* GPL HEADER END
*/
/*
- * Copyright (c) 2012, 2017 Intel Corporation.
+ * Copyright (c) 2017, Intel Corporation.
*/
/*
* lustre/mdt/mdt_io.c
#include <dt_object.h>
#include "mdt_internal.h"
-/* --------------- MDT grant code ---------------- */
+/* functions below are stubs for now, they will be implemented with
+ * grant support on MDT */
+static inline void mdt_io_counter_incr(struct obd_export *exp, int opcode,
+ char *jobid, long amount)
+{
+ return;
+}
-long mdt_grant_connect(const struct lu_env *env,
- struct obd_export *exp,
- u64 want, bool conservative)
+static inline void mdt_dom_read_lock(struct mdt_object *mo)
{
- struct mdt_device *mdt = mdt_exp2dev(exp);
- u64 left;
- long grant;
+ down_read(&mo->mot_dom_sem);
+}
- ENTRY;
+static inline void mdt_dom_read_unlock(struct mdt_object *mo)
+{
+ up_read(&mo->mot_dom_sem);
+}
- dt_statfs(env, mdt->mdt_bottom, &mdt->mdt_osfs);
+static inline void mdt_dom_write_lock(struct mdt_object *mo)
+{
+ down_write(&mo->mot_dom_sem);
+}
- left = (mdt->mdt_osfs.os_bavail * mdt->mdt_osfs.os_bsize) / 2;
+static inline void mdt_dom_write_unlock(struct mdt_object *mo)
+{
+ up_write(&mo->mot_dom_sem);
+}
- grant = left;
+/**
+ * Lock prolongation for Data-on-MDT.
+ * This is similar to OFD code but for DOM ibits lock.
+ */
+static inline time64_t prolong_timeout(struct ptlrpc_request *req)
+{
+ struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
+ time64_t req_timeout;
+
+ if (AT_OFF)
+ return obd_timeout / 2;
+
+ req_timeout = req->rq_deadline - req->rq_arrival_time.tv_sec;
+ return max_t(time64_t, at_est2timeout(at_get(&svcpt->scp_at_estimate)),
+ req_timeout);
+}
- CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %ld want: %llu left: %llu\n",
- exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
- exp, grant, want, left);
+static void mdt_prolong_dom_lock(struct tgt_session_info *tsi,
+ struct ldlm_prolong_args *data)
+{
+ struct obdo *oa = &tsi->tsi_ost_body->oa;
+ struct ldlm_lock *lock;
+
+ ENTRY;
- return grant;
+ data->lpa_timeout = prolong_timeout(tgt_ses_req(tsi));
+ data->lpa_export = tsi->tsi_exp;
+ data->lpa_resid = tsi->tsi_resid;
+
+ CDEBUG(D_RPCTRACE, "Prolong DOM lock for req %p with x%llu\n",
+ tgt_ses_req(tsi), tgt_ses_req(tsi)->rq_xid);
+
+ if (oa->o_valid & OBD_MD_FLHANDLE) {
+ /* mostly a request should be covered by only one lock, try
+ * fast path. */
+ lock = ldlm_handle2lock(&oa->o_handle);
+ if (lock != NULL) {
+ LASSERT(lock->l_export == data->lpa_export);
+ ldlm_lock_prolong_one(lock, data);
+ lock->l_last_used = ktime_get();
+ LDLM_LOCK_PUT(lock);
+ RETURN_EXIT;
+ }
+ }
+ EXIT;
}
-void mdt_grant_prepare_write(const struct lu_env *env,
- struct obd_export *exp, struct obdo *oa,
- struct niobuf_remote *rnb, int niocount)
+static int mdt_rw_hpreq_lock_match(struct ptlrpc_request *req,
+ struct ldlm_lock *lock)
{
- struct mdt_device *mdt = mdt_exp2dev(exp);
- u64 left;
+ struct obd_ioobj *ioo;
+ enum ldlm_mode mode;
+ __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
ENTRY;
- left = (mdt->mdt_osfs.os_bavail * mdt->mdt_osfs.os_bsize) / 2;
+ if (!(lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_DOM))
+ RETURN(0);
+
+ ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
+ LASSERT(ioo != NULL);
- /* grant more space back to the client if possible */
- oa->o_grant = left;
+ LASSERT(lock->l_resource != NULL);
+ if (!fid_res_name_eq(&ioo->ioo_oid.oi_fid, &lock->l_resource->lr_name))
+ RETURN(0);
+
+ /* a bulk write can only hold a reference on a PW extent lock. */
+ mode = LCK_PW;
+ if (opc == OST_READ)
+ /* whereas a bulk read can be protected by either a PR or PW
+ * extent lock */
+ mode |= LCK_PR;
+
+ if (!(lock->l_granted_mode & mode))
+ RETURN(0);
+
+ RETURN(1);
}
-/* ---------------- end of MDT grant code ---------------- */
-/* functions below are stubs for now, they will be implemented with
- * grant support on MDT */
-static inline void mdt_io_counter_incr(struct obd_export *exp, int opcode,
- char *jobid, long amount)
+static int mdt_rw_hpreq_check(struct ptlrpc_request *req)
{
- return;
+ struct tgt_session_info *tsi;
+ struct obd_ioobj *ioo;
+ struct niobuf_remote *rnb;
+ int opc;
+ struct ldlm_prolong_args pa = { 0 };
+
+ ENTRY;
+
+ /* Don't use tgt_ses_info() to get session info, because lock_match()
+ * can be called while request has no processing thread yet. */
+ tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
+
+ /*
+ * Use LASSERT below because malformed RPCs should have
+ * been filtered out in tgt_hpreq_handler().
+ */
+ opc = lustre_msg_get_opc(req->rq_reqmsg);
+ LASSERT(opc == OST_READ || opc == OST_WRITE);
+
+ ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
+ LASSERT(ioo != NULL);
+
+ rnb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
+ LASSERT(rnb != NULL);
+ LASSERT(!(rnb->rnb_flags & OBD_BRW_SRVLOCK));
+
+ pa.lpa_mode = LCK_PW;
+ if (opc == OST_READ)
+ pa.lpa_mode |= LCK_PR;
+
+ DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks: "DFID"\n",
+ tgt_name(tsi->tsi_tgt), current->comm, PFID(&tsi->tsi_fid));
+
+ mdt_prolong_dom_lock(tsi, &pa);
+
+ if (pa.lpa_blocks_cnt > 0) {
+ CDEBUG(D_DLMTRACE,
+ "%s: refreshed %u locks timeout for req %p.\n",
+ tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
+ RETURN(1);
+ }
+
+ RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
}
-void mdt_grant_prepare_read(const struct lu_env *env,
- struct obd_export *exp, struct obdo *oa)
+static void mdt_rw_hpreq_fini(struct ptlrpc_request *req)
{
- return;
+ mdt_rw_hpreq_check(req);
}
-void mdt_grant_commit(struct obd_export *exp, unsigned long pending,
- int rc)
+static struct ptlrpc_hpreq_ops mdt_hpreq_rw = {
+ .hpreq_lock_match = mdt_rw_hpreq_lock_match,
+ .hpreq_check = mdt_rw_hpreq_check,
+ .hpreq_fini = mdt_rw_hpreq_fini
+};
+
+/**
+ * Assign high priority operations to an IO request.
+ *
+ * Check if the incoming request is a candidate for
+ * high-priority processing. If it is, assign it a high
+ * priority operations table.
+ *
+ * \param[in] tsi target session environment for this request
+ */
+void mdt_hp_brw(struct tgt_session_info *tsi)
{
- return;
+ struct niobuf_remote *rnb;
+ struct obd_ioobj *ioo;
+
+ ENTRY;
+ ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
+ LASSERT(ioo != NULL); /* must exist after request preprocessing */
+ if (ioo->ioo_bufcnt > 0) {
+ rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
+ LASSERT(rnb != NULL); /* must exist after preprocessing */
+
+ /* no high priority if server lock is needed */
+ if (rnb->rnb_flags & OBD_BRW_SRVLOCK ||
+ (lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) &
+ MSG_REPLAY))
+ return;
+ }
+ tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_rw;
}
-static inline void mdt_dom_read_lock(struct mdt_object *mo)
+static int mdt_punch_hpreq_lock_match(struct ptlrpc_request *req,
+ struct ldlm_lock *lock)
{
- down_read(&mo->mot_dom_sem);
+ struct tgt_session_info *tsi;
+ struct obdo *oa;
+
+ ENTRY;
+
+ /* Don't use tgt_ses_info() to get session info, because lock_match()
+ * can be called while request has no processing thread yet. */
+ tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
+
+ /*
+ * Use LASSERT below because malformed RPCs should have
+ * been filtered out in tgt_hpreq_handler().
+ */
+ LASSERT(tsi->tsi_ost_body != NULL);
+ if (tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLHANDLE &&
+ tsi->tsi_ost_body->oa.o_handle.cookie == lock->l_handle.h_cookie)
+ RETURN(1);
+
+ oa = &tsi->tsi_ost_body->oa;
+
+ LASSERT(lock->l_resource != NULL);
+ if (!fid_res_name_eq(&oa->o_oi.oi_fid, &lock->l_resource->lr_name))
+ RETURN(0);
+
+ if (!(lock->l_granted_mode & LCK_PW))
+ RETURN(0);
+
+ RETURN(1);
}
-static inline void mdt_dom_read_unlock(struct mdt_object *mo)
+/**
+ * Implementation of ptlrpc_hpreq_ops::hpreq_lock_check for OST_PUNCH request.
+ *
+ * High-priority queue request check for whether the given punch request
+ * (\a req) is blocking an LDLM lock cancel. Also checks whether the request is
+ * covered by an LDLM lock.
+ *
+
+ *
+ * \param[in] req the incoming request
+ *
+ * \retval 1 if \a req is blocking an LDLM lock cancel
+ * \retval 0 if it is not
+ * \retval -ESTALE if lock is not found
+ */
+static int mdt_punch_hpreq_check(struct ptlrpc_request *req)
{
- up_read(&mo->mot_dom_sem);
+ struct tgt_session_info *tsi;
+ struct obdo *oa;
+ struct ldlm_prolong_args pa = { 0 };
+
+ ENTRY;
+
+ /* Don't use tgt_ses_info() to get session info, because lock_match()
+ * can be called while request has no processing thread yet. */
+ tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
+ LASSERT(tsi != NULL);
+ oa = &tsi->tsi_ost_body->oa;
+
+ LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS &&
+ oa->o_flags & OBD_FL_SRVLOCK));
+
+ pa.lpa_mode = LCK_PW;
+
+ CDEBUG(D_DLMTRACE, "%s: refresh DOM lock for "DFID"\n",
+ tgt_name(tsi->tsi_tgt), PFID(&tsi->tsi_fid));
+
+ mdt_prolong_dom_lock(tsi, &pa);
+
+
+ if (pa.lpa_blocks_cnt > 0) {
+ CDEBUG(D_DLMTRACE,
+ "%s: refreshed %u locks timeout for req %p.\n",
+ tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
+ RETURN(1);
+ }
+
+ RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
}
-static inline void mdt_dom_write_lock(struct mdt_object *mo)
+/**
+ * Implementation of ptlrpc_hpreq_ops::hpreq_lock_fini for OST_PUNCH request.
+ *
+ * Called after the request has been handled. It refreshes lock timeout again
+ * so that client has more time to send lock cancel RPC.
+ *
+ * \param[in] req request which is being processed.
+ */
+static void mdt_punch_hpreq_fini(struct ptlrpc_request *req)
{
- down_write(&mo->mot_dom_sem);
+ mdt_punch_hpreq_check(req);
}
-static inline void mdt_dom_write_unlock(struct mdt_object *mo)
+static struct ptlrpc_hpreq_ops mdt_hpreq_punch = {
+ .hpreq_lock_match = mdt_punch_hpreq_lock_match,
+ .hpreq_check = mdt_punch_hpreq_check,
+ .hpreq_fini = mdt_punch_hpreq_fini
+};
+
+void mdt_hp_punch(struct tgt_session_info *tsi)
{
- up_write(&mo->mot_dom_sem);
+ LASSERT(tsi->tsi_ost_body != NULL); /* must exists if we are here */
+ /* no high-priority if server lock is needed */
+ if ((tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
+ tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK) ||
+ tgt_conn_flags(tsi) & OBD_CONNECT_MDS ||
+ lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) & MSG_REPLAY)
+ return;
+ tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_punch;
}
static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp,
/* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
* space back if possible */
- mdt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
+ tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
mdt_dom_read_lock(mo);
if (!mdt_object_exists(mo)) {
if (unlikely(rc < 0))
GOTO(err, rc);
/* correct index for local buffers to continue with */
- for (k = 0; k < rc; k++)
- lnb[j+k].lnb_flags = rnb[i].rnb_flags;
+ for (k = 0; k < rc; k++) {
+ lnb[j + k].lnb_flags = rnb[i].rnb_flags;
+ if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
+ lnb[j + k].lnb_rc = -ENOSPC;
+ }
j += rc;
*nr_local += rc;
tot_bytes += rnb[i].rnb_len;
unlock:
mdt_dom_read_unlock(mo);
/* tgt_grant_prepare_write() was called, so we must commit */
- mdt_grant_commit(exp, oa->o_grant_used, rc);
+ tgt_grant_commit(exp, oa->o_grant_used, rc);
/* let's still process incoming grant information packed in the oa,
* but without enforcing grant since we won't proceed with the write.
* Just like a read request actually. */
- mdt_grant_prepare_read(env, exp, oa);
+ tgt_grant_prepare_read(env, exp, oa);
return rc;
}
objcount, obj, rnb, nr_local, lnb,
jobid);
} else if (cmd == OBD_BRW_READ) {
- mdt_grant_prepare_read(env, exp, oa);
+ tgt_grant_prepare_read(env, exp, oa);
rc = mdt_preprw_read(env, exp, mdt, mo, la,
obj->ioo_bufcnt, rnb, nr_local, lnb,
jobid);
GOTO(out_stop, rc);
}
+ tgt_vbr_obj_set(env, dob);
rc = dt_trans_start(env, dt, th);
if (rc)
GOTO(out_stop, rc);
if (rc == -ENOSPC)
th->th_sync = 1;
+
+ if (rc == 0 && granted > 0) {
+ if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
+ granted = 0;
+ }
+
th->th_result = rc;
dt_trans_stop(env, dt, th);
if (rc == -ENOSPC && retries++ < 3) {
out:
dt_bufs_put(env, dob, lnb, niocount);
mdt_dom_read_unlock(mo);
- mdt_grant_commit(exp, granted, old_rc);
+ if (granted > 0)
+ tgt_grant_commit(exp, granted, old_rc);
RETURN(rc);
}
+void mdt_dom_obj_lvb_update(const struct lu_env *env, struct mdt_object *mo,
+ bool increase_only)
+{
+ struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
+ struct ldlm_res_id resid;
+ struct ldlm_resource *res;
+
+ fid_build_reg_res_name(mdt_object_fid(mo), &resid);
+ res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
+ LDLM_IBITS, 1);
+ if (IS_ERR(res))
+ return;
+
+ /* Update lvbo data if exists. */
+ if (mdt_dom_lvb_is_valid(res))
+ mdt_dom_disk_lvbo_update(env, mo, res, increase_only);
+ ldlm_resource_putref(res);
+}
+
int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
struct obdo *oa, int objcount, struct obd_ioobj *obj,
struct niobuf_remote *rnb, int npages,
else
obdo_from_la(oa, la, LA_GID | LA_UID);
+ mdt_dom_obj_lvb_update(env, mo, false);
/* don't report overquota flag if we failed before reaching
* commit */
if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
OBD_MD_FLGRPQUOTA;
}
} else if (cmd == OBD_BRW_READ) {
+ /* If oa != NULL then mdt_preprw_read updated the inode
+ * atime and we should update the lvb so that other glimpses
+ * will also get the updated value. bug 5972 */
+ if (oa)
+ mdt_dom_obj_lvb_update(env, mo, true);
rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb);
if (old_rc)
rc = old_rc;
if (IS_ERR(mo))
GOTO(out_unlock, rc = PTR_ERR(mo));
- mdt_dom_write_lock(mo);
if (!mdt_object_exists(mo))
GOTO(out_put, rc = -ENOENT);
+
+ /* Shouldn't happen on dirs */
+ if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
+ rc = -EPERM;
+ CERROR("%s: Truncate on dir "DFID": rc = %d\n",
+ exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
+ GOTO(out_put, rc);
+ }
+
+ mdt_dom_write_lock(mo);
dob = mdt_obj2dt(mo);
la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
if (rc)
GOTO(out_put, rc);
+ mdt_dom_obj_lvb_update(tsi->tsi_env, mo, false);
mdt_io_counter_incr(tsi->tsi_exp, LPROC_MDT_IO_PUNCH,
tsi->tsi_jobid, 1);
EXIT;
lu_object_put(tsi->tsi_env, &mo->mot_obj);
out_unlock:
if (srvlock)
- mdt_save_lock(info, &lh, LCK_PW, rc);
+ tgt_extent_unlock(&lh, LCK_PW);
out:
mdt_thread_info_fini(info);
- if (rc == 0) {
- struct ldlm_resource *res;
-
- /* we do not call this before to avoid lu_object_find() in
- * ->lvbo_update() holding another reference on the object.
- * otherwise concurrent destroy can make the object unavailable
- * for 2nd lu_object_find() waiting for the first reference
- * to go... deadlock! */
- res = ldlm_resource_get(ns, NULL, &tsi->tsi_resid,
- LDLM_IBITS, 0);
- if (!IS_ERR(res)) {
- ldlm_res_lvbo_update(res, NULL, 0);
- ldlm_resource_putref(res);
- }
+ return rc;
+}
+
+/**
+ * MDT glimpse for Data-on-MDT
+ *
+ * If there is write lock on client then function issues glimpse_ast to get
+ * an actual size from that client.
+ *
+ */
+int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
+ struct ldlm_resource *res)
+{
+ union ldlm_policy_data policy;
+ struct lustre_handle lockh;
+ enum ldlm_mode mode;
+ struct ldlm_lock *lock;
+ struct ldlm_glimpse_work *gl_work;
+ struct list_head gl_list;
+ int rc;
+
+ ENTRY;
+
+ /* There can be only one write lock covering data, try to match it. */
+ policy.l_inodebits.bits = MDS_INODELOCK_DOM;
+ mode = ldlm_lock_match(ns, LDLM_FL_TEST_LOCK,
+ &res->lr_name, LDLM_IBITS, &policy,
+ LCK_PW, &lockh, 0);
+
+ /* There is no PW lock on this object; finished. */
+ if (mode == 0)
+ RETURN(0);
+
+ lock = ldlm_handle2lock(&lockh);
+ if (lock == NULL)
+ RETURN(0);
+
+ /*
+ * This check is for lock taken in mdt_reint_unlink() that does
+ * not have l_glimpse_ast set. So the logic is: if there is a lock
+ * with no l_glimpse_ast set, this object is being destroyed already.
+ * Hence, if you are grabbing DLM locks on the server, always set
+ * non-NULL glimpse_ast (e.g., ldlm_request.c::ldlm_glimpse_ast()).
+ */
+ if (lock->l_glimpse_ast == NULL) {
+ LDLM_DEBUG(lock, "no l_glimpse_ast");
+ GOTO(out, rc = -ENOENT);
+ }
+
+ OBD_SLAB_ALLOC_PTR_GFP(gl_work, ldlm_glimpse_work_kmem, GFP_ATOMIC);
+ if (!gl_work)
+ GOTO(out, rc = -ENOMEM);
+
+ /* Populate the gl_work structure.
+ * Grab additional reference on the lock which will be released in
+ * ldlm_work_gl_ast_lock() */
+ gl_work->gl_lock = LDLM_LOCK_GET(lock);
+ /* The glimpse callback is sent to one single IO lock. As a result,
+ * the gl_work list is just composed of one element */
+ INIT_LIST_HEAD(&gl_list);
+ list_add_tail(&gl_work->gl_list, &gl_list);
+ /* There is actually no need for a glimpse descriptor when glimpsing
+ * IO locks */
+ gl_work->gl_desc = NULL;
+ /* the ldlm_glimpse_work structure is allocated on the stack */
+ gl_work->gl_flags = LDLM_GL_WORK_SLAB_ALLOCATED;
+
+ ldlm_glimpse_locks(res, &gl_list); /* this will update the LVB */
+
+ /* If the list is not empty, we failed to glimpse a lock and
+ * must clean it up. Usually due to a race with unlink.*/
+ if (!list_empty(&gl_list)) {
+ LDLM_LOCK_RELEASE(lock);
+ OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
+ }
+ rc = 0;
+ EXIT;
+out:
+ LDLM_LOCK_PUT(lock);
+ return rc;
+}
+
+static void mdt_lvb2body(struct ldlm_resource *res, struct mdt_body *mb)
+{
+ struct ost_lvb *res_lvb;
+
+ lock_res(res);
+ res_lvb = res->lr_lvb_data;
+ mb->mbo_dom_size = res_lvb->lvb_size;
+ mb->mbo_dom_blocks = res_lvb->lvb_blocks;
+ mb->mbo_mtime = res_lvb->lvb_mtime;
+ mb->mbo_ctime = res_lvb->lvb_ctime;
+ mb->mbo_atime = res_lvb->lvb_atime;
+
+ CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
+
+ mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME | OBD_MD_FLMTIME |
+ OBD_MD_DOM_SIZE;
+ unlock_res(res);
+}
+
+/**
+ * MDT glimpse for Data-on-MDT
+ *
+ * This function is called when MDT get attributes for the DoM object.
+ * If there is write lock on client then function issues glimpse_ast to get
+ * an actual size from that client.
+ */
+int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt,
+ const struct lu_fid *fid, struct mdt_body *mb,
+ bool dom_lock)
+{
+ struct ldlm_res_id resid;
+ struct ldlm_resource *res;
+ int rc = 0;
+
+ ENTRY;
+
+ fid_build_reg_res_name(fid, &resid);
+ res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
+ LDLM_IBITS, 1);
+ if (IS_ERR(res))
+ RETURN(-ENOENT);
+
+ /* Update lvbo data if DoM lock returned or if LVB is not yet valid. */
+ if (dom_lock || !mdt_dom_lvb_is_valid(res))
+ mdt_dom_lvbo_update(res, NULL, NULL, false);
+
+ mdt_lvb2body(res, mb);
+ ldlm_resource_putref(res);
+ RETURN(rc);
+}
+
+/**
+ * MDT DoM lock intent policy (glimpse)
+ *
+ * Intent policy is called when lock has an intent, for DoM file that
+ * means glimpse lock and policy fills Lock Value Block (LVB).
+ *
+ * If already granted lock is found it will be placed in \a lockp and
+ * returned back to caller function.
+ *
+ * \param[in] tsi session info
+ * \param[in,out] lockp pointer to the lock
+ * \param[in] flags LDLM flags
+ *
+ * \retval ELDLM_LOCK_REPLACED if already granted lock was found
+ * and placed in \a lockp
+ * \retval ELDLM_LOCK_ABORTED in other cases except error
+ * \retval negative value on error
+ */
+int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
+ struct ldlm_lock **lockp, __u64 flags)
+{
+ struct ldlm_lock *lock = *lockp;
+ struct ldlm_resource *res = lock->l_resource;
+ ldlm_processing_policy policy;
+ struct ldlm_reply *rep;
+ struct mdt_body *mbo;
+ int rc;
+
+ ENTRY;
+
+ policy = ldlm_get_processing_policy(res);
+ LASSERT(policy != NULL);
+
+ req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
+ req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
+ rc = req_capsule_server_pack(mti->mti_pill);
+ if (rc)
+ RETURN(err_serious(rc));
+
+ rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
+ if (rep == NULL)
+ RETURN(-EPROTO);
+
+ mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
+ if (mbo == NULL)
+ RETURN(-EPROTO);
+
+ lock_res(res);
+ /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
+ * lock was found by ldlm_handle_enqueue(); if so no need to grant
+ * it again. */
+ if (flags & LDLM_FL_RESENT) {
+ rc = LDLM_ITER_CONTINUE;
+ } else {
+ __u64 tmpflags = 0;
+ enum ldlm_error err;
+
+ rc = policy(lock, &tmpflags, LDLM_PROCESS_RESCAN, &err, NULL);
+ check_res_locked(res);
+ }
+ unlock_res(res);
+
+ /* The lock met with no resistance; we're finished. */
+ if (rc == LDLM_ITER_CONTINUE) {
+ GOTO(fill_mbo, rc = ELDLM_LOCK_REPLACED);
+ } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
+ /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
+ * callback for glimpse size. The real size user will trigger
+ * the glimpse callback when necessary. */
+ GOTO(fill_mbo, rc = ELDLM_LOCK_ABORTED);
+ }
+
+ rc = mdt_do_glimpse(mti->mti_env, ns, res);
+ if (rc == -ENOENT) {
+ /* We are racing with unlink(); just return -ENOENT */
+ rep->lock_policy_res2 = ptlrpc_status_hton(-ENOENT);
+ rc = 0;
+ } else if (rc == -EINVAL) {
+ /* this is possible is client lock has been cancelled but
+ * still exists on server. If that lock was found on server
+ * as only conflicting lock then the client has already
+ * size authority and glimpse is not needed. */
+ CDEBUG(D_DLMTRACE, "Glimpse from the client owning lock\n");
+ rc = 0;
+ } else if (rc < 0) {
+ RETURN(rc);
+ }
+ rc = ELDLM_LOCK_ABORTED;
+fill_mbo:
+ /* LVB can be without valid data in case of DOM */
+ if (!mdt_dom_lvb_is_valid(res))
+ mdt_dom_lvbo_update(res, lock, NULL, false);
+ mdt_lvb2body(res, mbo);
+ RETURN(rc);
+}
+
+int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
+ struct ldlm_lock **lockp, __u64 flags)
+{
+ struct tgt_session_info *tsi = tgt_ses_info(mti->mti_env);
+ struct lu_fid *fid = &tsi->tsi_fid;
+ struct ldlm_lock *lock = *lockp;
+ struct ldlm_resource *res = lock->l_resource;
+ struct ldlm_reply *rep;
+ struct mdt_body *mbo;
+ struct mdt_lock_handle *lhc = &mti->mti_lh[MDT_LH_RMT];
+ struct mdt_object *mo;
+ int rc = 0;
+
+ ENTRY;
+
+ /* Get lock from request for possible resent case. */
+ mdt_intent_fixup_resent(mti, *lockp, lhc, flags);
+ req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
+ req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
+ rc = req_capsule_server_pack(mti->mti_pill);
+ if (rc)
+ RETURN(err_serious(rc));
+
+ rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
+ if (rep == NULL)
+ RETURN(-EPROTO);
+
+ mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
+ if (mbo == NULL)
+ RETURN(-EPROTO);
+
+ fid_extract_from_res_name(fid, &res->lr_name);
+ mo = mdt_object_find(mti->mti_env, mti->mti_mdt, fid);
+ if (unlikely(IS_ERR(mo)))
+ RETURN(PTR_ERR(mo));
+
+ if (!mdt_object_exists(mo))
+ GOTO(out, rc = -ENOENT);
+
+ if (mdt_object_remote(mo))
+ GOTO(out, rc = -EPROTO);
+
+ /* resent case */
+ if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
+ mdt_lock_handle_init(lhc);
+ mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
+ /* This will block MDT thread but it should be fine until
+ * client caches small amount of data for DoM, which should be
+ * smaller than one BRW RPC and should be able to be
+ * piggybacked by lock cancel RPC.
+ * If the client could hold the lock too long, this code can be
+ * revised to call mdt_object_lock_try(). And if fails, it will
+ * return ELDLM_OK here and fall back into normal lock enqueue
+ * process.
+ */
+ rc = mdt_object_lock(mti, mo, lhc, MDS_INODELOCK_DOM);
+ if (rc)
+ GOTO(out, rc);
+ }
+
+ if (!mdt_dom_lvb_is_valid(res)) {
+ rc = mdt_dom_lvb_alloc(res);
+ if (rc)
+ GOTO(out_fail, rc);
+ mdt_dom_disk_lvbo_update(mti->mti_env, mo, res, false);
}
+ mdt_lvb2body(res, mbo);
+out_fail:
+ rep->lock_policy_res2 = clear_serious(rc);
+ if (rep->lock_policy_res2) {
+ lhc->mlh_reg_lh.cookie = 0ull;
+ GOTO(out, rc = ELDLM_LOCK_ABORTED);
+ }
+
+ rc = mdt_intent_lock_replace(mti, lockp, lhc, flags, rc);
+out:
+ mdt_object_put(mti->mti_env, mo);
+ RETURN(rc);
+}
+
+void mdt_dom_discard_data(struct mdt_thread_info *info,
+ const struct lu_fid *fid)
+{
+ struct mdt_device *mdt = info->mti_mdt;
+ union ldlm_policy_data *policy = &info->mti_policy;
+ struct ldlm_res_id *res_id = &info->mti_res_id;
+ struct lustre_handle dom_lh;
+ __u64 flags = LDLM_FL_AST_DISCARD_DATA;
+ int rc = 0;
+
+ policy->l_inodebits.bits = MDS_INODELOCK_DOM;
+ policy->l_inodebits.try_bits = 0;
+ fid_build_reg_res_name(fid, res_id);
+
+ /* Tell the clients that the object is gone now and that they should
+ * throw away any cached pages. */
+ rc = ldlm_cli_enqueue_local(mdt->mdt_namespace, res_id, LDLM_IBITS,
+ policy, LCK_PW, &flags, ldlm_blocking_ast,
+ ldlm_completion_ast, NULL, NULL, 0,
+ LVB_T_NONE, NULL, &dom_lh);
+
+ /* We only care about the side-effects, just drop the lock. */
+ if (rc == ELDLM_OK)
+ ldlm_lock_decref(&dom_lh, LCK_PW);
+}
+
+/* check if client has already DoM lock for given resource */
+bool mdt_dom_client_has_lock(struct mdt_thread_info *info,
+ const struct lu_fid *fid)
+{
+ struct mdt_device *mdt = info->mti_mdt;
+ union ldlm_policy_data *policy = &info->mti_policy;
+ struct ldlm_res_id *res_id = &info->mti_res_id;
+ struct lustre_handle lockh;
+ enum ldlm_mode mode;
+ struct ldlm_lock *lock;
+ bool rc;
+
+ policy->l_inodebits.bits = MDS_INODELOCK_DOM;
+ fid_build_reg_res_name(fid, res_id);
+
+ mode = ldlm_lock_match(mdt->mdt_namespace, LDLM_FL_BLOCK_GRANTED |
+ LDLM_FL_TEST_LOCK, res_id, LDLM_IBITS, policy,
+ LCK_PW, &lockh, 0);
+
+ /* There is no other PW lock on this object; finished. */
+ if (mode == 0)
+ return false;
+
+ lock = ldlm_handle2lock(&lockh);
+ if (lock == 0)
+ return false;
+
+ /* check if lock from the same client */
+ rc = (lock->l_export->exp_handle.h_cookie ==
+ info->mti_exp->exp_handle.h_cookie);
+ LDLM_LOCK_PUT(lock);
return rc;
}
+/**
+ * MDT request handler for OST_GETATTR RPC.
+ *
+ * This is data-specific request to get object and layout versions under
+ * IO lock. It is reliable only for Data-on-MDT files.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+int mdt_data_version_get(struct tgt_session_info *tsi)
+{
+ struct mdt_thread_info *mti = mdt_th_info(tsi->tsi_env);
+ struct mdt_device *mdt = mti->mti_mdt;
+ struct mdt_body *repbody;
+ struct mdt_object *mo = mti->mti_object;
+ struct lov_comp_md_v1 *comp;
+ struct lustre_handle lh = { 0 };
+ __u64 flags = 0;
+ __s64 version;
+ enum ldlm_mode lock_mode = LCK_PR;
+ bool srvlock;
+ int rc;
+
+ ENTRY;
+
+ req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD, RCL_SERVER, 0);
+ req_capsule_set_size(tsi->tsi_pill, &RMF_ACL, RCL_SERVER, 0);
+ rc = req_capsule_server_pack(tsi->tsi_pill);
+ if (unlikely(rc != 0))
+ RETURN(err_serious(rc));
+
+ repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
+ if (repbody == NULL)
+ RETURN(-ENOMEM);
+
+ srvlock = tsi->tsi_mdt_body->mbo_valid & OBD_MD_FLFLAGS &&
+ tsi->tsi_mdt_body->mbo_flags & OBD_FL_SRVLOCK;
+
+ if (srvlock) {
+ if (unlikely(tsi->tsi_mdt_body->mbo_flags & OBD_FL_FLUSH))
+ lock_mode = LCK_PW;
+
+ fid_build_reg_res_name(&tsi->tsi_fid, &tsi->tsi_resid);
+ rc = tgt_mdt_data_lock(mdt->mdt_namespace, &tsi->tsi_resid,
+ &lh, lock_mode, &flags);
+ if (rc != 0)
+ RETURN(rc);
+ }
+
+ if (!mdt_object_exists(mo))
+ GOTO(out, rc = -ENOENT);
+ if (mdt_object_remote(mo))
+ GOTO(out, rc = -EREMOTE);
+ if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
+ GOTO(out, rc = -EBADF);
+
+ /* Get version first */
+ version = dt_version_get(tsi->tsi_env, mdt_obj2dt(mo));
+ if (version && version != -EOPNOTSUPP) {
+ repbody->mbo_valid |= OBD_MD_FLDATAVERSION;
+ /* re-use mbo_ioepoch to transfer version */
+ repbody->mbo_version = version;
+ }
+
+ /* Read layout to get its version */
+ rc = mdt_big_xattr_get(mti, mo, XATTR_NAME_LOV);
+ if (rc == -ENODATA) /* File has no layout yet */
+ GOTO(out, rc = 0);
+ else if (rc < 0)
+ GOTO(out, rc);
+
+ comp = mti->mti_buf.lb_buf;
+ if (le32_to_cpu(comp->lcm_magic) != LOV_MAGIC_COMP_V1) {
+ CDEBUG(D_INFO, DFID" has no composite layout",
+ PFID(&tsi->tsi_fid));
+ GOTO(out, rc = -ESTALE);
+ }
+
+ CDEBUG(D_INODE, DFID": layout version: %u\n",
+ PFID(&tsi->tsi_fid), le32_to_cpu(comp->lcm_layout_gen));
+
+ repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
+ /* re-use mbo_rdev for that */
+ repbody->mbo_layout_gen = le32_to_cpu(comp->lcm_layout_gen);
+ rc = 0;
+out:
+ if (srvlock)
+ tgt_mdt_data_unlock(&lh, lock_mode);
+
+ repbody->mbo_valid |= OBD_MD_FLFLAGS;
+ repbody->mbo_flags = OBD_FL_FLUSH;
+ RETURN(rc);
+}