* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2012, 2014, Intel Corporation.
+ * Copyright (c) 2012, 2015, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
static int osp_sync_id_traction_init(struct osp_device *d);
static void osp_sync_id_traction_fini(struct osp_device *d);
-static __u32 osp_sync_id_get(struct osp_device *d, __u32 id);
+static __u64 osp_sync_id_get(struct osp_device *d, __u64 id);
static void osp_sync_remove_from_tracker(struct osp_device *d);
/*
struct osp_job_req_args {
/** bytes reserved for ptlrpc_replay_req() */
struct ptlrpc_replay_async_args jra_raa;
- struct list_head jra_link;
+ struct list_head jra_committed_link;
+ struct list_head jra_inflight_link;
__u32 jra_magic;
};
|| (d->opd_syn_prev_done == 0);
}
+static inline int osp_sync_inflight_conflict(struct osp_device *d,
+ struct llog_rec_hdr *h)
+{
+ struct osp_job_req_args *jra;
+ struct ost_id ostid;
+ int conflict = 0;
+
+ if (h == NULL || h->lrh_type == LLOG_GEN_REC ||
+ list_empty(&d->opd_syn_inflight_list))
+ return conflict;
+
+ memset(&ostid, 0, sizeof(ostid));
+ switch (h->lrh_type) {
+ case MDS_UNLINK_REC:
+ ostid_set_seq(&ostid, ((struct llog_unlink_rec *)h)->lur_oseq);
+ ostid_set_id(&ostid, ((struct llog_unlink_rec *)h)->lur_oid);
+ break;
+ case MDS_UNLINK64_REC:
+ fid_to_ostid(&((struct llog_unlink64_rec *)h)->lur_fid, &ostid);
+ break;
+ case MDS_SETATTR64_REC:
+ ostid = ((struct llog_setattr64_rec *)h)->lsr_oi;
+ break;
+ default:
+ LBUG();
+ }
+
+ spin_lock(&d->opd_syn_lock);
+ list_for_each_entry(jra, &d->opd_syn_inflight_list, jra_inflight_link) {
+ struct ptlrpc_request *req;
+ struct ost_body *body;
+
+ LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
+
+ req = container_of((void *)jra, struct ptlrpc_request,
+ rq_async_args);
+ body = req_capsule_client_get(&req->rq_pill,
+ &RMF_OST_BODY);
+ LASSERT(body);
+
+ if (memcmp(&ostid, &body->oa.o_oi, sizeof(ostid)) == 0) {
+ conflict = 1;
+ break;
+ }
+ }
+ spin_unlock(&d->opd_syn_lock);
+
+ return conflict;
+}
+
static inline int osp_sync_low_in_progress(struct osp_device *d)
{
return d->opd_syn_rpc_in_progress < d->opd_syn_max_rpc_in_progress;
osp_sync_check_for_work(d);
}
+static inline __u64 osp_sync_correct_id(struct osp_device *d,
+ struct llog_rec_hdr *rec)
+{
+ /*
+ * llog use cyclic store with 32 bit lrh_id
+ * so overflow lrh_id is possible. Range between
+ * last_processed and last_committed is less than
+ * 64745 ^ 2 and less than 2^32 - 1
+ */
+ __u64 correct_id = d->opd_syn_last_committed_id;
+
+ if ((correct_id & 0xffffffffULL) < rec->lrh_id)
+ correct_id -= 0x100000000ULL;
+
+ correct_id &= ~0xffffffffULL;
+ correct_id |= rec->lrh_id;
+
+ return correct_id;
+}
/**
* Check and return ready-for-new status.
*
if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
return 0;
+ if (unlikely(osp_sync_inflight_conflict(d, rec)))
+ return 0;
if (!osp_sync_low_in_progress(d))
return 0;
if (!osp_sync_low_in_flight(d))
return 1;
if (d->opd_syn_changes == 0)
return 0;
- if (rec == NULL || rec->lrh_id <= d->opd_syn_last_committed_id)
+ if (rec == NULL ||
+ osp_sync_correct_id(d, rec) <= d->opd_syn_last_committed_id)
return 1;
return 0;
}
LASSERT(txn);
txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
- osi->osi_hdr.lrh_id = txn->oti_current_id;
-
+ osi->osi_hdr.lrh_id = (txn->oti_current_id & 0xffffffffULL);
ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
if (ctxt == NULL)
RETURN(-ENOMEM);
jra = ptlrpc_req_async_args(req);
LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
- LASSERT(list_empty(&jra->jra_link));
+ LASSERT(list_empty(&jra->jra_committed_link));
ptlrpc_request_addref(req);
spin_lock(&d->opd_syn_lock);
- list_add(&jra->jra_link, &d->opd_syn_committed_there);
+ list_add(&jra->jra_committed_link, &d->opd_syn_committed_there);
spin_unlock(&d->opd_syn_lock);
/* XXX: some batching wouldn't hurt */
* but object doesn't exist anymore - cancell llog record
*/
LASSERT(req->rq_transno == 0);
- LASSERT(list_empty(&jra->jra_link));
+ LASSERT(list_empty(&jra->jra_committed_link));
ptlrpc_request_addref(req);
spin_lock(&d->opd_syn_lock);
- list_add(&jra->jra_link, &d->opd_syn_committed_there);
+ list_add(&jra->jra_committed_link, &d->opd_syn_committed_there);
spin_unlock(&d->opd_syn_lock);
wake_up(&d->opd_syn_waitq);
LASSERT(d->opd_syn_rpc_in_flight > 0);
spin_lock(&d->opd_syn_lock);
d->opd_syn_rpc_in_flight--;
+ list_del_init(&jra->jra_inflight_link);
spin_unlock(&d->opd_syn_lock);
if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
wake_up(&d->opd_syn_barrier_waitq);
jra = ptlrpc_req_async_args(req);
jra->jra_magic = OSP_JOB_MAGIC;
- INIT_LIST_HEAD(&jra->jra_link);
+ INIT_LIST_HEAD(&jra->jra_committed_link);
+ spin_lock(&d->opd_syn_lock);
+ list_add_tail(&jra->jra_inflight_link, &d->opd_syn_inflight_list);
+ spin_unlock(&d->opd_syn_lock);
- ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
+ ptlrpcd_add_req(req);
}
/* Prepare the request */
imp = d->opd_obd->u.cli.cl_import;
LASSERT(imp);
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_ENOMEM))
+ RETURN(ERR_PTR(-ENOMEM));
+
req = ptlrpc_request_alloc(imp, format);
if (req == NULL)
RETURN(ERR_PTR(-ENOMEM));
* \param[in] h llog record
*
* \retval 0 on success
+ * \retval 1 on invalid record
* \retval negative negated errno on error
*/
static int osp_sync_new_setattr_job(struct osp_device *d,
ENTRY;
LASSERT(h->lrh_type == MDS_SETATTR64_REC);
+ if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_INVALID_REC))
+ RETURN(1);
/* lsr_valid can only be 0 or have OBD_MD_{FLUID,FLGID} set,
* so no bits other than these should be set. */
if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID)) != 0) {
CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
d->opd_obd->obd_name, rec->lsr_valid);
- /* return 0 so that sync thread can continue processing
- * other records. */
- RETURN(0);
+ /* return 1 on invalid record */
+ RETURN(1);
}
req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
body->oa.o_valid |= rec->lsr_valid;
osp_sync_send_new_rpc(d, req);
- RETURN(1);
+ RETURN(0);
}
/**
body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
osp_sync_send_new_rpc(d, req);
- RETURN(1);
+ RETURN(0);
}
/**
* use OUT for OST as well, this will allow batching and better code
* unification.
*
- * \param[in] env LU environment provided by the caller
* \param[in] d OSP device
* \param[in] llh llog handle where the record is stored
* \param[in] h llog record
* \retval 0 on success
* \retval negative negated errno on error
*/
-static int osp_sync_new_unlink64_job(const struct lu_env *env,
- struct osp_device *d,
+static int osp_sync_new_unlink64_job(struct osp_device *d,
struct llog_handle *llh,
struct llog_rec_hdr *h)
{
body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
OBD_MD_FLOBJCOUNT;
osp_sync_send_new_rpc(d, req);
- RETURN(1);
+ RETURN(0);
}
/**
* \param[in] d OSP device
* \param[in] llh llog handle where the record is stored
* \param[in] rec llog record
- *
- * \retval 0 on success
- * \retval negative negated errno on error
*/
-static int osp_sync_process_record(const struct lu_env *env,
- struct osp_device *d,
- struct llog_handle *llh,
- struct llog_rec_hdr *rec)
+static void osp_sync_process_record(const struct lu_env *env,
+ struct osp_device *d,
+ struct llog_handle *llh,
+ struct llog_rec_hdr *rec)
{
+ struct llog_handle *cathandle = llh->u.phd.phd_cat_handle;
struct llog_cookie cookie;
int rc = 0;
+ ENTRY;
+
cookie.lgc_lgl = llh->lgh_id;
cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
cookie.lgc_index = rec->lrh_index;
}
/* cancel any generation record */
- rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
- 1, &cookie);
+ rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
- return rc;
+ RETURN_EXIT;
}
/*
rc = osp_sync_new_unlink_job(d, llh, rec);
break;
case MDS_UNLINK64_REC:
- rc = osp_sync_new_unlink64_job(env, d, llh, rec);
+ rc = osp_sync_new_unlink64_job(d, llh, rec);
break;
case MDS_SETATTR64_REC:
rc = osp_sync_new_setattr_job(d, llh, rec);
default:
CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
rec->lrh_type);
- /* we should continue processing */
+ /* treat "unknown record type" as "invalid" */
+ rc = 1;
+ break;
}
- /* rc > 0 means sync RPC being added to the queue */
- if (likely(rc > 0)) {
- spin_lock(&d->opd_syn_lock);
- if (d->opd_syn_prev_done) {
- LASSERT(d->opd_syn_changes > 0);
- LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
- /*
- * NOTE: it's possible to meet same id if
- * OST stores few stripes of same file
- */
- if (rec->lrh_id > d->opd_syn_last_processed_id) {
- d->opd_syn_last_processed_id = rec->lrh_id;
- wake_up(&d->opd_syn_barrier_waitq);
- }
+ spin_lock(&d->opd_syn_lock);
- d->opd_syn_changes--;
+ /* For all kinds of records, not matter successful or not,
+ * we should decrease changes and bump last_processed_id.
+ */
+ if (d->opd_syn_prev_done) {
+ __u64 correct_id = osp_sync_correct_id(d, rec);
+ LASSERT(d->opd_syn_changes > 0);
+ LASSERT(correct_id <= d->opd_syn_last_committed_id);
+ /* NOTE: it's possible to meet same id if
+ * OST stores few stripes of same file
+ */
+ if (correct_id > d->opd_syn_last_processed_id) {
+ d->opd_syn_last_processed_id = correct_id;
+ wake_up(&d->opd_syn_barrier_waitq);
}
- CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
- d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
- d->opd_syn_rpc_in_progress);
- spin_unlock(&d->opd_syn_lock);
- rc = 0;
- } else {
- spin_lock(&d->opd_syn_lock);
+ d->opd_syn_changes--;
+ }
+ if (rc != 0) {
d->opd_syn_rpc_in_flight--;
d->opd_syn_rpc_in_progress--;
- spin_unlock(&d->opd_syn_lock);
}
+ CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
+ d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
+ d->opd_syn_rpc_in_progress);
- CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
- rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
- return rc;
+ spin_unlock(&d->opd_syn_lock);
+
+ /* Delete the invalid record */
+ if (rc == 1) {
+ rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
+ if (rc != 0)
+ CERROR("%s: can't delete invalid record: "
+ "fid = "DFID", rec_id = %u, rc = %d\n",
+ d->opd_obd->obd_name,
+ PFID(lu_object_fid(&cathandle->lgh_obj->do_lu)),
+ rec->lrh_id, rc);
+ }
+
+ CDEBUG(D_HA, "found record %x, %d, idx %u, id %u\n",
+ rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id);
+
+ RETURN_EXIT;
}
/**
spin_unlock(&d->opd_syn_lock);
while (!list_empty(&list)) {
- struct llog_cookie *lcookie = NULL;
struct osp_job_req_args *jra;
- jra = list_entry(list.next, struct osp_job_req_args, jra_link);
+ jra = list_entry(list.next, struct osp_job_req_args,
+ jra_committed_link);
LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
- list_del_init(&jra->jra_link);
+ list_del_init(&jra->jra_committed_link);
req = container_of((void *)jra, struct ptlrpc_request,
rq_async_args);
body = req_capsule_client_get(&req->rq_pill,
&RMF_OST_BODY);
LASSERT(body);
- lcookie = &body->oa.o_lcookie;
/* import can be closing, thus all commit cb's are
* called we can check committness directly */
- if (req->rq_transno <= imp->imp_peer_committed_transno) {
- rc = llog_cat_cancel_records(env, llh, 1, lcookie);
+ if (req->rq_import_generation == imp->imp_generation) {
+ rc = llog_cat_cancel_records(env, llh, 1,
+ &body->oa.o_lcookie);
if (rc)
CERROR("%s: can't cancel record: %d\n",
obd->obd_name, rc);
} else {
- DEBUG_REQ(D_HA, req, "not committed");
+ DEBUG_REQ(D_OTHER, req, "imp_committed = "LPU64,
+ imp->imp_peer_committed_transno);
}
-
ptlrpc_req_finished(req);
done++;
}
void *data)
{
struct osp_device *d = data;
- int rc;
do {
struct l_wait_info lwi = { 0 };
d->opd_syn_rpc_in_flight);
return 0;
}
-
- /*
- * try to send, in case of disconnection, suspend
- * processing till we can send this request
- */
- do {
- rc = osp_sync_process_record(env, d, llh, rec);
- /*
- * XXX: probably different handling is needed
- * for some bugs, like immediate exit or if
- * OSP gets inactive
- */
- if (rc) {
- CERROR("can't send: %d\n", rc);
- l_wait_event(d->opd_syn_waitq,
- !osp_sync_running(d) ||
- osp_sync_has_work(d),
- &lwi);
- }
- } while (rc != 0 && osp_sync_running(d));
-
+ osp_sync_process_record(env, d, llh, rec);
llh = NULL;
rec = NULL;
}
LASSERT(lgh != NULL);
ctxt->loc_handle = lgh;
- rc = llog_cat_init_and_process(env, lgh);
+ rc = llog_init_handle(env, lgh, LLOG_F_IS_CAT, NULL);
if (rc)
GOTO(out_close, rc);
init_waitqueue_head(&d->opd_syn_waitq);
init_waitqueue_head(&d->opd_syn_barrier_waitq);
init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
+ INIT_LIST_HEAD(&d->opd_syn_inflight_list);
INIT_LIST_HEAD(&d->opd_syn_committed_there);
task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
spin_lock(&tr->otr_lock);
if (likely(txn->oti_current_id > tr->otr_committed_id)) {
- CDEBUG(D_OTHER, "committed: %u -> %u\n",
+ CDEBUG(D_OTHER, "committed: "LPU64" -> "LPU64"\n",
tr->otr_committed_id, txn->oti_current_id);
tr->otr_committed_id = txn->oti_current_id;
* Generates a new ID using the tracker associated with the given OSP device
* \a d, if the given ID \a id is non-zero. Unconditially adds OSP device to
* the wakeup list, so OSP won't miss when a transaction using the ID is
- * committed. Notice ID is 32bit, but llog doesn't support >2^32 records anyway.
+ * committed.
*
* \param[in] d OSP device
* \param[in] id 0 or ID generated previously
*
* \retval ID the caller should use
*/
-static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
+static __u64 osp_sync_id_get(struct osp_device *d, __u64 id)
{
struct osp_id_tracker *tr;
/* XXX: we can improve this introducing per-cpu preallocated ids? */
spin_lock(&tr->otr_lock);
+ if (OBD_FAIL_CHECK(OBD_FAIL_MDS_TRACK_OVERFLOW))
+ tr->otr_next_id = 0xfffffff0;
+
if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
spin_unlock(&tr->otr_lock);
- CERROR("%s: next %u, last synced %lu\n",
+ CERROR("%s: next "LPU64", last synced "LPU64"\n",
d->opd_obd->obd_name, tr->otr_next_id,
d->opd_syn_last_used_id);
LBUG();
if (list_empty(&d->opd_syn_ontrack))
list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
spin_unlock(&tr->otr_lock);
- CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
+ CDEBUG(D_OTHER, "new id "LPU64"\n", id);
return id;
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff