#include "mdt_internal.h"
#include <lustre_nodemap.h>
-/* we do nothing because we do not have refcount now */
-static void mdt_mfd_get(void *mfdp)
-{
-}
-
-static const struct portals_handle_ops mfd_open_handle_ops = {
- .hop_addref = mdt_mfd_get,
- .hop_free = NULL,
-};
+static const char mfd_open_handle_owner[] = "mdt";
/* Create a new mdt_file_data struct, initialize it,
* and insert it to global hash table */
OBD_ALLOC_PTR(mfd);
if (mfd != NULL) {
- INIT_LIST_HEAD_RCU(&mfd->mfd_open_handle.h_link);
+ refcount_set(&mfd->mfd_open_handle.h_ref, 1);
+ INIT_HLIST_NODE(&mfd->mfd_open_handle.h_link);
mfd->mfd_owner = med;
INIT_LIST_HEAD(&mfd->mfd_list);
- class_handle_hash(&mfd->mfd_open_handle, &mfd_open_handle_ops);
+ class_handle_hash(&mfd->mfd_open_handle, mfd_open_handle_owner);
}
RETURN(mfd);
ENTRY;
LASSERT(open_handle != NULL);
- mfd = class_handle2object(open_handle->cookie, &mfd_open_handle_ops);
+ mfd = class_handle2object(open_handle->cookie, mfd_open_handle_owner);
+ if (mfd)
+ refcount_dec(&mfd->mfd_open_handle.h_ref);
+
/* during dw/setattr replay the mfd can be found by old handle */
if ((!mfd || mfd->mfd_owner != med) && is_replay_or_resent) {
list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
/* free mfd */
void mdt_mfd_free(struct mdt_file_data *mfd)
{
+ LASSERT(refcount_read(&mfd->mfd_open_handle.h_ref) == 1);
LASSERT(list_empty(&mfd->mfd_list));
- OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_open_handle);
+ OBD_FREE_PRE(mfd, sizeof(*mfd), "rcu");
+ kfree_rcu(mfd, mfd_open_handle.h_rcu);
}
static int mdt_create_data(struct mdt_thread_info *info,
struct mdt_object *obj,
struct md_attr *ma)
{
- LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
+ if (ma->ma_lmv || ma->ma_lmm) {
+ CDEBUG(D_INFO, DFID " %s already set.\n",
+ PFID(mdt_object_fid(obj)),
+ ma->ma_lmv ? (ma->ma_lmm ? "ma_lmv and ma_lmm"
+ : "ma_lmv")
+ : "ma_lmm");
+ return;
+ }
+
if (S_ISDIR(obj->mot_header.loh_attr)) {
ma->ma_lmv = req_capsule_server_get(info->mti_pill,
&RMF_MDT_MD);
bool try_layout = false;
bool create_layout = false;
int rc = 0;
- int dom_stripes = LMM_NO_DOM;
- bool dom_lock = false;
+ __u32 dom_stripe = 0;
+ unsigned int dom_only = 0;
+ unsigned int dom_lock = 0;
ENTRY;
ma->ma_need & MA_LOV)
try_layout = true;
- /* DoM files can have just MDT stripe or combined MDT + OST
- * stripes.
- * - In the first case the open for read/write will do IO to
- * the MDT stripe and it makes sense to take IO lock in
- * advance along with OPEN even if it is blocking lock.
- * - In the second case it is just size of MDT stripe and it
- * is quite unlikely that client will write into it, though
- * it may read it. So IO lock will be taken optionally if it
- * is non-blocking one.
+ /* DoM files can take IO lock at OPEN when it makes sense,
+ * check if file has DoM stripe and ask for lock if client
+ * no lock on that resource yet.
*/
if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
- dom_stripes = mdt_lmm_dom_entry(ma->ma_lmm);
-
- if (dom_stripes == LMM_DOM_ONLY &&
- info->mti_mdt->mdt_opts.mo_dom_lock > 0 &&
+ dom_stripe = mdt_lmm_dom_entry_check(ma->ma_lmm,
+ &dom_only);
+ /* If only DOM stripe is being used then we can expect IO
+ * to it after OPEN and will return corresponding DOM ibit
+ * using default strategy from mdt_opts.mo_dom_lock.
+ * Otherwise trylock mode is used always and DOM ibit will
+ * be returned optionally.
+ */
+ if (dom_stripe &&
!mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
- dom_lock = true;
+ dom_lock = !dom_only ? TRYLOCK_DOM_ON_OPEN :
+ info->mti_mdt->mdt_opts.mo_dom_lock;
}
if (acq_lease) {
lhc = &info->mti_lh[MDT_LH_LOCAL];
} else if (dom_lock) {
lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
- if (info->mti_mdt->mdt_opts.mo_dom_lock ==
- TRYLOCK_DOM_ON_OPEN) {
- trybits |= MDS_INODELOCK_DOM |
- MDS_INODELOCK_LAYOUT;
- } else {
- /* mo_dom_lock == ALWAYS_DOM_LOCK_ON_OPEN */
- *ibits = MDS_INODELOCK_DOM;
- if (info->mti_mdt->mdt_opts.mo_dom_read_open) {
- trybits |= MDS_INODELOCK_LAYOUT;
- }
- }
+ trybits |= MDS_INODELOCK_DOM | MDS_INODELOCK_LAYOUT;
}
CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
if (rc != 0)
GOTO(out, rc);
- mdt_pack_secctx_in_reply(info, o);
+ rc = mdt_pack_secctx_in_reply(info, o);
+ if (unlikely(rc))
+ GOTO(out, rc);
+
+ rc = mdt_pack_encctx_in_reply(info, o);
+ if (unlikely(rc))
+ GOTO(out, rc);
rc = mdt_finish_open(info, NULL, o, open_flags, 0, rep);
} else {
int result, rc;
int created = 0;
int object_locked = 0;
+ enum ldlm_mode lock_mode = LCK_PR;
u32 msg_flags;
+ ktime_t kstart = ktime_get();
ENTRY;
OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
(obd_timeout + 1) / 4);
- mdt_counter_incr(req, LPROC_MDT_OPEN);
repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
ma->ma_need = MA_INODE;
if (result < 0)
GOTO(out, result);
-again:
- lh = &info->mti_lh[MDT_LH_PARENT];
- mdt_lock_pdo_init(lh, (open_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
- &rr->rr_name);
-
parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
if (IS_ERR(parent))
GOTO(out, result = PTR_ERR(parent));
- result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
- if (result != 0) {
+ /* get and check version of parent */
+ result = mdt_version_get_check(info, parent, 0);
+ if (result) {
mdt_object_put(info->mti_env, parent);
GOTO(out, result);
}
- /* get and check version of parent */
- result = mdt_version_get_check(info, parent, 0);
- if (result)
- GOTO(out_parent, result);
+ OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
+again_pw:
+ lh = &info->mti_lh[MDT_LH_PARENT];
+ mdt_lock_pdo_init(lh, lock_mode, &rr->rr_name);
+ result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
+ if (result != 0) {
+ mdt_object_put(info->mti_env, parent);
+ GOTO(out, result);
+ }
fid_zero(child_fid);
result = -ENOENT;
PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
PFID(child_fid));
- if (result != 0 && result != -ENOENT && result != -ESTALE)
+ if (result != 0 && result != -ENOENT)
GOTO(out_parent, result);
- if (result == -ENOENT || result == -ESTALE) {
- /* If the object is dead, let's check if the object
- * is being migrated to a new object */
- if (result == -ESTALE) {
- struct lu_buf lmv_buf;
-
- lmv_buf.lb_buf = info->mti_xattr_buf;
- lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
- rc = mo_xattr_get(info->mti_env,
- mdt_object_child(parent),
- &lmv_buf, XATTR_NAME_LMV);
- if (rc > 0) {
- struct lmv_mds_md_v1 *lmv;
-
- lmv = lmv_buf.lb_buf;
- if (le32_to_cpu(lmv->lmv_hash_type) &
- LMV_HASH_FLAG_MIGRATION) {
- /* Get the new parent FID and retry */
- mdt_object_unlock_put(info, parent,
- lh, 1);
- mdt_lock_handle_init(lh);
- fid_le_to_cpu(
- (struct lu_fid *)rr->rr_fid1,
- &lmv->lmv_stripe_fids[1]);
- goto again;
- }
- }
- }
+ OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
+ if (result == -ENOENT) {
mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
- if (result == -ESTALE) {
- /*
- * -ESTALE means the parent is a dead(unlinked) dir, so
- * it should return -ENOENT to in accordance with the
- * original mds implementaion.
- */
- GOTO(out_parent, result = -ENOENT);
- }
-
if (!(open_flags & MDS_OPEN_CREAT))
GOTO(out_parent, result);
if (mdt_rdonly(req->rq_export))
GOTO(out_parent, result = -EROFS);
+
+ if (lock_mode == LCK_PR) {
+ /* first pass: get write lock and restart */
+ mdt_object_unlock(info, parent, lh, 1);
+ mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
+ mdt_lock_handle_init(lh);
+ lock_mode = LCK_PW;
+ goto again_pw;
+ }
+
*child_fid = *info->mti_rr.rr_fid2;
LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
PFID(child_fid));
GOTO(out_child, result);
}
created = 1;
- mdt_counter_incr(req, LPROC_MDT_MKNOD);
+ mdt_counter_incr(req, LPROC_MDT_MKNOD,
+ ktime_us_delta(ktime_get(), kstart));
} else {
/*
* The object is on remote node, return its FID for remote open.
result = -MDT_EREMOTE_OPEN;
GOTO(out_child, result);
} else if (mdt_object_exists(child)) {
+ /* Check early for MDS_OPEN_DIRECTORY/O_DIRECTORY to
+ * avoid opening regular files from lfs getstripe
+ * since doing so breaks the leases used by lfs
+ * mirror. See LU-13693. */
+ if (open_flags & MDS_OPEN_DIRECTORY &&
+ S_ISREG(lu_object_attr(&child->mot_obj)))
+ GOTO(out_child, result = -ENOTDIR);
+
/* We have to get attr & LOV EA & HSM for this
* object. */
mdt_prep_ma_buf_from_rep(info, child, ma);
}
}
- mdt_pack_secctx_in_reply(info, child);
+ repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
+ repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
+
+ rc = mdt_pack_secctx_in_reply(info, child);
+ if (unlikely(rc))
+ GOTO(out_child, result = rc);
+
+ rc = mdt_pack_encctx_in_reply(info, child);
+ if (unlikely(rc))
+ GOTO(out_child, result = rc);
rc = mdt_check_resent_lock(info, child, lhc);
if (rc < 0) {
mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
}
}
+
+ mdt_counter_incr(req, LPROC_MDT_OPEN,
+ ktime_us_delta(ktime_get(), kstart));
+
EXIT;
out_child_unlock:
if (object_locked)
uc = lu_ucred(env);
uc_cap_save = uc->uc_cap;
- uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
+ uc->uc_cap |= BIT(CAP_DAC_OVERRIDE);
rc = mdo_create(env, mdt_object_child(local_root), &lname,
mdt_object_child(obj), spec, attr);
uc->uc_cap = uc_cap_save;
/* The orphan has root ownership so we need to raise
* CAP_FOWNER to set the HSM attributes. */
cap = uc->uc_cap;
- uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
+ uc->uc_cap |= MD_CAP_TO_MASK(CAP_FOWNER);
rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
XATTR_NAME_HSM, 0);
uc->uc_cap = cap;
struct mdt_lock_handle *lh2 = &info->mti_lh[MDT_LH_OLD];
struct close_data *data;
struct ldlm_lock *lease;
- struct mdt_object *o1 = o, *o2;
+ struct mdt_object *o1 = o, *o2 = NULL;
bool lease_broken;
bool swap_objects;
int rc;
if (IS_ERR(o2))
GOTO(out_lease, rc = PTR_ERR(o2));
- if (!S_ISREG(lu_object_attr(&o2->mot_obj))) {
- swap_objects = false; /* not swapped yet */
+ if (!mdt_object_exists(o2))
+ GOTO(out_obj, rc = -ENOENT);
+
+ if (!S_ISREG(lu_object_attr(&o2->mot_obj)))
GOTO(out_obj, rc = -EINVAL);
- }
if (swap_objects)
swap(o1, o2);
}
out_obj:
- mdt_object_put(info->mti_env, swap_objects ? o1 : o2);
+ /* Callee takes care of o, we must put the other one. We know
+ * that o1 != o2 from check of lu_fid_cmp() above. */
+ mdt_object_put(info->mti_env, o1 != o ? o1 : o2);
ldlm_reprocess_all(lease->l_resource, lease);
RCL_CLIENT))
GOTO(out_unlock, rc = -EPROTO);
- OBD_ALLOC(resync_ids, resync_count * sizeof(__u32));
+ OBD_ALLOC_PTR_ARRAY(resync_ids, resync_count);
if (!resync_ids)
GOTO(out_unlock, rc = -ENOMEM);
}
if (resync_ids)
- OBD_FREE(resync_ids, resync_count * sizeof(__u32));
+ OBD_FREE_PTR_ARRAY(resync_ids, resync_count);
out_reprocess:
ldlm_reprocess_all(lease->l_resource, lease);
else if (open_flags & MDS_FMODE_EXEC)
mdt_write_allow(o);
- /* Update atime on close only. */
+ /* Update atime|mtime|ctime on close. */
if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
- (ma->ma_attr.la_valid & LA_ATIME)) {
- /* Set the atime only. */
- ma->ma_valid = MA_INODE;
- ma->ma_attr.la_valid = LA_ATIME;
- rc = mo_attr_set(info->mti_env, next, ma);
- }
+ (ma->ma_attr.la_valid & LA_ATIME ||
+ ma->ma_attr.la_valid & LA_MTIME ||
+ ma->ma_attr.la_valid & LA_CTIME)) {
+ ma->ma_valid = MA_INODE;
+ ma->ma_attr_flags |= MDS_CLOSE_UPDATE_TIMES;
+ ma->ma_attr.la_valid &= (LA_ATIME | LA_MTIME | LA_CTIME);
+
+ if (ma->ma_attr.la_valid & LA_MTIME) {
+ rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
+ if (!rc)
+ ma->ma_valid |= MA_PFID;
+ }
+
+ rc = mo_attr_set(info->mti_env, next, ma);
+ }
/* If file data is modified, add the dirty flag. */
if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
struct ptlrpc_request *req = tgt_ses_req(tsi);
struct md_attr *ma = &info->mti_attr;
struct mdt_body *repbody = NULL;
+ ktime_t kstart = ktime_get();
int rc, ret = 0;
ENTRY;
- mdt_counter_incr(req, LPROC_MDT_CLOSE);
/* Close may come with the Size-on-MDS update. Unpack it. */
rc = mdt_close_unpack(info);
if (rc)
tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
out:
mdt_thread_info_fini(info);
+ if (rc == 0)
+ mdt_counter_incr(req, LPROC_MDT_CLOSE,
+ ktime_us_delta(ktime_get(), kstart));
RETURN(rc ? rc : ret);
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff