/*
* Handler for: getattr, lookup and revalidate cases.
*/
int lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
void *lmm, int lmmsize, struct lookup_intent *it,
int flags, struct ptlrpc_request **reqp,
ldlm_blocking_callback cb_blocking,
__u64 extra_lock_flags)
{
struct obd_device *obd = exp->exp_obd;
struct lmv_obd *lmv = &obd->u.lmv;
struct lmv_tgt_desc *tgt = NULL;
struct mdt_body *body;
int rc = 0;
tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
if (IS_ERR(tgt))
return PTR_ERR(tgt);
if (!fid_is_sane(&op_data->op_fid2))
fid_zero(&op_data->op_fid2);
CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
", name='%s' -> mds #%d\n", PFID(&op_data->op_fid1),
PFID(&op_data->op_fid2),
op_data->op_name ? op_data->op_name : "<NULL>",
tgt->ltd_idx);
op_data->op_bias &= ~MDS_CROSS_REF;
rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it,
flags, reqp, cb_blocking, extra_lock_flags);
if (rc < 0 || *reqp == NULL)
return rc;
/*
* MDS has returned success. Probably name has been resolved in
* remote inode. Let's check this.
*/
body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
return -EPROTO;
/* Not cross-ref case, just get out of here. */
if (likely(!(body->valid & OBD_MD_MDS)))
return 0;
rc = lmv_intent_remote(exp, lmm, lmmsize, it, NULL, flags, reqp,
cb_blocking, extra_lock_flags);
return rc;
}
static int llu_lookup_it(struct inode *parent, struct pnode *pnode,
struct lookup_intent *it, int flags)
{
struct md_op_data op_data = {{ 0 }};
struct ptlrpc_request *req = NULL;
struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
__u32 opc;
int rc;
ENTRY;
if (pnode->p_base->pb_name.len > EXT2_NAME_LEN)
RETURN(-ENAMETOOLONG);
if (!it) {
it = &lookup_it;
it->it_op_release = ll_intent_release;
}
if (it->it_op & IT_CREAT)
opc = LUSTRE_OPC_CREATE;
else
opc = LUSTRE_OPC_ANY;
llu_prep_md_op_data(&op_data, parent, NULL,
pnode->p_base->pb_name.name,
pnode->p_base->pb_name.len, flags, opc);
rc = md_intent_lock(llu_i2mdexp(parent), &op_data, NULL, 0, it,
flags, &req, llu_md_blocking_ast,
LDLM_FL_CANCEL_ON_BLOCK);
if (rc < 0)
GOTO(out, rc);
rc = lookup_it_finish(req, DLM_REPLY_REC_OFF, it, parent, pnode);
if (rc != 0) {
ll_intent_release(it);
GOTO(out, rc);
}
llu_lookup_finish_locks(it, pnode);
out:
if (req)
ptlrpc_req_finished(req);
return rc;
}
int ll_revalidate_it(struct dentry *de, int lookup_flags,
struct lookup_intent *it)
{
struct md_op_data *op_data;
struct ptlrpc_request *req = NULL;
struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
struct obd_export *exp;
struct inode *parent = de->d_parent->d_inode;
int rc, first = 0;
ENTRY;
CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
LL_IT2STR(it));
if (de->d_inode == NULL) {
/* We can only use negative dentries if this is stat or lookup,
for opens and stuff we do need to query server. */
/* If there is IT_CREAT in intent op set, then we must throw
away this negative dentry and actually do the request to
kernel to create whatever needs to be created (if possible)*/
if (it && (it->it_op & IT_CREAT))
RETURN(0);
if (de->d_flags & DCACHE_LUSTRE_INVALID)
RETURN(0);
rc = ll_have_md_lock(parent, MDS_INODELOCK_UPDATE, LCK_MINMODE);
GOTO(out_sa, rc);
}
/* Never execute intents for mount points.
* Attributes will be fixed up in ll_inode_revalidate_it */
if (d_mountpoint(de))
GOTO(out_sa, rc = 1);
/* need to get attributes in case root got changed from other client */
if (de == de->d_sb->s_root) {
rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
if (rc == 0)
rc = 1;
GOTO(out_sa, rc);
}
exp = ll_i2mdexp(de->d_inode);
OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
ll_frob_intent(&it, &lookup_it);
LASSERT(it);
if (it->it_op == IT_LOOKUP && !(de->d_flags & DCACHE_LUSTRE_INVALID))
GOTO(out_sa, rc = 1);
op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
de->d_name.name, de->d_name.len,
0, LUSTRE_OPC_ANY, NULL);
if (IS_ERR(op_data))
RETURN(PTR_ERR(op_data));
if ((it->it_op == IT_OPEN) && de->d_inode) {
struct inode *inode = de->d_inode;
struct ll_inode_info *lli = ll_i2info(inode);
struct obd_client_handle **och_p;
__u64 *och_usecount;
/*
* We used to check for MDS_INODELOCK_OPEN here, but in fact
* just having LOOKUP lock is enough to justify inode is the
* same. And if inode is the same and we have suitable
* openhandle, then there is no point in doing another OPEN RPC
* just to throw away newly received openhandle. There are no
* security implications too, if file owner or access mode is
* change, LOOKUP lock is revoked.
*/
if (it->it_flags & FMODE_WRITE) {
och_p = &lli->lli_mds_write_och;
och_usecount = &lli->lli_open_fd_write_count;
} else if (it->it_flags & FMODE_EXEC) {
och_p = &lli->lli_mds_exec_och;
och_usecount = &lli->lli_open_fd_exec_count;
} else {
och_p = &lli->lli_mds_read_och;
och_usecount = &lli->lli_open_fd_read_count;
}
/* Check for the proper lock. */
if (!ll_have_md_lock(inode, MDS_INODELOCK_LOOKUP, LCK_MINMODE))
goto do_lock;
cfs_down(&lli->lli_och_sem);
if (*och_p) { /* Everything is open already, do nothing */
/*(*och_usecount)++; Do not let them steal our open
handle from under us */
/* XXX The code above was my original idea, but in case
we have the handle, but we cannot use it due to later
checks (e.g. O_CREAT|O_EXCL flags set), nobody
would decrement counter increased here. So we just
hope the lock won't be invalidated in between. But
if it would be, we'll reopen the open request to
MDS later during file open path */
cfs_up(&lli->lli_och_sem);
ll_finish_md_op_data(op_data);
RETURN(1);
} else {
cfs_up(&lli->lli_och_sem);
}
}
if (it->it_op == IT_GETATTR) {
first = ll_statahead_enter(parent, &de, 0);
if (first == 1) {
ll_statahead_exit(parent, de, 1);
ll_finish_md_op_data(op_data);
GOTO(out, rc = 1);
}
}
do_lock:
it->it_create_mode &= ~current->fs->umask;
it->it_create_mode |= M_CHECK_STALE;
rc = md_intent_lock(exp, op_data, NULL, 0, it,
lookup_flags,
&req, ll_md_blocking_ast, 0);
it->it_create_mode &= ~M_CHECK_STALE;
ll_finish_md_op_data(op_data);
if (it->it_op == IT_GETATTR && !first)
/* If there are too many locks on client-side, then some
* locks taken by statahead maybe dropped automatically
* before the real "revalidate" using them. */
ll_statahead_exit(parent, de, req == NULL ? rc : 0);
else if (first == -EEXIST)
ll_statahead_mark(parent, de);
/* If req is NULL, then md_intent_lock only tried to do a lock match;
* if all was well, it will return 1 if it found locks, 0 otherwise. */
if (req == NULL && rc >= 0) {
if (!rc)
goto do_lookup;
GOTO(out, rc);
}
if (rc < 0) {
if (rc != -ESTALE) {
CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
"%d\n", rc, it->d.lustre.it_status);
}
GOTO(out, rc = 0);
}
revalidate_finish:
rc = ll_revalidate_it_finish(req, it, de);
if (rc != 0) {
if (rc != -ESTALE && rc != -ENOENT)
ll_intent_release(it);
GOTO(out, rc = 0);
}
if ((it->it_op & IT_OPEN) && de->d_inode &&
!S_ISREG(de->d_inode->i_mode) &&
!S_ISDIR(de->d_inode->i_mode)) {
ll_release_openhandle(de, it);
}
rc = 1;
/* unfortunately ll_intent_lock may cause a callback and revoke our
* dentry */
cfs_spin_lock(&ll_lookup_lock);
spin_lock(&dcache_lock);
lock_dentry(de);
__d_drop(de);
unlock_dentry(de);
d_rehash_cond(de, 0);
spin_unlock(&dcache_lock);
cfs_spin_unlock(&ll_lookup_lock);
out:
/* We do not free request as it may be reused during following lookup
* (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
* be freed in ll_lookup_it or in ll_intent_release. But if
* request was not completed, we need to free it. (bug 5154, 9903) */
if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
ptlrpc_req_finished(req);
if (rc == 0) {
ll_unhash_aliases(de->d_inode);
/* done in ll_unhash_aliases()
dentry->d_flags |= DCACHE_LUSTRE_INVALID; */
} else {
CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
"inode %p refc %d\n", de->d_name.len,
de->d_name.name, de, de->d_parent, de->d_inode,
atomic_read(&de->d_count));
if (first != 1) {
if (de->d_flags & DCACHE_LUSTRE_INVALID) {
lock_dentry(de);
de->d_flags &= ~DCACHE_LUSTRE_INVALID;
unlock_dentry(de);
}
ll_lookup_finish_locks(it, de);
}
}
RETURN(rc);
/*
* This part is here to combat evil-evil race in real_lookup on 2.6
* kernels. The race details are: We enter do_lookup() looking for some
* name, there is nothing in dcache for this name yet and d_lookup()
* returns NULL. We proceed to real_lookup(), and while we do this,
* another process does open on the same file we looking up (most simple
* reproducer), open succeeds and the dentry is added. Now back to
* us. In real_lookup() we do d_lookup() again and suddenly find the
* dentry, so we call d_revalidate on it, but there is no lock, so
* without this code we would return 0, but unpatched real_lookup just
* returns -ENOENT in such a case instead of retrying the lookup. Once
* this is dealt with in real_lookup(), all of this ugly mess can go and
* we can just check locks in ->d_revalidate without doing any RPCs
* ever.
*/
do_lookup:
if (it != &lookup_it) {
/* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
if (it->it_op == IT_GETATTR)
lookup_it.it_op = IT_GETATTR;
ll_lookup_finish_locks(it, de);
it = &lookup_it;
}
/* Do real lookup here. */
op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
de->d_name.len, 0, (it->it_op & IT_CREAT ?
LUSTRE_OPC_CREATE :
LUSTRE_OPC_ANY), NULL);
if (IS_ERR(op_data))
RETURN(PTR_ERR(op_data));
rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
ll_md_blocking_ast, 0);
if (rc >= 0) {
struct mdt_body *mdt_body;
struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
if (de->d_inode)
fid = *ll_inode2fid(de->d_inode);
/* see if we got same inode, if not - return error */
if (lu_fid_eq(&fid, &mdt_body->fid1)) {
ll_finish_md_op_data(op_data);
op_data = NULL;
goto revalidate_finish;
}
ll_intent_release(it);
}
ll_finish_md_op_data(op_data);
GOTO(out, rc = 0);
out_sa:
/*
* For rc == 1 case, should not return directly to prevent losing
* statahead windows; for rc == 0 case, the "lookup" will be done later.
*/
if (it && it->it_op == IT_GETATTR && rc == 1) {
first = ll_statahead_enter(parent, &de, 0);
if (first >= 0)
ll_statahead_exit(parent, de, 1);
else if (first == -EEXIST)
ll_statahead_mark(parent, de);
}
return rc;
}
#if 0
static void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
{
struct inode *inode= de->d_inode;
struct ll_sb_info *sbi = ll_i2sbi(inode);
struct ll_dentry_data *ldd = ll_d2d(de);
struct obd_client_handle *handle;
struct obd_capa *oc;
int rc = 0;
ENTRY;
LASSERT(ldd);
cfs_lock_kernel();
/* Strictly speaking this introduces an additional race: the
* increments should wait until the rpc has returned.
* However, given that at present the function is void, this
* issue is moot. */
if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
cfs_unlock_kernel();
EXIT;
return;
}
if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
cfs_unlock_kernel();
EXIT;
return;
}
cfs_unlock_kernel();
handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
oc = ll_mdscapa_get(inode);
rc = obd_pin(sbi->ll_md_exp, ll_inode2fid(inode), oc, handle, flag);
capa_put(oc);
if (rc) {
cfs_lock_kernel();
memset(handle, 0, sizeof(*handle));
if (flag == 0)
ldd->lld_cwd_count--;
else
ldd->lld_mnt_count--;
cfs_unlock_kernel();
}
EXIT;
return;
}
static void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
{
struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
struct ll_dentry_data *ldd = ll_d2d(de);
struct obd_client_handle handle;
int count, rc = 0;
ENTRY;
LASSERT(ldd);
cfs_lock_kernel();
/* Strictly speaking this introduces an additional race: the
* increments should wait until the rpc has returned.
* However, given that at present the function is void, this
* issue is moot. */
handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
/* the "pin" failed */
cfs_unlock_kernel();
EXIT;
return;
}
if (flag)
count = --ldd->lld_mnt_count;
else
count = --ldd->lld_cwd_count;
cfs_unlock_kernel();
if (count != 0) {
EXIT;
return;
}
rc = obd_unpin(sbi->ll_md_exp, &handle, flag);
EXIT;
return;
}
#endif
#ifdef HAVE_VFS_INTENT_PATCHES
int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
{
int rc;
ENTRY;
if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
rc = ll_revalidate_it(dentry, nd->flags, &nd->intent);
else
rc = ll_revalidate_it(dentry, 0, NULL);
RETURN(rc);
}
static int llu_pb_revalidate(struct pnode *pnode, int flags,
struct lookup_intent *it)
{
struct pnode_base *pb = pnode->p_base;
struct md_op_data op_data = {{ 0 }};
struct ptlrpc_request *req = NULL;
struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
struct obd_export *exp;
int rc;
ENTRY;
CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s,intent=%x\n",
(int)pb->pb_name.len, pb->pb_name.name, it ? it->it_op : 0);
/* We don't want to cache negative dentries, so return 0 immediately.
* We believe that this is safe, that negative dentries cannot be
* pinned by someone else */
if (pb->pb_ino == NULL) {
CDEBUG(D_INODE, "negative pb\n");
RETURN(0);
}
/* This is due to bad interaction with libsysio. remove this when we
* switched to libbsdio XXX
*/
{
struct llu_inode_info *lli = llu_i2info(pb->pb_ino);
struct intnl_stat *st = llu_i2stat(pb->pb_ino);
if (lli->lli_it) {
CDEBUG(D_INODE, "inode %llu still have intent "
"%p(opc 0x%x), release it\n",
(long long) st->st_ino, lli->lli_it,
lli->lli_it->it_op);
ll_intent_release(lli->lli_it);
OBD_FREE(lli->lli_it, sizeof(*lli->lli_it));
lli->lli_it = NULL;
}
}
exp = llu_i2mdexp(pb->pb_ino);
if (!it) {
it = &lookup_it;
it->it_op_release = ll_intent_release;
}
llu_prep_md_op_data(&op_data, pnode->p_parent->p_base->pb_ino,
pb->pb_ino, pb->pb_name.name, pb->pb_name.len,
0, LUSTRE_OPC_ANY);
rc = md_intent_lock(exp, &op_data, NULL, 0, it, flags,
&req, llu_md_blocking_ast,
LDLM_FL_CANCEL_ON_BLOCK);
/* If req is NULL, then md_intent_lock only tried to do a lock match;
* if all was well, it will return 1 if it found locks, 0 otherwise. */
if (req == NULL && rc >= 0)
GOTO(out, rc);
if (rc < 0)
GOTO(out, rc = 0);
rc = pnode_revalidate_finish(req, it, pnode);
if (rc != 0) {
ll_intent_release(it);
GOTO(out, rc = 0);
}
rc = 1;
/* Note: ll_intent_lock may cause a callback, check this! */
if (it->it_op & IT_OPEN)
LL_SAVE_INTENT(pb->pb_ino, it);
out:
if (req && rc == 1)
ptlrpc_req_finished(req);
if (rc == 0) {
LASSERT(pb->pb_ino);
I_RELE(pb->pb_ino);
pb->pb_ino = NULL;
} else {
llu_lookup_finish_locks(it, pnode);
}
RETURN(rc);
}
static int lmv_intent_remote(struct obd_export *exp, struct lookup_intent *it,
const struct lu_fid *parent_fid,
struct ptlrpc_request **reqp,
ldlm_blocking_callback cb_blocking,
__u64 extra_lock_flags)
{
struct obd_device *obd = exp->exp_obd;
struct lmv_obd *lmv = &obd->u.lmv;
struct ptlrpc_request *req = NULL;
struct lustre_handle plock;
struct md_op_data *op_data;
struct lmv_tgt_desc *tgt;
struct mdt_body *body;
int pmode;
int rc = 0;
ENTRY;
body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
RETURN(-EPROTO);
LASSERT((body->mbo_valid & OBD_MD_MDS));
/*
* Unfortunately, we have to lie to MDC/MDS to retrieve
* attributes llite needs and provideproper locking.
*/
if (it->it_op & IT_LOOKUP)
it->it_op = IT_GETATTR;
/*
* We got LOOKUP lock, but we really need attrs.
*/
pmode = it->d.lustre.it_lock_mode;
if (pmode) {
plock.cookie = it->d.lustre.it_lock_handle;
it->d.lustre.it_lock_mode = 0;
it->d.lustre.it_data = NULL;
}
LASSERT(fid_is_sane(&body->mbo_fid1));
tgt = lmv_find_target(lmv, &body->mbo_fid1);
if (IS_ERR(tgt))
GOTO(out, rc = PTR_ERR(tgt));
OBD_ALLOC_PTR(op_data);
if (op_data == NULL)
GOTO(out, rc = -ENOMEM);
op_data->op_fid1 = body->mbo_fid1;
/* Sent the parent FID to the remote MDT */
if (parent_fid != NULL) {
/* The parent fid is only for remote open to
* check whether the open is from OBF,
* see mdt_cross_open */
LASSERT(it->it_op & IT_OPEN);
op_data->op_fid2 = *parent_fid;
}
op_data->op_bias = MDS_CROSS_REF;
CDEBUG(D_INODE, "REMOTE_INTENT with fid="DFID" -> mds #%u\n",
PFID(&body->mbo_fid1), tgt->ltd_idx);
rc = md_intent_lock(tgt->ltd_exp, op_data, it, &req, cb_blocking,
extra_lock_flags);
if (rc)
GOTO(out_free_op_data, rc);
/*
* LLite needs LOOKUP lock to track dentry revocation in order to
* maintain dcache consistency. Thus drop UPDATE|PERM lock here
* and put LOOKUP in request.
*/
if (it->d.lustre.it_lock_mode != 0) {
it->d.lustre.it_remote_lock_handle =
it->d.lustre.it_lock_handle;
it->d.lustre.it_remote_lock_mode = it->d.lustre.it_lock_mode;
}
if (pmode) {
it->d.lustre.it_lock_handle = plock.cookie;
it->d.lustre.it_lock_mode = pmode;
}
EXIT;
out_free_op_data:
OBD_FREE_PTR(op_data);
out:
if (rc && pmode)
ldlm_lock_decref(&plock, pmode);
ptlrpc_req_finished(*reqp);
*reqp = req;
return rc;
}
/*
* Handler for: getattr, lookup and revalidate cases.
*/
static int
lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
struct lookup_intent *it, struct ptlrpc_request **reqp,
ldlm_blocking_callback cb_blocking,
__u64 extra_lock_flags)
{
struct obd_device *obd = exp->exp_obd;
struct lmv_obd *lmv = &obd->u.lmv;
struct lmv_tgt_desc *tgt = NULL;
struct mdt_body *body;
struct lmv_stripe_md *lsm = op_data->op_mea1;
int rc = 0;
ENTRY;
/* If it returns ERR_PTR(-EBADFD) then it is an unknown hash type
* it will try all stripes to locate the object */
tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
if (IS_ERR(tgt) && (PTR_ERR(tgt) != -EBADFD))
RETURN(PTR_ERR(tgt));
/* Both migrating dir and unknown hash dir need to try
* all of sub-stripes */
if (lsm != NULL && !lmv_is_known_hash_type(lsm->lsm_md_hash_type)) {
struct lmv_oinfo *oinfo;
oinfo = &lsm->lsm_md_oinfo[0];
op_data->op_fid1 = oinfo->lmo_fid;
op_data->op_mds = oinfo->lmo_mds;
tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
if (IS_ERR(tgt))
RETURN(PTR_ERR(tgt));
}
if (!fid_is_sane(&op_data->op_fid2))
fid_zero(&op_data->op_fid2);
CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
", name='%s' -> mds #%u lsm=%p lsm_magic=%x\n",
PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
op_data->op_name ? op_data->op_name : "<NULL>",
tgt->ltd_idx, lsm, lsm == NULL ? -1 : lsm->lsm_md_magic);
op_data->op_bias &= ~MDS_CROSS_REF;
rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
extra_lock_flags);
if (rc < 0)
RETURN(rc);
if (*reqp == NULL) {
/* If RPC happens, lsm information will be revalidated
* during update_inode process (see ll_update_lsm_md) */
if (op_data->op_mea2 != NULL) {
rc = lmv_revalidate_slaves(exp, op_data->op_mea2,
cb_blocking,
extra_lock_flags);
if (rc != 0)
RETURN(rc);
}
RETURN(rc);
} else if (it_disposition(it, DISP_LOOKUP_NEG) && lsm != NULL &&
lmv_need_try_all_stripes(lsm)) {
/* For migrating and unknown hash type directory, it will
* try to target the entry on other stripes */
int stripe_index;
for (stripe_index = 1;
stripe_index < lsm->lsm_md_stripe_count &&
it_disposition(it, DISP_LOOKUP_NEG); stripe_index++) {
struct lmv_oinfo *oinfo;
/* release the previous request */
ptlrpc_req_finished(*reqp);
it->d.lustre.it_data = NULL;
*reqp = NULL;
oinfo = &lsm->lsm_md_oinfo[stripe_index];
tgt = lmv_find_target(lmv, &oinfo->lmo_fid);
if (IS_ERR(tgt))
RETURN(PTR_ERR(tgt));
CDEBUG(D_INODE, "Try other stripes " DFID"\n",
PFID(&oinfo->lmo_fid));
op_data->op_fid1 = oinfo->lmo_fid;
it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp,
cb_blocking, extra_lock_flags);
if (rc != 0)
RETURN(rc);
}
}
/*
* MDS has returned success. Probably name has been resolved in
* remote inode. Let's check this.
*/
body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
RETURN(-EPROTO);
/* Not cross-ref case, just get out of here. */
if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
extra_lock_flags);
if (rc != 0)
RETURN(rc);
body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
RETURN(-EPROTO);
}
RETURN(rc);
}
/*
* IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
* may be split dir.
*/
static int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
struct lookup_intent *it,
struct ptlrpc_request **reqp,
ldlm_blocking_callback cb_blocking,
__u64 extra_lock_flags)
{
struct obd_device *obd = exp->exp_obd;
struct lmv_obd *lmv = &obd->u.lmv;
struct lmv_tgt_desc *tgt;
struct mdt_body *body;
int rc;
ENTRY;
if (it->it_flags & MDS_OPEN_BY_FID) {
LASSERT(fid_is_sane(&op_data->op_fid2));
/* for striped directory, we can't know parent stripe fid
* without name, but we can set it to child fid, and MDT
* will obtain it from linkea in open in such case. */
if (op_data->op_mea1 != NULL)
op_data->op_fid1 = op_data->op_fid2;
tgt = lmv_find_target(lmv, &op_data->op_fid2);
if (IS_ERR(tgt))
RETURN(PTR_ERR(tgt));
op_data->op_mds = tgt->ltd_idx;
} else {
LASSERT(fid_is_sane(&op_data->op_fid1));
LASSERT(fid_is_zero(&op_data->op_fid2));
LASSERT(op_data->op_name != NULL);
tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
if (IS_ERR(tgt))
RETURN(PTR_ERR(tgt));
}
/* If it is ready to open the file by FID, do not need
* allocate FID at all, otherwise it will confuse MDT */
if ((it->it_op & IT_CREAT) && !(it->it_flags & MDS_OPEN_BY_FID)) {
/*
* For lookup(IT_CREATE) cases allocate new fid and setup FLD
* for it.
*/
rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
if (rc != 0)
RETURN(rc);
}
CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
" name='%s' -> mds #%u\n", PFID(&op_data->op_fid1),
PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
extra_lock_flags);
if (rc != 0)
RETURN(rc);
/*
* Nothing is found, do not access body->fid1 as it is zero and thus
* pointless.
*/
if ((it->d.lustre.it_disposition & DISP_LOOKUP_NEG) &&
!(it->d.lustre.it_disposition & DISP_OPEN_CREATE) &&
!(it->d.lustre.it_disposition & DISP_OPEN_OPEN))
RETURN(rc);
body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
RETURN(-EPROTO);
/* Not cross-ref case, just get out of here. */
if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
cb_blocking, extra_lock_flags);
if (rc != 0)
RETURN(rc);
body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
RETURN(-EPROTO);
}
RETURN(rc);
}
int lmv_revalidate_slaves(struct obd_export *exp,
const struct lmv_stripe_md *lsm,
ldlm_blocking_callback cb_blocking,
int extra_lock_flags)
{
struct obd_device *obd = exp->exp_obd;
struct lmv_obd *lmv = &obd->u.lmv;
struct ptlrpc_request *req = NULL;
struct mdt_body *body;
struct md_op_data *op_data;
int i;
int rc = 0;
ENTRY;
/**
* revalidate slaves has some problems, temporarily return,
* we may not need that
*/
OBD_ALLOC_PTR(op_data);
if (op_data == NULL)
RETURN(-ENOMEM);
/**
* Loop over the stripe information, check validity and update them
* from MDS if needed.
*/
for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
struct lu_fid fid;
struct lookup_intent it = { .it_op = IT_GETATTR };
struct lustre_handle *lockh = NULL;
struct lmv_tgt_desc *tgt = NULL;
struct inode *inode;
fid = lsm->lsm_md_oinfo[i].lmo_fid;
inode = lsm->lsm_md_oinfo[i].lmo_root;
/*
* Prepare op_data for revalidating. Note that @fid2 shluld be
* defined otherwise it will go to server and take new lock
* which is not needed here.
*/
memset(op_data, 0, sizeof(*op_data));
op_data->op_fid1 = fid;
op_data->op_fid2 = fid;
tgt = lmv_locate_mds(lmv, op_data, &fid);
if (IS_ERR(tgt))
GOTO(cleanup, rc = PTR_ERR(tgt));
CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%u\n",
PFID(&fid), tgt->ltd_idx);
if (req != NULL) {
ptlrpc_req_finished(req);
req = NULL;
}
rc = md_intent_lock(tgt->ltd_exp, op_data, &it, &req,
cb_blocking, extra_lock_flags);
if (rc < 0)
GOTO(cleanup, rc);
lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
if (rc > 0 && req == NULL) {
/* slave inode is still valid */
CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
PFID(&fid));
rc = 0;
} else {
/* refresh slave from server */
body = req_capsule_server_get(&req->rq_pill,
&RMF_MDT_BODY);
if (body == NULL) {
if (it.d.lustre.it_lock_mode && lockh) {
ldlm_lock_decref(lockh,
it.d.lustre.it_lock_mode);
it.d.lustre.it_lock_mode = 0;
}
GOTO(cleanup, rc = -ENOENT);
}
i_size_write(inode, body->mbo_size);
inode->i_blocks = body->mbo_blocks;
set_nlink(inode, body->mbo_nlink);
LTIME_S(inode->i_atime) = body->mbo_atime;
LTIME_S(inode->i_ctime) = body->mbo_ctime;
LTIME_S(inode->i_mtime) = body->mbo_mtime;
}
md_set_lock_data(tgt->ltd_exp, &lockh->cookie, inode, NULL);
if (it.d.lustre.it_lock_mode != 0 && lockh != NULL) {
ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
it.d.lustre.it_lock_mode = 0;
}
}
cleanup:
if (req != NULL)
ptlrpc_req_finished(req);
OBD_FREE_PTR(op_data);
RETURN(rc);
}
/*
* IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
* may be split dir.
*/
int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
void *lmm, int lmmsize, struct lookup_intent *it,
int flags, struct ptlrpc_request **reqp,
ldlm_blocking_callback cb_blocking,
__u64 extra_lock_flags)
{
struct obd_device *obd = exp->exp_obd;
struct lmv_obd *lmv = &obd->u.lmv;
struct lmv_tgt_desc *tgt;
struct mdt_body *body;
int rc;
tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
if (IS_ERR(tgt))
return PTR_ERR(tgt);
/* If it is ready to open the file by FID, do not need
* allocate FID at all, otherwise it will confuse MDT */
if ((it->it_op & IT_CREAT) &&
!(it->it_flags & MDS_OPEN_BY_FID)) {
/*
* For open with IT_CREATE and for IT_CREATE cases allocate new
* fid and setup FLD for it.
*/
op_data->op_fid3 = op_data->op_fid2;
rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
if (rc != 0)
return rc;
}
CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
" name='%s' -> mds #%d\n", PFID(&op_data->op_fid1),
PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it, flags,
reqp, cb_blocking, extra_lock_flags);
if (rc != 0)
return rc;
/*
* Nothing is found, do not access body->fid1 as it is zero and thus
* pointless.
*/
if ((it->d.lustre.it_disposition & DISP_LOOKUP_NEG) &&
!(it->d.lustre.it_disposition & DISP_OPEN_CREATE) &&
!(it->d.lustre.it_disposition & DISP_OPEN_OPEN))
return rc;
body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
return -EPROTO;
/*
* Not cross-ref case, just get out of here.
*/
if (likely(!(body->valid & OBD_MD_MDS)))
return 0;
/*
* Okay, MDS has returned success. Probably name has been resolved in
* remote inode.
*/
rc = lmv_intent_remote(exp, lmm, lmmsize, it, &op_data->op_fid1, flags,
reqp, cb_blocking, extra_lock_flags);
if (rc != 0) {
LASSERT(rc < 0);
/*
* This is possible, that some userspace application will try to
* open file as directory and we will have -ENOTDIR here. As
* this is normal situation, we should not print error here,
* only debug info.
*/
CDEBUG(D_INODE, "Can't handle remote %s: dir "DFID"("DFID"):"
"%*s: %d\n", LL_IT2STR(it), PFID(&op_data->op_fid2),
PFID(&op_data->op_fid1), op_data->op_namelen,
op_data->op_name, rc);
return rc;
}
return rc;
}
/*
* Handler for: getattr, lookup and revalidate cases.
*/
int lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
void *lmm, int lmmsize, struct lookup_intent *it,
int flags, struct ptlrpc_request **reqp,
ldlm_blocking_callback cb_blocking,
__u64 extra_lock_flags)
{
struct obd_device *obd = exp->exp_obd;
struct lmv_obd *lmv = &obd->u.lmv;
struct lmv_tgt_desc *tgt = NULL;
struct mdt_body *body;
struct lmv_stripe_md *lsm = op_data->op_mea1;
int rc = 0;
ENTRY;
tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
if (IS_ERR(tgt))
RETURN(PTR_ERR(tgt));
if (!fid_is_sane(&op_data->op_fid2))
fid_zero(&op_data->op_fid2);
CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
", name='%s' -> mds #%d lsm=%p lsm_magic=%x\n",
PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
op_data->op_name ? op_data->op_name : "<NULL>",
tgt->ltd_idx, lsm, lsm == NULL ? -1 : lsm->lsm_md_magic);
op_data->op_bias &= ~MDS_CROSS_REF;
rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it,
flags, reqp, cb_blocking, extra_lock_flags);
if (rc < 0)
RETURN(rc);
if (*reqp == NULL) {
/* If RPC happens, lsm information will be revalidated
* during update_inode process (see ll_update_lsm_md) */
if (op_data->op_mea2 != NULL) {
rc = lmv_revalidate_slaves(exp, NULL, op_data->op_mea2,
cb_blocking,
extra_lock_flags);
if (rc != 0)
RETURN(rc);
}
RETURN(rc);
} else if (it_disposition(it, DISP_LOOKUP_NEG) &&
lsm != NULL && lsm->lsm_md_magic == LMV_MAGIC_MIGRATE) {
/* For migrating directory, if it can not find the child in
* the source directory(master stripe), try the targeting
* directory(stripe 1) */
tgt = lmv_find_target(lmv, &lsm->lsm_md_oinfo[1].lmo_fid);
if (IS_ERR(tgt))
RETURN(PTR_ERR(tgt));
ptlrpc_req_finished(*reqp);
it->d.lustre.it_data = NULL;
*reqp = NULL;
CDEBUG(D_INODE, "For migrating dir, try target dir "DFID"\n",
PFID(&lsm->lsm_md_oinfo[1].lmo_fid));
op_data->op_fid1 = lsm->lsm_md_oinfo[1].lmo_fid;
it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it,
flags, reqp, cb_blocking, extra_lock_flags);
}
/*
* MDS has returned success. Probably name has been resolved in
* remote inode. Let's check this.
*/
body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
RETURN(-EPROTO);
/* Not cross-ref case, just get out of here. */
if (unlikely((body->valid & OBD_MD_MDS))) {
rc = lmv_intent_remote(exp, lmm, lmmsize, it, NULL, flags,
reqp, cb_blocking, extra_lock_flags);
if (rc != 0)
RETURN(rc);
body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
RETURN(-EPROTO);
}
RETURN(rc);
}
/*
* IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
* may be split dir.
*/
int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
void *lmm, int lmmsize, struct lookup_intent *it,
int flags, struct ptlrpc_request **reqp,
ldlm_blocking_callback cb_blocking,
__u64 extra_lock_flags)
{
struct obd_device *obd = exp->exp_obd;
struct lmv_obd *lmv = &obd->u.lmv;
struct lmv_tgt_desc *tgt;
struct mdt_body *body;
int rc;
ENTRY;
/* Note: client might open with some random flags(sanity 33b), so we can
* not make sure op_fid2 is being initialized with BY_FID flag */
if (it->it_flags & MDS_OPEN_BY_FID && fid_is_sane(&op_data->op_fid2)) {
if (op_data->op_mea1 != NULL) {
struct lmv_stripe_md *lsm = op_data->op_mea1;
const struct lmv_oinfo *oinfo;
oinfo = lsm_name_to_stripe_info(lsm, op_data->op_name,
op_data->op_namelen);
if (IS_ERR(oinfo))
RETURN(PTR_ERR(oinfo));
op_data->op_fid1 = oinfo->lmo_fid;
}
tgt = lmv_find_target(lmv, &op_data->op_fid2);
if (IS_ERR(tgt))
RETURN(PTR_ERR(tgt));
op_data->op_mds = tgt->ltd_idx;
} else {
tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
if (IS_ERR(tgt))
RETURN(PTR_ERR(tgt));
}
/* If it is ready to open the file by FID, do not need
* allocate FID at all, otherwise it will confuse MDT */
if ((it->it_op & IT_CREAT) &&
!(it->it_flags & MDS_OPEN_BY_FID)) {
/*
* For open with IT_CREATE and for IT_CREATE cases allocate new
* fid and setup FLD for it.
*/
op_data->op_fid3 = op_data->op_fid2;
rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
if (rc != 0)
RETURN(rc);
}
CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
" name='%s' -> mds #%d\n", PFID(&op_data->op_fid1),
PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
rc = md_intent_lock(tgt->ltd_exp, op_data, lmm, lmmsize, it, flags,
reqp, cb_blocking, extra_lock_flags);
if (rc != 0)
RETURN(rc);
/*
* Nothing is found, do not access body->fid1 as it is zero and thus
* pointless.
*/
if ((it->d.lustre.it_disposition & DISP_LOOKUP_NEG) &&
!(it->d.lustre.it_disposition & DISP_OPEN_CREATE) &&
!(it->d.lustre.it_disposition & DISP_OPEN_OPEN))
RETURN(rc);
body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
RETURN(-EPROTO);
/* Not cross-ref case, just get out of here. */
if (unlikely((body->valid & OBD_MD_MDS))) {
rc = lmv_intent_remote(exp, lmm, lmmsize, it, &op_data->op_fid1,
flags, reqp, cb_blocking,
extra_lock_flags);
if (rc != 0)
RETURN(rc);
body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
if (body == NULL)
RETURN(-EPROTO);
}
RETURN(rc);
}
int lmv_revalidate_slaves(struct obd_export *exp, struct mdt_body *mbody,
struct lmv_stripe_md *lsm,
ldlm_blocking_callback cb_blocking,
int extra_lock_flags)
{
struct obd_device *obd = exp->exp_obd;
struct lmv_obd *lmv = &obd->u.lmv;
struct mdt_body *body;
struct md_op_data *op_data;
unsigned long size = 0;
unsigned long nlink = 0;
obd_time atime = 0;
obd_time ctime = 0;
obd_time mtime = 0;
int i;
int rc = 0;
ENTRY;
/**
* revalidate slaves has some problems, temporarily return,
* we may not need that
*/
if (lsm->lsm_md_stripe_count <= 1)
RETURN(0);
OBD_ALLOC_PTR(op_data);
if (op_data == NULL)
RETURN(-ENOMEM);
/**
* Loop over the stripe information, check validity and update them
* from MDS if needed.
*/
for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
struct lu_fid fid;
struct lookup_intent it = { .it_op = IT_GETATTR };
struct ptlrpc_request *req = NULL;
struct lustre_handle *lockh = NULL;
struct lmv_tgt_desc *tgt = NULL;
struct inode *inode;
fid = lsm->lsm_md_oinfo[i].lmo_fid;
inode = lsm->lsm_md_oinfo[i].lmo_root;
if (i == 0) {
if (mbody != NULL) {
body = mbody;
goto update;
} else {
goto release_lock;
}
}
/*
* Prepare op_data for revalidating. Note that @fid2 shluld be
* defined otherwise it will go to server and take new lock
* which is not needed here.
*/
memset(op_data, 0, sizeof(*op_data));
op_data->op_fid1 = fid;
op_data->op_fid2 = fid;
tgt = lmv_locate_mds(lmv, op_data, &fid);
if (IS_ERR(tgt))
GOTO(cleanup, rc = PTR_ERR(tgt));
CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%d\n",
PFID(&fid), tgt->ltd_idx);
rc = md_intent_lock(tgt->ltd_exp, op_data, NULL, 0, &it, 0,
&req, cb_blocking, extra_lock_flags);
if (rc < 0)
GOTO(cleanup, rc);
lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
if (rc > 0 && req == NULL) {
/* slave inode is still valid */
CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
PFID(&fid));
rc = 0;
} else {
/* refresh slave from server */
body = req_capsule_server_get(&req->rq_pill,
&RMF_MDT_BODY);
LASSERT(body != NULL);
update:
if (unlikely(body->nlink < 2)) {
CERROR("%s: nlink %d < 2 corrupt stripe %d "DFID
":" DFID"\n", obd->obd_name, body->nlink,
i, PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
if (req != NULL)
ptlrpc_req_finished(req);
if (it.d.lustre.it_lock_mode && lockh) {
ldlm_lock_decref(lockh,
it.d.lustre.it_lock_mode);
it.d.lustre.it_lock_mode = 0;
}
GOTO(cleanup, rc = -EIO);
}
if (i != 0)
md_set_lock_data(tgt->ltd_exp, &lockh->cookie,
inode, NULL);
i_size_write(inode, body->size);
set_nlink(inode, body->nlink);
LTIME_S(inode->i_atime) = body->atime;
LTIME_S(inode->i_ctime) = body->ctime;
LTIME_S(inode->i_mtime) = body->mtime;
if (req != NULL)
ptlrpc_req_finished(req);
}
release_lock:
size += i_size_read(inode);
if (i != 0)
nlink += inode->i_nlink - 2;
else
nlink += inode->i_nlink;
atime = LTIME_S(inode->i_atime) > atime ?
LTIME_S(inode->i_atime) : atime;
ctime = LTIME_S(inode->i_ctime) > ctime ?
LTIME_S(inode->i_ctime) : ctime;
mtime = LTIME_S(inode->i_mtime) > mtime ?
LTIME_S(inode->i_mtime) : mtime;
if (it.d.lustre.it_lock_mode != 0 && lockh != NULL) {
ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
it.d.lustre.it_lock_mode = 0;
}
CDEBUG(D_INODE, "i %d "DFID" size %llu, nlink %u, atime "
"%lu, mtime %lu, ctime %lu.\n", i, PFID(&fid),
i_size_read(inode), inode->i_nlink,
LTIME_S(inode->i_atime), LTIME_S(inode->i_mtime),
LTIME_S(inode->i_ctime));
}
/*
* update attr of master request.
*/
CDEBUG(D_INODE, "Return refreshed attrs: size = %lu nlink %lu atime "
LPU64 "ctime "LPU64" mtime "LPU64" for " DFID"\n", size, nlink,
atime, ctime, mtime, PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
if (mbody != NULL) {
mbody->atime = atime;
mbody->ctime = ctime;
mbody->mtime = mtime;
}
cleanup:
OBD_FREE_PTR(op_data);
RETURN(rc);
}
int ll_revalidate_it(struct dentry *de, int lookup_flags,
struct lookup_intent *it)
{
struct md_op_data *op_data;
struct ptlrpc_request *req = NULL;
struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
struct obd_export *exp;
struct inode *parent = de->d_parent->d_inode;
int rc;
ENTRY;
CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
LL_IT2STR(it));
if (de->d_inode == NULL) {
__u64 ibits;
/* We can only use negative dentries if this is stat or lookup,
for opens and stuff we do need to query server. */
/* If there is IT_CREAT in intent op set, then we must throw
away this negative dentry and actually do the request to
kernel to create whatever needs to be created (if possible)*/
if (it && (it->it_op & IT_CREAT))
RETURN(0);
if (de->d_flags & DCACHE_LUSTRE_INVALID)
RETURN(0);
ibits = MDS_INODELOCK_UPDATE;
rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
GOTO(out_sa, rc);
}
/* Never execute intents for mount points.
* Attributes will be fixed up in ll_inode_revalidate_it */
if (d_mountpoint(de))
GOTO(out_sa, rc = 1);
/* need to get attributes in case root got changed from other client */
if (de == de->d_sb->s_root) {
rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
if (rc == 0)
rc = 1;
GOTO(out_sa, rc);
}
exp = ll_i2mdexp(de->d_inode);
OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
ll_frob_intent(&it, &lookup_it);
LASSERT(it);
if (it->it_op == IT_LOOKUP && !(de->d_flags & DCACHE_LUSTRE_INVALID))
RETURN(1);
if ((it->it_op == IT_OPEN) && de->d_inode) {
struct inode *inode = de->d_inode;
struct ll_inode_info *lli = ll_i2info(inode);
struct obd_client_handle **och_p;
__u64 *och_usecount;
__u64 ibits;
/*
* We used to check for MDS_INODELOCK_OPEN here, but in fact
* just having LOOKUP lock is enough to justify inode is the
* same. And if inode is the same and we have suitable
* openhandle, then there is no point in doing another OPEN RPC
* just to throw away newly received openhandle. There are no
* security implications too, if file owner or access mode is
* change, LOOKUP lock is revoked.
*/
if (it->it_flags & FMODE_WRITE) {
och_p = &lli->lli_mds_write_och;
och_usecount = &lli->lli_open_fd_write_count;
} else if (it->it_flags & FMODE_EXEC) {
och_p = &lli->lli_mds_exec_och;
och_usecount = &lli->lli_open_fd_exec_count;
} else {
och_p = &lli->lli_mds_read_och;
och_usecount = &lli->lli_open_fd_read_count;
}
/* Check for the proper lock. */
ibits = MDS_INODELOCK_LOOKUP;
if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
goto do_lock;
cfs_mutex_lock(&lli->lli_och_mutex);
if (*och_p) { /* Everything is open already, do nothing */
/*(*och_usecount)++; Do not let them steal our open
handle from under us */
SET_BUT_UNUSED(och_usecount);
/* XXX The code above was my original idea, but in case
we have the handle, but we cannot use it due to later
checks (e.g. O_CREAT|O_EXCL flags set), nobody
would decrement counter increased here. So we just
hope the lock won't be invalidated in between. But
if it would be, we'll reopen the open request to
MDS later during file open path */
cfs_mutex_unlock(&lli->lli_och_mutex);
RETURN(1);
} else {
cfs_mutex_unlock(&lli->lli_och_mutex);
}
}
if (it->it_op == IT_GETATTR) {
rc = ll_statahead_enter(parent, &de, 0);
if (rc == 1)
goto mark;
}
do_lock:
op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
de->d_name.name, de->d_name.len,
0, LUSTRE_OPC_ANY, NULL);
if (IS_ERR(op_data))
RETURN(PTR_ERR(op_data));
it->it_create_mode &= ~cfs_curproc_umask();
it->it_create_mode |= M_CHECK_STALE;
rc = md_intent_lock(exp, op_data, NULL, 0, it,
lookup_flags,
&req, ll_md_blocking_ast, 0);
it->it_create_mode &= ~M_CHECK_STALE;
ll_finish_md_op_data(op_data);
/* If req is NULL, then md_intent_lock only tried to do a lock match;
* if all was well, it will return 1 if it found locks, 0 otherwise. */
if (req == NULL && rc >= 0) {
if (!rc)
goto do_lookup;
GOTO(out, rc);
}
if (rc < 0) {
if (rc != -ESTALE) {
CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
"%d\n", rc, it->d.lustre.it_status);
}
GOTO(out, rc = 0);
}
revalidate_finish:
rc = ll_revalidate_it_finish(req, it, de);
if (rc != 0) {
if (rc != -ESTALE && rc != -ENOENT)
ll_intent_release(it);
GOTO(out, rc = 0);
}
if ((it->it_op & IT_OPEN) && de->d_inode &&
!S_ISREG(de->d_inode->i_mode) &&
!S_ISDIR(de->d_inode->i_mode)) {
ll_release_openhandle(de, it);
}
rc = 1;
/* unfortunately ll_intent_lock may cause a callback and revoke our
* dentry */
ll_dentry_rehash(de, 0);
out:
/* We do not free request as it may be reused during following lookup
* (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
* be freed in ll_lookup_it or in ll_intent_release. But if
* request was not completed, we need to free it. (bug 5154, 9903) */
if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
ptlrpc_req_finished(req);
if (rc == 0) {
ll_unhash_aliases(de->d_inode);
/* done in ll_unhash_aliases()
dentry->d_flags |= DCACHE_LUSTRE_INVALID; */
} else {
__u64 bits = 0;
CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
"inode %p refc %d\n", de->d_name.len,
de->d_name.name, de, de->d_parent, de->d_inode,
atomic_read(&de->d_count));
ll_set_lock_data(exp, de->d_inode, it, &bits);
ll_dentry_reset_flags(de, bits);
ll_lookup_finish_locks(it, de);
}
mark:
if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
ll_statahead_mark(parent, de);
RETURN(rc);
/*
* This part is here to combat evil-evil race in real_lookup on 2.6
* kernels. The race details are: We enter do_lookup() looking for some
* name, there is nothing in dcache for this name yet and d_lookup()
* returns NULL. We proceed to real_lookup(), and while we do this,
* another process does open on the same file we looking up (most simple
* reproducer), open succeeds and the dentry is added. Now back to
* us. In real_lookup() we do d_lookup() again and suddenly find the
* dentry, so we call d_revalidate on it, but there is no lock, so
* without this code we would return 0, but unpatched real_lookup just
* returns -ENOENT in such a case instead of retrying the lookup. Once
* this is dealt with in real_lookup(), all of this ugly mess can go and
* we can just check locks in ->d_revalidate without doing any RPCs
* ever.
*/
do_lookup:
if (it != &lookup_it) {
/* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
if (it->it_op == IT_GETATTR)
lookup_it.it_op = IT_GETATTR;
ll_lookup_finish_locks(it, de);
it = &lookup_it;
}
/* Do real lookup here. */
op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
de->d_name.len, 0, (it->it_op & IT_CREAT ?
LUSTRE_OPC_CREATE :
LUSTRE_OPC_ANY), NULL);
if (IS_ERR(op_data))
RETURN(PTR_ERR(op_data));
rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
ll_md_blocking_ast, 0);
if (rc >= 0) {
struct mdt_body *mdt_body;
struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
if (de->d_inode)
fid = *ll_inode2fid(de->d_inode);
/* see if we got same inode, if not - return error */
if (lu_fid_eq(&fid, &mdt_body->fid1)) {
ll_finish_md_op_data(op_data);
op_data = NULL;
goto revalidate_finish;
}
ll_intent_release(it);
}
ll_finish_md_op_data(op_data);
GOTO(out, rc = 0);
out_sa:
/*
* For rc == 1 case, should not return directly to prevent losing
* statahead windows; for rc == 0 case, the "lookup" will be done later.
*/
if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
ll_statahead_enter(parent, &de, 1);
goto mark;
}
int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
{
int rc;
ENTRY;
if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
struct lookup_intent *it;
it = ll_convert_intent(&nd->intent.open, nd->flags);
if (IS_ERR(it))
RETURN(0);
if (it->it_op == (IT_OPEN|IT_CREAT) &&
nd->intent.open.flags & O_EXCL) {
CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
rc = 0;
goto out_it;
}
rc = ll_revalidate_it(dentry, nd->flags, it);
if (rc && (nd->flags & LOOKUP_OPEN) &&
it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
// XXX Code duplication with ll_lookup_nd
if (S_ISFIFO(dentry->d_inode->i_mode)) {
// We cannot call open here as it would
// deadlock.
ptlrpc_req_finished(
(struct ptlrpc_request *)
it->d.lustre.it_data);
} else {
#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,17))
/* 2.6.1[456] have a bug in open_namei() that forgets to check
* nd->intent.open.file for error, so we need to return it as lookup's result
* instead */
struct file *filp;
nd->intent.open.file->private_data = it;
filp = lookup_instantiate_filp(nd, dentry,NULL);
if (IS_ERR(filp)) {
rc = PTR_ERR(filp);
}
#else
nd->intent.open.file->private_data = it;
(void)lookup_instantiate_filp(nd, dentry,NULL);
#endif
}
}
if (!rc && (nd->flags & LOOKUP_CREATE) &&
it_disposition(it, DISP_OPEN_CREATE)) {
/* We created something but we may only return
* negative dentry here, so save request in dentry,
* if lookup will be called later on, it will
* pick the request, otherwise it would be freed
* with dentry */
ll_d2d(dentry)->lld_it = it;
it = NULL; /* avoid freeing */
}
out_it:
if (it) {
ll_intent_release(it);
OBD_FREE(it, sizeof(*it));
}
} else {
rc = ll_revalidate_it(dentry, 0, NULL);
}
RETURN(rc);
}