static inline int out_check_resent(const struct lu_env *env,
struct dt_device *dt,
struct dt_object *obj,
struct ptlrpc_request *req,
out_reconstruct_t reconstruct,
struct object_update_reply *reply,
int index)
{
if (likely(!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT)))
return 0;
if (req_xid_is_last(req)) {
struct lsd_client_data *lcd;
/* XXX this does not support mulitple transactions yet, i.e.
* only 1 update RPC each time betwee MDTs */
lcd = req->rq_export->exp_target_data.ted_lcd;
req->rq_transno = lcd->lcd_last_transno;
req->rq_status = lcd->lcd_last_result;
if (req->rq_status != 0)
req->rq_transno = 0;
lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
lustre_msg_set_status(req->rq_repmsg, req->rq_status);
DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"status %d",
req->rq_transno, req->rq_status);
reconstruct(env, dt, obj, reply, index);
return 1;
}
DEBUG_REQ(D_HA, req, "no reply for RESENT req (have "LPD64")",
req->rq_export->exp_target_data.ted_lcd->lcd_last_xid);
return 0;
}
static int seq_handler(struct tgt_session_info *tsi)
{
struct lu_seq_range *out, *tmp;
struct lu_site *site;
int rc;
__u32 *opc;
ENTRY;
LASSERT(!(lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) & MSG_REPLAY));
site = tsi->tsi_exp->exp_obd->obd_lu_dev->ld_site;
LASSERT(site != NULL);
opc = req_capsule_client_get(tsi->tsi_pill, &RMF_SEQ_OPC);
if (opc != NULL) {
out = req_capsule_server_get(tsi->tsi_pill, &RMF_SEQ_RANGE);
if (out == NULL)
RETURN(err_serious(-EPROTO));
tmp = req_capsule_client_get(tsi->tsi_pill, &RMF_SEQ_RANGE);
/* seq client passed mdt id, we need to pass that using out
* range parameter */
out->lsr_index = tmp->lsr_index;
out->lsr_flags = tmp->lsr_flags;
rc = seq_server_handle(site, tsi->tsi_env, *opc, out);
} else {
rc = err_serious(-EPROTO);
}
RETURN(rc);
}
static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
{
struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
struct ptlrpc_service *svc = svcpt->scp_service;
int service_time = max_t(int, cfs_time_current_sec() -
req->rq_arrival_time.tv_sec, 1);
if (!(flags & PTLRPC_REPLY_EARLY) &&
(req->rq_type != PTL_RPC_MSG_ERR) &&
(req->rq_reqmsg != NULL) &&
!(lustre_msg_get_flags(req->rq_reqmsg) &
(MSG_RESENT | MSG_REPLAY |
MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
/* early replies, errors and recovery requests don't count
* toward our service time estimate */
int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
if (oldse != 0) {
DEBUG_REQ(D_ADAPTTO, req,
"svc %s changed estimate from %d to %d",
svc->srv_name, oldse,
at_get(&svcpt->scp_at_estimate));
}
}
/* Report actual service time for client latency calc */
lustre_msg_set_service_time(req->rq_repmsg, service_time);
/* Report service time estimate for future client reqs, but report 0
* (to be ignored by client) if it's a error reply during recovery.
* (bz15815) */
if (req->rq_type == PTL_RPC_MSG_ERR &&
(req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
lustre_msg_set_timeout(req->rq_repmsg, 0);
else
lustre_msg_set_timeout(req->rq_repmsg,
at_get(&svcpt->scp_at_estimate));
if (req->rq_reqmsg &&
!(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
"req_flags=%#x magic=%d:%x/%x len=%d\n",
flags, lustre_msg_get_flags(req->rq_reqmsg),
lustre_msg_is_v1(req->rq_reqmsg),
lustre_msg_get_magic(req->rq_reqmsg),
lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
}
}
static inline int out_check_resent(const struct lu_env *env,
struct dt_device *dt,
struct dt_object *obj,
struct ptlrpc_request *req,
out_reconstruct_t reconstruct,
struct update_reply *reply,
int index)
{
if (likely(!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT)))
return 0;
if (req_xid_is_last(req)) {
reconstruct(env, dt, obj, reply, index);
return 1;
}
DEBUG_REQ(D_HA, req, "no reply for RESENT req (have "LPD64")",
req->rq_export->exp_target_data.ted_lcd->lcd_last_xid);
return 0;
}
static int seq_req_handle(struct ptlrpc_request *req,
const struct lu_env *env,
struct seq_thread_info *info)
{
struct lu_seq_range *out, *tmp;
struct lu_site *site;
int rc = -EPROTO;
__u32 *opc;
ENTRY;
LASSERT(!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY));
site = req->rq_export->exp_obd->obd_lu_dev->ld_site;
LASSERT(site != NULL);
rc = req_capsule_server_pack(info->sti_pill);
if (rc)
RETURN(err_serious(rc));
opc = req_capsule_client_get(info->sti_pill, &RMF_SEQ_OPC);
if (opc != NULL) {
out = req_capsule_server_get(info->sti_pill, &RMF_SEQ_RANGE);
if (out == NULL)
RETURN(err_serious(-EPROTO));
tmp = req_capsule_client_get(info->sti_pill, &RMF_SEQ_RANGE);
/* seq client passed mdt id, we need to pass that using out
* range parameter */
out->lsr_index = tmp->lsr_index;
out->lsr_flags = tmp->lsr_flags;
rc = seq_server_handle(site, env, *opc, out);
} else
rc = err_serious(-EPROTO);
RETURN(rc);
}
/**
* Send request \a request.
* if \a noreply is set, don't expect any reply back and don't set up
* reply buffers.
* Returns 0 on success or error code.
*/
int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
{
int rc;
int rc2;
int mpflag = 0;
struct ptlrpc_connection *connection;
lnet_handle_me_t reply_me_h;
lnet_md_t reply_md;
struct obd_device *obd = request->rq_import->imp_obd;
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
return 0;
LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
LASSERT(request->rq_wait_ctx == 0);
/* If this is a re-transmit, we're required to have disengaged
* cleanly from the previous attempt */
LASSERT(!request->rq_receiving_reply);
LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
(request->rq_import->imp_state == LUSTRE_IMP_FULL)));
if (unlikely(obd != NULL && obd->obd_fail)) {
CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
obd->obd_name);
/* this prevents us from waiting in ptlrpc_queue_wait */
spin_lock(&request->rq_lock);
request->rq_err = 1;
spin_unlock(&request->rq_lock);
request->rq_status = -ENODEV;
return -ENODEV;
}
connection = request->rq_import->imp_connection;
lustre_msg_set_handle(request->rq_reqmsg,
&request->rq_import->imp_remote_handle);
lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
lustre_msg_set_conn_cnt(request->rq_reqmsg,
request->rq_import->imp_conn_cnt);
lustre_msghdr_set_flags(request->rq_reqmsg,
request->rq_import->imp_msghdr_flags);
if (request->rq_resend)
lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
if (request->rq_memalloc)
mpflag = cfs_memory_pressure_get_and_set();
rc = sptlrpc_cli_wrap_request(request);
if (rc)
goto out;
/* bulk register should be done after wrap_request() */
if (request->rq_bulk != NULL) {
rc = ptlrpc_register_bulk(request);
if (rc != 0)
goto out;
}
if (!noreply) {
LASSERT(request->rq_replen != 0);
if (request->rq_repbuf == NULL) {
LASSERT(request->rq_repdata == NULL);
LASSERT(request->rq_repmsg == NULL);
rc = sptlrpc_cli_alloc_repbuf(request,
request->rq_replen);
if (rc) {
/* this prevents us from looping in
* ptlrpc_queue_wait */
spin_lock(&request->rq_lock);
request->rq_err = 1;
spin_unlock(&request->rq_lock);
request->rq_status = rc;
goto cleanup_bulk;
}
} else {
request->rq_repdata = NULL;
request->rq_repmsg = NULL;
}
rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
connection->c_peer, request->rq_xid, 0,
LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
if (rc != 0) {
CERROR("LNetMEAttach failed: %d\n", rc);
LASSERT(rc == -ENOMEM);
rc = -ENOMEM;
goto cleanup_bulk;
}
}
spin_lock(&request->rq_lock);
/* If the MD attach succeeds, there _will_ be a reply_in callback */
request->rq_receiving_reply = !noreply;
request->rq_req_unlink = 1;
/* We are responsible for unlinking the reply buffer */
request->rq_reply_unlink = !noreply;
/* Clear any flags that may be present from previous sends. */
request->rq_replied = 0;
request->rq_err = 0;
request->rq_timedout = 0;
request->rq_net_err = 0;
request->rq_resend = 0;
request->rq_restart = 0;
request->rq_reply_truncate = 0;
spin_unlock(&request->rq_lock);
if (!noreply) {
reply_md.start = request->rq_repbuf;
reply_md.length = request->rq_repbuf_len;
/* Allow multiple early replies */
reply_md.threshold = LNET_MD_THRESH_INF;
/* Manage remote for early replies */
reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
LNET_MD_MANAGE_REMOTE |
LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */
reply_md.user_ptr = &request->rq_reply_cbid;
reply_md.eq_handle = ptlrpc_eq_h;
/* We must see the unlink callback to unset rq_reply_unlink,
so we can't auto-unlink */
rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
&request->rq_reply_md_h);
if (rc != 0) {
CERROR("LNetMDAttach failed: %d\n", rc);
LASSERT(rc == -ENOMEM);
spin_lock(&request->rq_lock);
/* ...but the MD attach didn't succeed... */
request->rq_receiving_reply = 0;
spin_unlock(&request->rq_lock);
rc = -ENOMEM;
goto cleanup_me;
}
CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
request->rq_repbuf_len, request->rq_xid,
request->rq_reply_portal);
}
/* add references on request for request_out_callback */
ptlrpc_request_addref(request);
if (obd != NULL && obd->obd_svc_stats != NULL)
lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
atomic_read(&request->rq_import->imp_inflight));
OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
ktime_get_real_ts64(&request->rq_arrival_time);
request->rq_sent = ktime_get_real_seconds();
/* We give the server rq_timeout secs to process the req, and
add the network latency for our local timeout. */
request->rq_deadline = request->rq_sent + request->rq_timeout +
ptlrpc_at_get_net_latency(request);
ptlrpc_pinger_sending_on_import(request->rq_import);
DEBUG_REQ(D_INFO, request, "send flg=%x",
lustre_msg_get_flags(request->rq_reqmsg));
rc = ptl_send_buf(&request->rq_req_md_h,
request->rq_reqbuf, request->rq_reqdata_len,
LNET_NOACK_REQ, &request->rq_req_cbid,
connection,
request->rq_request_portal,
request->rq_xid, 0);
if (rc == 0)
goto out;
ptlrpc_req_finished(request);
if (noreply)
goto out;
cleanup_me:
/* MEUnlink is safe; the PUT didn't even get off the ground, and
* nobody apart from the PUT's target has the right nid+XID to
* access the reply buffer. */
rc2 = LNetMEUnlink(reply_me_h);
LASSERT(rc2 == 0);
/* UNLINKED callback called synchronously */
LASSERT(!request->rq_receiving_reply);
cleanup_bulk:
/* We do sync unlink here as there was no real transfer here so
* the chance to have long unlink to sluggish net is smaller here. */
ptlrpc_unregister_bulk(request, 0);
out:
if (request->rq_memalloc)
cfs_memory_pressure_restore(mpflag);
return rc;
}
/*
* Client's incoming reply callback
*/
void reply_in_callback(lnet_event_t *ev)
{
struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
struct ptlrpc_request *req = cbid->cbid_arg;
DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
LASSERT(ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
LASSERT(ev->md.start == req->rq_repbuf);
LASSERT(ev->offset + ev->mlength <= req->rq_repbuf_len);
/* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
for adaptive timeouts' early reply. */
LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
spin_lock(&req->rq_lock);
req->rq_receiving_reply = 0;
req->rq_early = 0;
if (ev->unlinked)
req->rq_must_unlink = 0;
if (ev->status)
goto out_wake;
if (ev->type == LNET_EVENT_UNLINK) {
LASSERT(ev->unlinked);
DEBUG_REQ(D_NET, req, "unlink");
goto out_wake;
}
if (ev->mlength < ev->rlength) {
CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
req->rq_replen, ev->rlength, ev->offset);
req->rq_reply_truncate = 1;
req->rq_replied = 1;
req->rq_status = -EOVERFLOW;
req->rq_nob_received = ev->rlength + ev->offset;
goto out_wake;
}
if ((ev->offset == 0) &&
((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
/* Early reply */
DEBUG_REQ(D_ADAPTTO, req,
"Early reply received: mlen=%u offset=%d replen=%d "
"replied=%d unlinked=%d", ev->mlength, ev->offset,
req->rq_replen, req->rq_replied, ev->unlinked);
req->rq_early_count++; /* number received, client side */
if (req->rq_replied) /* already got the real reply */
goto out_wake;
req->rq_early = 1;
req->rq_reply_off = ev->offset;
req->rq_nob_received = ev->mlength;
/* And we're still receiving */
req->rq_receiving_reply = 1;
} else {
/* Real reply */
req->rq_rep_swab_mask = 0;
req->rq_replied = 1;
req->rq_reply_off = ev->offset;
req->rq_nob_received = ev->mlength;
/* LNetMDUnlink can't be called under the LNET_LOCK,
so we must unlink in ptlrpc_unregister_reply */
DEBUG_REQ(D_INFO, req,
"reply in flags=%x mlen=%u offset=%d replen=%d",
lustre_msg_get_flags(req->rq_reqmsg),
ev->mlength, ev->offset, req->rq_replen);
}
req->rq_import->imp_last_reply_time = cfs_time_current_sec();
out_wake:
/* NB don't unlock till after wakeup; req can disappear under us
* since we don't have our own ref */
ptlrpc_client_wake_req(req);
spin_unlock(&req->rq_lock);
}
/**
* Send request \a request.
* if \a noreply is set, don't expect any reply back and don't set up
* reply buffers.
* Returns 0 on success or error code.
*/
int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
{
int rc;
int rc2;
int mpflag = 0;
struct ptlrpc_connection *connection;
lnet_handle_me_t reply_me_h;
lnet_md_t reply_md;
struct obd_import *imp = request->rq_import;
struct obd_device *obd = imp->imp_obd;
ENTRY;
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
RETURN(0);
LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
LASSERT(request->rq_wait_ctx == 0);
/* If this is a re-transmit, we're required to have disengaged
* cleanly from the previous attempt */
LASSERT(!request->rq_receiving_reply);
LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
(imp->imp_state == LUSTRE_IMP_FULL)));
if (unlikely(obd != NULL && obd->obd_fail)) {
CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
obd->obd_name);
/* this prevents us from waiting in ptlrpc_queue_wait */
spin_lock(&request->rq_lock);
request->rq_err = 1;
spin_unlock(&request->rq_lock);
request->rq_status = -ENODEV;
RETURN(-ENODEV);
}
connection = imp->imp_connection;
lustre_msg_set_handle(request->rq_reqmsg,
&imp->imp_remote_handle);
lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
lustre_msg_set_conn_cnt(request->rq_reqmsg,
imp->imp_conn_cnt);
lustre_msghdr_set_flags(request->rq_reqmsg,
imp->imp_msghdr_flags);
/* If it's the first time to resend the request for EINPROGRESS,
* we need to allocate a new XID (see after_reply()), it's different
* from the resend for reply timeout. */
if (request->rq_nr_resend != 0 &&
list_empty(&request->rq_unreplied_list)) {
__u64 min_xid = 0;
/* resend for EINPROGRESS, allocate new xid to avoid reply
* reconstruction */
spin_lock(&imp->imp_lock);
ptlrpc_assign_next_xid_nolock(request);
request->rq_mbits = request->rq_xid;
min_xid = ptlrpc_known_replied_xid(imp);
spin_unlock(&imp->imp_lock);
lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
DEBUG_REQ(D_RPCTRACE, request, "Allocating new xid for "
"resend on EINPROGRESS");
} else if (request->rq_bulk != NULL) {
ptlrpc_set_bulk_mbits(request);
lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
}
if (list_empty(&request->rq_unreplied_list) ||
request->rq_xid <= imp->imp_known_replied_xid) {
DEBUG_REQ(D_ERROR, request, "xid: "LPU64", replied: "LPU64", "
"list_empty:%d\n", request->rq_xid,
imp->imp_known_replied_xid,
list_empty(&request->rq_unreplied_list));
LBUG();
}
/** For enabled AT all request should have AT_SUPPORT in the
* FULL import state when OBD_CONNECT_AT is set */
LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
(imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
!(imp->imp_connect_data.ocd_connect_flags &
OBD_CONNECT_AT));
if (request->rq_resend) {
lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
if (request->rq_resend_cb != NULL)
request->rq_resend_cb(request, &request->rq_async_args);
}
if (request->rq_memalloc)
mpflag = cfs_memory_pressure_get_and_set();
rc = sptlrpc_cli_wrap_request(request);
if (rc == -ENOMEM)
/* set rq_sent so that this request is treated
* as a delayed send in the upper layers */
request->rq_sent = cfs_time_current_sec();
if (rc)
GOTO(out, rc);
/* bulk register should be done after wrap_request() */
if (request->rq_bulk != NULL) {
rc = ptlrpc_register_bulk (request);
if (rc != 0)
GOTO(out, rc);
}
if (!noreply) {
LASSERT (request->rq_replen != 0);
if (request->rq_repbuf == NULL) {
LASSERT(request->rq_repdata == NULL);
LASSERT(request->rq_repmsg == NULL);
rc = sptlrpc_cli_alloc_repbuf(request,
request->rq_replen);
if (rc) {
/* this prevents us from looping in
* ptlrpc_queue_wait */
spin_lock(&request->rq_lock);
request->rq_err = 1;
spin_unlock(&request->rq_lock);
request->rq_status = rc;
GOTO(cleanup_bulk, rc);
}
} else {
request->rq_repdata = NULL;
request->rq_repmsg = NULL;
}
rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
connection->c_peer, request->rq_xid, 0,
LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
if (rc != 0) {
CERROR("LNetMEAttach failed: %d\n", rc);
LASSERT (rc == -ENOMEM);
GOTO(cleanup_bulk, rc = -ENOMEM);
}
}
spin_lock(&request->rq_lock);
/* We are responsible for unlinking the reply buffer */
request->rq_reply_unlinked = noreply;
request->rq_receiving_reply = !noreply;
/* Clear any flags that may be present from previous sends. */
request->rq_req_unlinked = 0;
request->rq_replied = 0;
request->rq_err = 0;
request->rq_timedout = 0;
request->rq_net_err = 0;
request->rq_resend = 0;
request->rq_restart = 0;
request->rq_reply_truncated = 0;
spin_unlock(&request->rq_lock);
if (!noreply) {
reply_md.start = request->rq_repbuf;
reply_md.length = request->rq_repbuf_len;
/* Allow multiple early replies */
reply_md.threshold = LNET_MD_THRESH_INF;
/* Manage remote for early replies */
reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
LNET_MD_MANAGE_REMOTE |
LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
reply_md.user_ptr = &request->rq_reply_cbid;
reply_md.eq_handle = ptlrpc_eq_h;
/* We must see the unlink callback to set rq_reply_unlinked,
* so we can't auto-unlink */
rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
&request->rq_reply_md_h);
if (rc != 0) {
CERROR("LNetMDAttach failed: %d\n", rc);
LASSERT (rc == -ENOMEM);
spin_lock(&request->rq_lock);
/* ...but the MD attach didn't succeed... */
request->rq_receiving_reply = 0;
spin_unlock(&request->rq_lock);
GOTO(cleanup_me, rc = -ENOMEM);
}
CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
", portal %u\n",
request->rq_repbuf_len, request->rq_xid,
request->rq_reply_portal);
}
/* add references on request for request_out_callback */
ptlrpc_request_addref(request);
if (obd != NULL && obd->obd_svc_stats != NULL)
lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
atomic_read(&imp->imp_inflight));
OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
do_gettimeofday(&request->rq_sent_tv);
request->rq_sent = cfs_time_current_sec();
/* We give the server rq_timeout secs to process the req, and
add the network latency for our local timeout. */
request->rq_deadline = request->rq_sent + request->rq_timeout +
ptlrpc_at_get_net_latency(request);
ptlrpc_pinger_sending_on_import(imp);
DEBUG_REQ(D_INFO, request, "send flg=%x",
lustre_msg_get_flags(request->rq_reqmsg));
rc = ptl_send_buf(&request->rq_req_md_h,
request->rq_reqbuf, request->rq_reqdata_len,
LNET_NOACK_REQ, &request->rq_req_cbid,
connection,
request->rq_request_portal,
request->rq_xid, 0);
if (likely(rc == 0))
GOTO(out, rc);
request->rq_req_unlinked = 1;
ptlrpc_req_finished(request);
if (noreply)
GOTO(out, rc);
cleanup_me:
/* MEUnlink is safe; the PUT didn't even get off the ground, and
* nobody apart from the PUT's target has the right nid+XID to
* access the reply buffer. */
rc2 = LNetMEUnlink(reply_me_h);
LASSERT (rc2 == 0);
/* UNLINKED callback called synchronously */
LASSERT(!request->rq_receiving_reply);
cleanup_bulk:
/* We do sync unlink here as there was no real transfer here so
* the chance to have long unlink to sluggish net is smaller here. */
ptlrpc_unregister_bulk(request, 0);
out:
if (request->rq_memalloc)
cfs_memory_pressure_restore(mpflag);
return rc;
}
int llog_origin_handle_cancel(struct ptlrpc_request *req)
{
struct obd_device *obd = req->rq_export->exp_obd;
int num_cookies, rc = 0, err, i, failed = 0;
struct obd_device *disk_obd;
struct llog_cookie *logcookies;
struct llog_ctxt *ctxt = NULL;
struct lvfs_run_ctxt saved;
struct llog_handle *cathandle;
struct inode *inode;
void *handle;
ENTRY;
logcookies = req_capsule_client_get(&req->rq_pill, &RMF_LOGCOOKIES);
num_cookies = req_capsule_get_size(&req->rq_pill, &RMF_LOGCOOKIES,
RCL_CLIENT) / sizeof(*logcookies);
if (logcookies == NULL || num_cookies == 0) {
DEBUG_REQ(D_HA, req, "No llog cookies sent");
RETURN(-EFAULT);
}
ctxt = llog_get_context(obd, logcookies->lgc_subsys);
if (ctxt == NULL)
RETURN(-ENODEV);
disk_obd = ctxt->loc_exp->exp_obd;
push_ctxt(&saved, &disk_obd->obd_lvfs_ctxt, NULL);
for (i = 0; i < num_cookies; i++, logcookies++) {
cathandle = ctxt->loc_handle;
LASSERT(cathandle != NULL);
inode = cathandle->lgh_file->f_dentry->d_inode;
handle = fsfilt_start_log(disk_obd, inode,
FSFILT_OP_CANCEL_UNLINK, NULL, 1);
if (IS_ERR(handle)) {
CERROR("fsfilt_start_log() failed: %ld\n",
PTR_ERR(handle));
GOTO(pop_ctxt, rc = PTR_ERR(handle));
}
rc = llog_cat_cancel_records(cathandle, 1, logcookies);
/*
* Do not raise -ENOENT errors for resent rpcs. This rec already
* might be killed.
*/
if (rc == -ENOENT &&
(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT)) {
/*
* Do not change this message, reply-single.sh test_59b
* expects to find this in log.
*/
CDEBUG(D_RPCTRACE, "RESENT cancel req %p - ignored\n",
req);
rc = 0;
} else if (rc == 0) {
CDEBUG(D_RPCTRACE, "Canceled %d llog-records\n",
num_cookies);
}
err = fsfilt_commit(disk_obd, inode, handle, 0);
if (err) {
CERROR("Error committing transaction: %d\n", err);
if (!rc)
rc = err;
failed++;
GOTO(pop_ctxt, rc);
} else if (rc)
failed++;
}
GOTO(pop_ctxt, rc);
pop_ctxt:
pop_ctxt(&saved, &disk_obd->obd_lvfs_ctxt, NULL);
if (rc)
CERROR("Cancel %d of %d llog-records failed: %d\n",
failed, num_cookies, rc);
llog_ctxt_put(ctxt);
return rc;
}
/*
* VBR: save parent version in reply and child version getting by its name.
* Version of child is getting and checking during its lookup. If
*/
static int mdt_reint_unlink(struct mdt_thread_info *info,
struct mdt_lock_handle *lhc)
{
struct mdt_reint_record *rr = &info->mti_rr;
struct ptlrpc_request *req = mdt_info_req(info);
struct md_attr *ma = &info->mti_attr;
struct lu_fid *child_fid = &info->mti_tmp_fid1;
struct mdt_object *mp;
struct mdt_object *mc;
struct mdt_lock_handle *parent_lh;
struct mdt_lock_handle *child_lh;
struct lu_name *lname;
int rc;
int no_name = 0;
ENTRY;
DEBUG_REQ(D_INODE, req, "unlink "DFID"/%s", PFID(rr->rr_fid1),
rr->rr_name);
if (info->mti_dlm_req)
ldlm_request_cancel(req, info->mti_dlm_req, 0);
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
RETURN(err_serious(-ENOENT));
if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
RETURN(-EPERM);
/*
* step 1: Found the parent.
*/
mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
if (IS_ERR(mp)) {
rc = PTR_ERR(mp);
GOTO(out, rc);
}
parent_lh = &info->mti_lh[MDT_LH_PARENT];
lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
if (mdt_object_remote(mp)) {
mdt_lock_reg_init(parent_lh, LCK_EX);
rc = mdt_remote_object_lock(info, mp, &parent_lh->mlh_rreg_lh,
parent_lh->mlh_rreg_mode,
MDS_INODELOCK_UPDATE);
if (rc != ELDLM_OK)
GOTO(put_parent, rc);
} else {
mdt_lock_pdo_init(parent_lh, LCK_PW, rr->rr_name,
rr->rr_namelen);
rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
MDT_LOCAL_LOCK);
if (rc)
GOTO(put_parent, rc);
rc = mdt_version_get_check_save(info, mp, 0);
if (rc)
GOTO(unlock_parent, rc);
}
/* step 2: find & lock the child */
/* lookup child object along with version checking */
fid_zero(child_fid);
rc = mdt_lookup_version_check(info, mp, lname, child_fid, 1);
if (rc != 0) {
/* Name might not be able to find during resend of
* remote unlink, considering following case.
* dir_A is a remote directory, the name entry of
* dir_A is on MDT0, the directory is on MDT1,
*
* 1. client sends unlink req to MDT1.
* 2. MDT1 sends name delete update to MDT0.
* 3. name entry is being deleted in MDT0 synchronously.
* 4. MDT1 is restarted.
* 5. client resends unlink req to MDT1. So it can not
* find the name entry on MDT0 anymore.
* In this case, MDT1 only needs to destory the local
* directory.
* */
if (mdt_object_remote(mp) && rc == -ENOENT &&
!fid_is_zero(rr->rr_fid2) &&
lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
no_name = 1;
*child_fid = *rr->rr_fid2;
} else {
GOTO(unlock_parent, rc);
}
}
if (fid_is_obf(child_fid) || fid_is_dot_lustre(child_fid))
GOTO(unlock_parent, rc = -EPERM);
mdt_reint_init_ma(info, ma);
/* We will lock the child regardless it is local or remote. No harm. */
mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
if (IS_ERR(mc))
GOTO(unlock_parent, rc = PTR_ERR(mc));
child_lh = &info->mti_lh[MDT_LH_CHILD];
mdt_lock_reg_init(child_lh, LCK_EX);
if (mdt_object_remote(mc)) {
struct mdt_body *repbody;
if (!fid_is_zero(rr->rr_fid2)) {
CDEBUG(D_INFO, "%s: name %s can not find "DFID"\n",
mdt_obd_name(info->mti_mdt),
(char *)rr->rr_name, PFID(mdt_object_fid(mc)));
GOTO(put_child, rc = -ENOENT);
}
CDEBUG(D_INFO, "%s: name %s: "DFID" is another MDT\n",
mdt_obd_name(info->mti_mdt),
(char *)rr->rr_name, PFID(mdt_object_fid(mc)));
if (!mdt_is_dne_client(req->rq_export))
/* Return -EIO for old client */
GOTO(put_child, rc = -EIO);
if (info->mti_spec.sp_rm_entry) {
struct lu_ucred *uc = mdt_ucred(info);
if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
CERROR("%s: unlink remote entry is only "
"permitted for administrator: rc = %d\n",
mdt_obd_name(info->mti_mdt),
-EPERM);
GOTO(put_child, rc = -EPERM);
}
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
NULL, lname, ma, no_name);
GOTO(put_child, rc);
}
/* Revoke the LOOKUP lock of the remote object granted by
* this MDT. Since the unlink will happen on another MDT,
* it will release the LOOKUP lock right away. Then What
* would happen if another client try to grab the LOOKUP
* lock at the same time with unlink XXX */
mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
MDT_CROSS_LOCK);
repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
LASSERT(repbody != NULL);
repbody->fid1 = *mdt_object_fid(mc);
repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
GOTO(unlock_child, rc = -EREMOTE);
} else if (info->mti_spec.sp_rm_entry) {
rc = -EPERM;
CDEBUG(D_INFO, "%s: no rm_entry on local dir '%s': rc = %d\n",
mdt_obd_name(info->mti_mdt), (char *)rr->rr_name, rc);
GOTO(put_child, rc);
}
/* We used to acquire MDS_INODELOCK_FULL here but we can't do
* this now because a running HSM restore on the child (unlink
* victim) will hold the layout lock. See LU-4002. */
rc = mdt_object_lock(info, mc, child_lh,
MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
MDT_CROSS_LOCK);
if (rc != 0)
GOTO(put_child, rc);
mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
OBD_FAIL_MDS_REINT_UNLINK_WRITE);
/* save version when object is locked */
mdt_version_get_save(info, mc, 1);
/*
* Now we can only make sure we need MA_INODE, in mdd layer, will check
* whether need MA_LOV and MA_COOKIE.
*/
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
mutex_lock(&mc->mot_lov_mutex);
rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
mdt_object_child(mc), lname, ma, no_name);
mutex_unlock(&mc->mot_lov_mutex);
if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
rc = mdt_attr_get_complex(info, mc, ma);
if (rc == 0)
mdt_handle_last_unlink(info, mc, ma);
if (ma->ma_valid & MA_INODE) {
switch (ma->ma_attr.la_mode & S_IFMT) {
case S_IFDIR:
mdt_counter_incr(req, LPROC_MDT_RMDIR);
break;
case S_IFREG:
case S_IFLNK:
case S_IFCHR:
case S_IFBLK:
case S_IFIFO:
case S_IFSOCK:
mdt_counter_incr(req, LPROC_MDT_UNLINK);
break;
default:
LASSERTF(0, "bad file type %o unlinking\n",
ma->ma_attr.la_mode);
}
}
EXIT;
unlock_child:
mdt_object_unlock(info, mc, child_lh, rc);
put_child:
mdt_object_put(info->mti_env, mc);
unlock_parent:
mdt_object_unlock(info, mp, parent_lh, rc);
put_parent:
mdt_object_put(info->mti_env, mp);
out:
return rc;
}
