* GPL HEADER END
*/
/*
- * Copyright (c) 2011, 2014, Intel Corporation.
+ * Copyright (c) 2012, 2015, Intel Corporation.
*/
/*
* lustre/target/tgt_handler.c
{
int serious = 0;
int rc;
+ __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
ENTRY;
+
+ /* When dealing with sec context requests, no export is associated yet,
+ * because these requests are sent before *_CONNECT requests.
+ * A NULL req->rq_export means the normal *_common_slice handlers will
+ * not be called, because there is no reference to the target.
+ * So deal with them by hand and jump directly to target_send_reply().
+ */
+ switch (opc) {
+ case SEC_CTX_INIT:
+ case SEC_CTX_INIT_CONT:
+ case SEC_CTX_FINI:
+ CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
+ GOTO(out, rc = 0);
+ }
+
/*
* Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
* to put same checks into handlers like mdt_close(), mdt_reint(),
*/
if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
RETURN(0);
+ if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
+ OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
+ RETURN(0);
rc = tgt_request_preprocess(tsi, h, req);
/* pack reply if reply format is fixed */
if (likely(rc == 0 && req->rq_export))
target_committed_to_req(req);
+out:
target_send_reply(req, rc, tsi->tsi_reply_fail_id);
RETURN(0);
}
/* sanity check: if the xid matches, the request must be marked as a
* resent or replayed */
- if (req_xid_is_last(req)) {
+ if (req_can_reconstruct(req, NULL)) {
if (!(lustre_msg_get_flags(req->rq_reqmsg) &
(MSG_RESENT | MSG_REPLAY))) {
DEBUG_REQ(D_WARNING, req, "rq_xid "LPU64" matches "
- "last_xid, expected REPLAY or RESENT flag "
+ "saved xid, expected REPLAY or RESENT flag "
"(%x)", req->rq_xid,
lustre_msg_get_flags(req->rq_reqmsg));
req->rq_status = -ENOTCONN;
RETURN(h);
}
+static int process_req_last_xid(struct ptlrpc_request *req)
+{
+ __u64 last_xid;
+ ENTRY;
+
+ /* check request's xid is consistent with export's last_xid */
+ last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
+ if (last_xid > req->rq_export->exp_last_xid)
+ req->rq_export->exp_last_xid = last_xid;
+
+ if (req->rq_xid == 0 ||
+ (req->rq_xid <= req->rq_export->exp_last_xid)) {
+ DEBUG_REQ(D_ERROR, req, "Unexpected xid %llx vs. "
+ "last_xid %llx\n", req->rq_xid,
+ req->rq_export->exp_last_xid);
+ /* Some request is allowed to be sent during replay,
+ * such as OUT update requests, FLD requests, so it
+ * is possible that replay requests has smaller XID
+ * than the exp_last_xid.
+ *
+ * Some non-replay requests may have smaller XID as
+ * well:
+ *
+ * - Client send a no_resend RPC, like statfs;
+ * - The RPC timedout (or some other error) on client,
+ * then it's removed from the unreplied list;
+ * - Client send some other request to bump the
+ * exp_last_xid on server;
+ * - The former RPC got chance to be processed;
+ */
+ if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
+ RETURN(-EPROTO);
+ }
+
+ /* try to release in-memory reply data */
+ if (tgt_is_multimodrpcs_client(req->rq_export)) {
+ tgt_handle_received_xid(req->rq_export,
+ lustre_msg_get_last_xid(req->rq_reqmsg));
+ if (!(lustre_msg_get_flags(req->rq_reqmsg) &
+ (MSG_RESENT | MSG_REPLAY)))
+ tgt_handle_tag(req->rq_export,
+ lustre_msg_get_tag(req->rq_reqmsg));
+ }
+ RETURN(0);
+}
+
int tgt_request_handle(struct ptlrpc_request *req)
{
struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
struct lu_target *tgt;
int request_fail_id = 0;
__u32 opc = lustre_msg_get_opc(msg);
+ struct obd_device *obd;
int rc;
-
+ bool is_connect = false;
ENTRY;
/* Refill the context, to make sure all thread keys are allocated */
* target, otherwise that should be connect operation */
if (opc == MDS_CONNECT || opc == OST_CONNECT ||
opc == MGS_CONNECT) {
+ is_connect = true;
req_capsule_set(&req->rq_pill, &RQF_CONNECT);
rc = target_handle_connect(req);
if (rc != 0) {
}
if (unlikely(!class_connected_export(req->rq_export))) {
+ if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
+ opc == SEC_CTX_FINI) {
+ /* sec context initialization has to be handled
+ * by hand in tgt_handle_request0() */
+ tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
+ h = NULL;
+ GOTO(handle_recov, rc = 0);
+ }
CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
opc, libcfs_id2str(req->rq_peer));
req->rq_status = -ENOTCONN;
GOTO(out, rc);
}
+ /* Skip last_xid processing for the recovery thread, otherwise, the
+ * last_xid on same request could be processed twice: first time when
+ * processing the incoming request, second time when the request is
+ * being processed by recovery thread. */
+ obd = class_exp2obd(req->rq_export);
+ if (is_connect) {
+ /* reset the exp_last_xid on each connection. */
+ req->rq_export->exp_last_xid = 0;
+ } else if (obd->obd_recovery_data.trd_processing_task !=
+ current_pid()) {
+ rc = process_req_last_xid(req);
+ if (rc) {
+ req->rq_status = rc;
+ rc = ptlrpc_error(req);
+ GOTO(out, rc);
+ }
+ }
+
request_fail_id = tgt->lut_request_fail_id;
tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
GOTO(out, rc);
}
+ LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
+ h->th_opc, opc);
+
if (CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
GOTO(out, rc = 0);
GOTO(out, rc);
}
+handle_recov:
rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
if (likely(rc == 1)) {
- LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
- h->th_opc, opc);
rc = tgt_handle_request0(tsi, h, req);
if (rc)
GOTO(out, rc);
spin_lock(&exp->exp_lock);
exp->exp_sp_peer = req->rq_sp_from;
exp->exp_flvr = flvr;
+
+ /* when on mgs, if no restriction is set, or if client
+ * is loopback, allow any flavor */
+ if ((strcmp(exp->exp_obd->obd_type->typ_name,
+ LUSTRE_MGS_NAME) == 0) &&
+ (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
+ LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
+ == LOLND))
+ exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
+
if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
CERROR("%s: unauthorized rpc flavor %x from %s, "
return err_serious(-EOPNOTSUPP);
}
+int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
+{
+ struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
+ struct ptlrpc_request *req = tgt_ses_req(tsi);
+ struct obd_export *exp = req->rq_export;
+ struct ptlrpc_bulk_desc *desc;
+ struct l_wait_info *lwi = &tti->tti_u.update.tti_wait_info;
+ int i;
+ int rc;
+
+ ENTRY;
+
+ desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
+ PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
+ MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
+ if (desc == NULL)
+ RETURN(-ENOMEM);
+
+ for (i = 0; i < rdbuf->rb_nbufs; i++)
+ desc->bd_frag_ops->add_iov_frag(desc,
+ rdbuf->rb_bufs[i].lb_buf,
+ rdbuf->rb_bufs[i].lb_len);
+
+ rc = target_bulk_io(exp, desc, lwi);
+ ptlrpc_free_bulk(desc);
+ RETURN(rc);
+}
+EXPORT_SYMBOL(tgt_send_buffer);
+
int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
{
struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
ENTRY;
- desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1, BULK_PUT_SOURCE,
- MDS_BULK_PORTAL);
+ desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
+ PTLRPC_BULK_PUT_SOURCE |
+ PTLRPC_BULK_BUF_KIOV,
+ MDS_BULK_PORTAL,
+ &ptlrpc_bulk_kiov_pin_ops);
if (desc == NULL)
RETURN(-ENOMEM);
for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
i++, tmpcount -= tmpsize) {
tmpsize = min_t(int, tmpcount, PAGE_CACHE_SIZE);
- ptlrpc_prep_bulk_page_pin(desc, rdpg->rp_pages[i], 0, tmpsize);
+ desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
+ tmpsize);
}
LASSERT(desc->bd_nob == nob);
rc = target_bulk_io(exp, desc, lwi);
- ptlrpc_free_bulk_pin(desc);
+ ptlrpc_free_bulk(desc);
RETURN(rc);
}
EXPORT_SYMBOL(tgt_sendpage);
tgt->lut_obd->obd_last_committed) {
rc = dt_object_sync(env, obj, start, end);
}
+ atomic_inc(&tgt->lut_sync_count);
RETURN(rc);
}
* Unified target DLM handlers.
*/
-/* Ensure that data and metadata are synced to the disk when lock is cancelled
- * (if requested) */
+/**
+ * Unified target BAST
+ *
+ * Ensure data and metadata are synced to disk when lock is canceled if Sync on
+ * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
+ * but if it's cross-MDT lock, because remote object version is not set, a
+ * filesystem sync is needed.
+ *
+ * \param lock server side lock
+ * \param desc lock desc
+ * \param data ldlm_cb_set_arg
+ * \param flag indicates whether this cancelling or blocking callback
+ * \retval 0 on success
+ * \retval negative number on error
+ */
static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
void *data, int flag)
{
struct lu_env env;
struct lu_target *tgt;
- struct dt_object *obj;
+ struct dt_object *obj = NULL;
struct lu_fid fid;
int rc = 0;
}
if (flag == LDLM_CB_CANCELING &&
- (lock->l_granted_mode & (LCK_PW | LCK_GROUP)) &&
+ (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
(tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
(tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
- lock->l_flags & LDLM_FL_CBPENDING))) {
+ ldlm_is_cbpending(lock))) &&
+ ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
+ lock->l_resource->lr_type == LDLM_EXTENT)) {
__u64 start = 0;
__u64 end = OBD_OBJECT_EOF;
ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
tgt->lut_lsd.lsd_osd_index);
- obj = dt_locate(&env, tgt->lut_bottom, &fid);
- if (IS_ERR(obj))
- GOTO(err_env, rc = PTR_ERR(obj));
-
- if (!dt_object_exists(obj))
- GOTO(err_put, rc = -ENOENT);
if (lock->l_resource->lr_type == LDLM_EXTENT) {
+ obj = dt_locate(&env, tgt->lut_bottom, &fid);
+ if (IS_ERR(obj))
+ GOTO(err_env, rc = PTR_ERR(obj));
+
+ if (!dt_object_exists(obj))
+ GOTO(err_put, rc = -ENOENT);
+
start = lock->l_policy_data.l_extent.start;
end = lock->l_policy_data.l_extent.end;
}
lock->l_policy_data.l_extent.end, rc);
}
err_put:
- lu_object_put(&env, &obj->do_lu);
+ if (obj != NULL)
+ lu_object_put(&env, &obj->do_lu);
err_env:
lu_env_fini(&env);
}
__u64 start, __u64 end, struct lustre_handle *lh,
int mode, __u64 *flags)
{
- ldlm_policy_data_t policy;
- int rc;
+ union ldlm_policy_data policy;
+ int rc;
ENTRY;
}
EXPORT_SYMBOL(tgt_extent_lock);
-void tgt_extent_unlock(struct lustre_handle *lh, ldlm_mode_t mode)
+void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
{
LASSERT(lustre_handle_is_used(lh));
ldlm_lock_decref(lh, mode);
int tgt_brw_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
struct obd_ioobj *obj, struct niobuf_remote *nb,
- struct lustre_handle *lh, int mode)
+ struct lustre_handle *lh, enum ldlm_mode mode)
{
__u64 flags = 0;
int nrbufs = obj->ioo_bufcnt;
}
void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
- struct lustre_handle *lh, int mode)
+ struct lustre_handle *lh, enum ldlm_mode mode)
{
ENTRY;
unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
__u32 cksum;
+ LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
+
hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
if (IS_ERR(hdesc)) {
CERROR("%s: unable to initialize checksum hash %s\n",
* simulate a client->OST data error */
if (i == 0 && opc == OST_WRITE &&
OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
- int off = desc->bd_iov[i].kiov_offset & ~PAGE_MASK;
- int len = desc->bd_iov[i].kiov_len;
+ int off = BD_GET_KIOV(desc, i).kiov_offset &
+ ~PAGE_MASK;
+ int len = BD_GET_KIOV(desc, i).kiov_len;
struct page *np = tgt_page_to_corrupt;
- char *ptr = kmap(desc->bd_iov[i].kiov_page) + off;
+ char *ptr = kmap(BD_GET_KIOV(desc, i).kiov_page) + off;
if (np) {
char *ptr2 = kmap(np) + off;
memcpy(ptr2, ptr, len);
memcpy(ptr2, "bad3", min(4, len));
kunmap(np);
- desc->bd_iov[i].kiov_page = np;
+ BD_GET_KIOV(desc, i).kiov_page = np;
} else {
CERROR("%s: can't alloc page for corruption\n",
tgt_name(tgt));
}
}
- cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page,
- desc->bd_iov[i].kiov_offset & ~PAGE_MASK,
- desc->bd_iov[i].kiov_len);
+ cfs_crypto_hash_update_page(hdesc,
+ BD_GET_KIOV(desc, i).kiov_page,
+ BD_GET_KIOV(desc, i).kiov_offset &
+ ~PAGE_MASK,
+ BD_GET_KIOV(desc, i).kiov_len);
/* corrupt the data after we compute the checksum, to
* simulate an OST->client data error */
if (i == 0 && opc == OST_READ &&
OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
- int off = desc->bd_iov[i].kiov_offset & ~PAGE_MASK;
- int len = desc->bd_iov[i].kiov_len;
+ int off = BD_GET_KIOV(desc, i).kiov_offset
+ & ~PAGE_MASK;
+ int len = BD_GET_KIOV(desc, i).kiov_len;
struct page *np = tgt_page_to_corrupt;
- char *ptr = kmap(desc->bd_iov[i].kiov_page) + off;
+ char *ptr =
+ kmap(BD_GET_KIOV(desc, i).kiov_page) + off;
if (np) {
char *ptr2 = kmap(np) + off;
memcpy(ptr2, ptr, len);
memcpy(ptr2, "bad4", min(4, len));
kunmap(np);
- desc->bd_iov[i].kiov_page = np;
+ BD_GET_KIOV(desc, i).kiov_page = np;
} else {
CERROR("%s: can't alloc page for corruption\n",
tgt_name(tgt));
GOTO(out_lock, rc);
desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
- BULK_PUT_SOURCE, OST_BULK_PORTAL);
+ PTLRPC_BULK_PUT_SOURCE |
+ PTLRPC_BULK_BUF_KIOV,
+ OST_BULK_PORTAL,
+ &ptlrpc_bulk_kiov_nopin_ops);
if (desc == NULL)
GOTO(out_commitrw, rc = -ENOMEM);
nob += page_rc;
if (page_rc != 0) { /* some data! */
LASSERT(local_nb[i].lnb_page != NULL);
- ptlrpc_prep_bulk_page_nopin(desc, local_nb[i].lnb_page,
- local_nb[i].lnb_page_offset,
- page_rc);
+ desc->bd_frag_ops->add_kiov_frag
+ (desc, local_nb[i].lnb_page,
+ local_nb[i].lnb_page_offset,
+ page_rc);
}
if (page_rc != local_nb[i].lnb_len) { /* short read */
break;
}
}
+ if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
+ nob != cfs_fail_val)
+ rc = -E2BIG;
if (body->oa.o_valid & OBD_MD_FLCKSUM) {
cksum_type_t cksum_type =
tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
- ptlrpc_free_bulk_nopin(desc);
+ ptlrpc_free_bulk(desc);
LASSERT(rc <= 0);
if (rc == 0) {
lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
l_wait_event(waitq, 0, &lwi1);
target_bulk_io(exp, desc, &lwi);
- ptlrpc_free_bulk_nopin(desc);
+ ptlrpc_free_bulk(desc);
}
RETURN(rc);
GOTO(out_lock, rc);
desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
- BULK_GET_SINK, OST_BULK_PORTAL);
+ PTLRPC_BULK_GET_SINK | PTLRPC_BULK_BUF_KIOV,
+ OST_BULK_PORTAL,
+ &ptlrpc_bulk_kiov_nopin_ops);
if (desc == NULL)
GOTO(skip_transfer, rc = -ENOMEM);
/* NB Having prepped, we must commit... */
for (i = 0; i < npages; i++)
- ptlrpc_prep_bulk_page_nopin(desc, local_nb[i].lnb_page,
- local_nb[i].lnb_page_offset,
- local_nb[i].lnb_len);
+ desc->bd_frag_ops->add_kiov_frag(desc,
+ local_nb[i].lnb_page,
+ local_nb[i].lnb_page_offset,
+ local_nb[i].lnb_len);
rc = sptlrpc_svc_prep_bulk(req, desc);
if (rc != 0)
out_lock:
tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
if (desc)
- ptlrpc_free_bulk_nopin(desc);
+ ptlrpc_free_bulk(desc);
out:
if (no_reply) {
req->rq_no_reply = 1;
RETURN(rc);
}
EXPORT_SYMBOL(tgt_brw_write);
+
+/* Check if request can be reconstructed from saved reply data
+ * A copy of the reply data is returned in @trd if the pointer is not NULL
+ */
+bool req_can_reconstruct(struct ptlrpc_request *req,
+ struct tg_reply_data *trd)
+{
+ struct tg_export_data *ted = &req->rq_export->exp_target_data;
+ struct lsd_client_data *lcd = ted->ted_lcd;
+ bool found;
+
+ if (tgt_is_multimodrpcs_client(req->rq_export))
+ return tgt_lookup_reply(req, trd);
+
+ mutex_lock(&ted->ted_lcd_lock);
+ found = req->rq_xid == lcd->lcd_last_xid ||
+ req->rq_xid == lcd->lcd_last_close_xid;
+
+ if (found && trd != NULL) {
+ if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
+ trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
+ trd->trd_reply.lrd_transno =
+ lcd->lcd_last_close_transno;
+ trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
+ } else {
+ trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
+ trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
+ trd->trd_reply.lrd_result = lcd->lcd_last_result;
+ trd->trd_reply.lrd_data = lcd->lcd_last_data;
+ trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
+ trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
+ trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
+ trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
+ }
+ }
+ mutex_unlock(&ted->ted_lcd_lock);
+
+ return found;
+}
+EXPORT_SYMBOL(req_can_reconstruct);
+
git://git.whamcloud.com / fs / lustre-release.git / blobdiff