/**
* lock the part of the directory according to the hash of the name
* (lh->mlh_pdo_hash) in parallel directory lock.
*/
static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
struct mdt_lock_handle *lh,
struct mdt_object *obj, __u64 ibits)
{
struct ldlm_res_id *res_id = &info->mti_res_id;
struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
ldlm_policy_data_t *policy = &info->mti_policy;
int rc;
/*
* Finish res_id initializing by name hash marking part of
* directory which is taking modification.
*/
LASSERT(lh->mlh_pdo_hash != 0);
fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res_id);
memset(policy, 0, sizeof(*policy));
policy->l_inodebits.bits = ibits;
/*
* Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
* going to be sent to client. If it is - mdt_intent_policy() path will
* fix it up and turn FL_LOCAL flag off.
*/
rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
res_id, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
&info->mti_exp->exp_handle.h_cookie);
return rc;
}
/* if object is dying, pack the lov/llog data,
* parameter info->mti_attr should be valid at this point! */
int mdt_handle_last_unlink(struct mdt_thread_info *info, struct mdt_object *mo,
const struct md_attr *ma)
{
struct mdt_body *repbody;
const struct lu_attr *la = &ma->ma_attr;
int rc;
ENTRY;
repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
LASSERT(repbody != NULL);
if (ma->ma_valid & MA_INODE)
mdt_pack_attr2body(info, repbody, la, mdt_object_fid(mo));
if (ma->ma_valid & MA_LOV) {
__u32 mode;
if (mdt_object_exists(mo) < 0)
/* If it is a remote object, and we do not retrieve
* EA back unlink reg file*/
mode = S_IFREG;
else
mode = lu_object_attr(&mo->mot_obj.mo_lu);
LASSERT(ma->ma_lmm_size);
mdt_dump_lmm(D_INFO, ma->ma_lmm);
repbody->eadatasize = ma->ma_lmm_size;
if (S_ISREG(mode))
repbody->valid |= OBD_MD_FLEASIZE;
else if (S_ISDIR(mode))
repbody->valid |= OBD_MD_FLDIREA;
else
LBUG();
}
if (ma->ma_cookie_size && (ma->ma_valid & MA_COOKIE)) {
repbody->aclsize = ma->ma_cookie_size;
repbody->valid |= OBD_MD_FLCOOKIE;
}
if (info->mti_mdt->mdt_opts.mo_oss_capa &&
info->mti_exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
repbody->valid & OBD_MD_FLEASIZE) {
struct lustre_capa *capa;
capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
LASSERT(capa);
capa->lc_opc = CAPA_OPC_OSS_DESTROY;
rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
if (rc)
RETURN(rc);
repbody->valid |= OBD_MD_FLOSSCAPA;
}
RETURN(0);
}
/**
* Get version of object by fid.
*
* Return real version or ENOENT_VERSION if object doesn't exist
*/
static void mdt_obj_version_get(struct mdt_thread_info *info,
struct mdt_object *o, __u64 *version)
{
LASSERT(o);
if (mdt_object_exists(o) && !mdt_object_remote(o) &&
!mdt_object_obf(o))
*version = dt_version_get(info->mti_env, mdt_obj2dt(o));
else
*version = ENOENT_VERSION;
CDEBUG(D_INODE, "FID "DFID" version is "LPX64"\n",
PFID(mdt_object_fid(o)), *version);
}
/**
* Check HSM flags and add HS_DIRTY flag if relevant.
*
* A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
* and is not RELEASED.
*/
int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
struct md_attr *ma)
{
int rc;
ENTRY;
/* If the file was modified, add the dirty flag */
ma->ma_need = MA_HSM;
rc = mdt_attr_get_complex(info, mo, ma);
if (rc) {
CERROR("file attribute read error for "DFID": %d.\n",
PFID(mdt_object_fid(mo)), rc);
RETURN(rc);
}
/* If an up2date copy exists in the backend, add dirty flag */
if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
&& !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
ma->ma_hsm.mh_flags |= HS_DIRTY;
mdt_lock_reg_init(lh, LCK_PW);
rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
MDT_LOCAL_LOCK);
if (rc != 0)
RETURN(rc);
rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
if (rc)
CERROR("file attribute change error for "DFID": %d\n",
PFID(mdt_object_fid(mo)), rc);
mdt_object_unlock(info, mo, lh, rc);
}
RETURN(rc);
}
/* Partial request to create object only */
static int mdt_md_mkobj(struct mdt_thread_info *info)
{
struct mdt_device *mdt = info->mti_mdt;
struct mdt_object *o;
struct mdt_body *repbody;
struct md_attr *ma = &info->mti_attr;
int rc;
ENTRY;
DEBUG_REQ(D_INODE, mdt_info_req(info), "Partial create "DFID"",
PFID(info->mti_rr.rr_fid2));
repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
o = mdt_object_find(info->mti_env, mdt, info->mti_rr.rr_fid2,
MDT_OBJ_MAY_NOT_EXIST);
if (!IS_ERR(o)) {
struct md_object *next = mdt_object_child(o);
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
/*
* Cross-ref create can encounter already created obj in case of
* recovery, just get attr in that case.
*/
if (mdt_object_exists(o) == 1) {
rc = mo_attr_get(info->mti_env, next, ma);
} else {
/*
* Here, NO permission check for object_create,
* such check has been done on the original MDS.
*/
rc = mo_object_create(info->mti_env, next,
&info->mti_spec, ma);
}
if (rc == 0) {
/* Return fid & attr to client. */
if (ma->ma_valid & MA_INODE)
mdt_pack_attr2body(info, repbody, &ma->ma_attr,
mdt_object_fid(o));
}
mdt_object_put(info->mti_env, o);
} else
rc = PTR_ERR(o);
mdt_create_pack_capa(info, rc, o, repbody);
RETURN(rc);
}
/* if object is dying, pack the lov/llog data,
* parameter info->mti_attr should be valid at this point! */
int mdt_handle_last_unlink(struct mdt_thread_info *info, struct mdt_object *mo,
const struct md_attr *ma)
{
struct mdt_body *repbody;
const struct lu_attr *la = &ma->ma_attr;
int rc;
ENTRY;
repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
LASSERT(repbody != NULL);
if (ma->ma_valid & MA_INODE)
mdt_pack_attr2body(info, repbody, la, mdt_object_fid(mo));
if (ma->ma_valid & MA_LOV) {
CERROR("No need in LOV EA upon unlink\n");
dump_stack();
}
repbody->eadatasize = 0;
if (ma->ma_cookie_size && (ma->ma_valid & MA_COOKIE)) {
repbody->aclsize = ma->ma_cookie_size;
repbody->valid |= OBD_MD_FLCOOKIE;
}
if (info->mti_mdt->mdt_opts.mo_oss_capa &&
exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
repbody->valid & OBD_MD_FLEASIZE) {
struct lustre_capa *capa;
capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
LASSERT(capa);
capa->lc_opc = CAPA_OPC_OSS_DESTROY;
rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
if (rc)
RETURN(rc);
repbody->valid |= OBD_MD_FLOSSCAPA;
}
RETURN(0);
}
/*
* 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;
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));
/*
* 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);
}
if (mdt_object_obf(mp))
GOTO(put_parent, rc = -EPERM);
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)
GOTO(unlock_parent, rc);
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",
mdt2obd_dev(info->mti_mdt)->obd_name,
(char *)rr->rr_name, PFID(mdt_object_fid(mc)));
GOTO(unlock_parent, rc = -ENOENT);
}
CDEBUG(D_INFO, "%s: name %s: "DFID" is another MDT\n",
mdt2obd_dev(info->mti_mdt)->obd_name,
(char *)rr->rr_name, PFID(mdt_object_fid(mc)));
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",
mdt2obd_dev(info->mti_mdt)->obd_name,
-EPERM);
GOTO(unlock_parent, 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);
mdt_object_put(info->mti_env, mc);
GOTO(unlock_parent, 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);
mdt_object_unlock_put(info, mc, child_lh, rc);
GOTO(unlock_parent, rc = -EREMOTE);
} else if (info->mti_spec.sp_rm_entry) {
CERROR("%s: lfs rmdir should not be used on local dir %s\n",
mdt2obd_dev(info->mti_mdt)->obd_name,
(char *)rr->rr_name);
mdt_object_put(info->mti_env, mc);
GOTO(unlock_parent, rc = -EPERM);
}
rc = mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_FULL,
MDT_CROSS_LOCK);
if (rc != 0) {
mdt_object_put(info->mti_env, mc);
GOTO(unlock_parent, 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);
rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
mdt_object_child(mc), lname, ma);
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;
mdt_object_unlock_put(info, mc, child_lh, rc);
unlock_parent:
mdt_object_unlock(info, mp, parent_lh, rc);
put_parent:
mdt_object_put(info->mti_env, mp);
out:
return rc;
}
static int mdt_reint_setattr(struct mdt_thread_info *info,
struct mdt_lock_handle *lhc)
{
struct md_attr *ma = &info->mti_attr;
struct mdt_reint_record *rr = &info->mti_rr;
struct ptlrpc_request *req = mdt_info_req(info);
struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
struct mdt_file_data *mfd;
struct mdt_object *mo;
struct mdt_body *repbody;
int som_au, rc, rc2;
ENTRY;
DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
(unsigned int)ma->ma_attr.la_valid);
if (info->mti_dlm_req)
ldlm_request_cancel(req, info->mti_dlm_req, 0);
repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
if (IS_ERR(mo))
GOTO(out, rc = PTR_ERR(mo));
if (mdt_object_obf(mo))
GOTO(out_put, rc = -EPERM);
/* start a log jounal handle if needed */
if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
if ((ma->ma_attr.la_valid & LA_SIZE) ||
(rr->rr_flags & MRF_OPEN_TRUNC)) {
/* Check write access for the O_TRUNC case */
if (mdt_write_read(mo) < 0)
GOTO(out_put, rc = -ETXTBSY);
}
} else if (info->mti_ioepoch &&
(info->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
/* Truncate case. IOEpoch is opened. */
rc = mdt_write_get(mo);
if (rc)
GOTO(out_put, rc);
mfd = mdt_mfd_new();
if (mfd == NULL) {
mdt_write_put(mo);
GOTO(out_put, rc = -ENOMEM);
}
mdt_ioepoch_open(info, mo, 0);
repbody->ioepoch = mo->mot_ioepoch;
mdt_object_get(info->mti_env, mo);
mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC);
mfd->mfd_object = mo;
mfd->mfd_xid = req->rq_xid;
spin_lock(&med->med_open_lock);
cfs_list_add(&mfd->mfd_list, &med->med_open_head);
spin_unlock(&med->med_open_lock);
repbody->handle.cookie = mfd->mfd_handle.h_cookie;
}
som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE;
if (som_au) {
/* SOM Attribute update case. Find the proper mfd and update
* SOM attributes on the proper object. */
LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
LASSERT(info->mti_ioepoch);
spin_lock(&med->med_open_lock);
mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
if (mfd == NULL) {
spin_unlock(&med->med_open_lock);
CDEBUG(D_INODE, "no handle for file close: "
"fid = "DFID": cookie = "LPX64"\n",
PFID(info->mti_rr.rr_fid1),
info->mti_ioepoch->handle.cookie);
GOTO(out_put, rc = -ESTALE);
}
LASSERT(mfd->mfd_mode == MDS_FMODE_SOM);
LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE));
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 if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
LASSERT((ma->ma_valid & MA_LOV) == 0);
rc = mdt_attr_set(info, mo, ma, rr->rr_flags);
if (rc)
GOTO(out_put, rc);
} else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) {
struct lu_buf *buf = &info->mti_buf;
LASSERT(ma->ma_attr.la_valid == 0);
buf->lb_buf = ma->ma_lmm;
buf->lb_len = ma->ma_lmm_size;
rc = mo_xattr_set(info->mti_env, mdt_object_child(mo),
buf, XATTR_NAME_LOV, 0);
if (rc)
GOTO(out_put, rc);
} else
LBUG();
/* If file data is modified, add the dirty flag */
if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
rc = mdt_add_dirty_flag(info, mo, ma);
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
rc = mdt_attr_get_complex(info, mo, ma);
if (rc != 0)
GOTO(out_put, rc);
mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
if (info->mti_mdt->mdt_opts.mo_oss_capa &&
exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
S_ISREG(lu_object_attr(&mo->mot_obj.mo_lu)) &&
(ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
struct lustre_capa *capa;
capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
LASSERT(capa);
capa->lc_opc = CAPA_OPC_OSS_DEFAULT | CAPA_OPC_OSS_TRUNC;
rc = mo_capa_get(info->mti_env, mdt_object_child(mo), capa, 0);
if (rc)
GOTO(out_put, rc);
repbody->valid |= OBD_MD_FLOSSCAPA;
}
EXIT;
out_put:
mdt_object_put(info->mti_env, mo);
out:
if (rc == 0)
mdt_counter_incr(req, LPROC_MDT_SETATTR);
mdt_client_compatibility(info);
rc2 = mdt_fix_reply(info);
if (rc == 0)
rc = rc2;
return rc;
}
/*
* VBR: we save three versions in reply:
* 0 - parent. Check that parent version is the same during replay.
* 1 - name. Version of 'name' if file exists with the same name or
* ENOENT_VERSION, it is needed because file may appear due to missed replays.
* 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
* check.
*/
static int mdt_md_create(struct mdt_thread_info *info)
{
struct mdt_device *mdt = info->mti_mdt;
struct mdt_object *parent;
struct mdt_object *child;
struct mdt_lock_handle *lh;
struct mdt_body *repbody;
struct md_attr *ma = &info->mti_attr;
struct mdt_reint_record *rr = &info->mti_rr;
struct lu_name *lname;
int rc;
ENTRY;
DEBUG_REQ(D_INODE, mdt_info_req(info), "Create (%s->"DFID") in "DFID,
rr->rr_name, PFID(rr->rr_fid2), PFID(rr->rr_fid1));
repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
lh = &info->mti_lh[MDT_LH_PARENT];
mdt_lock_pdo_init(lh, LCK_PW, rr->rr_name, rr->rr_namelen);
parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
MDS_INODELOCK_UPDATE);
if (IS_ERR(parent))
RETURN(PTR_ERR(parent));
if (mdt_object_obf(parent))
GOTO(out_put_parent, rc = -EPERM);
rc = mdt_version_get_check_save(info, parent, 0);
if (rc)
GOTO(out_put_parent, rc);
/*
* Check child name version during replay.
* During create replay a file may exist with same name.
*/
lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
rc = mdt_lookup_version_check(info, parent, lname,
&info->mti_tmp_fid1, 1);
if (rc == 0)
GOTO(out_put_parent, rc = -EEXIST);
/* -ENOENT is expected here */
if (rc != -ENOENT)
GOTO(out_put_parent, rc);
/* save version of file name for replay, it must be ENOENT here */
mdt_enoent_version_save(info, 1);
child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
if (likely(!IS_ERR(child))) {
struct md_object *next = mdt_object_child(parent);
if (mdt_object_remote(child)) {
struct seq_server_site *ss;
struct lu_ucred *uc = mdt_ucred(info);
if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
CERROR("%s: Creating remote dir is only "
"permitted for administrator: rc = %d\n",
mdt2obd_dev(mdt)->obd_name, -EPERM);
GOTO(out_put_child, rc = -EPERM);
}
ss = mdt_seq_site(mdt);
if (ss->ss_node_id != 0 &&
mdt->mdt_enable_remote_dir == 0) {
CERROR("%s: remote dir is only permitted on"
" MDT0 or set_param"
" mdt.*.enable_remote_dir=1\n",
mdt2obd_dev(mdt)->obd_name);
GOTO(out_put_child, rc = -EPERM);
}
}
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
/* capa for cross-ref will be stored here */
ma->ma_capa = req_capsule_server_get(info->mti_pill,
&RMF_CAPA1);
LASSERT(ma->ma_capa);
mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
OBD_FAIL_MDS_REINT_CREATE_WRITE);
/* Version of child will be updated on disk. */
info->mti_mos = child;
rc = mdt_version_get_check_save(info, child, 2);
if (rc)
GOTO(out_put_child, rc);
/* Let lower layer know current lock mode. */
info->mti_spec.sp_cr_mode =
mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
/*
* Do not perform lookup sanity check. We know that name does
* not exist.
*/
info->mti_spec.sp_cr_lookup = 0;
info->mti_spec.sp_feat = &dt_directory_features;
rc = mdo_create(info->mti_env, next, lname,
mdt_object_child(child),
&info->mti_spec, ma);
if (rc == 0)
rc = mdt_attr_get_complex(info, child, ma);
if (rc == 0) {
/* Return fid & attr to client. */
if (ma->ma_valid & MA_INODE)
mdt_pack_attr2body(info, repbody, &ma->ma_attr,
mdt_object_fid(child));
}
out_put_child:
mdt_object_put(info->mti_env, child);
} else {
rc = PTR_ERR(child);
}
mdt_create_pack_capa(info, rc, child, repbody);
out_put_parent:
mdt_object_unlock_put(info, parent, lh, rc);
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;
}
/*
* VBR: we save three versions in reply:
* 0 - parent. Check that parent version is the same during replay.
* 1 - name. Version of 'name' if file exists with the same name or
* ENOENT_VERSION, it is needed because file may appear due to missed replays.
* 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
* check.
*/
static int mdt_md_create(struct mdt_thread_info *info)
{
struct mdt_device *mdt = info->mti_mdt;
struct mdt_object *parent;
struct mdt_object *child;
struct mdt_lock_handle *lh;
struct mdt_body *repbody;
struct md_attr *ma = &info->mti_attr;
struct mdt_reint_record *rr = &info->mti_rr;
struct lu_name *lname;
int rc;
ENTRY;
DEBUG_REQ(D_INODE, mdt_info_req(info), "Create (%s->"DFID") in "DFID,
rr->rr_name, PFID(rr->rr_fid2), PFID(rr->rr_fid1));
repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
lh = &info->mti_lh[MDT_LH_PARENT];
mdt_lock_pdo_init(lh, LCK_PW, rr->rr_name, rr->rr_namelen);
parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
MDS_INODELOCK_UPDATE, MDT_OBJ_MUST_EXIST);
if (IS_ERR(parent))
RETURN(PTR_ERR(parent));
rc = mdt_version_get_check_save(info, parent, 0);
if (rc)
GOTO(out_put_parent, rc);
/*
* Check child name version during replay.
* During create replay a file may exist with same name.
*/
lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
rc = mdt_lookup_version_check(info, parent, lname,
&info->mti_tmp_fid1, 1);
/* -ENOENT is expected here */
if (rc != 0 && rc != -ENOENT)
GOTO(out_put_parent, rc);
/* save version of file name for replay, it must be ENOENT here */
mdt_enoent_version_save(info, 1);
child = mdt_object_find(info->mti_env, mdt, rr->rr_fid2,
MDT_OBJ_MAY_NOT_EXIST);
if (likely(!IS_ERR(child))) {
struct md_object *next = mdt_object_child(parent);
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
/* capa for cross-ref will be stored here */
ma->ma_capa = req_capsule_server_get(info->mti_pill,
&RMF_CAPA1);
LASSERT(ma->ma_capa);
mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
OBD_FAIL_MDS_REINT_CREATE_WRITE);
/* Version of child will be updated on disk. */
info->mti_mos = child;
rc = mdt_version_get_check_save(info, child, 2);
if (rc)
GOTO(out_put_child, rc);
/* Let lower layer know current lock mode. */
info->mti_spec.sp_cr_mode =
mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
/*
* Do perform lookup sanity check. We do not know if name exists
* or not.
*/
info->mti_spec.sp_cr_lookup = 1;
info->mti_spec.sp_feat = &dt_directory_features;
rc = mdo_create(info->mti_env, next, lname,
mdt_object_child(child),
&info->mti_spec, ma);
if (rc == 0) {
/* Return fid & attr to client. */
if (ma->ma_valid & MA_INODE)
mdt_pack_attr2body(info, repbody, &ma->ma_attr,
mdt_object_fid(child));
}
out_put_child:
mdt_object_put(info->mti_env, child);
} else {
rc = PTR_ERR(child);
}
mdt_create_pack_capa(info, rc, child, repbody);
out_put_parent:
mdt_object_unlock_put(info, parent, lh, rc);
RETURN(rc);
}
