if (opc == OST_READ)
pa.lpa_mode |= LCK_PR;
- DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks: "DFID"\n",
+ DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks for " DFID,
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",
+ "%s: refreshed %u locks timeout for req %p",
tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
RETURN(1);
}
{
struct dt_object *dob;
int i, j, rc, tot_bytes = 0;
+ int maxlnb = *nr_local;
+ int level;
ENTRY;
mdt_dom_read_lock(mo);
- if (!mdt_object_exists(mo))
- GOTO(unlock, rc = -ENOENT);
+ *nr_local = 0;
+ /* the only valid case when READ can find object is missing or stale
+ * when export is just evicted and open files are closed forcefully
+ * on server while client's READ can be in progress.
+ * This should not happen on healthy export, object can't be missing
+ * or dying because both states means it was finally destroyed.
+ */
+ level = exp->exp_failed ? D_INFO : D_ERROR;
+ if (!mdt_object_exists(mo)) {
+ CDEBUG_LIMIT(level,
+ "%s: READ IO to missing obj "DFID": rc = %d\n",
+ exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
+ -ENOENT);
+ /* return 0 and continue with empty commit to skip such READ
+ * without more BRW errors.
+ */
+ RETURN(0);
+ }
+ if (lu_object_is_dying(&mo->mot_header)) {
+ CDEBUG_LIMIT(level,
+ "%s: READ IO to stale obj "DFID": rc = %d\n",
+ exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
+ -ESTALE);
+ /* return 0 and continue with empty commit to skip such READ
+ * without more BRW errors.
+ */
+ RETURN(0);
+ }
dob = mdt_obj2dt(mo);
/* parse remote buffers to local buffers and prepare the latter */
- *nr_local = 0;
for (i = 0, j = 0; i < niocount; i++) {
- rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 0);
+ rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 0);
if (unlikely(rc < 0))
GOTO(buf_put, rc);
/* correct index for local buffers to continue with */
j += rc;
+ maxlnb -= rc;
*nr_local += rc;
tot_bytes += rnb[i].rnb_len;
}
RETURN(0);
buf_put:
dt_bufs_put(env, dob, lnb, *nr_local);
-unlock:
mdt_dom_read_unlock(mo);
return rc;
}
{
struct dt_object *dob;
int i, j, k, rc = 0, tot_bytes = 0;
+ int maxlnb = *nr_local;
ENTRY;
tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
mdt_dom_read_lock(mo);
+ *nr_local = 0;
+ /* don't report error in cases with failed export */
if (!mdt_object_exists(mo)) {
- CDEBUG(D_ERROR, "%s: BRW to missing obj "DFID"\n",
- exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)));
- GOTO(unlock, rc = -ENOENT);
+ int level = exp->exp_failed ? D_INFO : D_ERROR;
+
+ rc = -ENOENT;
+ CDEBUG_LIMIT(level,
+ "%s: WRITE IO to missing obj "DFID": rc = %d\n",
+ exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
+ rc);
+ /* exit with no data written, note nr_local = 0 above */
+ GOTO(unlock, rc);
+ }
+ if (lu_object_is_dying(&mo->mot_header)) {
+ /* This is possible race between object destroy followed by
+ * discard BL AST and client cache flushing. Object is
+ * referenced until discard finish.
+ */
+ CDEBUG(D_INODE, "WRITE IO to stale object "DFID"\n",
+ PFID(mdt_object_fid(mo)));
+ /* Note: continue with no error here to don't cause BRW errors
+ * but skip transaction in commitrw silently so no data is
+ * written.
+ */
}
dob = mdt_obj2dt(mo);
/* parse remote buffers to local buffers and prepare the latter */
- *nr_local = 0;
for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
- rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 1);
+ rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 1);
if (unlikely(rc < 0))
GOTO(err, rc);
/* correct index for local buffers to continue with */
lnb[j + k].lnb_rc = -ENOSPC;
}
j += rc;
+ maxlnb -= rc;
*nr_local += rc;
tot_bytes += rnb[i].rnb_len;
}
ENTRY;
- LASSERT(niocount > 0);
-
dob = mdt_obj2dt(mo);
- dt_bufs_put(env, dob, lnb, niocount);
+ if (niocount)
+ dt_bufs_put(env, dob, lnb, niocount);
mdt_dom_read_unlock(mo);
RETURN(rc);
retry:
if (!dt_object_exists(dob))
GOTO(out, rc = -ENOENT);
+ if (lu_object_is_dying(&mo->mot_header)) {
+ /* Commit to stale object can be just skipped silently. */
+ CDEBUG(D_INODE, "skip commit to stale object "DFID"\n",
+ PFID(mdt_object_fid(mo)));
+ GOTO(out, rc = 0);
+ }
+
+ if (niocount == 0) {
+ rc = -EPROTO;
+ DEBUG_REQ(D_WARNING, tgt_ses_req(tgt_ses_info(env)),
+ "%s: commit with no pages for "DFID": rc = %d\n",
+ exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)), rc);
+ GOTO(out, rc);
+ }
th = dt_trans_create(env, dt);
if (IS_ERR(th))
__u64 valid;
int rc = 0;
- if (npages == 0) {
- CERROR("%s: no pages to commit\n",
- exp->exp_obd->obd_name);
- rc = -EPROTO;
- }
-
LASSERT(mo);
if (cmd == OBD_BRW_WRITE) {
* doesn't already exist so we can store the reservation handle
* there. */
valid = OBD_MD_FLUID | OBD_MD_FLGID;
- valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
+ if (tgt_fmd_check(exp, mdt_object_fid(mo),
+ mdt_info_req(info)->rq_xid))
+ valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
+ OBD_MD_FLCTIME;
la_from_obdo(la, oa, valid);
oa->o_flags = OBD_FL_NO_GRPQUOTA;
}
+ if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
+ if (oa->o_valid & OBD_MD_FLFLAGS)
+ oa->o_flags |= OBD_FL_NO_PRJQUOTA;
+ else
+ oa->o_flags = OBD_FL_NO_PRJQUOTA;
+ }
+
oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
- OBD_MD_FLGRPQUOTA;
+ OBD_MD_FLGRPQUOTA | OBD_MD_FLPRJQUOTA;
}
} else if (cmd == OBD_BRW_READ) {
/* If oa != NULL then mdt_preprw_read updated the inode
} else {
rc = -EPROTO;
}
- /* this put is pair to object_get in ofd_preprw_write */
mdt_thread_info_fini(info);
RETURN(rc);
}
ENTRY;
/* check that we do support OBD_CONNECT_TRUNCLOCK. */
- CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
+ BUILD_BUG_ON(!(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK));
if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
la->la_size = start;
la->la_valid |= LA_SIZE;
+ /* MDT supports FMD for Data-on-MDT needs */
+ if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
+ tgt_fmd_update(tsi->tsi_exp, &tsi->tsi_fid,
+ tgt_ses_req(tsi)->rq_xid);
+
rc = mdt_object_punch(tsi->tsi_env, mdt->mdt_bottom, dob,
start, end, la);
mdt_dom_write_unlock(mo);
/* 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(env, res, NULL, NULL, false);
+ mdt_dom_lvbo_update(res, NULL, NULL, false);
mdt_lvb2body(res, mb);
ldlm_resource_putref(res);
fill_mbo:
/* LVB can be without valid data in case of DOM */
if (!mdt_dom_lvb_is_valid(res))
- mdt_dom_lvbo_update(mti->mti_env, res, lock, NULL, false);
+ mdt_dom_lvbo_update(res, lock, NULL, false);
mdt_lvb2body(res, mbo);
RETURN(rc);
}
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(info->mti_env, 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_and_cancel(&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 niobuf_remote *rnb = NULL;
struct niobuf_local *lnb;
int rc;
- int max_reply_len;
loff_t offset;
unsigned int len, copied = 0;
int lnbs, nr_local, i;
if (mbo->mbo_dom_size == 0)
RETURN(0);
- /* check the maximum size available in reply */
- max_reply_len =
- req->rq_rqbd->rqbd_svcpt->scp_service->srv_max_reply_size;
-
- CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d/%d\n",
- mbo->mbo_dom_size, max_reply_len, req->rq_reqmsg->lm_repsize,
- req->rq_replen);
+ CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d\n",
+ mbo->mbo_dom_size, req->rq_reqmsg->lm_repsize, req->rq_replen);
len = req->rq_reqmsg->lm_repsize - req->rq_replen;
- max_reply_len -= req->rq_replen;
/* NB: at this moment we have the following sizes:
* - req->rq_replen: used data in reply
* - req->rq_reqmsg->lm_repsize: total allocated reply buffer at client
- * - max_reply_len: maximum reply size allowed by protocol
*
* Ideal case when file size fits in allocated reply buffer,
* that mean we can return whole data in reply. We can also fit more
*
* At the moment the following strategy is used:
* 1) try to fit into the buffer we have
- * 2) respond with bigger buffer so client will re-allocate it and
- * resend (up to srv_max_reply_size value).
- * 3) return just file tail otherwise.
+ * 2) return just file tail otherwise.
*/
if (mbo->mbo_dom_size <= len) {
/* can fit whole data */
len = mbo->mbo_dom_size;
offset = 0;
- } else if (mbo->mbo_dom_size <= max_reply_len) {
- /* It is worth to make this tunable ON/OFF because this will
- * cause buffer re-allocation and resend
- */
- len = mbo->mbo_dom_size;
- offset = 0;
} else {
- int tail = mbo->mbo_dom_size % PAGE_SIZE;
+ int tail, pgbits;
+
+ /* File tail offset must be aligned with larger page size
+ * between client and server, so the maximum page size is
+ * used here to align offset.
+ *
+ * NB: DOM feature was introduced when server supports pagebits
+ * already, so it should be always non-zero value. Report error
+ * if it is not for some reason.
+ */
+ if (!req->rq_export->exp_target_data.ted_pagebits) {
+ CERROR("%s: client page bits are not saved on server\n",
+ mdt_obd_name(mdt));
+ RETURN(0);
+ }
+ pgbits = max_t(int, PAGE_SHIFT,
+ req->rq_export->exp_target_data.ted_pagebits);
+ tail = mbo->mbo_dom_size % (1 << pgbits);
- /* no tail or tail can't fit in reply */
+ /* no partial tail or tail can't fit in reply */
if (tail == 0 || len < tail)
RETURN(0);
GOTO(out, rc = -E2BIG);
}
- /* re-take MDT_BODY buffer after the buffer growing above */
+ /* re-take MDT_BODY and NIOBUF_INLINE buffers after the buffer grow */
mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
fid = &mbo->mbo_fid1;
if (!fid_is_sane(fid))
- RETURN(0);
+ GOTO(out, rc = -EINVAL);
rnb = req_capsule_server_get(tsi->tsi_pill, &RMF_NIOBUF_INLINE);
if (rnb == NULL)
GOTO(out, rc = -EPROTO);
+
buf = (char *)rnb + sizeof(*rnb);
rnb->rnb_len = len;
rnb->rnb_offset = offset;
mo = dt_locate(env, dt, fid);
if (IS_ERR(mo))
- GOTO(out, rc = PTR_ERR(mo));
+ GOTO(out_rnb, rc = PTR_ERR(mo));
LASSERT(mo != NULL);
dt_read_lock(env, mo, 0);
if (lnb == NULL)
GOTO(unlock, rc = -ENOMEM);
- rc = dt_bufs_get(env, mo, rnb, lnb, 0);
+ rc = dt_bufs_get(env, mo, rnb, lnb, lnbs, 0);
if (unlikely(rc < 0))
GOTO(free, rc);
LASSERT(rc <= lnbs);
}
CDEBUG(D_INFO, "Read %i (wanted %u) bytes from %llu\n", copied,
len, offset);
- if (copied < len)
+ if (copied < len) {
CWARN("%s: read %i bytes for "DFID
" but wanted %u, is size wrong?\n",
tsi->tsi_exp->exp_obd->obd_name, copied,
PFID(&tsi->tsi_fid), len);
+ /* Ignore partially copied data */
+ copied = 0;
+ }
EXIT;
buf_put:
dt_bufs_put(env, mo, lnb, nr_local);
unlock:
dt_read_unlock(env, mo);
lu_object_put(env, &mo->do_lu);
+out_rnb:
+ rnb->rnb_len = copied;
out:
- if (rnb != NULL)
- rnb->rnb_len = copied;
+ /* Don't fail OPEN request if read-on-open is failed, but drop
+ * a message in log about the error.
+ */
+ if (rc)
+ CDEBUG(D_INFO, "Read-on-open is failed, rc = %d", rc);
+
+ RETURN(0);
+}
+
+/**
+ * Completion AST for DOM discard locks:
+ *
+ * CP AST an DOM discard lock is called always right after enqueue or from
+ * reprocess if lock was blocked, in the latest case l_ast_data is set to
+ * the mdt_object which is kept while there are pending locks on it.
+ */
+int ldlm_dom_discard_cp_ast(struct ldlm_lock *lock, __u64 flags, void *data)
+{
+ struct mdt_object *mo;
+ struct lustre_handle dom_lh;
+ struct lu_env *env;
+
+ ENTRY;
+
+ /* l_ast_data is set when lock was not granted immediately
+ * in mdt_dom_discard_data() below but put into waiting list,
+ * so this CP callback means we are finished and corresponding
+ * MDT object should be released finally as well as lock itself.
+ */
+ lock_res_and_lock(lock);
+ if (!lock->l_ast_data) {
+ unlock_res_and_lock(lock);
+ RETURN(0);
+ }
+
+ mo = lock->l_ast_data;
+ lock->l_ast_data = NULL;
+ unlock_res_and_lock(lock);
+
+ ldlm_lock2handle(lock, &dom_lh);
+ ldlm_lock_decref(&dom_lh, LCK_PW);
+
+ env = lu_env_find();
+ LASSERT(env);
+ mdt_object_put(env, mo);
+
RETURN(0);
}
+void mdt_dom_discard_data(struct mdt_thread_info *info,
+ struct mdt_object *mo)
+{
+ struct ptlrpc_request *req = mdt_info_req(info);
+ struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
+ union ldlm_policy_data policy;
+ struct ldlm_res_id res_id;
+ struct lustre_handle dom_lh;
+ struct ldlm_lock *lock;
+ __u64 flags = LDLM_FL_AST_DISCARD_DATA;
+ int rc = 0;
+ bool old_client;
+
+ ENTRY;
+
+ if (req && req_is_replay(req))
+ RETURN_EXIT;
+
+ policy.l_inodebits.bits = MDS_INODELOCK_DOM;
+ policy.l_inodebits.try_bits = 0;
+ fid_build_reg_res_name(mdt_object_fid(mo), &res_id);
+
+ /* Keep blocking version of discard for an old client to avoid
+ * crashes on non-patched clients. LU-11359.
+ */
+ old_client = req && !(exp_connect_flags2(req->rq_export) &
+ OBD_CONNECT2_ASYNC_DISCARD);
+
+ /* Tell the clients that the object is gone now and that they should
+ * throw away any cached pages. */
+ rc = ldlm_cli_enqueue_local(info->mti_env, mdt->mdt_namespace, &res_id,
+ LDLM_IBITS, &policy, LCK_PW, &flags,
+ ldlm_blocking_ast, old_client ?
+ ldlm_completion_ast :
+ ldlm_dom_discard_cp_ast,
+ NULL, NULL, 0, LVB_T_NONE, NULL, &dom_lh);
+ if (rc != ELDLM_OK) {
+ CDEBUG(D_DLMTRACE,
+ "Failed to issue discard lock, rc = %d\n", rc);
+ RETURN_EXIT;
+ }
+
+ lock = ldlm_handle2lock(&dom_lh);
+ lock_res_and_lock(lock);
+ /* if lock is not granted then there are BL ASTs in progress and
+ * lock will be granted in result of reprocessing with CP callback
+ * notifying about that. The mdt object has to be kept until that and
+ * it is saved in l_ast_data of the lock. Lock reference is kept too
+ * until that to prevent it from canceling.
+ */
+ if (!is_granted_or_cancelled_nolock(lock)) {
+ mdt_object_get(info->mti_env, mo);
+ lock->l_ast_data = mo;
+ unlock_res_and_lock(lock);
+ } else {
+ unlock_res_and_lock(lock);
+ ldlm_lock_decref_and_cancel(&dom_lh, LCK_PW);
+ }
+ LDLM_LOCK_PUT(lock);
+
+ RETURN_EXIT;
+}