/**
* Lookup with version checking.
*
* This checks version of 'name'. Many reint functions uses 'name' for child not
* FID, therefore we need to get object by name and check its version.
*/
int mdt_lookup_version_check(struct mdt_thread_info *info,
struct mdt_object *p, struct lu_name *lname,
struct lu_fid *fid, int idx)
{
int rc, vbrc;
rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
&info->mti_spec);
/* Check version only during replay */
if (!req_is_replay(mdt_info_req(info)))
return rc;
info->mti_ver[idx] = ENOENT_VERSION;
if (rc == 0) {
struct mdt_object *child;
child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
if (likely(!IS_ERR(child))) {
mdt_obj_version_get(info, child, &info->mti_ver[idx]);
mdt_object_put(info->mti_env, child);
}
}
vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
return vbrc ? vbrc : rc;
}
static int mdt_rename_unpack(struct mdt_thread_info *info)
{
struct lu_ucred *uc = mdt_ucred(info);
struct mdt_rec_rename *rec;
struct md_attr *ma = &info->mti_attr;
struct lu_attr *attr = &info->mti_attr.ma_attr;
struct mdt_reint_record *rr = &info->mti_rr;
struct req_capsule *pill = info->mti_pill;
int rc;
ENTRY;
CLASSERT(sizeof(struct mdt_rec_rename) == sizeof(struct mdt_rec_reint));
rec = req_capsule_client_get(pill, &RMF_REC_REINT);
if (rec == NULL)
RETURN(-EFAULT);
/* This prior initialization is needed for old_init_ucred_reint() */
uc->uc_fsuid = rec->rn_fsuid;
uc->uc_fsgid = rec->rn_fsgid;
uc->uc_cap = rec->rn_cap;
uc->uc_suppgids[0] = rec->rn_suppgid1;
uc->uc_suppgids[1] = rec->rn_suppgid2;
attr->la_uid = rec->rn_fsuid;
attr->la_gid = rec->rn_fsgid;
rr->rr_fid1 = &rec->rn_fid1;
rr->rr_fid2 = &rec->rn_fid2;
attr->la_ctime = rec->rn_time;
attr->la_mtime = rec->rn_time;
/* rename_tgt contains the mode already */
attr->la_mode = rec->rn_mode;
attr->la_valid = LA_UID | LA_GID | LA_CTIME | LA_MTIME | LA_MODE;
if (req_capsule_get_size(pill, &RMF_CAPA1, RCL_CLIENT))
mdt_set_capainfo(info, 0, rr->rr_fid1,
req_capsule_client_get(pill, &RMF_CAPA1));
if (req_capsule_get_size(pill, &RMF_CAPA2, RCL_CLIENT))
mdt_set_capainfo(info, 1, rr->rr_fid2,
req_capsule_client_get(pill, &RMF_CAPA2));
rc = mdt_name_unpack(pill, &RMF_NAME, &rr->rr_name, 0);
if (rc < 0)
RETURN(rc);
rc = mdt_name_unpack(pill, &RMF_SYMTGT, &rr->rr_tgt_name, 0);
if (rc < 0)
RETURN(rc);
if (rec->rn_bias & MDS_VTX_BYPASS)
ma->ma_attr_flags |= MDS_VTX_BYPASS;
else
ma->ma_attr_flags &= ~MDS_VTX_BYPASS;
info->mti_spec.no_create = !!req_is_replay(mdt_info_req(info));
rc = mdt_dlmreq_unpack(info);
RETURN(rc);
}
/**
* Check version is correct.
*
* Should be called only during replay.
*/
static int mdt_version_check(struct ptlrpc_request *req,
__u64 version, int idx)
{
__u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
ENTRY;
if (!exp_connect_vbr(req->rq_export))
RETURN(0);
LASSERT(req_is_replay(req));
/** VBR: version is checked always because costs nothing */
LASSERT(idx < PTLRPC_NUM_VERSIONS);
/** Sanity check for malformed buffers */
if (pre_ver == NULL) {
CERROR("No versions in request buffer\n");
spin_lock(&req->rq_export->exp_lock);
req->rq_export->exp_vbr_failed = 1;
spin_unlock(&req->rq_export->exp_lock);
RETURN(-EOVERFLOW);
} else if (pre_ver[idx] != version) {
CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n",
pre_ver[idx], version);
spin_lock(&req->rq_export->exp_lock);
req->rq_export->exp_vbr_failed = 1;
spin_unlock(&req->rq_export->exp_lock);
RETURN(-EOVERFLOW);
}
RETURN(0);
}
static int mdt_rename_unpack(struct mdt_thread_info *info)
{
struct md_ucred *uc = mdt_ucred(info);
struct mdt_rec_rename *rec;
struct md_attr *ma = &info->mti_attr;
struct lu_attr *attr = &info->mti_attr.ma_attr;
struct mdt_reint_record *rr = &info->mti_rr;
struct req_capsule *pill = info->mti_pill;
int rc;
ENTRY;
CLASSERT(sizeof(struct mdt_rec_rename) == sizeof(struct mdt_rec_reint));
rec = req_capsule_client_get(pill, &RMF_REC_REINT);
if (rec == NULL)
RETURN(-EFAULT);
uc->mu_fsuid = rec->rn_fsuid;
uc->mu_fsgid = rec->rn_fsgid;
uc->mu_cap = rec->rn_cap;
uc->mu_suppgids[0] = rec->rn_suppgid1;
uc->mu_suppgids[1] = rec->rn_suppgid2;
attr->la_uid = rec->rn_fsuid;
attr->la_gid = rec->rn_fsgid;
rr->rr_fid1 = &rec->rn_fid1;
rr->rr_fid2 = &rec->rn_fid2;
attr->la_ctime = rec->rn_time;
attr->la_mtime = rec->rn_time;
/* rename_tgt contains the mode already */
attr->la_mode = rec->rn_mode;
attr->la_valid = LA_UID | LA_GID | LA_CTIME | LA_MTIME | LA_MODE;
if (req_capsule_get_size(pill, &RMF_CAPA1, RCL_CLIENT))
mdt_set_capainfo(info, 0, rr->rr_fid1,
req_capsule_client_get(pill, &RMF_CAPA1));
if (req_capsule_get_size(pill, &RMF_CAPA2, RCL_CLIENT))
mdt_set_capainfo(info, 1, rr->rr_fid2,
req_capsule_client_get(pill, &RMF_CAPA2));
info->mti_spec.sp_ck_split = !!(rec->rn_bias & MDS_CHECK_SPLIT);
info->mti_cross_ref = !!(rec->rn_bias & MDS_CROSS_REF);
rr->rr_name = req_capsule_client_get(pill, &RMF_NAME);
rr->rr_tgt = req_capsule_client_get(pill, &RMF_SYMTGT);
if (rr->rr_name == NULL || rr->rr_tgt == NULL)
RETURN(-EFAULT);
rr->rr_namelen = req_capsule_get_size(pill, &RMF_NAME, RCL_CLIENT) - 1;
rr->rr_tgtlen = req_capsule_get_size(pill, &RMF_SYMTGT, RCL_CLIENT) - 1;
if (!info->mti_cross_ref)
LASSERT(rr->rr_namelen > 0 && rr->rr_tgtlen > 0);
if (rec->rn_bias & MDS_VTX_BYPASS)
ma->ma_attr_flags |= MDS_VTX_BYPASS;
else
ma->ma_attr_flags &= ~MDS_VTX_BYPASS;
info->mti_spec.no_create = !!req_is_replay(mdt_info_req(info));
rc = mdt_dlmreq_unpack(info);
RETURN(rc);
}
/**
* Save enoent version, it is needed when it is obvious that object doesn't
* exist, e.g. child during create.
*/
static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
{
/* save version of file name for replay, it must be ENOENT here */
if (!req_is_replay(mdt_info_req(info))) {
info->mti_ver[idx] = ENOENT_VERSION;
mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
}
}
/**
* Get version from disk and save in reply buffer.
*
* Versions are saved in reply only during normal operations not replays.
*/
void mdt_version_get_save(struct mdt_thread_info *info,
struct mdt_object *mto, int idx)
{
/* don't save versions during replay */
if (!req_is_replay(mdt_info_req(info))) {
mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
}
}
/**
* Get version from disk and check it, no save in reply.
*/
int mdt_version_get_check(struct mdt_thread_info *info,
struct mdt_object *mto, int idx)
{
/* only check versions during replay */
if (!req_is_replay(mdt_info_req(info)))
return 0;
mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
}
/**
* Get version from disk and check if recovery or just save.
*/
int mdt_version_get_check_save(struct mdt_thread_info *info,
struct mdt_object *mto, int idx)
{
int rc = 0;
mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
if (req_is_replay(mdt_info_req(info)))
rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
idx);
else
mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
return rc;
}
/**
* Save pre-versions in reply.
*/
static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
int idx)
{
__u64 *reply_ver;
if (!exp_connect_vbr(req->rq_export))
return;
LASSERT(!req_is_replay(req));
LASSERT(req->rq_repmsg != NULL);
reply_ver = lustre_msg_get_versions(req->rq_repmsg);
if (reply_ver)
reply_ver[idx] = version;
}
/*
* please see comment above LOV_MAGIC_V1_DEF
*/
static void mdt_fix_lov_magic(struct mdt_thread_info *info)
{
struct mdt_reint_record *rr = &info->mti_rr;
struct lov_user_md_v1 *v1;
v1 = (void *)rr->rr_eadata;
LASSERT(v1);
if (unlikely(req_is_replay(mdt_info_req(info)))) {
if (v1->lmm_magic == LOV_USER_MAGIC_V1) {
v1->lmm_magic = LOV_MAGIC_V1_DEF;
} else if (v1->lmm_magic == __swab32(LOV_USER_MAGIC_V1)) {
v1->lmm_magic = __swab32(LOV_MAGIC_V1_DEF);
} else if (v1->lmm_magic == LOV_USER_MAGIC_V3) {
v1->lmm_magic = LOV_MAGIC_V3_DEF;
} else if (v1->lmm_magic == __swab32(LOV_USER_MAGIC_V3)) {
v1->lmm_magic = __swab32(LOV_MAGIC_V3_DEF);
}
}
}
static int mdt_open_unpack(struct mdt_thread_info *info)
{
struct md_ucred *uc = mdt_ucred(info);
struct mdt_rec_create *rec;
struct lu_attr *attr = &info->mti_attr.ma_attr;
struct req_capsule *pill = info->mti_pill;
struct mdt_reint_record *rr = &info->mti_rr;
struct ptlrpc_request *req = mdt_info_req(info);
struct md_op_spec *sp = &info->mti_spec;
ENTRY;
CLASSERT(sizeof(struct mdt_rec_create) == sizeof(struct mdt_rec_reint));
rec = req_capsule_client_get(pill, &RMF_REC_REINT);
if (rec == NULL)
RETURN(-EFAULT);
uc->mu_fsuid = rec->cr_fsuid;
uc->mu_fsgid = rec->cr_fsgid;
uc->mu_cap = rec->cr_cap;
uc->mu_suppgids[0] = rec->cr_suppgid1;
uc->mu_suppgids[1] = rec->cr_suppgid2;
rr->rr_fid1 = &rec->cr_fid1;
rr->rr_fid2 = &rec->cr_fid2;
rr->rr_handle = &rec->cr_old_handle;
attr->la_mode = rec->cr_mode;
attr->la_rdev = rec->cr_rdev;
attr->la_uid = rec->cr_fsuid;
attr->la_gid = rec->cr_fsgid;
attr->la_ctime = rec->cr_time;
attr->la_mtime = rec->cr_time;
attr->la_atime = rec->cr_time;
attr->la_valid = LA_MODE | LA_RDEV | LA_UID | LA_GID |
LA_CTIME | LA_MTIME | LA_ATIME;
memset(&info->mti_spec.u, 0, sizeof(info->mti_spec.u));
info->mti_spec.sp_cr_flags = get_mrc_cr_flags(rec);
/* Do not trigger ASSERTION if client miss to set such flags. */
if (unlikely(info->mti_spec.sp_cr_flags == 0))
RETURN(-EPROTO);
info->mti_replayepoch = rec->cr_ioepoch;
info->mti_spec.sp_ck_split = !!(rec->cr_bias & MDS_CHECK_SPLIT);
info->mti_cross_ref = !!(rec->cr_bias & MDS_CROSS_REF);
if (req_capsule_get_size(pill, &RMF_CAPA1, RCL_CLIENT))
mdt_set_capainfo(info, 0, rr->rr_fid1,
req_capsule_client_get(pill, &RMF_CAPA1));
if (req_is_replay(req) &&
req_capsule_get_size(pill, &RMF_CAPA2, RCL_CLIENT)) {
#if 0
mdt_set_capainfo(info, 1, rr->rr_fid2,
req_capsule_client_get(pill, &RMF_CAPA2));
#else
/*
* FIXME: capa in replay open request might have expired,
* bypass capa check. Security hole?
*/
mdt_set_capainfo(info, 0, rr->rr_fid1, BYPASS_CAPA);
mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
#endif
}
rr->rr_name = req_capsule_client_get(pill, &RMF_NAME);
if (rr->rr_name == NULL)
RETURN(-EFAULT);
rr->rr_namelen = req_capsule_get_size(pill, &RMF_NAME, RCL_CLIENT) - 1;
sp->u.sp_ea.eadatalen = req_capsule_get_size(pill, &RMF_EADATA,
RCL_CLIENT);
if (sp->u.sp_ea.eadatalen) {
sp->u.sp_ea.eadata = req_capsule_client_get(pill, &RMF_EADATA);
sp->no_create = !!req_is_replay(req);
}
RETURN(0);
}
static int mdt_open_unpack(struct mdt_thread_info *info)
{
struct lu_ucred *uc = mdt_ucred(info);
struct mdt_rec_create *rec;
struct lu_attr *attr = &info->mti_attr.ma_attr;
struct req_capsule *pill = info->mti_pill;
struct mdt_reint_record *rr = &info->mti_rr;
struct ptlrpc_request *req = mdt_info_req(info);
struct md_op_spec *sp = &info->mti_spec;
ENTRY;
CLASSERT(sizeof(struct mdt_rec_create) == sizeof(struct mdt_rec_reint));
rec = req_capsule_client_get(pill, &RMF_REC_REINT);
if (rec == NULL)
RETURN(-EFAULT);
/* This prior initialization is needed for old_init_ucred_reint() */
uc->uc_fsuid = rec->cr_fsuid;
uc->uc_fsgid = rec->cr_fsgid;
uc->uc_cap = rec->cr_cap;
uc->uc_suppgids[0] = rec->cr_suppgid1;
uc->uc_suppgids[1] = rec->cr_suppgid2;
uc->uc_umask = rec->cr_umask;
rr->rr_fid1 = &rec->cr_fid1;
rr->rr_fid2 = &rec->cr_fid2;
rr->rr_handle = &rec->cr_old_handle;
attr->la_mode = rec->cr_mode;
attr->la_rdev = rec->cr_rdev;
attr->la_uid = rec->cr_fsuid;
attr->la_gid = rec->cr_fsgid;
attr->la_ctime = rec->cr_time;
attr->la_mtime = rec->cr_time;
attr->la_atime = rec->cr_time;
attr->la_valid = LA_MODE | LA_RDEV | LA_UID | LA_GID |
LA_CTIME | LA_MTIME | LA_ATIME;
memset(&info->mti_spec.u, 0, sizeof(info->mti_spec.u));
info->mti_spec.sp_cr_flags = get_mrc_cr_flags(rec);
/* Do not trigger ASSERTION if client miss to set such flags. */
if (unlikely(info->mti_spec.sp_cr_flags == 0))
RETURN(-EPROTO);
info->mti_replayepoch = rec->cr_ioepoch;
info->mti_cross_ref = !!(rec->cr_bias & MDS_CROSS_REF);
if (req_capsule_get_size(pill, &RMF_CAPA1, RCL_CLIENT))
mdt_set_capainfo(info, 0, rr->rr_fid1,
req_capsule_client_get(pill, &RMF_CAPA1));
if (req_is_replay(req) &&
req_capsule_get_size(pill, &RMF_CAPA2, RCL_CLIENT)) {
#if 0
mdt_set_capainfo(info, 1, rr->rr_fid2,
req_capsule_client_get(pill, &RMF_CAPA2));
#else
/*
* FIXME: capa in replay open request might have expired,
* bypass capa check. Security hole?
*/
mdt_set_capainfo(info, 0, rr->rr_fid1, BYPASS_CAPA);
mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
#endif
}
mdt_name_unpack(pill, &RMF_NAME, &rr->rr_name, MNF_FIX_ANON);
if (req_capsule_field_present(pill, &RMF_EADATA, RCL_CLIENT)) {
rr->rr_eadatalen = req_capsule_get_size(pill, &RMF_EADATA,
RCL_CLIENT);
if (rr->rr_eadatalen > 0) {
rr->rr_eadata = req_capsule_client_get(pill,
&RMF_EADATA);
sp->u.sp_ea.eadatalen = rr->rr_eadatalen;
sp->u.sp_ea.eadata = rr->rr_eadata;
sp->no_create = !!req_is_replay(req);
mdt_fix_lov_magic(info);
}
/*
* Client default md_size may be 0 right after client start,
* until all osc are connected, set here just some reasonable
* value to prevent misbehavior.
*/
if (rr->rr_eadatalen == 0 &&
!(info->mti_spec.sp_cr_flags & MDS_OPEN_DELAY_CREATE))
rr->rr_eadatalen = MIN_MD_SIZE;
}
RETURN(0);
}
/**
* Object updates between Targets. Because all the updates has been
* dis-assemblied into object updates at sender side, so OUT will
* call OSD API directly to execute these updates.
*
* In DNE phase I all of the updates in the request need to be executed
* in one transaction, and the transaction has to be synchronously.
*
* Please refer to lustre/include/lustre/lustre_idl.h for req/reply
* format.
*/
int out_handle(struct tgt_session_info *tsi)
{
const struct lu_env *env = tsi->tsi_env;
struct tgt_thread_info *tti = tgt_th_info(env);
struct thandle_exec_args *ta = &tti->tti_tea;
struct req_capsule *pill = tsi->tsi_pill;
struct dt_device *dt = tsi->tsi_tgt->lut_bottom;
struct object_update_request *ureq;
struct object_update *update;
struct object_update_reply *reply;
int bufsize;
int count;
int old_batchid = -1;
int i;
int rc = 0;
int rc1 = 0;
ENTRY;
req_capsule_set(pill, &RQF_OUT_UPDATE);
ureq = req_capsule_client_get(pill, &RMF_OUT_UPDATE);
if (ureq == NULL) {
CERROR("%s: No buf!: rc = %d\n", tgt_name(tsi->tsi_tgt),
-EPROTO);
RETURN(err_serious(-EPROTO));
}
bufsize = req_capsule_get_size(pill, &RMF_OUT_UPDATE, RCL_CLIENT);
if (bufsize != object_update_request_size(ureq)) {
CERROR("%s: invalid bufsize %d: rc = %d\n",
tgt_name(tsi->tsi_tgt), bufsize, -EPROTO);
RETURN(err_serious(-EPROTO));
}
if (ureq->ourq_magic != UPDATE_REQUEST_MAGIC) {
CERROR("%s: invalid update buffer magic %x expect %x: "
"rc = %d\n", tgt_name(tsi->tsi_tgt), ureq->ourq_magic,
UPDATE_REQUEST_MAGIC, -EPROTO);
RETURN(err_serious(-EPROTO));
}
count = ureq->ourq_count;
if (count <= 0) {
CERROR("%s: empty update: rc = %d\n", tgt_name(tsi->tsi_tgt),
-EPROTO);
RETURN(err_serious(-EPROTO));
}
req_capsule_set_size(pill, &RMF_OUT_UPDATE_REPLY, RCL_SERVER,
OUT_UPDATE_REPLY_SIZE);
rc = req_capsule_server_pack(pill);
if (rc != 0) {
CERROR("%s: Can't pack response: rc = %d\n",
tgt_name(tsi->tsi_tgt), rc);
RETURN(rc);
}
/* Prepare the update reply buffer */
reply = req_capsule_server_get(pill, &RMF_OUT_UPDATE_REPLY);
if (reply == NULL)
RETURN(err_serious(-EPROTO));
object_update_reply_init(reply, count);
tti->tti_u.update.tti_update_reply = reply;
rc = out_tx_start(env, dt, ta, tsi->tsi_exp);
if (rc != 0)
RETURN(rc);
tti->tti_mult_trans = !req_is_replay(tgt_ses_req(tsi));
/* Walk through updates in the request to execute them synchronously */
for (i = 0; i < count; i++) {
struct tgt_handler *h;
struct dt_object *dt_obj;
update = object_update_request_get(ureq, i, NULL);
if (update == NULL)
GOTO(out, rc = -EPROTO);
if (ptlrpc_req_need_swab(pill->rc_req))
lustre_swab_object_update(update);
if (old_batchid == -1) {
old_batchid = update->ou_batchid;
} else if (old_batchid != update->ou_batchid) {
/* Stop the current update transaction,
* create a new one */
rc = out_tx_end(env, ta);
if (rc < 0)
RETURN(rc);
rc = out_tx_start(env, dt, ta, tsi->tsi_exp);
if (rc != 0)
RETURN(rc);
old_batchid = update->ou_batchid;
}
if (!fid_is_sane(&update->ou_fid)) {
CERROR("%s: invalid FID "DFID": rc = %d\n",
tgt_name(tsi->tsi_tgt), PFID(&update->ou_fid),
-EPROTO);
GOTO(out, rc = err_serious(-EPROTO));
}
dt_obj = dt_locate(env, dt, &update->ou_fid);
if (IS_ERR(dt_obj))
GOTO(out, rc = PTR_ERR(dt_obj));
if (dt->dd_record_fid_accessed) {
lfsck_pack_rfa(&tti->tti_lr,
lu_object_fid(&dt_obj->do_lu));
tgt_lfsck_in_notify(env, dt, &tti->tti_lr);
}
tti->tti_u.update.tti_dt_object = dt_obj;
tti->tti_u.update.tti_update = update;
tti->tti_u.update.tti_update_reply_index = i;
h = out_handler_find(update->ou_type);
if (likely(h != NULL)) {
/* For real modification RPC, check if the update
* has been executed */
if (h->th_flags & MUTABOR) {
struct ptlrpc_request *req = tgt_ses_req(tsi);
if (out_check_resent(env, dt, dt_obj, req,
out_reconstruct, reply, i))
GOTO(next, rc);
}
rc = h->th_act(tsi);
} else {
CERROR("%s: The unsupported opc: 0x%x\n",
tgt_name(tsi->tsi_tgt), update->ou_type);
lu_object_put(env, &dt_obj->do_lu);
GOTO(out, rc = -ENOTSUPP);
}
next:
lu_object_put(env, &dt_obj->do_lu);
if (rc < 0)
GOTO(out, rc);
}
out:
rc1 = out_tx_end(env, ta);
if (rc == 0)
rc = rc1;
RETURN(rc);
}
/**
* Object updates between Targets. Because all the updates has been
* dis-assemblied into object updates at sender side, so OUT will
* call OSD API directly to execute these updates.
*
* In DNE phase I all of the updates in the request need to be executed
* in one transaction, and the transaction has to be synchronously.
*
* Please refer to lustre/include/lustre/lustre_idl.h for req/reply
* format.
*/
int out_handle(struct tgt_session_info *tsi)
{
const struct lu_env *env = tsi->tsi_env;
struct tgt_thread_info *tti = tgt_th_info(env);
struct thandle_exec_args *ta = &tti->tti_tea;
struct req_capsule *pill = tsi->tsi_pill;
struct dt_device *dt = tsi->tsi_tgt->lut_bottom;
struct out_update_header *ouh;
struct out_update_buffer *oub = NULL;
struct object_update *update;
struct object_update_reply *reply;
struct ptlrpc_bulk_desc *desc = NULL;
struct l_wait_info lwi;
void **update_bufs;
int current_batchid = -1;
__u32 update_buf_count;
unsigned int i;
unsigned int reply_index = 0;
int rc = 0;
int rc1 = 0;
int ouh_size, reply_size;
int updates;
ENTRY;
req_capsule_set(pill, &RQF_OUT_UPDATE);
ouh_size = req_capsule_get_size(pill, &RMF_OUT_UPDATE_HEADER,
RCL_CLIENT);
if (ouh_size <= 0)
RETURN(err_serious(-EPROTO));
ouh = req_capsule_client_get(pill, &RMF_OUT_UPDATE_HEADER);
if (ouh == NULL)
RETURN(err_serious(-EPROTO));
if (ouh->ouh_magic != OUT_UPDATE_HEADER_MAGIC) {
CERROR("%s: invalid update buffer magic %x expect %x: "
"rc = %d\n", tgt_name(tsi->tsi_tgt), ouh->ouh_magic,
UPDATE_REQUEST_MAGIC, -EPROTO);
RETURN(err_serious(-EPROTO));
}
update_buf_count = ouh->ouh_count;
if (update_buf_count == 0)
RETURN(err_serious(-EPROTO));
OBD_ALLOC(update_bufs, sizeof(*update_bufs) * update_buf_count);
if (update_bufs == NULL)
RETURN(err_serious(-ENOMEM));
if (ouh->ouh_inline_length > 0) {
update_bufs[0] = ouh->ouh_inline_data;
} else {
struct out_update_buffer *tmp;
oub = req_capsule_client_get(pill, &RMF_OUT_UPDATE_BUF);
if (oub == NULL)
GOTO(out_free, rc = err_serious(-EPROTO));
desc = ptlrpc_prep_bulk_exp(pill->rc_req, update_buf_count,
PTLRPC_BULK_OPS_COUNT,
PTLRPC_BULK_GET_SINK |
PTLRPC_BULK_BUF_KVEC,
MDS_BULK_PORTAL,
&ptlrpc_bulk_kvec_ops);
if (desc == NULL)
GOTO(out_free, rc = err_serious(-ENOMEM));
tmp = oub;
for (i = 0; i < update_buf_count; i++, tmp++) {
if (tmp->oub_size >= OUT_MAXREQSIZE)
GOTO(out_free, rc = err_serious(-EPROTO));
OBD_ALLOC(update_bufs[i], tmp->oub_size);
if (update_bufs[i] == NULL)
GOTO(out_free, rc = err_serious(-ENOMEM));
desc->bd_frag_ops->add_iov_frag(desc, update_bufs[i],
tmp->oub_size);
}
pill->rc_req->rq_bulk_write = 1;
rc = sptlrpc_svc_prep_bulk(pill->rc_req, desc);
if (rc != 0)
GOTO(out_free, rc = err_serious(rc));
rc = target_bulk_io(pill->rc_req->rq_export, desc, &lwi);
if (rc < 0)
GOTO(out_free, rc = err_serious(rc));
}
/* validate the request and calculate the total update count and
* set it to reply */
reply_size = 0;
updates = 0;
for (i = 0; i < update_buf_count; i++) {
struct object_update_request *our;
int j;
our = update_bufs[i];
if (ptlrpc_req_need_swab(pill->rc_req))
lustre_swab_object_update_request(our);
if (our->ourq_magic != UPDATE_REQUEST_MAGIC) {
CERROR("%s: invalid update buffer magic %x"
" expect %x: rc = %d\n",
tgt_name(tsi->tsi_tgt), our->ourq_magic,
UPDATE_REQUEST_MAGIC, -EPROTO);
GOTO(out_free, rc = err_serious(-EPROTO));
}
updates += our->ourq_count;
/* need to calculate reply size */
for (j = 0; j < our->ourq_count; j++) {
update = object_update_request_get(our, j, NULL);
if (update == NULL)
GOTO(out, rc = err_serious(-EPROTO));
if (ptlrpc_req_need_swab(pill->rc_req))
lustre_swab_object_update(update);
if (!fid_is_sane(&update->ou_fid)) {
CERROR("%s: invalid FID "DFID": rc = %d\n",
tgt_name(tsi->tsi_tgt),
PFID(&update->ou_fid), -EPROTO);
GOTO(out, rc = err_serious(-EPROTO));
}
/* XXX: what ou_result_size can be considered safe? */
reply_size += sizeof(reply->ourp_lens[0]);
reply_size += sizeof(struct object_update_result);
reply_size += update->ou_result_size;
}
}
reply_size += sizeof(*reply);
if (unlikely(reply_size > ouh->ouh_reply_size)) {
CERROR("%s: too small reply buf %u for %u, need %u at least\n",
tgt_name(tsi->tsi_tgt), ouh->ouh_reply_size,
updates, reply_size);
GOTO(out_free, rc = err_serious(-EPROTO));
}
req_capsule_set_size(pill, &RMF_OUT_UPDATE_REPLY, RCL_SERVER,
ouh->ouh_reply_size);
rc = req_capsule_server_pack(pill);
if (rc != 0) {
CERROR("%s: Can't pack response: rc = %d\n",
tgt_name(tsi->tsi_tgt), rc);
GOTO(out_free, rc = err_serious(-EPROTO));
}
/* Prepare the update reply buffer */
reply = req_capsule_server_get(pill, &RMF_OUT_UPDATE_REPLY);
if (reply == NULL)
GOTO(out_free, rc = -EPROTO);
reply->ourp_magic = UPDATE_REPLY_MAGIC;
reply->ourp_count = updates;
tti->tti_u.update.tti_update_reply = reply;
tti->tti_mult_trans = !req_is_replay(tgt_ses_req(tsi));
/* Walk through updates in the request to execute them */
for (i = 0; i < update_buf_count; i++) {
struct tgt_handler *h;
struct dt_object *dt_obj;
int update_count;
struct object_update_request *our;
int j;
our = update_bufs[i];
update_count = our->ourq_count;
for (j = 0; j < update_count; j++) {
update = object_update_request_get(our, j, NULL);
dt_obj = dt_locate(env, dt, &update->ou_fid);
if (IS_ERR(dt_obj))
GOTO(out, rc = PTR_ERR(dt_obj));
if (dt->dd_record_fid_accessed) {
lfsck_pack_rfa(&tti->tti_lr,
lu_object_fid(&dt_obj->do_lu),
LE_FID_ACCESSED,
LFSCK_TYPE_LAYOUT);
tgt_lfsck_in_notify(env, dt, &tti->tti_lr,
NULL);
}
tti->tti_u.update.tti_dt_object = dt_obj;
tti->tti_u.update.tti_update = update;
tti->tti_u.update.tti_update_reply_index = reply_index;
h = out_handler_find(update->ou_type);
if (unlikely(h == NULL)) {
CERROR("%s: unsupported opc: 0x%x\n",
tgt_name(tsi->tsi_tgt), update->ou_type);
GOTO(next, rc = -ENOTSUPP);
}
/* Check resend case only for modifying RPC */
if (h->th_flags & MUTABOR) {
struct ptlrpc_request *req = tgt_ses_req(tsi);
if (out_check_resent(env, dt, dt_obj, req,
out_reconstruct, reply,
reply_index))
GOTO(next, rc = 0);
}
/* start transaction for modification RPC only */
if (h->th_flags & MUTABOR && current_batchid == -1) {
current_batchid = update->ou_batchid;
rc = out_tx_start(env, dt, ta, tsi->tsi_exp);
if (rc != 0)
GOTO(next, rc);
if (update->ou_flags & UPDATE_FL_SYNC)
ta->ta_handle->th_sync = 1;
}
/* Stop the current update transaction, if the update
* has different batchid, or read-only update */
if (((current_batchid != update->ou_batchid) ||
!(h->th_flags & MUTABOR)) &&
ta->ta_handle != NULL) {
rc = out_tx_end(env, ta, rc);
current_batchid = -1;
if (rc != 0)
GOTO(next, rc);
/* start a new transaction if needed */
if (h->th_flags & MUTABOR) {
rc = out_tx_start(env, dt, ta,
tsi->tsi_exp);
if (rc != 0)
GOTO(next, rc);
if (update->ou_flags & UPDATE_FL_SYNC)
ta->ta_handle->th_sync = 1;
current_batchid = update->ou_batchid;
}
}
rc = h->th_act(tsi);
next:
reply_index++;
lu_object_put(env, &dt_obj->do_lu);
if (rc < 0)
GOTO(out, rc);
}
}
out:
if (current_batchid != -1) {
rc1 = out_tx_end(env, ta, rc);
if (rc == 0)
rc = rc1;
}
out_free:
if (update_bufs != NULL) {
if (oub != NULL) {
for (i = 0; i < update_buf_count; i++, oub++) {
if (update_bufs[i] != NULL)
OBD_FREE(update_bufs[i], oub->oub_size);
}
}
OBD_FREE(update_bufs, sizeof(*update_bufs) * update_buf_count);
}
if (desc != NULL)
ptlrpc_free_bulk(desc);
RETURN(rc);
}
/*
* VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
* name.
*/
static int mdt_reint_link(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 mdt_object *ms;
struct mdt_object *mp;
struct mdt_lock_handle *lhs;
struct mdt_lock_handle *lhp;
struct lu_name *lname;
int rc;
ENTRY;
DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/%s",
PFID(rr->rr_fid1), PFID(rr->rr_fid2), rr->rr_name);
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
RETURN(err_serious(-ENOENT));
if (info->mti_dlm_req)
ldlm_request_cancel(req, info->mti_dlm_req, 0);
/* Invalid case so return error immediately instead of
* processing it */
if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
RETURN(-EPERM);
/* step 1: find & lock the target parent dir */
lhp = &info->mti_lh[MDT_LH_PARENT];
mdt_lock_pdo_init(lhp, LCK_PW, rr->rr_name,
rr->rr_namelen);
mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
MDS_INODELOCK_UPDATE);
if (IS_ERR(mp))
RETURN(PTR_ERR(mp));
if (mdt_object_obf(mp))
GOTO(out_unlock_parent, rc = -EPERM);
rc = mdt_version_get_check_save(info, mp, 0);
if (rc)
GOTO(out_unlock_parent, rc);
/* step 2: find & lock the source */
lhs = &info->mti_lh[MDT_LH_CHILD];
mdt_lock_reg_init(lhs, LCK_EX);
ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
if (IS_ERR(ms))
GOTO(out_unlock_parent, rc = PTR_ERR(ms));
if (mdt_object_remote(ms)) {
mdt_object_put(info->mti_env, ms);
CERROR("Target directory "DFID" is on another MDT\n",
PFID(rr->rr_fid1));
GOTO(out_unlock_parent, rc = -EXDEV);
}
rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE,
MDT_CROSS_LOCK);
if (rc != 0) {
mdt_object_put(info->mti_env, ms);
GOTO(out_unlock_parent, rc);
}
/* step 3: link it */
mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
OBD_FAIL_MDS_REINT_LINK_WRITE);
info->mti_mos = ms;
rc = mdt_version_get_check_save(info, ms, 1);
if (rc)
GOTO(out_unlock_child, rc);
lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
/** check target version by name during replay */
rc = mdt_lookup_version_check(info, mp, lname, &info->mti_tmp_fid1, 2);
if (rc != 0 && rc != -ENOENT)
GOTO(out_unlock_child, rc);
/* save version of file name for replay, it must be ENOENT here */
if (!req_is_replay(mdt_info_req(info))) {
info->mti_ver[2] = ENOENT_VERSION;
mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
}
rc = mdo_link(info->mti_env, mdt_object_child(mp),
mdt_object_child(ms), lname, ma);
if (rc == 0)
mdt_counter_incr(req, LPROC_MDT_LINK);
EXIT;
out_unlock_child:
mdt_object_unlock_put(info, ms, lhs, rc);
out_unlock_parent:
mdt_object_unlock_put(info, mp, lhp, rc);
return rc;
}
/**
* Object updates between Targets. Because all the updates has been
* dis-assemblied into object updates at sender side, so OUT will
* call OSD API directly to execute these updates.
*
* In DNE phase I all of the updates in the request need to be executed
* in one transaction, and the transaction has to be synchronously.
*
* Please refer to lustre/include/lustre/lustre_idl.h for req/reply
* format.
*/
int out_handle(struct tgt_session_info *tsi)
{
const struct lu_env *env = tsi->tsi_env;
struct tgt_thread_info *tti = tgt_th_info(env);
struct thandle_exec_args *ta = &tti->tti_tea;
struct req_capsule *pill = tsi->tsi_pill;
struct dt_device *dt = tsi->tsi_tgt->lut_bottom;
struct update_buf *ubuf;
struct update *update;
struct update_reply *update_reply;
int bufsize;
int count;
int old_batchid = -1;
unsigned off;
int i;
int rc = 0;
int rc1 = 0;
ENTRY;
req_capsule_set(pill, &RQF_UPDATE_OBJ);
bufsize = req_capsule_get_size(pill, &RMF_UPDATE, RCL_CLIENT);
if (bufsize != UPDATE_BUFFER_SIZE) {
CERROR("%s: invalid bufsize %d: rc = %d\n",
tgt_name(tsi->tsi_tgt), bufsize, -EPROTO);
RETURN(err_serious(-EPROTO));
}
ubuf = req_capsule_client_get(pill, &RMF_UPDATE);
if (ubuf == NULL) {
CERROR("%s: No buf!: rc = %d\n", tgt_name(tsi->tsi_tgt),
-EPROTO);
RETURN(err_serious(-EPROTO));
}
if (le32_to_cpu(ubuf->ub_magic) != UPDATE_BUFFER_MAGIC) {
CERROR("%s: invalid magic %x expect %x: rc = %d\n",
tgt_name(tsi->tsi_tgt), le32_to_cpu(ubuf->ub_magic),
UPDATE_BUFFER_MAGIC, -EPROTO);
RETURN(err_serious(-EPROTO));
}
count = le32_to_cpu(ubuf->ub_count);
if (count <= 0) {
CERROR("%s: No update!: rc = %d\n",
tgt_name(tsi->tsi_tgt), -EPROTO);
RETURN(err_serious(-EPROTO));
}
req_capsule_set_size(pill, &RMF_UPDATE_REPLY, RCL_SERVER,
UPDATE_BUFFER_SIZE);
rc = req_capsule_server_pack(pill);
if (rc != 0) {
CERROR("%s: Can't pack response: rc = %d\n",
tgt_name(tsi->tsi_tgt), rc);
RETURN(rc);
}
/* Prepare the update reply buffer */
update_reply = req_capsule_server_get(pill, &RMF_UPDATE_REPLY);
if (update_reply == NULL)
RETURN(err_serious(-EPROTO));
update_init_reply_buf(update_reply, count);
tti->tti_u.update.tti_update_reply = update_reply;
rc = out_tx_start(env, dt, ta);
if (rc != 0)
RETURN(rc);
tti->tti_mult_trans = !req_is_replay(tgt_ses_req(tsi));
/* Walk through updates in the request to execute them synchronously */
off = cfs_size_round(offsetof(struct update_buf, ub_bufs[0]));
for (i = 0; i < count; i++) {
struct tgt_handler *h;
struct dt_object *dt_obj;
update = (struct update *)((char *)ubuf + off);
if (old_batchid == -1) {
old_batchid = update->u_batchid;
} else if (old_batchid != update->u_batchid) {
/* Stop the current update transaction,
* create a new one */
rc = out_tx_end(env, ta);
if (rc != 0)
RETURN(rc);
rc = out_tx_start(env, dt, ta);
if (rc != 0)
RETURN(rc);
old_batchid = update->u_batchid;
}
fid_le_to_cpu(&update->u_fid, &update->u_fid);
if (!fid_is_sane(&update->u_fid)) {
CERROR("%s: invalid FID "DFID": rc = %d\n",
tgt_name(tsi->tsi_tgt), PFID(&update->u_fid),
-EPROTO);
GOTO(out, rc = err_serious(-EPROTO));
}
dt_obj = dt_locate(env, dt, &update->u_fid);
if (IS_ERR(dt_obj))
GOTO(out, rc = PTR_ERR(dt_obj));
tti->tti_u.update.tti_dt_object = dt_obj;
tti->tti_u.update.tti_update = update;
tti->tti_u.update.tti_update_reply_index = i;
h = out_handler_find(update->u_type);
if (likely(h != NULL)) {
/* For real modification RPC, check if the update
* has been executed */
if (h->th_flags & MUTABOR) {
struct ptlrpc_request *req = tgt_ses_req(tsi);
if (out_check_resent(env, dt, dt_obj, req,
out_reconstruct,
update_reply, i))
GOTO(next, rc);
}
rc = h->th_act(tsi);
} else {
CERROR("%s: The unsupported opc: 0x%x\n",
tgt_name(tsi->tsi_tgt), update->u_type);
lu_object_put(env, &dt_obj->do_lu);
GOTO(out, rc = -ENOTSUPP);
}
next:
lu_object_put(env, &dt_obj->do_lu);
if (rc < 0)
GOTO(out, rc);
off += cfs_size_round(update_size(update));
}
out:
rc1 = out_tx_end(env, ta);
if (rc == 0)
rc = rc1;
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));
/* 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(med);
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(med, &info->mti_ioepoch->handle,
req_is_replay(req));
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_lut.lut_oss_capa &&
exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
S_ISREG(lu_object_attr(&mo->mot_obj)) &&
(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;
}
