* GPL HEADER END
*/
/*
- * Copyright (c) 2014, 2016, Intel Corporation.
+ * Copyright (c) 2014, 2017, Intel Corporation.
*/
/*
* lustre/osp/osp_trans.c
osp_update_interpreter_t ouc_interpreter;
};
-static struct object_update_request *object_update_request_alloc(size_t size)
-{
- struct object_update_request *ourq;
-
- OBD_ALLOC_LARGE(ourq, size);
- if (ourq == NULL)
- return ERR_PTR(-ENOMEM);
-
- ourq->ourq_magic = UPDATE_REQUEST_MAGIC;
- ourq->ourq_count = 0;
-
- return ourq;
-}
-
/**
* Allocate new update request
*
size_t size)
{
struct osp_update_request_sub *ours;
+ struct object_update_request *ourq;
OBD_ALLOC_PTR(ours);
if (ours == NULL)
return -ENOMEM;
- if (size < OUT_UPDATE_INIT_BUFFER_SIZE)
- size = OUT_UPDATE_INIT_BUFFER_SIZE;
-
- ours->ours_req = object_update_request_alloc(size);
-
- if (IS_ERR(ours->ours_req)) {
+ /* The object update request will be added to an SG list for
+ * bulk transfer. Some IB HW cannot handle partial pages in SG
+ * lists (since they create gaps in memory regions) so we
+ * round the size up to the next multiple of PAGE_SIZE. See
+ * LU-9983. */
+ LASSERT(size > 0);
+ size = round_up(size, PAGE_SIZE);
+ OBD_ALLOC_LARGE(ourq, size);
+ if (ourq == NULL) {
OBD_FREE_PTR(ours);
return -ENOMEM;
}
+ ourq->ourq_magic = UPDATE_REQUEST_MAGIC;
+ ourq->ourq_count = 0;
+ ours->ours_req = ourq;
ours->ours_req_size = size;
INIT_LIST_HEAD(&ours->ours_list);
list_add_tail(&ours->ours_list, &our->our_req_list);
INIT_LIST_HEAD(&our->our_invalidate_cb_list);
spin_lock_init(&our->our_list_lock);
- rc = osp_object_update_request_create(our, OUT_UPDATE_INIT_BUFFER_SIZE);
+ rc = osp_object_update_request_create(our, PAGE_SIZE);
if (rc != 0) {
OBD_FREE_PTR(our);
return ERR_PTR(rc);
list_del_init(&obj->opo_invalidate_cb_list);
spin_unlock(&obj->opo_lock);
- lu_object_put(env, &obj->opo_obj.do_lu);
+ dt_object_put(env, &obj->opo_obj);
}
if (env == &lenv)
buf_count++;
}
repsize += sizeof(*reply);
- repsize = (repsize + OUT_UPDATE_REPLY_SIZE - 1) &
- ~(OUT_UPDATE_REPLY_SIZE - 1);
+ if (repsize < OUT_UPDATE_REPLY_SIZE)
+ repsize = OUT_UPDATE_REPLY_SIZE;
LASSERT(buf_count > 0);
req = ptlrpc_request_alloc(imp, &RQF_OUT_UPDATE);
list_del_init(&obj->opo_invalidate_cb_list);
spin_unlock(&obj->opo_lock);
- lu_object_put(env, &obj->opo_obj.do_lu);
+ dt_object_put(env, &obj->opo_obj);
}
}
oth->ot_magic = OSP_THANDLE_MAGIC;
th = &oth->ot_super;
th->th_dev = d;
- th->th_tags = LCT_TX_HANDLE;
atomic_set(&oth->ot_refcount, 1);
INIT_LIST_HEAD(&oth->ot_commit_dcb_list);
if (!osp->opd_connect_mdt) {
down_read(&osp->opd_async_updates_rwsem);
args->oaua_count = &osp->opd_async_updates_count;
- args->oaua_waitq = &osp->opd_syn_barrier_waitq;
+ args->oaua_waitq = &osp->opd_sync_barrier_waitq;
up_read(&osp->opd_async_updates_rwsem);
atomic_inc(args->oaua_count);
}
rc = ptlrpc_queue_wait(req);
if (osp->opd_connect_mdt)
osp_put_rpc_lock(osp);
- if ((rc == -ENOMEM && req->rq_set == NULL) ||
+
+ /* We use rq_queued_time to distinguish between local
+ * and remote -ENOMEM. */
+ if ((rc == -ENOMEM && req->rq_queued_time == 0) ||
(req->rq_transno == 0 && !req->rq_committed)) {
if (args->oaua_update != NULL) {
/* If osp_update_interpret is not being called,
* Get local thandle for osp_thandle
*
* Get the local OSD thandle from the OSP thandle. Currently, there
- * are a few OSP API (osp_object_create() and osp_sync_add()) needs
+ * are a few OSP API (osp_create() and osp_sync_add()) needs
* to update the object on local OSD device.
*
* If the osp_thandle comes from normal stack (MDD->LOD->OSP), then
*
* Set the version for the transaction and add the request to
* the sending list, then after transaction stop, the request
- * will be picked in the order of version, by sending thread.
+ * will be sent in the order of version by the sending thread.
*
* \param [in] oth osp thandle to be set version.
*
/* Assign the version and add it to the sending list */
osp_thandle_get(oth);
oth->ot_our->our_version = ou->ou_version++;
+ oth->ot_our->our_generation = ou->ou_generation;
list_add_tail(&oth->ot_our->our_list,
&osp->opd_update->ou_list);
oth->ot_our->our_req_ready = 0;
spin_unlock(&ou->ou_lock);
LASSERT(oth->ot_super.th_wait_submit == 1);
- CDEBUG(D_INFO, "%s: version %llu oth:version %p:%llu\n",
- osp->opd_obd->obd_name, ou->ou_version, oth,
+ CDEBUG(D_INFO, "%s: version %llu gen %llu oth:version %p:%llu\n",
+ osp->opd_obd->obd_name, ou->ou_version, ou->ou_generation, oth,
oth->ot_our->our_version);
return 0;
if (rc < 0) {
CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
rc);
+
+ spin_lock(&ou->ou_lock);
+ ou->ou_generation++;
+ spin_unlock(&ou->ou_lock);
+
return;
}
our);
}
+ /* Increase the generation, then the update request with old generation
+ * will fail with -EIO. */
+ ou->ou_generation++;
spin_unlock(&ou->ou_lock);
/* invalidate all of request in the sending list */
osp_trans_callback(&env, our->our_th,
our->our_th->ot_super.th_result);
rc = our->our_th->ot_super.th_result;
- } else if (OBD_FAIL_CHECK(OBD_FAIL_INVALIDATE_UPDATE)) {
+ } else if (ou->ou_generation != our->our_generation ||
+ OBD_FAIL_CHECK(OBD_FAIL_INVALIDATE_UPDATE)) {
rc = -EIO;
osp_trans_callback(&env, our->our_th, rc);
} else {