struct obd_device *obd = ctxt->loc_obd;
struct obd_export *exp = obd->obd_self_export;
struct llog_cookie cookie;
+ struct llog_gen_rec *lgr;
struct llog_unlink_rec *lur;
struct obdo *oa;
- struct obd_trans_info oti = { 0 };
obd_id oid;
int rc = 0;
ENTRY;
- if (!le32_to_cpu(llh->lgh_hdr->llh_flags) & LLOG_F_IS_PLAIN) {
+ if (!(le32_to_cpu(llh->lgh_hdr->llh_flags) & LLOG_F_IS_PLAIN)) {
CERROR("log is not plain\n");
RETURN(-EINVAL);
}
- if (rec->lrh_type != MDS_UNLINK_REC) {
+ if (rec->lrh_type != MDS_UNLINK_REC &&
+ rec->lrh_type != LLOG_GEN_REC) {
CERROR("log record type error\n");
RETURN(-EINVAL);
}
cookie.lgc_lgl = llh->lgh_id;
cookie.lgc_subsys = LLOG_UNLINK_ORIG_CTXT;
cookie.lgc_index = le32_to_cpu(rec->lrh_index);
-
+
+ if (rec->lrh_type == LLOG_GEN_REC) {
+ lgr = (struct llog_gen_rec *)rec;
+ if (llog_gen_lt(lgr->lgr_gen, ctxt->loc_gen))
+ rc = 0;
+ else
+ rc = LLOG_PROC_BREAK;
+ CWARN("fetch generation log, send cookie\n");
+ llog_cancel(ctxt, NULL, 1, &cookie, 0);
+ RETURN(rc);
+ }
+
lur = (struct llog_unlink_rec *)rec;
-
oa = obdo_alloc();
if (oa == NULL)
RETURN(-ENOMEM);
memcpy(obdo_logcookie(oa), &cookie, sizeof(cookie));
oid = oa->o_id;
- rc = obd_destroy(exp, oa, NULL, &oti);
+ rc = obd_destroy(exp, oa, NULL, NULL);
obdo_free(oa);
if (rc == -ENOENT) {
- CWARN("object already removed: send cookie\n");
+ CDEBUG(D_HA, "object already removed, send cookie\n");
llog_cancel(ctxt, NULL, 1, &cookie, 0);
RETURN(0);
}
if (rc == 0)
- CWARN("object: "LPU64" in record is destroyed\n", oid);
+ CDEBUG(D_HA, "object: "LPU64" in record is destroyed\n", oid);
RETURN(rc);
}
#include <linux/lustre_log.h>
#include "ptlrpc_internal.h"
+#define LLCD_SIZE 4096
+
#ifdef __KERNEL__
static struct llog_commit_master lustre_lcm;
struct llog_canceld_ctxt *llcd;
int offset = offsetof(struct llog_canceld_ctxt, llcd_cookies);
- OBD_ALLOC(llcd, PAGE_SIZE + offset);
+ OBD_ALLOC(llcd, LLCD_SIZE + offset);
if (llcd == NULL)
return -ENOMEM;
int offset = offsetof(struct llog_canceld_ctxt, llcd_cookies);
if (atomic_read(&lcm->lcm_llcd_numfree) >= lcm->lcm_llcd_maxfree) {
- OBD_FREE(llcd, PAGE_SIZE + offset);
+ OBD_FREE(llcd, LLCD_SIZE + offset);
} else {
spin_lock(&lcm->lcm_llcd_lock);
list_add(&llcd->llcd_list, &lcm->lcm_llcd_free);
EXPORT_SYMBOL(llcd_send);
/* deleted objects have a commit callback that cancels the MDS
- * log record for the deletion. The commit callback calls this
- * function
+ * log record for the deletion. The commit callback calls this
+ * function
*/
int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
struct lov_stripe_md *lsm, int count,
cookies->lgc_lgl.lgl_ogen, cookies->lgc_index);
GOTO(out, rc = -ENOMEM);
}
- llcd->llcd_import = ctxt->loc_imp;
- llcd->llcd_gen = ctxt->loc_gen;
+ llcd->llcd_ctxt = ctxt;
ctxt->loc_llcd = llcd;
}
llcd->llcd_cookiebytes += sizeof(*cookies);
send_now:
- if ((PAGE_SIZE - llcd->llcd_cookiebytes < sizeof(*cookies) ||
+ if ((LLCD_SIZE - llcd->llcd_cookiebytes < sizeof(*cookies) ||
flags & OBD_LLOG_FL_SENDNOW)) {
- CDEBUG(D_HA, "send llcd: %p\n", llcd);
+ CDEBUG(D_HA, "send llcd %p:%p\n", llcd, llcd->llcd_ctxt);
ctxt->loc_llcd = NULL;
llcd_send(llcd);
}
int rc = 0;
ENTRY;
- LASSERT(ctxt->loc_llcd);
-
if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
- CWARN("import will be destroyed, put llcd %p\n",
- ctxt->loc_llcd);
- llcd_put(ctxt->loc_llcd);
- ctxt->loc_llcd = NULL;
+ down(&ctxt->loc_sem);
+ if (ctxt->loc_llcd != NULL) {
+ CWARN("import will be destroyed, put "
+ "llcd %p:%p\n", ctxt->loc_llcd, ctxt);
+ llcd_put(ctxt->loc_llcd);
+ ctxt->loc_llcd = NULL;
+ ctxt->loc_imp = NULL;
+ }
up(&ctxt->loc_sem);
} else {
- up(&ctxt->loc_sem);
rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW);
}
llcd = list_entry(lcd->lcd_llcd_list.next,
typeof(*llcd), llcd_list);
LASSERT(llcd->llcd_lcm == lcm);
- import = llcd->llcd_import;
+ import = llcd->llcd_ctxt->loc_imp;
}
list_for_each_entry_safe(llcd, n, sending_list, llcd_list) {
LASSERT(llcd->llcd_lcm == lcm);
- if (import == llcd->llcd_import)
+ if (import == llcd->llcd_ctxt->loc_imp)
list_move_tail(&llcd->llcd_list,
&lcd->lcd_llcd_list);
}
list_for_each_entry_safe(llcd, n, &lcm->lcm_llcd_resend,
llcd_list) {
LASSERT(llcd->llcd_lcm == lcm);
- if (import == llcd->llcd_import)
+ if (import == llcd->llcd_ctxt->loc_imp)
list_move_tail(&llcd->llcd_list,
&lcd->lcd_llcd_list);
}
/* We are the only one manipulating our local list - no lock */
list_for_each_entry_safe(llcd,n, &lcd->lcd_llcd_list,llcd_list){
char *bufs[1] = {(char *)llcd->llcd_cookies};
- struct obd_device *obd = import->imp_obd;
- struct llog_ctxt *ctxt;
list_del(&llcd->llcd_list);
if (llcd->llcd_cookiebytes == 0) {
- CDEBUG(D_HA, "just put empty llcd %p\n", llcd);
+ CDEBUG(D_HA, "put empty llcd %p:%p\n",
+ llcd, llcd->llcd_ctxt);
llcd_put(llcd);
continue;
}
- /* check whether the cookies are new. if new then send, otherwise
- * just put llcd */
- ctxt = llog_get_context(obd, llcd->llcd_cookies[0].lgc_subsys + 1);
- LASSERT(ctxt != NULL);
- down(&ctxt->loc_sem);
- if (log_gen_lt(llcd->llcd_gen, ctxt->loc_gen)) {
- up(&ctxt->loc_sem);
- CDEBUG(D_HA, "just put stale llcd %p\n", llcd);
+
+ down(&llcd->llcd_ctxt->loc_sem);
+ if (llcd->llcd_ctxt->loc_imp == NULL) {
+ up(&llcd->llcd_ctxt->loc_sem);
+ CWARN("import will be destroyed, put "
+ "llcd %p:%p\n", llcd, llcd->llcd_ctxt);
llcd_put(llcd);
continue;
}
- up(&ctxt->loc_sem);
request = ptlrpc_prep_req(import, OBD_LOG_CANCEL, 1,
&llcd->llcd_cookiebytes,
bufs);
+ up(&llcd->llcd_ctxt->loc_sem);
+
if (request == NULL) {
rc = -ENOMEM;
CERROR("error preparing commit: rc %d\n", rc);
}
request->rq_replen = lustre_msg_size(0, NULL);
+ down(&llcd->llcd_ctxt->loc_sem);
+ if (llcd->llcd_ctxt->loc_imp == NULL) {
+ up(&llcd->llcd_ctxt->loc_sem);
+ CWARN("import will be destroyed, put "
+ "llcd %p:%p\n", llcd, llcd->llcd_ctxt);
+ llcd_put(llcd);
+ ptlrpc_req_finished(request);
+ continue;
+ }
rc = ptlrpc_queue_wait(request);
ptlrpc_req_finished(request);
+ up(&llcd->llcd_ctxt->loc_sem);
/* If the RPC failed, we put this and the remaining
* messages onto the resend list for another time. */
continue;
}
-#if 0 /* FIXME just put llcd, not send it again */
+#if 0 /* FIXME just put llcd, not put it on resend list */
spin_lock(&lcm->lcm_llcd_lock);
list_splice(&lcd->lcd_llcd_list, &lcm->lcm_llcd_resend);
if (++llcd->llcd_tries < 5) {
} else {
spin_unlock(&lcm->lcm_llcd_lock);
#endif
- CERROR("commit %p dropped %d cookies: rc %d\n",
- llcd, (int)(llcd->llcd_cookiebytes /
- sizeof(*llcd->llcd_cookies)),
- rc);
+ CERROR("commit %p:%p drop %d cookies: rc %d\n",
+ llcd, llcd->llcd_ctxt,
+ (int)(llcd->llcd_cookiebytes /
+ sizeof(*llcd->llcd_cookies)), rc);
llcd_put(llcd);
-// }
+#if 0
+ }
break;
+#endif
}
if (rc == 0) {
EXPORT_SYMBOL(llog_start_commit_thread);
static struct llog_process_args {
- struct semaphore llpa_sem;
+ struct semaphore llpa_sem;
struct llog_ctxt *llpa_ctxt;
void *llpa_cb;
void *llpa_arg;
} llpa;
+
int llog_init_commit_master(void)
{
INIT_LIST_HEAD(&lcm->lcm_thread_busy);
return 0;
}
-
static int log_process_thread(void *args)
{
struct llog_process_args *data = args;
unsigned long flags;
int rc;
ENTRY;
-
+
up(&data->llpa_sem);
lock_kernel();
ptlrpc_daemonize(); /* thread never needs to do IO */
-
+
SIGNAL_MASK_LOCK(current, flags);
sigfillset(¤t->blocked);
RECALC_SIGPENDING;
SIGNAL_MASK_UNLOCK(current, flags);
unlock_kernel();
-
+
rc = llog_create(ctxt, &llh, &logid, NULL);
if (rc) {
CERROR("llog_create failed %d\n", rc);
CERROR("llog_init_handle failed %d\n", rc);
GOTO(out, rc);
}
-
+
if (cb) {
rc = llog_cat_process(llh, (llog_cb_t)cb, NULL);
- if (rc)
+ if (rc != LLOG_PROC_BREAK)
CERROR("llog_cat_process failed %d\n", rc);
- } else
+ } else {
CWARN("no callback function for recovery\n");
+ }
- CDEBUG(D_HA, "send to llcd :%p forcibly\n", ctxt->loc_llcd);
+ CDEBUG(D_HA, "send llcd %p:%p forcibly after recovery\n",
+ ctxt->loc_llcd, ctxt);
llog_sync(ctxt, NULL);
out:
rc = llog_cat_put(llh);
if (rc)
CERROR("llog_cat_put failed %d\n", rc);
-
+
RETURN(rc);
}
-static int llog_recovery_generic(struct llog_ctxt *ctxt,
- void *handle,
- void *arg)
+
+static int llog_recovery_generic(struct llog_ctxt *ctxt, void *handle,void *arg)
{
int rc;
ENTRY;
RETURN(rc);
}
+
int llog_repl_connect(struct llog_ctxt *ctxt, int count,
- struct llog_logid *logid, struct llog_ctxt_gen *gen)
+ struct llog_logid *logid, struct llog_gen *gen)
{
struct llog_canceld_ctxt *llcd;
int rc;
ENTRY;
-
+
+ /* send back llcd before recovery from llog */
+ if (ctxt->loc_llcd != NULL) {
+ CWARN("llcd %p:%p not empty\n", ctxt->loc_llcd, ctxt);
+ llog_sync(ctxt, NULL);
+ }
+
down(&ctxt->loc_sem);
ctxt->loc_gen = *gen;
- llcd = ctxt->loc_llcd;
- if (llcd) {
- CDEBUG(D_HA, "put current llcd when new connection arrives\n");
- llcd_put(llcd);
- }
llcd = llcd_grab();
if (llcd == NULL) {
CERROR("couldn't get an llcd\n");
+ up(&ctxt->loc_sem);
RETURN(-ENOMEM);
}
- llcd->llcd_import = ctxt->loc_imp;
- llcd->llcd_gen = ctxt->loc_gen;
+ llcd->llcd_ctxt = ctxt;
ctxt->loc_llcd = llcd;
up(&ctxt->loc_sem);
- rc = llog_recovery_generic(ctxt, ctxt->llog_proc_cb, logid);
+ rc = llog_recovery_generic(ctxt, ctxt->llog_proc_cb, logid);
if (rc != 0)
CERROR("error recovery process: %d\n", rc);
. ${CONFIG:=$LUSTRE/tests/cfg/local.sh}
# Skip these tests
-ALWAYS_EXCEPT="35"
+ALWAYS_EXCEPT=""
gen_config() {
test_35() {
touch $DIR/$tfile
- echo 0x80000119 > /proc/sys/lustre/fail_loc
+#define OBD_FAIL_MDS_REINT_NET_REP 0x119
+ do_facet mds "sysctl -w lustre.fail_loc=0x80000119"
rm -f $DIR/$tfile &
sleep 1
+ sync
+ sleep 1
# give a chance to remove from MDS
fail_abort mds
$CHECKSTAT -t file $DIR/$tfile && return 1 || true
}
run_test 37 "abort recovery before client does replay (test mds_cleanup_orphans for directories)"
+test_38() {
+ createmany -o $DIR/$tfile-%d 800
+ unlinkmany $DIR/$tfile-%d 0 400
+ replay_barrier mds
+ fail mds
+ unlinkmany $DIR/$tfile-%d 400 400
+ sleep 2
+ $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
+}
+run_test 38 "test recovery from unlink llog (test llog_gen_rec) "
+
+test_39() {
+ createmany -o $DIR/$tfile-%d 800
+ replay_barrier mds
+ unlinkmany $DIR/$tfile-%d 0 400
+ fail mds
+ unlinkmany $DIR/$tfile-%d 400 400
+ sleep 2
+ $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
+}
+run_test 39 "test recovery from unlink llog (test llog_gen_rec) "
+
+count_ost_writes() {
+ cat /proc/fs/lustre/osc/*/stats |
+ awk -vwrites=0 '/ost_write/ { writes += $2 } END { print writes; }'
+}
+
+#b=2477,2532
+test_40(){
+ $LCTL mark multiop $MOUNT/$tfile OS_c
+ multiop $MOUNT/$tfile OS_c &
+ PID=$!
+ writeme -s $MOUNT/${tfile}-2 &
+ WRITE_PID=$!
+ sleep 1
+ facet_failover mds
+#define OBD_FAIL_MDS_CONNECT_NET 0x117
+ do_facet mds "sysctl -w lustre.fail_loc=0x80000117"
+ kill -USR1 $PID
+ stat1=`count_ost_writes`
+ sleep $TIMEOUT
+ stat2=`count_ost_writes`
+ echo "$stat1, $stat2"
+ if [ $stat1 -lt $stat2 ]; then
+ echo "writes continuing during recovery"
+ RC=0
+ else
+ echo "writes not continuing during recovery, bug 2477"
+ RC=4
+ fi
+ echo "waiting for writeme $WRITE_PID"
+ kill $WRITE_PID
+ wait $WRITE_PID
+
+ echo "waiting for multiop $PID"
+ wait $PID || return 2
+ do_facet client munlink $MOUNT/$tfile || return 3
+ do_facet client munlink $MOUNT/${tfile}-2 || return 3
+ return $RC
+}
+run_test 40 "cause recovery in ptlrpc, ensure IO continues"
+
equals_msg test complete, cleaning up
$CLEANUP
+
git://git.whamcloud.com / fs / lustre-release.git / commitdiff