X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_io.c;h=bf80be4ef6dc27331ae653eb0c37d9445404ecc2;hb=b0db02d74458959845566e354e4e75359bf5ef00;hp=6924296cd680a612c4564320a0498b7f959360b5;hpb=238a309ea6081163c1928630414ee9767746beaa;p=fs%2Flustre-release.git diff --git a/lustre/mdt/mdt_io.c b/lustre/mdt/mdt_io.c index 6924296..bf80be4 100644 --- a/lustre/mdt/mdt_io.c +++ b/lustre/mdt/mdt_io.c @@ -20,7 +20,7 @@ * GPL HEADER END */ /* - * Copyright (c) 2012, 2017 Intel Corporation. + * Copyright (c) 2017, Intel Corporation. */ /* * lustre/mdt/mdt_io.c @@ -31,16 +31,13 @@ #define DEBUG_SUBSYSTEM S_FILTER #include +#include +#include + #include "mdt_internal.h" /* 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; -} - static inline void mdt_dom_read_lock(struct mdt_object *mo) { down_read(&mo->mot_dom_sem); @@ -61,30 +58,360 @@ static inline void mdt_dom_write_unlock(struct mdt_object *mo) up_write(&mo->mot_dom_sem); } +static void mdt_dom_resource_prolong(struct ldlm_prolong_args *arg) +{ + struct ldlm_resource *res; + struct ldlm_lock *lock; + + ENTRY; + + res = ldlm_resource_get(arg->lpa_export->exp_obd->obd_namespace, NULL, + &arg->lpa_resid, LDLM_IBITS, 0); + if (IS_ERR(res)) { + CDEBUG(D_DLMTRACE, + "Failed to get resource for resid %llu/%llu\n", + arg->lpa_resid.name[0], arg->lpa_resid.name[1]); + RETURN_EXIT; + } + + lock_res(res); + list_for_each_entry(lock, &res->lr_granted, l_res_link) { + if (ldlm_has_dom(lock)) { + LDLM_DEBUG(lock, "DOM lock to prolong "); + ldlm_lock_prolong_one(lock, arg); + /* only one PW or EX lock can be granted, + * no need to continue search + */ + if (lock->l_granted_mode & (LCK_PW | LCK_EX)) + break; + } + } + unlock_res(res); + ldlm_resource_putref(res); + + EXIT; +} + +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; + + 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); + if (data->lpa_locks_cnt > 0) + RETURN_EXIT; + } + } + mdt_dom_resource_prolong(data); + EXIT; +} + +static int mdt_rw_hpreq_lock_match(struct ptlrpc_request *req, + struct ldlm_lock *lock) +{ + struct obd_ioobj *ioo; + enum ldlm_mode mode; + __u32 opc = lustre_msg_get_opc(req->rq_reqmsg); + + ENTRY; + + 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); + + 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 | LCK_GROUP; + 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); +} + +static int mdt_rw_hpreq_check(struct ptlrpc_request *req) +{ + 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 | LCK_GROUP; + if (opc == OST_READ) + pa.lpa_mode |= LCK_PR; + + 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", + tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req); + RETURN(1); + } + + RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE); +} + +static void mdt_rw_hpreq_fini(struct ptlrpc_request *req) +{ + mdt_rw_hpreq_check(req); +} + +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) +{ + 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 int mdt_punch_hpreq_lock_match(struct ptlrpc_request *req, + struct ldlm_lock *lock) +{ + 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 | LCK_GROUP))) + RETURN(0); + + RETURN(1); +} + +/** + * 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) +{ + 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 | LCK_GROUP; + + 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); +} + +/** + * 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) +{ + mdt_punch_hpreq_check(req); +} + +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) +{ + 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, struct mdt_device *mdt, struct mdt_object *mo, struct lu_attr *la, int niocount, struct niobuf_remote *rnb, int *nr_local, struct niobuf_local *lnb, char *jobid) { + struct tgt_session_info *tsi = tgt_ses_info(env); + struct ptlrpc_request *req = tgt_ses_req(tsi); 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; } @@ -97,11 +424,10 @@ static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp, if (unlikely(rc)) GOTO(buf_put, rc); - mdt_io_counter_incr(exp, LPROC_MDT_IO_READ, jobid, tot_bytes); + mdt_counter_incr(req, LPROC_MDT_IO_READ, tot_bytes); RETURN(0); buf_put: dt_bufs_put(env, dob, lnb, *nr_local); -unlock: mdt_dom_read_unlock(mo); return rc; } @@ -113,8 +439,11 @@ static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp, struct niobuf_remote *rnb, int *nr_local, struct niobuf_local *lnb, char *jobid) { + struct tgt_session_info *tsi = tgt_ses_info(env); + struct ptlrpc_request *req = tgt_ses_req(tsi); struct dt_object *dob; int i, j, k, rc = 0, tot_bytes = 0; + int maxlnb = *nr_local; ENTRY; @@ -123,17 +452,36 @@ static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp, 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 */ @@ -143,6 +491,7 @@ static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp, lnb[j + k].lnb_rc = -ENOSPC; } j += rc; + maxlnb -= rc; *nr_local += rc; tot_bytes += rnb[i].rnb_len; } @@ -151,7 +500,7 @@ static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp, if (likely(rc)) GOTO(err, rc); - mdt_io_counter_incr(exp, LPROC_MDT_IO_WRITE, jobid, tot_bytes); + mdt_counter_incr(req, LPROC_MDT_IO_WRITE, tot_bytes); RETURN(0); err: dt_bufs_put(env, dob, lnb, *nr_local); @@ -232,11 +581,10 @@ static int mdt_commitrw_read(const struct lu_env *env, struct mdt_device *mdt, 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); @@ -244,16 +592,16 @@ static int mdt_commitrw_read(const struct lu_env *env, struct mdt_device *mdt, static int mdt_commitrw_write(const struct lu_env *env, struct obd_export *exp, struct mdt_device *mdt, struct mdt_object *mo, - struct lu_attr *la, int objcount, int niocount, - struct niobuf_local *lnb, unsigned long granted, - int old_rc) + struct lu_attr *la, struct obdo *oa, int objcount, + int niocount, struct niobuf_local *lnb, + unsigned long granted, int old_rc) { struct dt_device *dt = mdt->mdt_bottom; struct dt_object *dob; struct thandle *th; int rc = 0; int retries = 0; - int i; + int i, restart = 0; ENTRY; @@ -267,6 +615,16 @@ retry: if (!dt_object_exists(dob)) GOTO(out, rc = -ENOENT); + 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); + } + + CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_COMMITRW_DELAY, cfs_fail_val); + th = dt_trans_create(env, dt); if (IS_ERR(th)) GOTO(out, rc = PTR_ERR(th)); @@ -292,14 +650,23 @@ retry: GOTO(out_stop, rc); } + tgt_vbr_obj_set(env, dob); rc = dt_trans_start(env, dt, th); if (rc) GOTO(out_stop, rc); dt_write_lock(env, dob, 0); - rc = dt_write_commit(env, dob, lnb, niocount, th); - if (rc) + 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(unlock, rc = 0); + } + rc = dt_write_commit(env, dob, lnb, niocount, th, oa->o_size); + if (rc) { + restart = th->th_restart_tran; GOTO(unlock, rc); + } if (la->la_valid) { rc = dt_attr_set(env, dob, la, th); @@ -323,13 +690,23 @@ out_stop: granted = 0; } - th->th_result = rc; + th->th_result = restart ? 0 : rc; dt_trans_stop(env, dt, th); if (rc == -ENOSPC && retries++ < 3) { CDEBUG(D_INODE, "retry after force commit, retries:%d\n", retries); goto retry; } + if (restart) { + retries++; + restart = 0; + if (retries % 10000 == 0) + CERROR("%s: restart IO write too many times: %d\n", + exp->exp_obd->obd_name, retries); + CDEBUG(D_INODE, "retry transaction, retries:%d\n", + retries); + goto retry; + } out: dt_bufs_put(env, dob, lnb, niocount); @@ -340,7 +717,7 @@ out: } void mdt_dom_obj_lvb_update(const struct lu_env *env, struct mdt_object *mo, - bool increase_only) + struct obdo *oa, bool increase_only) { struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev); struct ldlm_res_id resid; @@ -353,8 +730,23 @@ void mdt_dom_obj_lvb_update(const struct lu_env *env, struct mdt_object *mo, return; /* Update lvbo data if exists. */ - if (mdt_dom_lvb_is_valid(res)) + if (mdt_dom_lvb_is_valid(res)) { mdt_dom_disk_lvbo_update(env, mo, res, increase_only); + if (oa) { + struct ost_lvb *res_lvb = res->lr_lvb_data; + + lock_res(res); + oa->o_valid |= OBD_MD_FLBLOCKS | OBD_MD_FLSIZE | + OBD_MD_FLMTIME | OBD_MD_FLATIME | + OBD_MD_FLCTIME; + oa->o_blocks = res_lvb->lvb_blocks; + oa->o_size = res_lvb->lvb_size; + oa->o_atime = res_lvb->lvb_atime; + oa->o_mtime = res_lvb->lvb_mtime; + oa->o_ctime = res_lvb->lvb_ctime; + unlock_res(res); + } + } ldlm_resource_putref(res); } @@ -369,16 +761,38 @@ int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, struct lu_attr *la = &info->mti_attr.ma_attr; __u64 valid; int rc = 0; - - if (npages == 0) { - CERROR("%s: no pages to commit\n", - exp->exp_obd->obd_name); - rc = -EPROTO; - } + int root_squash = 0; LASSERT(mo); if (cmd == OBD_BRW_WRITE) { + struct lu_nodemap *nodemap; + __u32 mapped_uid, mapped_gid, mapped_projid; + + nodemap = nodemap_get_from_exp(exp); + if (IS_ERR(nodemap)) + RETURN(PTR_ERR(nodemap)); + mapped_uid = nodemap_map_id(nodemap, NODEMAP_UID, + NODEMAP_FS_TO_CLIENT, + oa->o_uid); + mapped_gid = nodemap_map_id(nodemap, NODEMAP_GID, + NODEMAP_FS_TO_CLIENT, + oa->o_gid); + mapped_projid = nodemap_map_id(nodemap, NODEMAP_PROJID, + NODEMAP_FS_TO_CLIENT, + oa->o_projid); + if (!IS_ERR_OR_NULL(nodemap)) { + /* do not bypass quota enforcement if squashed uid */ + if (unlikely(mapped_uid == nodemap->nm_squash_uid)) { + int idx; + + for (idx = 0; idx < npages; idx++) + lnb[idx].lnb_flags &= + ~OBD_BRW_SYS_RESOURCE; + root_squash = 1; + } + nodemap_putref(nodemap); + } /* Don't update timestamps if this write is older than a * setattr which modifies the timestamps. b=10150 */ @@ -387,18 +801,21 @@ int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, * 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); - rc = mdt_commitrw_write(env, exp, mdt, mo, la, objcount, + rc = mdt_commitrw_write(env, exp, mdt, mo, la, oa, objcount, npages, lnb, oa->o_grant_used, old_rc); if (rc == 0) obdo_from_la(oa, la, VALID_FLAGS | LA_GID | LA_UID); else obdo_from_la(oa, la, LA_GID | LA_UID); - mdt_dom_obj_lvb_update(env, mo, false); + mdt_dom_obj_lvb_update(env, mo, NULL, false); /* don't report overquota flag if we failed before reaching * commit */ if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) { @@ -417,26 +834,211 @@ int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, 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; + } + + if (root_squash) + oa->o_flags |= OBD_FL_ROOT_SQUASH; + oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA | - OBD_MD_FLGRPQUOTA; + OBD_MD_FLGRPQUOTA | OBD_MD_FLPRJQUOTA; } + /* Convert back to client IDs. LU-9671. + * nodemap_get_from_exp() may fail due to nodemap deactivated, + * server ID will be returned back to client in that case. + */ + oa->o_uid = mapped_uid; + oa->o_gid = mapped_gid; + oa->o_projid = mapped_projid; } 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); + mdt_dom_obj_lvb_update(env, mo, NULL, true); rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb); if (old_rc) rc = old_rc; } else { rc = -EPROTO; } - /* this put is pair to object_get in ofd_preprw_write */ mdt_thread_info_fini(info); RETURN(rc); } +int mdt_object_fallocate(const struct lu_env *env, struct dt_device *dt, + struct dt_object *dob, __u64 start, __u64 end, + int mode, struct lu_attr *la) +{ + struct thandle *th; + int rc; + + ENTRY; + + if (!dt_object_exists(dob)) + RETURN(-ENOENT); + + th = dt_trans_create(env, dt); + if (IS_ERR(th)) + RETURN(PTR_ERR(th)); + + rc = dt_declare_attr_set(env, dob, la, th); + if (rc) + GOTO(stop, rc); + + rc = dt_declare_fallocate(env, dob, start, end, mode, th); + if (rc) + GOTO(stop, rc); + + tgt_vbr_obj_set(env, dob); + rc = dt_trans_start(env, dt, th); + if (rc) + GOTO(stop, rc); + + dt_write_lock(env, dob, 0); + rc = dt_falloc(env, dob, start, end, mode, th); + if (rc) + GOTO(unlock, rc); + rc = dt_attr_set(env, dob, la, th); + if (rc) + GOTO(unlock, rc); +unlock: + dt_write_unlock(env, dob); +stop: + th->th_result = rc; + dt_trans_stop(env, dt, th); + RETURN(rc); +} + +/** + * MDT request handler for OST_FALLOCATE RPC. + * + * This is part of request processing. Validate request fields, + * preallocate the given MDT object and pack reply. + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative value on error + */ +int mdt_fallocate_hdl(struct tgt_session_info *tsi) +{ + struct obdo *oa = &tsi->tsi_ost_body->oa; + struct ptlrpc_request *req = tgt_ses_req(tsi); + struct ost_body *repbody; + struct mdt_thread_info *info; + struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace; + struct obd_export *exp = tsi->tsi_exp; + struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev); + struct mdt_object *mo; + struct dt_object *dob; + struct lu_attr *la; + __u64 flags = 0; + struct lustre_handle lh = { 0, }; + int rc, mode; + __u64 start, end; + bool srvlock; + ktime_t kstart = ktime_get(); + + repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY); + if (repbody == NULL) + RETURN(err_serious(-ENOMEM)); + + /* + * fallocate start and end are passed in o_size, o_blocks + * on the wire. + */ + if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) != + (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) + RETURN(err_serious(-EPROTO)); + + start = oa->o_size; + end = oa->o_blocks; + mode = oa->o_falloc_mode; + + CDEBUG(D_INODE, + "fallocate: "DFID", mode = %#x, start = %lld, end = %lld\n", + PFID(&tsi->tsi_fid), mode, start, end); + + /* + * mode == 0 (which is standard prealloc) and PUNCH is supported + * Rest of mode options are not supported yet. + */ + if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE)) + RETURN(-EOPNOTSUPP); + + if (mode & FALLOC_FL_PUNCH_HOLE && !(mode & FALLOC_FL_KEEP_SIZE)) { + CWARN("%s: PUNCH mode misses KEEP_SIZE flag, setting it\n", + tsi->tsi_tgt->lut_obd->obd_name); + mode |= FALLOC_FL_KEEP_SIZE; + } + + info = tsi2mdt_info(tsi); + la = &info->mti_attr.ma_attr; + + repbody->oa.o_oi = oa->o_oi; + repbody->oa.o_valid = OBD_MD_FLID; + + srvlock = oa->o_valid & OBD_MD_FLFLAGS && + oa->o_flags & OBD_FL_SRVLOCK; + + if (srvlock) { + rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW, + &flags); + if (rc != 0) + GOTO(out, rc); + } + + mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid); + if (IS_ERR(mo)) + GOTO(out_unlock, rc = PTR_ERR(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: fallocate on dir "DFID": rc = %d\n", + exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc); + GOTO(out_put, rc); + } + + la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME); + + mdt_dom_write_lock(mo); + dob = mdt_obj2dt(mo); + + 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_fallocate(tsi->tsi_env, mdt->mdt_bottom, dob, start, + end, mode, la); + mdt_dom_write_unlock(mo); + if (rc) + GOTO(out_put, rc); + + mdt_dom_obj_lvb_update(tsi->tsi_env, mo, &repbody->oa, false); + + mdt_counter_incr(req, LPROC_MDT_FALLOCATE, + ktime_us_delta(ktime_get(), kstart)); + + EXIT; +out_put: + lu_object_put(tsi->tsi_env, &mo->mot_obj); +out_unlock: + if (srvlock) + tgt_data_unlock(&lh, LCK_PW); +out: + mdt_thread_info_fini(info); + return rc; +} + int mdt_object_punch(const struct lu_env *env, struct dt_device *dt, struct dt_object *dob, __u64 start, __u64 end, struct lu_attr *la) @@ -487,6 +1089,7 @@ stop: int mdt_punch_hdl(struct tgt_session_info *tsi) { const struct obdo *oa = &tsi->tsi_ost_body->oa; + struct ptlrpc_request *req = tgt_ses_req(tsi); struct ost_body *repbody; struct mdt_thread_info *info; struct lu_attr *la; @@ -497,15 +1100,13 @@ int mdt_punch_hdl(struct tgt_session_info *tsi) struct dt_object *dob; __u64 flags = 0; struct lustre_handle lh = { 0, }; + ktime_t kstart = ktime_get(); __u64 start, end; int rc; bool srvlock; ENTRY; - /* check that we do support OBD_CONNECT_TRUNCLOCK. */ - CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK); - if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) != (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) RETURN(err_serious(-EPROTO)); @@ -550,30 +1151,45 @@ int mdt_punch_hdl(struct tgt_session_info *tsi) 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); 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); 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); + mdt_dom_obj_lvb_update(tsi->tsi_env, mo, &repbody->oa, false); + + mdt_counter_incr(req, LPROC_MDT_IO_PUNCH, + ktime_us_delta(ktime_get(), kstart)); EXIT; out_put: lu_object_put(tsi->tsi_env, &mo->mot_obj); out_unlock: if (srvlock) - mdt_save_lock(info, &lh, LCK_PW, rc); + tgt_data_unlock(&lh, LCK_PW); out: mdt_thread_info_fini(info); return rc; @@ -594,16 +1210,16 @@ int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns, 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; /* 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_BLOCK_GRANTED | LDLM_FL_TEST_LOCK, + mode = ldlm_lock_match(ns, LDLM_FL_TEST_LOCK, &res->lr_name, LDLM_IBITS, &policy, - LCK_PW, &lockh, 0); + LCK_PW, &lockh); /* There is no PW lock on this object; finished. */ if (mode == 0) @@ -635,7 +1251,6 @@ int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns, 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 */ @@ -658,22 +1273,28 @@ out: return rc; } -static void mdt_lvb2body(struct ldlm_resource *res, struct mdt_body *mb) +static void mdt_lvb2reply(struct ldlm_resource *res, struct mdt_body *mb, + struct ost_lvb *lvb) { 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; + if (res_lvb) { + if (lvb) + *lvb = *res_lvb; + + if (mb) { + 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; + mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME | + OBD_MD_FLMTIME | OBD_MD_DOM_SIZE; + } + CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size); + } unlock_res(res); } @@ -704,7 +1325,7 @@ int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt, if (dom_lock || !mdt_dom_lvb_is_valid(res)) mdt_dom_lvbo_update(res, NULL, NULL, false); - mdt_lvb2body(res, mb); + mdt_lvb2reply(res, mb, NULL); ldlm_resource_putref(res); RETURN(rc); } @@ -735,6 +1356,8 @@ int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns, ldlm_processing_policy policy; struct ldlm_reply *rep; struct mdt_body *mbo; + struct ost_lvb *lvb; + bool old_client = !exp_connect_dom_lvb(mti->mti_exp); int rc; ENTRY; @@ -742,19 +1365,28 @@ int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns, 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); + if (unlikely(old_client)) { + 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); + } else { + req_capsule_set_size(mti->mti_pill, &RMF_DLM_LVB, RCL_SERVER, + sizeof(*lvb)); + } 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); + if (unlikely(old_client)) { + mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY); + LASSERT(mbo); + lvb = NULL; + } else { + lvb = req_capsule_server_get(mti->mti_pill, &RMF_DLM_LVB); + LASSERT(lvb); + mbo = NULL; + } lock_res(res); /* Check if this is a resend case (MSG_RESENT is set on RPC) and a @@ -763,7 +1395,7 @@ int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns, if (flags & LDLM_FL_RESENT) { rc = LDLM_ITER_CONTINUE; } else { - __u64 tmpflags = 0; + __u64 tmpflags = LDLM_FL_BLOCK_NOWAIT; enum ldlm_error err; rc = policy(lock, &tmpflags, LDLM_PROCESS_RESCAN, &err, NULL); @@ -785,14 +1417,12 @@ int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns, 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); } @@ -801,7 +1431,8 @@ 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); + mdt_lvb2reply(res, mbo, lvb); + RETURN(rc); } @@ -820,8 +1451,6 @@ int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns, 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); @@ -847,10 +1476,13 @@ int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns, if (mdt_object_remote(mo)) GOTO(out, rc = -EPROTO); + /* Get lock from request for possible resent case. */ + mdt_intent_fixup_resent(mti, *lockp, lhc, flags); /* 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); + mdt_lh_reg_init(lhc, *lockp); + /* 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 @@ -871,7 +1503,7 @@ int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns, GOTO(out_fail, rc); mdt_dom_disk_lvbo_update(mti->mti_env, mo, res, false); } - mdt_lvb2body(res, mbo); + mdt_lvb2reply(res, mbo, NULL); out_fail: rep->lock_policy_res2 = clear_serious(rc); if (rep->lock_policy_res2) { @@ -881,36 +1513,12 @@ out_fail: rc = mdt_intent_lock_replace(mti, lockp, lhc, flags, rc); out: + if (rc < 0) + lhc->mlh_reg_lh.cookie = 0ull; 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) @@ -918,17 +1526,21 @@ bool mdt_dom_client_has_lock(struct mdt_thread_info *info, 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; + __u64 open_flags = info->mti_spec.sp_cr_flags; struct lustre_handle lockh; enum ldlm_mode mode; struct ldlm_lock *lock; + enum ldlm_mode lm; bool rc; policy->l_inodebits.bits = MDS_INODELOCK_DOM; fid_build_reg_res_name(fid, res_id); + + lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR | LCK_PW; mode = ldlm_lock_match(mdt->mdt_namespace, LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK, res_id, LDLM_IBITS, policy, - LCK_PW, &lockh, 0); + lm, &lockh); /* There is no other PW lock on this object; finished. */ if (mode == 0) @@ -945,3 +1557,411 @@ bool mdt_dom_client_has_lock(struct mdt_thread_info *info, 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_data_unlock(&lh, lock_mode); + + repbody->mbo_valid |= OBD_MD_FLFLAGS; + repbody->mbo_flags = OBD_FL_FLUSH; + RETURN(rc); +} + +/* read file data to the buffer */ +int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt, + struct lustre_handle *lh) +{ + const struct lu_env *env = mti->mti_env; + struct tgt_session_info *tsi = tgt_ses_info(env); + struct req_capsule *pill = tsi->tsi_pill; + const struct lu_fid *fid; + struct ptlrpc_request *req = tgt_ses_req(tsi); + struct mdt_body *mbo; + struct dt_device *dt = mdt->mdt_bottom; + struct dt_object *mo; + void *buf; + struct niobuf_remote *rnb = NULL; + struct niobuf_local *lnb; + int rc; + loff_t offset; + unsigned int len, copied = 0; + __u64 real_dom_size; + int lnbs, nr_local, i; + bool dom_lock = false; + + ENTRY; + + if (!req_capsule_field_present(pill, &RMF_NIOBUF_INLINE, RCL_SERVER)) { + /* There is no reply buffers for this field, this means that + * client has no support for data in reply. + */ + RETURN(0); + } + + 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; + + lock = ldlm_handle2lock(lh); + if (lock) { + dom_lock = ldlm_has_dom(lock) && ldlm_has_layout(lock); + LDLM_LOCK_PUT(lock); + } + } + + /* return data along with open only along with DoM lock */ + if (!dom_lock || !mdt->mdt_opts.mo_dom_read_open) + RETURN(0); + + /* if DoM object holds encrypted content, we need to make sure we + * send whole encryption units, or client will read corrupted content + */ + if (mbo->mbo_valid & LA_FLAGS && mbo->mbo_flags & LUSTRE_ENCRYPT_FL && + mbo->mbo_dom_size & ~LUSTRE_ENCRYPTION_MASK) + real_dom_size = (mbo->mbo_dom_size & LUSTRE_ENCRYPTION_MASK) + + LUSTRE_ENCRYPTION_UNIT_SIZE; + else + real_dom_size = mbo->mbo_dom_size; + + CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d\n", + real_dom_size, req->rq_reqmsg->lm_repsize, req->rq_replen); + len = req->rq_reqmsg->lm_repsize - 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 + * + * Ideal case when file size fits in allocated reply buffer, + * that mean we can return whole data in reply. We can also fit more + * data up to max_reply_size in total reply size, but this will cause + * re-allocation on client and resend with larger buffer. This is still + * faster than separate READ IO. + * Third case if file is too big to fit even in maximum size, in that + * case we return just tail to optimize possible append. + * + * At the moment the following strategy is used: + * 1) try to fit into the buffer we have + * 2) return just file tail otherwise. + */ + if (real_dom_size <= len) { + /* can fit whole data */ + len = real_dom_size; + offset = 0; + } else if (real_dom_size < + mdt_lmm_dom_stripesize(mti->mti_attr.ma_lmm)) { + 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 = real_dom_size % (1 << pgbits); + + /* no partial tail or tail can't fit in reply */ + if (tail == 0 || len < tail) + RETURN(0); + + len = tail; + offset = real_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_MASK) == 0); + rc = req_capsule_server_grow(pill, &RMF_NIOBUF_INLINE, + sizeof(*rnb) + len); + if (rc != 0) { + /* failed to grow data buffer, just exit */ + GOTO(out, rc = -E2BIG); + } + + /* 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)) + 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_rnb, rc = PTR_ERR(mo)); + LASSERT(mo != NULL); + + dt_read_lock(env, mo, 0); + if (!dt_object_exists(mo)) + GOTO(unlock, rc = -ENOENT); + + /* parse remote buffers to local buffers and prepare the latter */ + lnbs = (len >> PAGE_SHIFT) + 1; + OBD_ALLOC_PTR_ARRAY(lnb, lnbs); + if (lnb == NULL) + GOTO(unlock, rc = -ENOMEM); + + rc = dt_bufs_get(env, mo, rnb, lnb, lnbs, 0); + if (unlikely(rc < 0)) + GOTO(free, rc); + LASSERT(rc <= lnbs); + nr_local = rc; + rc = dt_read_prep(env, mo, lnb, nr_local); + if (unlikely(rc)) + GOTO(buf_put, rc); + /* copy data to the buffer finally */ + for (i = 0; i < nr_local; i++) { + char *p = kmap(lnb[i].lnb_page); + long off; + + LASSERT(lnb[i].lnb_page_offset == 0); + off = lnb[i].lnb_len & ~PAGE_MASK; + if (off > 0) + memset(p + off, 0, PAGE_SIZE - off); + + memcpy(buf + (i << PAGE_SHIFT), p, lnb[i].lnb_len); + kunmap(lnb[i].lnb_page); + copied += lnb[i].lnb_len; + LASSERT(rc <= len); + } + CDEBUG(D_INFO, "Read %i (wanted %u) bytes from %llu\n", copied, + len, offset); + 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); +free: + OBD_FREE_PTR_ARRAY(lnb, lnbs); +unlock: + dt_read_unlock(env, mo); + lu_object_put(env, &mo->do_lu); +out_rnb: + rnb->rnb_len = copied; +out: + /* 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; +}