rc < 0)
RETURN(rc);
list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
if (st->st_sub_th == NULL)
continue;
if (th->th_sync)
st->st_sub_th->th_sync = th->th_sync;
if (th->th_local)
st->st_sub_th->th_local = th->th_local;
rc = dt_trans_start(env, st->st_sub_th->th_dev,
st->st_sub_th);
if (rc != 0)
GOTO(out, rc);
LASSERT(st->st_started == 0);
st->st_started = 1;
}
out:
th->th_result = rc;
RETURN(rc);
}
EXPORT_SYMBOL(top_trans_start);
/**
* top_check_write_updates() - Check whether we need write updates record
* @top_th: top thandle.
*
* Check if the updates for the top_thandle needs to be writen
* to all targets. Only if the transaction succeeds and the updates
* number > 2, it will write the updates,
*
* Return:
* * %true if it needs to write updates
* * %false if it does not need to write updates
*/
static bool top_check_write_updates(struct top_thandle *top_th)
{
struct top_multiple_thandle *tmt;
struct thandle_update_records *tur;
/* Do not write updates to records if the transaction fails */
if (top_th->tt_super.th_result != 0)
return false;
tmt = top_th->tt_multiple_thandle;
if (tmt == NULL)
return false;
tur = tmt->tmt_update_records;
if (tur == NULL)
return false;
/* Hmm, false update records, since the cross-MDT operation
* should includes both local and remote updates, so the
* updates count should >= 2 */
if (tur->tur_update_records == NULL ||
tur->tur_update_records->lur_update_rec.ur_update_count <= 1)
return false;
return true;
}
/**
* top_trans_is_stopped() - Check if top transaction is stopped
* @top_th: top thandle
*
* Check if top transaction is stopped, only if all sub transaction
* is stopped, then the top transaction is stopped.
*
* Return:
* * %true if the top transaction is stopped.
* * %false if the top transaction is not stopped.
*/
static bool top_trans_is_stopped(struct top_thandle *top_th)
{
struct top_multiple_thandle *tmt;
struct sub_thandle *st;
bool all_stopped = true;
tmt = top_th->tt_multiple_thandle;
list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
if (!st->st_stopped && st->st_sub_th != NULL) {
all_stopped = false;
break;
}
if (st->st_result != 0 &&
top_th->tt_super.th_result == 0)
top_th->tt_super.th_result = st->st_result;
}
return all_stopped;
}
/**
* top_trans_wait_result() - Wait result of top transaction
* @top_th: top thandle.
*
* Wait until all sub transaction get its result.
*
* Returns the result of top thandle.
*/
static int top_trans_wait_result(struct top_thandle *top_th)
{
wait_event_idle(top_th->tt_multiple_thandle->tmt_stop_waitq,
top_trans_is_stopped(top_th));
RETURN(top_th->tt_super.th_result);
}
/**
* top_trans_stop() - Stop the top transaction.
* @env: execution environment
* @master_dev: master_dev the top thandle will be created
* @th: top thandle
*
* Stop the transaction on the master device first, then stop transactions
* on other sub devices.
*
* Return:
* * %0 if stop transaction succeeds.
* * %negative errno if stop transaction fails.
*/
int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
struct thandle *th)
{
struct top_thandle *top_th = container_of(th, struct top_thandle,
tt_super);
struct sub_thandle *st;
struct sub_thandle *master_st;
struct top_multiple_thandle *tmt;
struct thandle_update_records *tur;
bool write_updates = false;
int rc = 0;
ENTRY;
if (likely(top_th->tt_multiple_thandle == NULL)) {
LASSERT(master_dev != NULL);
if (th->th_sync)
top_th->tt_master_sub_thandle->th_sync = th->th_sync;
if (th->th_local)
top_th->tt_master_sub_thandle->th_local = th->th_local;
rc = dt_trans_stop(env, master_dev,
top_th->tt_master_sub_thandle);
OBD_FREE_PTR(top_th);
RETURN(rc);
}
tmt = top_th->tt_multiple_thandle;
tur = tmt->tmt_update_records;
/* Note: we need stop the master thandle first, then the stop
* callback will fill the master transno in the update logs,
* then these update logs will be sent to other MDTs */
/* get the master sub thandle */
master_st = lookup_sub_thandle(tmt, tmt->tmt_master_sub_dt);
write_updates = top_check_write_updates(top_th);
/* Step 1: write the updates log on Master MDT */
if (master_st != NULL && master_st->st_sub_th != NULL &&
write_updates) {
struct llog_update_record *lur;
/* Merge the parameters and updates into one buffer */
rc = prepare_writing_updates(env, tmt);
if (rc < 0) {
CERROR("%s: cannot prepare updates: rc = %d\n",
master_dev->dd_lu_dev.ld_obd->obd_name, rc);
th->th_result = rc;
write_updates = false;
GOTO(stop_master_trans, rc);
}
lur = tur->tur_update_records;
/* Write updates to the master MDT */
rc = sub_updates_write(env, lur, master_st);
/* Cleanup the common parameters in the update records,
* master transno callback might add more parameters.
* and we need merge the update records again in the
* following */
if (tur->tur_update_params != NULL)
lur->lur_update_rec.ur_param_count = 0;
if (rc < 0) {
CERROR("%s: write updates failed: rc = %d\n",
master_dev->dd_lu_dev.ld_obd->obd_name, rc);
th->th_result = rc;
write_updates = false;
GOTO(stop_master_trans, rc);
}
}
stop_master_trans:
/* Step 2: Stop the transaction on the master MDT, and fill the
* master transno in the update logs to other MDT. */
if (master_st != NULL && master_st->st_sub_th != NULL) {
if (th->th_local)
master_st->st_sub_th->th_local = th->th_local;
if (th->th_sync)
master_st->st_sub_th->th_sync = th->th_sync;
master_st->st_sub_th->th_result = th->th_result;
rc = dt_trans_stop(env, master_st->st_dt, master_st->st_sub_th);
/* If it does not write_updates, then we call submit callback
* here, otherwise callback is done through
* osd(osp)_trans_commit_cb() */
if (!master_st->st_started &&
!list_empty(&tmt->tmt_commit_list))
sub_trans_commit_cb_internal(tmt,
master_st->st_sub_th, rc);
if (rc < 0) {
CERROR("%s: stop trans failed: rc = %d\n",
master_dev->dd_lu_dev.ld_obd->obd_name, rc);
th->th_result = rc;
GOTO(stop_other_trans, rc);
} else if (tur != NULL && tur->tur_update_records != NULL) {
struct llog_update_record *lur;
lur = tur->tur_update_records;
if (lur->lur_update_rec.ur_master_transno == 0)
/* Update master transno after master stop
* callback */
lur->lur_update_rec.ur_master_transno =
tgt_th_info(env)->tti_transno;
}
}
/* Step 3: write updates to other MDTs */
if (write_updates) {
struct llog_update_record *lur;
if (CFS_FAIL_PRECHECK(OBD_FAIL_OUT_OBJECT_MISS)) {
if (cfs_fail_val == 1) {
long timeout = cfs_time_seconds(1) / 10;
CFS_RACE(OBD_FAIL_OUT_OBJECT_MISS);
set_current_state(TASK_UNINTERRUPTIBLE);
schedule_timeout(schedule_timeout(timeout));
cfs_fail_loc = 0;
}
cfs_fail_val++;
}
/* Stop callback of master will add more updates and also update
* master transno, so merge the parameters and updates into one
* buffer again */
rc = prepare_writing_updates(env, tmt);
if (rc < 0) {
CERROR("%s: prepare updates failed: rc = %d\n",
master_dev->dd_lu_dev.ld_obd->obd_name, rc);
th->th_result = rc;
GOTO(stop_other_trans, rc);
}
lur = tur->tur_update_records;
list_for_each_entry(st, &tmt->tmt_sub_thandle_list,
st_sub_list) {
if (st->st_sub_th == NULL || st == master_st ||
st->st_sub_th->th_result < 0)
continue;
rc = sub_updates_write(env, lur, st);
if (rc < 0) {
CERROR("%s: write updates failed: rc = %d\n",
st->st_dt->dd_lu_dev.ld_obd->obd_name,
rc);
th->th_result = rc;
break;
}
}
}
stop_other_trans:
/* Step 4: Stop the transaction on other MDTs */
list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
if (st == master_st || st->st_sub_th == NULL)
continue;
if (th->th_sync)
st->st_sub_th->th_sync = th->th_sync;
if (th->th_local)
st->st_sub_th->th_local = th->th_local;
st->st_sub_th->th_result = th->th_result;
rc = dt_trans_stop(env, st->st_sub_th->th_dev,
st->st_sub_th);
if (rc < 0) {
CERROR("%s: stop trans failed: rc = %d\n",
st->st_dt->dd_lu_dev.ld_obd->obd_name, rc);
if (th->th_result == 0)
th->th_result = rc;
}
}
rc = top_trans_wait_result(top_th);
tmt->tmt_result = rc;
/* Balance for the refcount in top_trans_create, Note: if it is NOT
* multiple node transaction, the top transaction will be destroyed. */
top_multiple_thandle_put(tmt);
OBD_FREE_PTR(top_th);
RETURN(rc);
}
EXPORT_SYMBOL(top_trans_stop);
/**
* top_trans_create_tmt() - Create top_multiple_thandle for top_thandle
* @env: execution environment
* @top_th: the top thandle
*
* Create top_mutilple_thandle to manage the mutiple node transaction
* for top_thandle, and it also needs to add master sub thandle to the
* sub trans list now.
*
* Return:
* * %0 if creation succeeds
* * %negative errno if creation fails
*/
int top_trans_create_tmt(const struct lu_env *env,
struct top_thandle *top_th)
{
struct top_multiple_thandle *tmt;
OBD_ALLOC_PTR(tmt);
if (tmt == NULL)
return -ENOMEM;
tmt->tmt_magic = TOP_THANDLE_MAGIC;
INIT_LIST_HEAD(&tmt->tmt_sub_thandle_list);
INIT_LIST_HEAD(&tmt->tmt_commit_list);
kref_init(&tmt->tmt_refcount);
spin_lock_init(&tmt->tmt_sub_lock);
init_waitqueue_head(&tmt->tmt_stop_waitq);
top_th->tt_multiple_thandle = tmt;
return 0;
}
static struct sub_thandle *
create_sub_thandle_with_thandle(struct top_thandle *top_th,
struct thandle *sub_th)
{
struct sub_thandle *st;
/* create and init sub th to the top trans list */
st = create_sub_thandle(top_th->tt_multiple_thandle,
sub_th->th_dev);
if (IS_ERR(st))
return st;
st->st_sub_th = sub_th;
sub_th->th_top = &top_th->tt_super;
sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
return st;
}
/**
* thandle_get_sub_by_dt() - Get sub thandle.
* @env: execution environment
* @th: thandle on the top layer.
* @sub_dt: sub dt_device used to get sub transaction
*
* Get sub thandle from the top thandle according to the sub dt_device.
*
* Return:
* * %thandle of sub transaction if succeed
* * %PTR_ERR(errno) if failed
*/
struct thandle *thandle_get_sub_by_dt(const struct lu_env *env,
struct thandle *th,
struct dt_device *sub_dt)
{
struct sub_thandle *st = NULL;
struct sub_thandle *master_st = NULL;
struct top_thandle *top_th;
struct thandle *sub_th = NULL;
int rc = 0;
ENTRY;
top_th = container_of(th, struct top_thandle, tt_super);
if (likely(sub_dt == top_th->tt_master_sub_thandle->th_dev))
RETURN(top_th->tt_master_sub_thandle);
if (top_th->tt_multiple_thandle != NULL) {
st = lookup_sub_thandle(top_th->tt_multiple_thandle, sub_dt);
if (st != NULL)
RETURN(st->st_sub_th);
}
sub_th = dt_trans_create(env, sub_dt);
if (IS_ERR(sub_th))
RETURN(sub_th);
/* Create top_multiple_thandle if necessary */
if (top_th->tt_multiple_thandle == NULL) {
struct top_multiple_thandle *tmt;
rc = top_trans_create_tmt(env, top_th);
if (rc < 0)
GOTO(stop_trans, rc);
tmt = top_th->tt_multiple_thandle;
/* Add master sub th to the top trans list */
tmt->tmt_master_sub_dt =
top_th->tt_master_sub_thandle->th_dev;
master_st = create_sub_thandle_with_thandle(top_th,
top_th->tt_master_sub_thandle);
if (IS_ERR(master_st)) {
rc = PTR_ERR(master_st);
master_st = NULL;
GOTO(stop_trans, rc);
}
}
/* create and init sub th to the top trans list */
st = create_sub_thandle_with_thandle(top_th, sub_th);
if (IS_ERR(st)) {
rc = PTR_ERR(st);
st = NULL;
GOTO(stop_trans, rc);
}
st->st_sub_th->th_wait_submit = 1;
stop_trans:
if (rc < 0) {
if (master_st != NULL) {
list_del(&master_st->st_sub_list);
OBD_FREE_PTR(master_st);
}
sub_th->th_result = rc;
dt_trans_stop(env, sub_dt, sub_th);
sub_th = ERR_PTR(rc);
}
RETURN(sub_th);
}
EXPORT_SYMBOL(thandle_get_sub_by_dt);
/**
* top_multiple_thandle_destroy() - Top multiple thandle destroy
* @kref: Pointer to struct kref
*
* Destroy multiple thandle and all its sub thandle.
*/
void top_multiple_thandle_destroy(struct kref *kref)
{
struct top_multiple_thandle *tmt;
struct sub_thandle *st;
struct sub_thandle *tmp;
tmt = container_of(kref, struct top_multiple_thandle, tmt_refcount);
LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
list_for_each_entry_safe(st, tmp, &tmt->tmt_sub_thandle_list,
st_sub_list) {
struct sub_thandle_cookie *stc;
struct sub_thandle_cookie *tmp;
list_del(&st->st_sub_list);
list_for_each_entry_safe(stc, tmp, &st->st_cookie_list,
stc_list) {
list_del(&stc->stc_list);
OBD_FREE_PTR(stc);
}
OBD_FREE_PTR(st);
}
OBD_FREE_PTR(tmt);
}
EXPORT_SYMBOL(top_multiple_thandle_destroy);
/**
* distribute_txn_cancel_records() - Cancel the update log on MDTs
* @env: execution environment
* @tmt: the top multiple thandle whose updates records will be cancelled.
*
* Cancel the update log on MDTs then destroy the thandle.
*
* Return:
* * %0 if cancellation succeeds.
* * %negative errno if cancellation fails.
*/
static int distribute_txn_cancel_records(const struct lu_env *env,
struct top_multiple_thandle *tmt)
{
struct sub_thandle *st;
ENTRY;
if (CFS_FAIL_CHECK(OBD_FAIL_TGT_TXN_NO_CANCEL))
RETURN(0);
top_multiple_thandle_dump(tmt, D_INFO);
/* Cancel update logs on other MDTs */
list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
struct llog_ctxt *ctxt;
struct obd_device *obd;
struct llog_cookie *cookie;
struct sub_thandle_cookie *stc;
int rc;
obd = st->st_dt->dd_lu_dev.ld_obd;
ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
if (ctxt == NULL)
continue;
list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
cookie = &stc->stc_cookie;
if (fid_is_zero(&cookie->lgc_lgl.lgl_oi.oi_fid))
continue;
rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
cookie);
CDEBUG(D_HA, "%s: batchid %llu cancel update log "
DFID".%u: rc = %d\n", obd->obd_name,
tmt->tmt_batchid, PLOGID(&cookie->lgc_lgl),
cookie->lgc_index, rc);
}
llog_ctxt_put(ctxt);
}
RETURN(0);
}
struct distribute_txn_bid_data {
struct dt_txn_commit_cb dtbd_cb;
struct target_distribute_txn_data *dtbd_tdtd;
__u64 dtbd_batchid;
};
/**
* distribute_txn_batchid_cb() - callback of updating commit batchid
* @env: execution environment
* @th: thandle to updating commit batchid
* @cb: commit callback
* @err: result of thandle
*
* Updating commit batchid then wake up the commit thread to cancel the
* records.
*/
static void distribute_txn_batchid_cb(struct lu_env *env,
struct thandle *th,
struct dt_txn_commit_cb *cb,
int err)
{
struct distribute_txn_bid_data *dtbd = NULL;
struct target_distribute_txn_data *tdtd;
dtbd = container_of(cb, struct distribute_txn_bid_data, dtbd_cb);
tdtd = dtbd->dtbd_tdtd;
CDEBUG(D_HA, "%s: %llu batchid updated\n",
tdtd->tdtd_lut->lut_obd->obd_name, dtbd->dtbd_batchid);
spin_lock(&tdtd->tdtd_batchid_lock);
if (dtbd->dtbd_batchid > tdtd->tdtd_committed_batchid &&
!tdtd->tdtd_lut->lut_obd->obd_no_transno)
tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
spin_unlock(&tdtd->tdtd_batchid_lock);
if (atomic_dec_and_test(&tdtd->tdtd_refcount))
wake_up_process(tdtd->tdtd_commit_task);
OBD_FREE_PTR(dtbd);
}
/**
* distribute_txn_commit_batchid_update() - Update the commit batchid in disk
* @env: execution environment
* @tdtd: distribute transaction structure
* @batchid: commit batchid to be updated
*
* Update commit batchid in the disk, after this is committed, it can start
* to cancel the update records.
*
* Return:
* * %0 if update succeeds.
* * %negative errno if update fails.
*/
static int
distribute_txn_commit_batchid_update(const struct lu_env *env,
struct target_distribute_txn_data *tdtd,
__u64 batchid)
{
struct distribute_txn_bid_data *dtbd = NULL;
struct thandle *th;
struct lu_buf buf;
__u64 tmp;
__u64 off;
int rc;
ENTRY;
OBD_ALLOC_PTR(dtbd);
if (dtbd == NULL)
RETURN(-ENOMEM);
dtbd->dtbd_batchid = batchid;
dtbd->dtbd_tdtd = tdtd;
dtbd->dtbd_cb.dcb_func = distribute_txn_batchid_cb;
atomic_inc(&tdtd->tdtd_refcount);
th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
if (IS_ERR(th)) {
atomic_dec(&tdtd->tdtd_refcount);
OBD_FREE_PTR(dtbd);
RETURN(PTR_ERR(th));
}
tmp = cpu_to_le64(batchid);
buf.lb_buf = &tmp;
buf.lb_len = sizeof(tmp);
off = 0;
rc = dt_declare_record_write(env, tdtd->tdtd_batchid_obj, &buf, off,
th);
if (rc < 0)
GOTO(stop, rc);
rc = dt_trans_start_local(env, tdtd->tdtd_lut->lut_bottom, th);
if (rc < 0)
GOTO(stop, rc);
rc = dt_trans_cb_add(th, &dtbd->dtbd_cb);
if (rc < 0)
GOTO(stop, rc);
rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
&off, th);
CDEBUG(D_INFO, "%s: update batchid %llu: rc = %d\n",
tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
stop:
dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
if (rc < 0) {
atomic_dec(&tdtd->tdtd_refcount);
OBD_FREE_PTR(dtbd);
}
RETURN(rc);
}
/**
* distribute_txn_commit_batchid_init() - Init commit batchid for distribute
* transaction.
* @env: execution environment
* @tdtd: distribute transaction whose batchid is initialized.
*
* Initialize the batchid object and get commit batchid from the object.
*
* Return:
* * %0 if initialization succeeds.
* * %negative errno if initialization fails.
*/
static int
distribute_txn_commit_batchid_init(const struct lu_env *env,
struct target_distribute_txn_data *tdtd)
{
struct tgt_thread_info *tti = tgt_th_info(env);
struct lu_target *lut = tdtd->tdtd_lut;
struct lu_attr *attr = &tti->tti_attr;
struct lu_fid *fid = &tti->tti_fid1;
struct dt_object_format *dof = &tti->tti_u.update.tti_update_dof;
struct dt_object *dt_obj = NULL;
struct lu_buf buf;
__u64 tmp;
__u64 off;
int rc;
ENTRY;
memset(attr, 0, sizeof(*attr));
attr->la_valid = LA_MODE;
attr->la_mode = S_IFREG | 0644;
dof->dof_type = dt_mode_to_dft(S_IFREG);
lu_local_obj_fid(fid, BATCHID_COMMITTED_OID);
dt_obj = dt_find_or_create(env, lut->lut_bottom, fid, dof,
attr);
if (IS_ERR(dt_obj)) {
rc = PTR_ERR(dt_obj);
dt_obj = NULL;
GOTO(out_put, rc);
}
tdtd->tdtd_batchid_obj = dt_obj;
buf.lb_buf = &tmp;
buf.lb_len = sizeof(tmp);
off = 0;
rc = dt_read(env, dt_obj, &buf, &off);
if (rc < 0 || (rc < buf.lb_len && rc > 0)) {
CERROR("%s can't read last committed batchid: rc = %d\n",
tdtd->tdtd_lut->lut_obd->obd_name, rc);
if (rc > 0)
rc = -EINVAL;
GOTO(out_put, rc);
} else if (rc == buf.lb_len) {
tdtd->tdtd_committed_batchid = le64_to_cpu(tmp);
CDEBUG(D_HA, "%s: committed batchid %llu\n",
tdtd->tdtd_lut->lut_obd->obd_name,
tdtd->tdtd_committed_batchid);
rc = 0;
}
out_put:
if (rc < 0 && dt_obj != NULL) {
dt_object_put(env, dt_obj);
tdtd->tdtd_batchid_obj = NULL;
}
return rc;
}
#ifndef TASK_IDLE
#define TASK_IDLE TASK_INTERRUPTIBLE
#endif
/**
* distribute_txn_commit_thread() - manage the distribute transaction thread
* @_arg: argument for commit thread
*
* Distribute transaction are linked to the list, and once the distribute
* transaction is committed, it will update the last committed batchid first,
* after it is committed, it will cancel the records.
*
* Return:
* * %0 if thread is running successfully
* * %negative errno if the thread can not be run.
*/
static int distribute_txn_commit_thread(void *_arg)
{
struct target_distribute_txn_data *tdtd = _arg;
struct lu_env *env = &tdtd->tdtd_env;
LIST_HEAD(list);
int rc;
struct top_multiple_thandle *tmt;
struct top_multiple_thandle *tmp;
__u64 batchid = 0, committed;
ENTRY;
CDEBUG(D_HA, "%s: start commit thread committed batchid %llu\n",
tdtd->tdtd_lut->lut_obd->obd_name,
tdtd->tdtd_committed_batchid);
while (({set_current_state(TASK_IDLE);
!kthread_should_stop(); })) {
spin_lock(&tdtd->tdtd_batchid_lock);
list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
tmt_commit_list) {
if (tmt->tmt_committed == 0)
break;
/* Note: right now, replay is based on master MDT
* transno, but cancellation is based on batchid.
* so we do not try to cancel the update log until
* the recoverying is done, unless the update records
* batchid < committed_batchid. */
if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
__set_current_state(TASK_RUNNING);
list_move_tail(&tmt->tmt_commit_list, &list);
} else if (!test_bit(OBDF_RECOVERING, tdtd->tdtd_lut->lut_obd->obd_flags)) {
__set_current_state(TASK_RUNNING);
LASSERTF(tmt->tmt_batchid >= batchid,
"tmt %px tmt_batchid: %llu, batchid %llu\n",
tmt, tmt->tmt_batchid, batchid);
/* There are three types of distribution
* transaction result
*
* 1. If tmt_result < 0, it means the
* distribution transaction fails, which should
* be rare, because once declare phase succeeds,
* the operation should succeeds anyway. Note in
* this case, we will still update batchid so
* cancellation would be stopped.
*
* 2. If tmt_result == 0, it means the
* distribution transaction succeeds, and we
* will update batchid.
*
* 3. If tmt_result > 0, it means distribute
* transaction is not yet committed on every
* node, but we need release this tmt before
* that, which usuually happens during umount.
*/
if (tmt->tmt_result <= 0)
batchid = tmt->tmt_batchid;
list_move_tail(&tmt->tmt_commit_list, &list);
}
}
spin_unlock(&tdtd->tdtd_batchid_lock);
CDEBUG(D_HA, "%s: batchid: %llu committed batchid %llu\n",
tdtd->tdtd_lut->lut_obd->obd_name, batchid,
tdtd->tdtd_committed_batchid);
/* update globally committed on a storage */
if (batchid > tdtd->tdtd_committed_batchid) {
rc = distribute_txn_commit_batchid_update(env, tdtd,
batchid);
if (rc == 0)
batchid = 0;
}
/* cancel the records for committed batchid's */
/* XXX: should we postpone cancel's till the end of recovery? */
committed = tdtd->tdtd_committed_batchid;
list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
if (tmt->tmt_batchid > committed)
break;
__set_current_state(TASK_RUNNING);
list_del_init(&tmt->tmt_commit_list);
if (tmt->tmt_result <= 0)
distribute_txn_cancel_records(env, tmt);
top_multiple_thandle_put(tmt);
}
if (!task_is_running(current))
schedule();
if (CFS_FAIL_PRECHECK(OBD_FAIL_OUT_OBJECT_MISS)) {
set_current_state(TASK_UNINTERRUPTIBLE);
schedule_timeout(cfs_time_seconds(5));
}
}
while (({set_current_state(TASK_IDLE);
atomic_read(&tdtd->tdtd_refcount) != 0; }))
schedule();
__set_current_state(TASK_RUNNING);
spin_lock(&tdtd->tdtd_batchid_lock);
list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
tmt_commit_list)
list_move_tail(&tmt->tmt_commit_list, &list);
spin_unlock(&tdtd->tdtd_batchid_lock);
CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
tdtd->tdtd_lut->lut_obd->obd_name);
list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
list_del_init(&tmt->tmt_commit_list);
top_multiple_thandle_dump(tmt, D_HA);
top_multiple_thandle_put(tmt);
}
RETURN(0);
}
/**
* distribute_txn_init() - Start llog cancel thread
* @env: Lustre environment
* @lut: Lustre target (MDT or OST)
* @tdtd: cancel log thread to be started.
* @index: Used by kthread_create() as unique thread identifier
*
* Start llog cancel(master/slave) thread on LOD
*
* Return:
* * %0 if the thread is started successfully.
* * %negative errno if the thread is not being started.
*/
int distribute_txn_init(const struct lu_env *env,
struct lu_target *lut,
struct target_distribute_txn_data *tdtd,
__u32 index)
{
struct task_struct *task;
int rc;
ENTRY;
INIT_LIST_HEAD(&tdtd->tdtd_list);
INIT_LIST_HEAD(&tdtd->tdtd_replay_finish_list);
INIT_LIST_HEAD(&tdtd->tdtd_replay_list);
spin_lock_init(&tdtd->tdtd_batchid_lock);
spin_lock_init(&tdtd->tdtd_replay_list_lock);
tdtd->tdtd_replay_handler = distribute_txn_replay_handle;
tdtd->tdtd_replay_ready = 0;
tdtd->tdtd_batchid = lut->lut_last_transno + 1;
init_waitqueue_head(&tdtd->tdtd_recovery_threads_waitq);
atomic_set(&tdtd->tdtd_refcount, 0);
atomic_set(&tdtd->tdtd_recovery_threads_count, 0);
tdtd->tdtd_lut = lut;
if (lut->lut_bottom->dd_rdonly)
RETURN(0);
rc = distribute_txn_commit_batchid_init(env, tdtd);
if (rc != 0)
RETURN(rc);
rc = lu_env_init(&tdtd->tdtd_env, LCT_LOCAL | LCT_MD_THREAD);
if (rc)
RETURN(rc);
task = kthread_create(distribute_txn_commit_thread, tdtd, "dist_txn-%u",
index);
if (IS_ERR(task)) {
lu_env_fini(&tdtd->tdtd_env);
RETURN(PTR_ERR(task));
}
tdtd->tdtd_commit_task = task;
wake_up_process(task);
RETURN(0);
}
EXPORT_SYMBOL(distribute_txn_init);
/**
* distribute_txn_fini() - Stop llog cancel thread
* @env: Lustre environment
* @tdtd: cancel log thread to be stopped.
*
* Stop llog cancel(master/slave) thread on LOD and also destory
* all of transaction in the list.
*/
void distribute_txn_fini(const struct lu_env *env,
struct target_distribute_txn_data *tdtd)
{
struct top_multiple_thandle *tmt;
LIST_HEAD(list);
/* Stop cancel thread */
if (!tdtd->tdtd_commit_task)
return;
kthread_stop(tdtd->tdtd_commit_task);
tdtd->tdtd_commit_task = NULL;
spin_lock(&tdtd->tdtd_batchid_lock);
list_splice_init(&tdtd->tdtd_list, &list);
spin_unlock(&tdtd->tdtd_batchid_lock);
CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
tdtd->tdtd_lut->lut_obd->obd_name);
while ((tmt = list_first_entry_or_null(&list,
struct top_multiple_thandle,
tmt_commit_list)) != NULL) {
list_del_init(&tmt->tmt_commit_list);
top_multiple_thandle_dump(tmt, D_HA);
top_multiple_thandle_put(tmt);
}
lu_env_fini(&tdtd->tdtd_env);
dtrq_list_destroy(tdtd);
if (tdtd->tdtd_batchid_obj != NULL) {
dt_object_put(env, tdtd->tdtd_batchid_obj);
tdtd->tdtd_batchid_obj = NULL;
}
}
EXPORT_SYMBOL(distribute_txn_fini);
