/* Create a new mdt_file_data struct, initialize it,
* and insert it to global hash table */
-struct mdt_file_data *mdt_mfd_new(void)
+struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
{
- struct mdt_file_data *mfd;
- ENTRY;
+ struct mdt_file_data *mfd;
+ ENTRY;
- OBD_ALLOC_PTR(mfd);
- if (mfd != NULL) {
- CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
- CFS_INIT_LIST_HEAD(&mfd->mfd_list);
+ OBD_ALLOC_PTR(mfd);
+ if (mfd != NULL) {
+ CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
+ mfd->mfd_handle.h_owner = med;
+ CFS_INIT_LIST_HEAD(&mfd->mfd_list);
class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
- }
- RETURN(mfd);
+ }
+
+ RETURN(mfd);
}
/*
* Find the mfd pointed to by handle in global hash table.
* In case of replay the handle is obsoleted
- * but mfd can be found in mfd list by that handle
+ * but mfd can be found in mfd list by that handle.
+ * Callers need to be holding med_open_lock.
*/
-struct mdt_file_data *mdt_handle2mfd(struct mdt_thread_info *info,
- const struct lustre_handle *handle)
+struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
+ const struct lustre_handle *handle,
+ bool is_replay_or_resent)
{
- struct ptlrpc_request *req = mdt_info_req(info);
- struct mdt_file_data *mfd;
- ENTRY;
+ struct mdt_file_data *mfd;
+ ENTRY;
- LASSERT(handle != NULL);
- mfd = class_handle2object(handle->cookie);
- /* during dw/setattr replay the mfd can be found by old handle */
- if (mfd == NULL && req_is_replay(req)) {
- struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
- cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
- if (mfd->mfd_old_handle.cookie == handle->cookie)
- RETURN (mfd);
- }
- mfd = NULL;
- }
- RETURN (mfd);
+ LASSERT(handle != NULL);
+ mfd = class_handle2object(handle->cookie, med);
+ /* during dw/setattr replay the mfd can be found by old handle */
+ if (mfd == NULL && is_replay_or_resent) {
+ cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
+ if (mfd->mfd_old_handle.cookie == handle->cookie)
+ RETURN(mfd);
+ }
+ mfd = NULL;
+ }
+
+ RETURN(mfd);
}
/* free mfd */
int rc = 0;
ENTRY;
- if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
- !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
- RETURN(0);
+ if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
+ !S_ISREG(lu_object_attr(&o->mot_obj)))
+ RETURN(0);
mutex_lock(&o->mot_ioepoch_mutex);
if (mdt_ioepoch_opened(o)) {
*/
static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
{
- struct ptlrpc_request *req = mdt_info_req(info);
- ENTRY;
+ struct ptlrpc_request *req = mdt_info_req(info);
+ ENTRY;
- if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
- !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
- RETURN(0);
+ if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
+ !S_ISREG(lu_object_attr(&o->mot_obj)))
+ RETURN(0);
LASSERT(o->mot_ioepoch_count);
LASSERT(info->mti_ioepoch == NULL ||
RETURN(mdt_ioepoch_close_on_eviction(info, o));
if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
RETURN(mdt_ioepoch_close_on_replay(info, o));
- if (info->mti_ioepoch->flags & MF_EPOCH_CLOSE)
+ if (info->mti_ioepoch && (info->mti_ioepoch->flags & MF_EPOCH_CLOSE))
RETURN(mdt_ioepoch_close_reg(info, o));
/* IO epoch is not closed. */
RETURN(MDT_IOEPOCH_OPENED);
*/
int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
{
- struct ptlrpc_request *req = mdt_info_req(info);
- __u64 ioepoch = 0;
- int act = MDT_SOM_ENABLE;
- int rc = 0;
- ENTRY;
+ struct ptlrpc_request *req = mdt_info_req(info);
+ __u64 ioepoch = 0;
+ int act = MDT_SOM_ENABLE;
+ int rc = 0;
+ ENTRY;
- LASSERT(!req || info->mti_ioepoch);
- if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
- !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
- RETURN(0);
+ LASSERT(!req || info->mti_ioepoch);
+ if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
+ !S_ISREG(lu_object_attr(&o->mot_obj)))
+ RETURN(0);
/* No size whereas MF_SOM_CHANGE is set means client failed to
* obtain ost attributes, drop the SOM cache on disk if so. */
ioepoch = info->mti_ioepoch ?
info->mti_ioepoch->ioepoch : o->mot_ioepoch;
- if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
+ if (req != NULL
+ && !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
rc = mdt_som_attr_set(info, o, ioepoch, act);
mdt_object_som_enable(o, ioepoch);
}
EXIT;
}
-void mdt_mfd_set_mode(struct mdt_file_data *mfd, int mode)
+void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
{
- LASSERT(mfd != NULL);
+ LASSERT(mfd != NULL);
- CDEBUG(D_HA, "Change mfd %p mode 0x%x->0x%x\n",
- mfd, (unsigned int)mfd->mfd_mode, (unsigned int)mode);
+ CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
+ PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
- mfd->mfd_mode = mode;
+ mfd->mfd_mode = mode;
}
static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
isreg = S_ISREG(la->la_mode);
isdir = S_ISDIR(la->la_mode);
- if (isreg && !(ma->ma_valid & MA_LOV)) {
+ if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
/*
* No EA, check whether it is will set regEA and dirEA since in
* above attr get, these size might be zero, so reset it, to
repbody->ioepoch = o->mot_ioepoch;
}
} else if (flags & MDS_FMODE_EXEC) {
- /* if file is released, we can't deny write because we must
- * restore (write) it to access it.*/
- if ((ma->ma_valid & MA_HSM) &&
- (ma->ma_hsm.mh_flags & HS_RELEASED))
- rc = 0;
- else
- rc = mdt_write_deny(o);
+ rc = mdt_write_deny(o);
}
if (rc)
RETURN(rc);
if (rc)
GOTO(err_out, rc);
- mfd = mdt_mfd_new();
- if (mfd != NULL) {
- /*
- * Keep a reference on this object for this open, and is
- * released by mdt_mfd_close().
- */
- mdt_object_get(info->mti_env, o);
+ mfd = mdt_mfd_new(med);
+ if (mfd == NULL)
+ GOTO(err_out, rc = -ENOMEM);
- /*
- * @flags is always not zero. At least it should be FMODE_READ,
- * FMODE_WRITE or MDS_FMODE_EXEC.
- */
- LASSERT(flags != 0);
-
- /* Open handling. */
- mdt_mfd_set_mode(mfd, flags);
-
- mfd->mfd_object = o;
- mfd->mfd_xid = req->rq_xid;
-
- /* replay handle */
- if (req_is_replay(req)) {
- struct mdt_file_data *old_mfd;
- /* Check wheather old cookie already exist in
- * the list, becasue when do recovery, client
- * might be disconnected from server, and
- * restart replay, so there maybe some orphan
- * mfd here, we should remove them */
- LASSERT(info->mti_rr.rr_handle != NULL);
- old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle);
- if (old_mfd) {
- CDEBUG(D_HA, "del orph mfd %p fid=("DFID") "
- "cookie=" LPX64"\n", mfd,
- PFID(mdt_object_fid(mfd->mfd_object)),
- info->mti_rr.rr_handle->cookie);
- spin_lock(&med->med_open_lock);
- class_handle_unhash(&old_mfd->mfd_handle);
- cfs_list_del_init(&old_mfd->mfd_list);
- spin_unlock(&med->med_open_lock);
- /* no attr update for that close */
- la->la_valid = 0;
- ma->ma_valid |= MA_FLAGS;
- ma->ma_attr_flags |= MDS_RECOV_OPEN;
- mdt_mfd_close(info, old_mfd);
- ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
- ma->ma_valid &= ~MA_FLAGS;
- }
- CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
- info->mti_rr.rr_handle->cookie);
- mfd->mfd_old_handle.cookie =
- info->mti_rr.rr_handle->cookie;
- }
- repbody->handle.cookie = mfd->mfd_handle.h_cookie;
+ /*
+ * Keep a reference on this object for this open, and is
+ * released by mdt_mfd_close().
+ */
+ mdt_object_get(info->mti_env, o);
+ mfd->mfd_object = o;
+ mfd->mfd_xid = req->rq_xid;
- if (req->rq_export->exp_disconnected) {
- spin_lock(&med->med_open_lock);
- class_handle_unhash(&mfd->mfd_handle);
- cfs_list_del_init(&mfd->mfd_list);
+ /*
+ * @flags is always not zero. At least it should be FMODE_READ,
+ * FMODE_WRITE or MDS_FMODE_EXEC.
+ */
+ LASSERT(flags != 0);
+
+ /* Open handling. */
+ mdt_mfd_set_mode(mfd, flags);
+
+ atomic_inc(&o->mot_open_count);
+ if (flags & MDS_OPEN_LEASE)
+ atomic_inc(&o->mot_lease_count);
+
+ /* replay handle */
+ if (req_is_replay(req)) {
+ struct mdt_file_data *old_mfd;
+ /* Check wheather old cookie already exist in
+ * the list, becasue when do recovery, client
+ * might be disconnected from server, and
+ * restart replay, so there maybe some orphan
+ * mfd here, we should remove them */
+ LASSERT(info->mti_rr.rr_handle != NULL);
+ spin_lock(&med->med_open_lock);
+ old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
+ if (old_mfd != NULL) {
+ CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
+ "cookie = "LPX64"\n", mfd,
+ PFID(mdt_object_fid(mfd->mfd_object)),
+ info->mti_rr.rr_handle->cookie);
+ class_handle_unhash(&old_mfd->mfd_handle);
+ cfs_list_del_init(&old_mfd->mfd_list);
spin_unlock(&med->med_open_lock);
- mdt_mfd_close(info, mfd);
+ /* no attr update for that close */
+ la->la_valid = 0;
+ ma->ma_valid |= MA_FLAGS;
+ ma->ma_attr_flags |= MDS_RECOV_OPEN;
+ mdt_mfd_close(info, old_mfd);
+ ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
+ ma->ma_valid &= ~MA_FLAGS;
} else {
- spin_lock(&med->med_open_lock);
- cfs_list_add(&mfd->mfd_list, &med->med_open_head);
spin_unlock(&med->med_open_lock);
- }
+ CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
+ "cookie = "LPX64"\n",
+ PFID(mdt_object_fid(mfd->mfd_object)),
+ info->mti_rr.rr_handle->cookie);
+ }
- mdt_empty_transno(info, rc);
- } else {
- GOTO(err_out, rc = -ENOMEM);
- }
+ CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
+ info->mti_rr.rr_handle->cookie);
+
+ mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
+ }
+
+ repbody->handle.cookie = mfd->mfd_handle.h_cookie;
+
+ if (req->rq_export->exp_disconnected) {
+ spin_lock(&med->med_open_lock);
+ class_handle_unhash(&mfd->mfd_handle);
+ cfs_list_del_init(&mfd->mfd_list);
+ spin_unlock(&med->med_open_lock);
+ mdt_mfd_close(info, mfd);
+ } else {
+ spin_lock(&med->med_open_lock);
+ cfs_list_add(&mfd->mfd_list, &med->med_open_head);
+ spin_unlock(&med->med_open_lock);
+ }
+
+ mdt_empty_transno(info, rc);
RETURN(rc);
}
#endif
- if (info->mti_mdt->mdt_opts.mo_mds_capa &&
+ if (info->mti_mdt->mdt_lut.lut_mds_capa &&
exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
struct lustre_capa *capa;
RETURN(rc);
repbody->valid |= OBD_MD_FLMDSCAPA;
}
- if (info->mti_mdt->mdt_opts.mo_oss_capa &&
+ if (info->mti_mdt->mdt_lut.lut_oss_capa &&
exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
- S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) {
+ S_ISREG(lu_object_attr(&o->mot_obj))) {
struct lustre_capa *capa;
capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
struct lsd_client_data *lcd = ted->ted_lcd;
struct md_attr *ma = &info->mti_attr;
struct mdt_reint_record *rr = &info->mti_rr;
- __u32 flags = info->mti_spec.sp_cr_flags;
+ __u64 flags = info->mti_spec.sp_cr_flags;
struct ldlm_reply *ldlm_rep;
struct mdt_object *parent;
struct mdt_object *child;
LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
}
-int mdt_open_by_fid(struct mdt_thread_info* info,
- struct ldlm_reply *rep)
+int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
{
- __u32 flags = info->mti_spec.sp_cr_flags;
+ __u64 flags = info->mti_spec.sp_cr_flags;
struct mdt_reint_record *rr = &info->mti_rr;
struct md_attr *ma = &info->mti_attr;
struct mdt_object *o;
struct md_attr *ma = &info->mti_attr;
__u64 open_flags = info->mti_spec.sp_cr_flags;
ldlm_mode_t lm = LCK_CR;
+ bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
bool try_layout = false;
bool create_layout = false;
int rc = 0;
ENTRY;
*ibits = 0;
- if (open_flags & MDS_OPEN_LOCK) {
- if (open_flags & FMODE_WRITE)
- lm = LCK_CW;
- /* if file is released, we can't deny write because we must
- * restore (write) it to access it. */
- else if ((open_flags & MDS_FMODE_EXEC) &&
- !((ma->ma_valid & MA_HSM) &&
- (ma->ma_hsm.mh_flags & HS_RELEASED)))
- lm = LCK_PR;
- else
- lm = LCK_CR;
+ mdt_lock_handle_init(lhc);
- *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
- }
+ if (req_is_replay(mdt_info_req(info)))
+ RETURN(0);
- if (S_ISREG(lu_object_attr(&obj->mot_obj.mo_lu))) {
+ if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
md_should_create(open_flags))
create_layout = true;
try_layout = true;
}
- mdt_lock_handle_init(lhc);
+ if (acq_lease) {
+ /* lease open, acquire write mode of open sem */
+ down_write(&obj->mot_open_sem);
+
+ /* Lease exists and ask for new lease */
+ if (atomic_read(&obj->mot_lease_count) > 0) {
+ /* only exclusive open is supported, so lease
+ * are conflicted to each other */
+ GOTO(out, rc = -EBUSY);
+ }
+
+ /* Lease must be with open lock */
+ if (!(open_flags & MDS_OPEN_LOCK)) {
+ CERROR("Request lease for file:"DFID ", but open lock "
+ "is missed, open_flags = "LPO64".\n",
+ PFID(mdt_object_fid(obj)), open_flags);
+ GOTO(out, rc = -EPROTO);
+ }
+
+ /* XXX: only exclusive open is supported. */
+ lm = LCK_EX;
+ *ibits = MDS_INODELOCK_OPEN;
+
+ /* never grant LCK_EX layout lock to client */
+ try_layout = false;
+ } else { /* normal open */
+ /* normal open holds read mode of open sem */
+ down_read(&obj->mot_open_sem);
+
+ if (open_flags & MDS_OPEN_LOCK) {
+ if (open_flags & FMODE_WRITE)
+ lm = LCK_CW;
+ else if (open_flags & MDS_FMODE_EXEC)
+ lm = LCK_PR;
+ else
+ lm = LCK_CR;
+
+ *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
+ } else if (atomic_read(&obj->mot_lease_count) > 0) {
+ if (open_flags & FMODE_WRITE)
+ lm = LCK_CW;
+ else
+ lm = LCK_CR;
+
+ /* revoke lease */
+ *ibits = MDS_INODELOCK_OPEN;
+ try_layout = false;
+
+ lhc = &info->mti_lh[MDT_LH_LOCAL];
+ }
+ CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
+ PFID(mdt_object_fid(obj)),
+ atomic_read(&obj->mot_open_count), lm);
+ }
+
mdt_lock_reg_init(lhc, lm);
/* one problem to return layout lock on open is that it may result
OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
}
+ /* Check if there is any other open handles after acquiring
+ * open lock. At this point, caching open handles have been revoked
+ * by open lock.
+ * XXX: Now only exclusive open is supported. Need to check the
+ * type of open for generic lease support. */
+ if (rc == 0 && acq_lease) {
+ struct ptlrpc_request *req = mdt_info_req(info);
+ struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
+ struct mdt_file_data *mfd;
+ bool is_replay_or_resent;
+ int open_count = 0;
+
+ /* For lease: application can open a file and then apply lease,
+ * @handle contains original open handle in that case.
+ * In recovery, open REQ will be replayed and the lease REQ may
+ * be resent that means the open handle is already stale, so we
+ * need to fix it up here by finding new handle. */
+ is_replay_or_resent = req_is_replay(req) ||
+ lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
+
+ /* if the request is _not_ a replay request, rr_handle
+ * may be used to hold an openhandle which is issuing the
+ * lease request, so that this openhandle doesn't count. */
+ mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
+ is_replay_or_resent);
+ if (mfd != NULL)
+ ++open_count;
+
+ CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
+ PFID(mdt_object_fid(obj)),
+ atomic_read(&obj->mot_open_count), open_count);
+
+ if (atomic_read(&obj->mot_open_count) > open_count)
+ GOTO(out, rc = -EBUSY);
+ }
+ GOTO(out, rc);
+
+out:
RETURN(rc);
}
__u64 ibits, int rc)
{
__u64 open_flags = info->mti_spec.sp_cr_flags;
- struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
+ struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
+ ENTRY;
+
+ if (req_is_replay(mdt_info_req(info)))
+ RETURN_EXIT;
- /* Release local layout lock - the layout lock put in MDT_LH_LAYOUT
- * will never return to client side. */
+ /* Release local lock - the lock put in MDT_LH_LOCAL will never
+ * return to client side. */
+ if (lustre_handle_is_used(&ll->mlh_reg_lh))
+ mdt_object_unlock(info, obj, ll, 1);
+
+ ll = &info->mti_lh[MDT_LH_LAYOUT];
+ /* Release local layout lock, layout was created */
if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
mdt_object_unlock(info, obj, ll, 1);
}
+ if (open_flags & MDS_OPEN_LEASE)
+ up_write(&obj->mot_open_sem);
+ else
+ up_read(&obj->mot_open_sem);
+
/* Cross-ref case, the lock should be returned to the client */
if (ibits == 0 || rc == -EREMOTE)
- return;
+ RETURN_EXIT;
if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
/* for the open request, the lock will only return to client
mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
mdt_object_unlock(info, obj, lhc, 1);
}
+ RETURN_EXIT;
+}
+
+/**
+ * Check release is permitted for the current HSM flags.
+ */
+static bool mdt_hsm_release_allow(const struct md_attr *ma)
+{
+ if (!(ma->ma_valid & MA_HSM))
+ return false;
+
+ if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
+ return false;
+
+ if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
+ return false;
+
+ return true;
}
int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
{
const struct lu_env *env = info->mti_env;
struct mdt_device *mdt = info->mti_mdt;
- __u32 flags = info->mti_spec.sp_cr_flags;
+ __u64 flags = info->mti_spec.sp_cr_flags;
struct mdt_reint_record *rr = &info->mti_rr;
struct md_attr *ma = &info->mti_attr;
struct mdt_object *parent= NULL;
mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
+ if (flags & MDS_OPEN_RELEASE)
+ ma->ma_need |= MA_HSM;
rc = mdt_attr_get_complex(info, o, ma);
- if (rc)
- GOTO(out, rc);
+ if (rc)
+ GOTO(out, rc);
+
+ /* If a release request, check file flags are fine and ask for an
+ * exclusive open access. */
+ if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
+ GOTO(out, rc = -EPERM);
rc = mdt_object_open_lock(info, o, lhc, &ibits);
if (rc)
- GOTO(out, rc);
+ GOTO(out_unlock, rc);
if (ma->ma_valid & MA_PFID) {
parent = mdt_object_find(env, mdt, &ma->ma_pfid);
mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
if (flags & MDS_OPEN_LOCK)
mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
+ if (flags & MDS_OPEN_LEASE)
+ mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
}
- GOTO(out, rc);
+ GOTO(out_unlock, rc);
-out:
+out_unlock:
mdt_object_open_unlock(info, o, lhc, ibits, rc);
- mdt_object_put(env, o);
- if (parent != NULL)
- mdt_object_put(env, parent);
- return rc;
-}
-
-int mdt_pin(struct mdt_thread_info* info)
-{
- ENTRY;
- RETURN(err_serious(-EOPNOTSUPP));
+out:
+ mdt_object_put(env, o);
+ if (parent != NULL)
+ mdt_object_put(env, parent);
+ return rc;
}
/* Cross-ref request. Currently it can only be a pure open (w/o create) */
CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
- &o->mot_obj.mo_lu,
+ &o->mot_obj,
"Object isn't on this server! FLD error?\n");
rc = -EFAULT;
} else {
struct lu_fid *child_fid = &info->mti_tmp_fid1;
struct md_attr *ma = &info->mti_attr;
__u64 create_flags = info->mti_spec.sp_cr_flags;
- __u64 ibits;
+ __u64 ibits = 0;
struct mdt_reint_record *rr = &info->mti_rr;
struct lu_name *lname;
int result, rc;
}
if (!(create_flags & MDS_OPEN_CREAT))
GOTO(out_parent, result);
+ if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
+ GOTO(out_parent, result = -EROFS);
*child_fid = *info->mti_rr.rr_fid2;
LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
PFID(child_fid));
if (!req_is_replay(req)) {
rc = mdt_object_open_lock(info, child, lhc, &ibits);
if (rc != 0)
- GOTO(out_child, result = rc);
+ GOTO(out_child_unlock, result = rc);
else if (create_flags & MDS_OPEN_LOCK)
mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
}
result = rc;
/* openlock will be released if mdt_finish_open failed */
mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
+
+ if (created && create_flags & MDS_OPEN_VOLATILE) {
+ CERROR("%s: cannot open volatile file "DFID", orphan "
+ "file will be left in PENDING directory until "
+ "next reboot, rc = %d\n", mdt_obd_name(mdt),
+ PFID(mdt_object_fid(child)), rc);
+ GOTO(out_child_unlock, result);
+ }
+
if (created) {
ma->ma_need = 0;
ma->ma_valid = 0;
mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
}
}
- EXIT;
-out_child:
+ EXIT;
+out_child_unlock:
mdt_object_open_unlock(info, child, lhc, ibits, result);
- mdt_object_put(info->mti_env, child);
+out_child:
+ mdt_object_put(info->mti_env, child);
out_parent:
- mdt_object_unlock_put(info, parent, lh, result || !created);
+ mdt_object_unlock_put(info, parent, lh, result || !created);
out:
if (result)
lustre_msg_set_transno(req->rq_repmsg, 0);
return result;
}
+/**
+ * Create an orphan object use local root.
+ */
+static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
+ struct mdt_device *mdt,
+ const struct lu_fid *fid,
+ struct md_attr *attr, fmode_t fmode)
+{
+ const struct lu_env *env = info->mti_env;
+ struct md_op_spec *spec = &info->mti_spec;
+ struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
+ struct mdt_object *obj = NULL;
+ struct mdt_object *local_root;
+ static const char name[] = "i_am_nobody";
+ struct lu_name *lname;
+ struct lu_ucred *uc;
+ cfs_cap_t uc_cap_save;
+ int rc;
+ ENTRY;
+
+ rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
+ if (rc != 0)
+ RETURN(ERR_PTR(rc));
+
+ local_root = mdt_object_find(env, mdt, local_root_fid);
+ if (IS_ERR(local_root))
+ RETURN(local_root);
+
+ obj = mdt_object_new(env, mdt, fid);
+ if (IS_ERR(obj))
+ GOTO(out, rc = PTR_ERR(obj));
+
+ spec->sp_cr_lookup = 0;
+ spec->sp_feat = &dt_directory_features;
+ spec->sp_cr_mode = MDL_MINMODE; /* no lock */
+ spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
+ if (attr->ma_valid & MA_LOV) {
+ spec->u.sp_ea.eadata = attr->ma_lmm;
+ spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
+ spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
+ } else {
+ spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
+ }
+
+ lname = mdt_name(env, (char *)name, sizeof(name) - 1);
+
+ uc = lu_ucred(env);
+ uc_cap_save = uc->uc_cap;
+ uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
+ rc = mdo_create(env, mdt_object_child(local_root), lname,
+ mdt_object_child(obj), spec, attr);
+ uc->uc_cap = uc_cap_save;
+ if (rc < 0) {
+ CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
+ mdt_obd_name(mdt), PFID(fid), rc);
+ GOTO(out, rc);
+ }
+
+ rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
+ if (rc < 0)
+ CERROR("%s: cannot open volatile file "DFID", orphan "
+ "file will be left in PENDING directory until "
+ "next reboot, rc = %d\n", mdt_obd_name(mdt),
+ PFID(fid), rc);
+ GOTO(out, rc);
+
+out:
+ if (rc < 0) {
+ if (!IS_ERR(obj))
+ mdt_object_put(env, obj);
+ obj = ERR_PTR(rc);
+ }
+ mdt_object_put(env, local_root);
+ return obj;
+}
+
+static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
+ struct md_attr *ma)
+{
+ struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
+ struct close_data *data;
+ struct ldlm_lock *lease;
+ struct mdt_object *orphan;
+ struct md_attr *orp_ma;
+ struct lu_buf *buf;
+ bool lease_broken;
+ int rc;
+ int rc2;
+ ENTRY;
+
+ if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
+ RETURN(-EROFS);
+
+ data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
+ if (data == NULL)
+ RETURN(-EPROTO);
+
+ lease = ldlm_handle2lock(&data->cd_handle);
+ if (lease == NULL)
+ RETURN(-ESTALE);
+
+ /* try to hold open_sem so that nobody else can open the file */
+ if (!down_write_trylock(&o->mot_open_sem)) {
+ ldlm_lock_cancel(lease);
+ GOTO(out_reprocess, rc = -EBUSY);
+ }
+
+ /* Check if the lease open lease has already canceled */
+ lock_res_and_lock(lease);
+ lease_broken = ldlm_is_cancel(lease);
+ unlock_res_and_lock(lease);
+
+ LDLM_DEBUG(lease, DFID " lease broken? %d\n",
+ PFID(mdt_object_fid(o)), lease_broken);
+
+ /* Cancel server side lease. Client side counterpart should
+ * have been cancelled. It's okay to cancel it now as we've
+ * held mot_open_sem. */
+ ldlm_lock_cancel(lease);
+
+ if (lease_broken) /* don't perform release task */
+ GOTO(out_unlock, rc = -ESTALE);
+
+ if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
+ GOTO(out_unlock, rc = -EINVAL);
+
+ /* ma_need was set before but it seems fine to change it in order to
+ * avoid modifying the one from RPC */
+ ma->ma_need = MA_HSM;
+ rc = mdt_attr_get_complex(info, o, ma);
+ if (rc != 0)
+ GOTO(out_unlock, rc);
+
+ if (!mdt_hsm_release_allow(ma))
+ GOTO(out_unlock, rc = -EPERM);
+
+ /* already released? */
+ if (ma->ma_hsm.mh_flags & HS_RELEASED)
+ GOTO(out_unlock, rc = 0);
+
+ /* Compare on-disk and packed data_version */
+ if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
+ CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
+ " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
+ data->cd_data_version, ma->ma_hsm.mh_arch_ver);
+ GOTO(out_unlock, rc = -EPERM);
+ }
+
+ ma->ma_valid = MA_INODE;
+ ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
+ rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
+ if (rc < 0)
+ GOTO(out_unlock, rc);
+
+ ma->ma_need = MA_INODE | MA_LOV;
+ rc = mdt_attr_get_complex(info, o, ma);
+ if (rc < 0)
+ GOTO(out_unlock, rc);
+
+ if (!(ma->ma_valid & MA_LOV)) {
+ /* Even empty file are released */
+ memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
+ ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
+ ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
+ ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
+ ma->ma_lmm_size = sizeof(*ma->ma_lmm);
+ } else {
+ /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
+ * ma_lmm as lov_user_md, then it will be confused by union of
+ * layout_gen and stripe_offset. */
+ if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
+ ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
+ else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
+ ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
+ else
+ GOTO(out_unlock, rc = -EINVAL);
+ }
+
+ /* Set file as released */
+ ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
+
+ /* Hopefully it's not used in this call path */
+ orp_ma = &info->mti_u.som.attr;
+ orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
+ orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
+ orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
+ orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
+ orp_ma->ma_lmm = ma->ma_lmm;
+ orp_ma->ma_lmm_size = ma->ma_lmm_size;
+ orp_ma->ma_valid = MA_INODE | MA_LOV;
+ orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
+ FMODE_WRITE);
+ if (IS_ERR(orphan)) {
+ CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
+ mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
+ PTR_ERR(orphan));
+ GOTO(out_unlock, rc = PTR_ERR(orphan));
+ }
+
+ /* Set up HSM attribute for orphan object */
+ CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
+ buf = &info->mti_buf;
+ buf->lb_buf = info->mti_xattr_buf;
+ buf->lb_len = sizeof(struct hsm_attrs);
+ ma->ma_hsm.mh_flags |= HS_RELEASED;
+ lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
+ ma->ma_hsm.mh_flags &= ~HS_RELEASED;
+
+ mdt_lock_reg_init(lh, LCK_EX);
+ rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
+ MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
+ if (rc != 0)
+ GOTO(out_close, rc);
+
+ rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
+ XATTR_NAME_HSM, 0);
+
+ if (rc == 0)
+ /* Swap layout with orphan object */
+ rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
+ mdt_object_child(orphan),
+ SWAP_LAYOUTS_MDS_HSM);
+
+ /* Release exclusive LL */
+ mdt_object_unlock(info, o, lh, 1);
+
+ EXIT;
+
+out_close:
+ /* Close orphan object anyway */
+ rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
+ FMODE_WRITE);
+ if (rc2 < 0)
+ CERROR("%s: error closing volatile file "DFID": rc = %d\n",
+ mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
+ LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
+ "object closed\n");
+ mdt_object_put(info->mti_env, orphan);
+
+out_unlock:
+ up_write(&o->mot_open_sem);
+
+ if (rc == 0) { /* already released */
+ struct mdt_body *repbody;
+ repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
+ LASSERT(repbody != NULL);
+ repbody->valid |= OBD_MD_FLRELEASED;
+ }
+
+out_reprocess:
+ ldlm_reprocess_all(lease->l_resource);
+ LDLM_LOCK_PUT(lease);
+
+ ma->ma_valid = 0;
+ ma->ma_need = 0;
+
+ return rc;
+}
+
#define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
struct md_attr *ma = &info->mti_attr;
int ret = MDT_IOEPOCH_CLOSED;
int rc = 0;
- int mode;
+ __u64 mode;
ENTRY;
mode = mfd->mfd_mode;
+ if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
+ rc = mdt_hsm_release(info, o, ma);
+ if (rc < 0) {
+ CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
+ mdt_obd_name(info->mti_mdt),
+ PFID(mdt_object_fid(o)), rc);
+ /* continue to close even error occurred. */
+ }
+ }
+
if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
mdt_write_put(o);
ret = mdt_ioepoch_close(info, o);
rc = mo_attr_set(info->mti_env, next, ma);
}
- /* If file data is modified, add the dirty flag.
- *
- * If MDS_CLOSE_CLEANUP is set, this file is being closed due to an
- * eviction, file could have been modified and now dirty
- * regarding to HSM archive, check this!
- * The logic here is to mark a file dirty if there's a chance it was
- * dirtied before the client was evicted, so that we don't have to wait
- * for a release attempt before finding out the file was actually dirty
- * and fail the release. Aggressively marking it dirty here will cause
- * the policy engine to attempt to re-archive it; when rearchiving, we
- * can compare the current version to the LMA data_version and make the
- * archive request into a noop if it's not actually dirty.
- */
- if ((ma->ma_attr_flags & MDS_DATA_MODIFIED) ||
- ((ma->ma_attr_flags & MDS_CLOSE_CLEANUP) &&
- (mode & (FMODE_WRITE|MDS_FMODE_TRUNC))))
+ /* If file data is modified, add the dirty flag. */
+ if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
rc = mdt_add_dirty_flag(info, o, ma);
ma->ma_need |= MA_INODE;
"needed on "DFID"\n", PFID(mdt_object_fid(o)));
}
} else {
- mdt_mfd_free(mfd);
- mdt_object_put(info->mti_env, o);
- }
+ /* adjust open and lease count */
+ if (mode & MDS_OPEN_LEASE) {
+ LASSERT(atomic_read(&o->mot_lease_count) > 0);
+ atomic_dec(&o->mot_lease_count);
+ }
+ LASSERT(atomic_read(&o->mot_open_count) > 0);
+ atomic_dec(&o->mot_open_count);
- RETURN(rc ? rc : ret);
+ mdt_mfd_free(mfd);
+ mdt_object_put(info->mti_env, o);
+ }
+
+ RETURN(rc ? rc : ret);
}
-int mdt_close(struct mdt_thread_info *info)
+int mdt_close(struct tgt_session_info *tsi)
{
+ struct mdt_thread_info *info = tsi2mdt_info(tsi);
+ struct ptlrpc_request *req = tgt_ses_req(tsi);
struct mdt_export_data *med;
struct mdt_file_data *mfd;
struct mdt_object *o;
struct md_attr *ma = &info->mti_attr;
struct mdt_body *repbody = NULL;
- struct ptlrpc_request *req = mdt_info_req(info);
int rc, ret = 0;
ENTRY;
mdt_counter_incr(req, LPROC_MDT_CLOSE);
- /* Close may come with the Size-on-MDS update. Unpack it. */
- rc = mdt_close_unpack(info);
- if (rc)
- RETURN(err_serious(rc));
+ /* Close may come with the Size-on-MDS update. Unpack it. */
+ rc = mdt_close_unpack(info);
+ if (rc)
+ GOTO(out, rc = err_serious(rc));
LASSERT(info->mti_ioepoch);
if (rc == 0)
mdt_fix_reply(info);
mdt_exit_ucred(info);
- RETURN(lustre_msg_get_status(req->rq_repmsg));
+ GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
}
/* Continue to close handle even if we can not pack reply */
med = &req->rq_export->exp_mdt_data;
spin_lock(&med->med_open_lock);
- mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
+ mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
+ req_is_replay(req));
if (mdt_mfd_closed(mfd)) {
spin_unlock(&med->med_open_lock);
CDEBUG(D_INODE, "no handle for file close: fid = "DFID
}
mdt_exit_ucred(info);
- if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
- RETURN(err_serious(-ENOMEM));
+ if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
+ GOTO(out, rc = err_serious(-ENOMEM));
- if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
- OBD_FAIL_MDS_CLOSE_NET_REP))
- info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
- RETURN(rc ? rc : ret);
+ if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
+ OBD_FAIL_MDS_CLOSE_NET_REP))
+ tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
+out:
+ mdt_thread_info_fini(info);
+ RETURN(rc ? rc : ret);
}
/**
* and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
* skip attributes and reconstruct the reply here.
*/
-int mdt_done_writing(struct mdt_thread_info *info)
+int mdt_done_writing(struct tgt_session_info *tsi)
{
- struct ptlrpc_request *req = mdt_info_req(info);
+ struct ptlrpc_request *req = tgt_ses_req(tsi);
+ struct mdt_thread_info *info = tsi2mdt_info(tsi);
struct mdt_body *repbody = NULL;
struct mdt_export_data *med;
struct mdt_file_data *mfd;
int rc;
ENTRY;
- rc = req_capsule_server_pack(info->mti_pill);
- if (rc)
- RETURN(err_serious(rc));
+ rc = req_capsule_server_pack(tsi->tsi_pill);
+ if (rc)
+ GOTO(out, rc = err_serious(rc));
- repbody = req_capsule_server_get(info->mti_pill,
- &RMF_MDT_BODY);
- repbody->eadatasize = 0;
- repbody->aclsize = 0;
+ repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
+ repbody->eadatasize = 0;
+ repbody->aclsize = 0;
- /* Done Writing may come with the Size-on-MDS update. Unpack it. */
- rc = mdt_close_unpack(info);
- if (rc)
- RETURN(err_serious(rc));
+ /* Done Writing may come with the Size-on-MDS update. Unpack it. */
+ rc = mdt_close_unpack(info);
+ if (rc)
+ GOTO(out, rc = err_serious(rc));
if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
mdt_exit_ucred(info);
- RETURN(lustre_msg_get_status(req->rq_repmsg));
+ GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
}
- med = &info->mti_exp->exp_mdt_data;
+ med = &info->mti_exp->exp_mdt_data;
spin_lock(&med->med_open_lock);
- mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
+ mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
+ req_is_replay(req));
if (mfd == NULL) {
spin_unlock(&med->med_open_lock);
CDEBUG(D_INODE, "no handle for done write: fid = "DFID
mdt_empty_transno(info, rc);
error_ucred:
mdt_exit_ucred(info);
+out:
+ mdt_thread_info_fini(info);
RETURN(rc);
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff