up_write(&mo->mot_dom_sem);
}
-/**
- * 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);
-}
-
static void mdt_dom_resource_prolong(struct ldlm_prolong_args *arg)
{
struct ldlm_resource *res;
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\n",
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))
GOTO(out_stop, rc);
dt_write_lock(env, dob, 0);
- rc = dt_write_commit(env, dob, lnb, niocount, th);
+ rc = dt_write_commit(env, dob, lnb, niocount, th, 0);
if (rc)
GOTO(unlock, rc);
__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) {
} 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))
enum ldlm_mode mode;
struct ldlm_lock *lock;
struct ldlm_glimpse_work *gl_work;
- struct list_head gl_list;
+ LIST_HEAD(gl_list);
int rc;
ENTRY;
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 */
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)
}
mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
+ if (!(mbo->mbo_valid & OBD_MD_DOM_SIZE))
+ RETURN(0);
+
+ if (!mbo->mbo_dom_size)
+ RETURN(0);
if (lustre_handle_is_used(lh)) {
struct ldlm_lock *lock;
if (!dom_lock || !mdt->mdt_opts.mo_dom_read_open)
RETURN(0);
- if (!(mbo->mbo_valid & OBD_MD_DOM_SIZE))
- RETURN(0);
-
- if (mbo->mbo_dom_size == 0)
- RETURN(0);
-
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;
/* can fit whole data */
len = mbo->mbo_dom_size;
offset = 0;
- } else {
+ } else if (mbo->mbo_dom_size <
+ mdt_lmm_dom_stripesize(mti->mti_attr.ma_lmm)) {
int tail, pgbits;
/* File tail offset must be aligned with larger page size
len = tail;
offset = mbo->mbo_dom_size - len;
+ } else {
+ /* DOM stripe is fully written, so don't expect its tail
+ * will be used by append.
+ */
+ RETURN(0);
}
- LASSERT((offset % PAGE_SIZE) == 0);
+
+ LASSERT((offset & ~PAGE_MASK) == 0);
rc = req_capsule_server_grow(pill, &RMF_NIOBUF_INLINE,
sizeof(*rnb) + len);
if (rc != 0) {
/* parse remote buffers to local buffers and prepare the latter */
lnbs = (len >> PAGE_SHIFT) + 1;
- OBD_ALLOC(lnb, sizeof(*lnb) * lnbs);
+ OBD_ALLOC_PTR_ARRAY(lnb, lnbs);
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);
buf_put:
dt_bufs_put(env, mo, lnb, nr_local);
free:
- OBD_FREE(lnb, sizeof(*lnb) * lnbs);
+ OBD_FREE_PTR_ARRAY(lnb, lnbs);
unlock:
dt_read_unlock(env, mo);
lu_object_put(env, &mo->do_lu);
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;
+}