X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_handler.c;h=c2533e404e0dab05a39810486a774ea3e61aa29c;hp=633529b4d633509db73e61cc3b9bb366a997c430;hb=96d7758c5f4f5ec0cdd70b62844e0bcdab434c0d;hpb=4d1bb7a76ebef4e69f9aaeb28175e62f73afa002 diff --git a/lustre/mdt/mdt_handler.c b/lustre/mdt/mdt_handler.c index 633529b..c2533e4 100644 --- a/lustre/mdt/mdt_handler.c +++ b/lustre/mdt/mdt_handler.c @@ -50,19 +50,18 @@ #include #include #include -#include +#include #include #include -#include #include #include -#include +#include #include #include #include #include #include - +#include #include #include "mdt_internal.h" @@ -416,7 +415,8 @@ static int mdt_statfs(struct tgt_session_info *tsi) { struct ptlrpc_request *req = tgt_ses_req(tsi); struct mdt_thread_info *info = tsi2mdt_info(tsi); - struct md_device *next = info->mti_mdt->mdt_child; + struct mdt_device *mdt = info->mti_mdt; + struct tg_grants_data *tgd = &mdt->mdt_lut.lut_tgd; struct ptlrpc_service_part *svcpt; struct obd_statfs *osfs; int rc; @@ -441,24 +441,44 @@ static int mdt_statfs(struct tgt_session_info *tsi) if (!osfs) GOTO(out, rc = -EPROTO); - /** statfs information are cached in the mdt_device */ - if (cfs_time_before_64(info->mti_mdt->mdt_osfs_age, - cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) { - /** statfs data is too old, get up-to-date one */ - rc = next->md_ops->mdo_statfs(info->mti_env, next, osfs); - if (rc) - GOTO(out, rc); - spin_lock(&info->mti_mdt->mdt_lock); - info->mti_mdt->mdt_osfs = *osfs; - info->mti_mdt->mdt_osfs_age = cfs_time_current_64(); - spin_unlock(&info->mti_mdt->mdt_lock); - } else { - /** use cached statfs data */ - spin_lock(&info->mti_mdt->mdt_lock); - *osfs = info->mti_mdt->mdt_osfs; - spin_unlock(&info->mti_mdt->mdt_lock); - } + rc = tgt_statfs_internal(tsi->tsi_env, &mdt->mdt_lut, osfs, + cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS), + NULL); + if (unlikely(rc)) + GOTO(out, rc); + /* at least try to account for cached pages. its still racy and + * might be under-reporting if clients haven't announced their + * caches with brw recently */ + CDEBUG(D_SUPER | D_CACHE, "blocks cached %llu granted %llu" + " pending %llu free %llu avail %llu\n", + tgd->tgd_tot_dirty, tgd->tgd_tot_granted, + tgd->tgd_tot_pending, + osfs->os_bfree << tgd->tgd_blockbits, + osfs->os_bavail << tgd->tgd_blockbits); + + osfs->os_bavail -= min_t(u64, osfs->os_bavail, + ((tgd->tgd_tot_dirty + tgd->tgd_tot_pending + + osfs->os_bsize - 1) >> tgd->tgd_blockbits)); + + tgt_grant_sanity_check(mdt->mdt_lu_dev.ld_obd, __func__); + CDEBUG(D_CACHE, "%llu blocks: %llu free, %llu avail; " + "%llu objects: %llu free; state %x\n", + osfs->os_blocks, osfs->os_bfree, osfs->os_bavail, + osfs->os_files, osfs->os_ffree, osfs->os_state); + + if (!exp_grant_param_supp(tsi->tsi_exp) && + tgd->tgd_blockbits > COMPAT_BSIZE_SHIFT) { + /* clients which don't support OBD_CONNECT_GRANT_PARAM + * should not see a block size > page size, otherwise + * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12) + * block size which is the biggest block size known to work + * with all client's page size. */ + osfs->os_blocks <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT; + osfs->os_bfree <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT; + osfs->os_bavail <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT; + osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT; + } if (rc == 0) mdt_counter_incr(req, LPROC_MDT_STATFS); out: @@ -466,6 +486,41 @@ out: RETURN(rc); } +/** + * Pack size attributes into the reply. + */ +int mdt_pack_size2body(struct mdt_thread_info *info, + const struct lu_fid *fid, bool dom_lock) +{ + struct mdt_body *b; + struct md_attr *ma = &info->mti_attr; + int dom_stripe; + + ENTRY; + + LASSERT(ma->ma_attr.la_valid & LA_MODE); + + if (!S_ISREG(ma->ma_attr.la_mode) || + !(ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)) + RETURN(-ENODATA); + + dom_stripe = mdt_lmm_dom_entry(ma->ma_lmm); + /* no DoM stripe, no size in reply */ + if (dom_stripe == LMM_NO_DOM) + RETURN(-ENOENT); + + /* no DoM lock, no size in reply */ + if (!dom_lock) + RETURN(0); + + /* Either DoM lock exists or LMM has only DoM stripe then + * return size on body. */ + b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + + mdt_dom_object_size(info->mti_env, info->mti_mdt, fid, b, dom_lock); + RETURN(0); +} + #ifdef CONFIG_FS_POSIX_ACL /* * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap. @@ -483,14 +538,18 @@ int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody, const struct lu_env *env = info->mti_env; struct md_object *next = mdt_object_child(o); struct lu_buf *buf = &info->mti_buf; + struct mdt_device *mdt = info->mti_mdt; int rc; + ENTRY; + buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL); buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL, RCL_SERVER); if (buf->lb_len == 0) - return 0; + RETURN(0); +again: rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS); if (rc < 0) { if (rc == -ENODATA) { @@ -500,17 +559,49 @@ int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody, } else if (rc == -EOPNOTSUPP) { rc = 0; } else { + if (rc == -ERANGE && + exp_connect_large_acl(info->mti_exp) && + buf->lb_buf != info->mti_big_acl) { + if (info->mti_big_acl == NULL) { + OBD_ALLOC_LARGE(info->mti_big_acl, + mdt->mdt_max_ea_size); + if (info->mti_big_acl == NULL) { + CERROR("%s: unable to grow " + DFID" ACL buffer\n", + mdt_obd_name(mdt), + PFID(mdt_object_fid(o))); + RETURN(-ENOMEM); + } + + info->mti_big_aclsize = + mdt->mdt_max_ea_size; + } + + CDEBUG(D_INODE, "%s: grow the "DFID + " ACL buffer to size %d\n", + mdt_obd_name(mdt), + PFID(mdt_object_fid(o)), + mdt->mdt_max_ea_size); + + buf->lb_buf = info->mti_big_acl; + buf->lb_len = info->mti_big_aclsize; + + goto again; + } + CERROR("%s: unable to read "DFID" ACL: rc = %d\n", - mdt_obd_name(info->mti_mdt), - PFID(mdt_object_fid(o)), rc); + mdt_obd_name(mdt), PFID(mdt_object_fid(o)), rc); } } else { + if (buf->lb_buf == info->mti_big_acl) + info->mti_big_acl_used = 1; + rc = nodemap_map_acl(nodemap, buf->lb_buf, rc, NODEMAP_FS_TO_CLIENT); /* if all ACLs mapped out, rc is still >= 0 */ if (rc < 0) { CERROR("%s: nodemap_map_acl unable to parse "DFID - " ACL: rc = %d\n", mdt_obd_name(info->mti_mdt), + " ACL: rc = %d\n", mdt_obd_name(mdt), PFID(mdt_object_fid(o)), rc); } else { repbody->mbo_aclsize = rc; @@ -518,7 +609,8 @@ int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody, rc = 0; } } - return rc; + + RETURN(rc); } #endif @@ -629,17 +721,20 @@ void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b, /* if no object is allocated on osts, the size on mds is valid. * b=22272 */ b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS; - } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL && - mdt_hsm_is_released(ma->ma_lmm)) { - /* A released file stores its size on MDS. */ - /* But return 1 block for released file, unless tools like tar - * will consider it fully sparse. (LU-3864) - */ - if (unlikely(b->mbo_size == 0)) - b->mbo_blocks = 0; - else - b->mbo_blocks = 1; - b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS; + } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL) { + if (mdt_hsm_is_released(ma->ma_lmm)) { + /* A released file stores its size on MDS. */ + /* But return 1 block for released file, unless tools + * like tar will consider it fully sparse. (LU-3864) + */ + if (unlikely(b->mbo_size == 0)) + b->mbo_blocks = 0; + else + b->mbo_blocks = 1; + b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS; + } else if (info->mti_som_valid) { /* som is valid */ + b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS; + } } if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE)) @@ -899,6 +994,9 @@ int mdt_attr_get_complex(struct mdt_thread_info *info, rc = mo_attr_get(env, next, ma); if (rc) GOTO(out, rc); + + if (S_ISREG(mode)) + (void) mdt_get_som(info, o, &ma->ma_attr); ma->ma_valid |= MA_INODE; } @@ -1205,6 +1303,12 @@ static int mdt_getattr(struct tgt_session_info *tsi) req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc); + /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD + * by default. If the target object has more ACL entries, then + * enlarge the buffer when necessary. */ + req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER, + LUSTRE_POSIX_ACL_MAX_SIZE_OLD); + rc = req_capsule_server_pack(pill); if (unlikely(rc != 0)) GOTO(out, rc = err_serious(rc)); @@ -1237,32 +1341,33 @@ out: * * \param[in] info thread environment * \param[in] obj object - * \param[in] layout layout intent - * \param[in] buf buffer containing client's lovea, could be empty + * \param[in] layout layout change descriptor * * \retval 0 on success * \retval < 0 error code */ static int mdt_layout_change(struct mdt_thread_info *info, struct mdt_object *obj, - struct layout_intent *layout, - const struct lu_buf *buf) + struct md_layout_change *layout) { struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL]; + struct layout_intent *intent = layout->mlc_intent; int rc; ENTRY; CDEBUG(D_INFO, "got layout change request from client: " - "opc:%u flags:%#x extent[%#llx,%#llx)\n", - layout->li_opc, layout->li_flags, - layout->li_start, layout->li_end); - if (layout->li_start >= layout->li_end) { - CERROR("Recieved an invalid layout change range [%llu, %llu) " - "for "DFID"\n", layout->li_start, layout->li_end, - PFID(mdt_object_fid(obj))); + "opc:%u flags:%#x extent "DEXT"\n", + intent->li_opc, intent->li_flags, PEXT(&intent->li_extent)); + + if (intent->li_extent.e_start >= intent->li_extent.e_end) { + CERROR(DFID ":invalid range of layout change "DEXT"\n", + PFID(mdt_object_fid(obj)), PEXT(&intent->li_extent)); RETURN(-EINVAL); } + if (!mdt_object_exists(obj)) + GOTO(out, rc = -ENOENT); + if (!S_ISREG(lu_object_attr(&obj->mot_obj))) GOTO(out, rc = -EINVAL); @@ -1273,13 +1378,11 @@ static int mdt_layout_change(struct mdt_thread_info *info, /* take layout lock to prepare layout change */ mdt_lock_reg_init(lh, LCK_EX); - rc = mdt_object_lock(info, obj, lh, - MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR); + rc = mdt_object_lock(info, obj, lh, MDS_INODELOCK_LAYOUT); if (rc) GOTO(out, rc); - rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout, - buf); + rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout); mdt_object_unlock(info, obj, lh, 1); out: @@ -1439,18 +1542,19 @@ static int mdt_getattr_name_lock(struct mdt_thread_info *info, __u64 child_bits, struct ldlm_reply *ldlm_rep) { - struct ptlrpc_request *req = mdt_info_req(info); - struct mdt_body *reqbody = NULL; - struct mdt_object *parent = info->mti_object; - struct mdt_object *child; - struct lu_fid *child_fid = &info->mti_tmp_fid1; - struct lu_name *lname = NULL; - struct mdt_lock_handle *lhp = NULL; - struct ldlm_lock *lock; - bool is_resent; - bool try_layout; - int ma_need = 0; - int rc; + struct ptlrpc_request *req = mdt_info_req(info); + struct mdt_body *reqbody = NULL; + struct mdt_object *parent = info->mti_object; + struct mdt_object *child; + struct lu_fid *child_fid = &info->mti_tmp_fid1; + struct lu_name *lname = NULL; + struct mdt_lock_handle *lhp = NULL; + struct ldlm_lock *lock; + __u64 try_bits = 0; + bool is_resent; + int ma_need = 0; + int rc; + ENTRY; is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh); @@ -1612,19 +1716,18 @@ static int mdt_getattr_name_lock(struct mdt_thread_info *info, if (rc < 0) { GOTO(out_child, rc); } else if (rc > 0) { - mdt_lock_handle_init(lhc); + mdt_lock_handle_init(lhc); mdt_lock_reg_init(lhc, LCK_PR); - try_layout = false; if (!(child_bits & MDS_INODELOCK_UPDATE) && mdt_object_exists(child) && !mdt_object_remote(child)) { - struct md_attr *ma = &info->mti_attr; + struct md_attr *ma = &info->mti_attr; - ma->ma_valid = 0; - ma->ma_need = MA_INODE; + ma->ma_valid = 0; + ma->ma_need = MA_INODE; rc = mdt_attr_get_complex(info, child, ma); - if (unlikely(rc != 0)) - GOTO(out_child, rc); + if (unlikely(rc != 0)) + GOTO(out_child, rc); /* If the file has not been changed for some time, we * return not only a LOOKUP lock, but also an UPDATE @@ -1641,28 +1744,25 @@ static int mdt_getattr_name_lock(struct mdt_thread_info *info, /* layout lock must be granted in a best-effort way * for IT operations */ LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT)); - if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) && - exp_connect_layout(info->mti_exp) && - S_ISREG(lu_object_attr(&child->mot_obj)) && + if (S_ISREG(lu_object_attr(&child->mot_obj)) && !mdt_object_remote(child) && ldlm_rep != NULL) { - /* try to grant layout lock for regular file. */ - try_layout = true; + if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) && + exp_connect_layout(info->mti_exp)) { + /* try to grant layout lock for regular file. */ + try_bits = MDS_INODELOCK_LAYOUT; + } + /* Acquire DOM lock in advance for data-on-mdt file */ + if (child != parent) + try_bits |= MDS_INODELOCK_DOM; } - rc = 0; - if (try_layout) { - child_bits |= MDS_INODELOCK_LAYOUT; + if (try_bits != 0) { /* try layout lock, it may fail to be granted due to * contention at LOOKUP or UPDATE */ - if (!mdt_object_lock_try(info, child, lhc, - child_bits)) { - child_bits &= ~MDS_INODELOCK_LAYOUT; - LASSERT(child_bits != 0); - rc = mdt_object_lock(info, child, lhc, - child_bits); - } else { + rc = mdt_object_lock_try(info, child, lhc, &child_bits, + try_bits, false); + if (child_bits & MDS_INODELOCK_LAYOUT) ma_need |= MA_LOV; - } } else { /* Do not enqueue the UPDATE lock from MDT(cross-MDT), * client will enqueue the lock to the remote MDT */ @@ -1688,6 +1788,27 @@ static int mdt_getattr_name_lock(struct mdt_thread_info *info, "Lock res_id: "DLDLMRES", fid: "DFID"\n", PLDLMRES(lock->l_resource), PFID(mdt_object_fid(child))); + + if (S_ISREG(lu_object_attr(&child->mot_obj)) && + mdt_object_exists(child) && !mdt_object_remote(child) && + child != parent) { + LDLM_LOCK_PUT(lock); + mdt_object_put(info->mti_env, child); + /* NB: call the mdt_pack_size2body always after + * mdt_object_put(), that is why this speacial + * exit path is used. */ + rc = mdt_pack_size2body(info, child_fid, + child_bits & MDS_INODELOCK_DOM); + if (rc != 0 && child_bits & MDS_INODELOCK_DOM) { + /* DOM lock was taken in advance but this is + * not DoM file. Drop the lock. */ + lock_res_and_lock(lock); + ldlm_inodebits_drop(lock, MDS_INODELOCK_DOM); + unlock_res_and_lock(lock); + } + + GOTO(out_parent, rc = 0); + } } if (lock) LDLM_LOCK_PUT(lock); @@ -1919,6 +2040,13 @@ static int mdt_reint_internal(struct mdt_thread_info *info, if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER)) req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0); + /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD + * by default. If the target object has more ACL entries, then + * enlarge the buffer when necessary. */ + if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER)) + req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER, + LUSTRE_POSIX_ACL_MAX_SIZE_OLD); + rc = req_capsule_server_pack(pill); if (rc != 0) { CERROR("Can't pack response, rc %d\n", rc); @@ -2040,20 +2168,21 @@ static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt) } /* this should sync this object */ -static int mdt_object_sync(struct mdt_thread_info *info) +static int mdt_object_sync(const struct lu_env *env, struct obd_export *exp, + struct mdt_object *mo) { - struct md_object *next; int rc; + ENTRY; - if (!mdt_object_exists(info->mti_object)) { + if (!mdt_object_exists(mo)) { CWARN("%s: non existing object "DFID": rc = %d\n", - mdt_obd_name(info->mti_mdt), - PFID(mdt_object_fid(info->mti_object)), -ESTALE); + exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)), + -ESTALE); RETURN(-ESTALE); } - next = mdt_object_child(info->mti_object); - rc = mo_object_sync(info->mti_env, next); + + rc = mo_object_sync(env, mdt_object_child(mo)); RETURN(rc); } @@ -2076,7 +2205,8 @@ static int mdt_sync(struct tgt_session_info *tsi) struct mdt_thread_info *info = tsi2mdt_info(tsi); /* sync an object */ - rc = mdt_object_sync(info); + rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, + info->mti_object); if (rc == 0) { const struct lu_fid *fid; struct lu_attr *la = &info->mti_attr.ma_attr; @@ -2100,6 +2230,54 @@ static int mdt_sync(struct tgt_session_info *tsi) RETURN(rc); } +static int mdt_data_sync(struct tgt_session_info *tsi) +{ + struct mdt_thread_info *info; + struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp); + struct ost_body *body = tsi->tsi_ost_body; + struct ost_body *repbody; + struct mdt_object *mo = NULL; + struct md_attr *ma; + int rc = 0; + + ENTRY; + + repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY); + + /* if no fid is specified then do nothing, + * device sync is done via MDS_SYNC */ + if (fid_is_zero(&tsi->tsi_fid)) + RETURN(0); + + mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid); + if (IS_ERR(mo)) + RETURN(PTR_ERR(mo)); + + rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, mo); + if (rc) + GOTO(put, rc); + + repbody->oa.o_oi = body->oa.o_oi; + repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP; + + info = tsi2mdt_info(tsi); + ma = &info->mti_attr; + ma->ma_need = MA_INODE; + ma->ma_valid = 0; + rc = mdt_attr_get_complex(info, mo, ma); + if (rc == 0) + obdo_from_la(&repbody->oa, &ma->ma_attr, VALID_FLAGS); + else + rc = 0; + mdt_thread_info_fini(info); + + EXIT; +put: + if (mo != NULL) + mdt_object_put(tsi->tsi_env, mo); + return rc; +} + /* * Handle quota control requests to consult current usage/limit, but also * to configure quota enforcement @@ -2557,10 +2735,10 @@ int mdt_check_resent_lock(struct mdt_thread_info *info, return 1; } -int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o, - const struct lu_fid *fid, struct lustre_handle *lh, - enum ldlm_mode mode, __u64 ibits, bool nonblock, - bool cache) +int mdt_remote_object_lock_try(struct mdt_thread_info *mti, + struct mdt_object *o, const struct lu_fid *fid, + struct lustre_handle *lh, enum ldlm_mode mode, + __u64 *ibits, __u64 trybits, bool cache) { struct ldlm_enqueue_info *einfo = &mti->mti_einfo; union ldlm_policy_data *policy = &mti->mti_policy; @@ -2579,8 +2757,7 @@ int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o, einfo->ei_cb_cp = ldlm_completion_ast; einfo->ei_enq_slave = 0; einfo->ei_res_id = res_id; - if (nonblock) - einfo->ei_nonblock = 1; + if (cache) { /* * if we cache lock, couple lock with mdt_object, so that object @@ -2590,8 +2767,10 @@ int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o, einfo->ei_cbdata = o; } + memset(policy, 0, sizeof(*policy)); - policy->l_inodebits.bits = ibits; + policy->l_inodebits.bits = *ibits; + policy->l_inodebits.try_bits = trybits; rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo, policy); @@ -2599,13 +2778,30 @@ int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o, mdt_object_put(mti->mti_env, o); einfo->ei_cbdata = NULL; } + + /* Return successfully acquired bits to a caller */ + if (rc == 0) { + struct ldlm_lock *lock = ldlm_handle2lock(lh); + + LASSERT(lock); + *ibits = lock->l_policy_data.l_inodebits.bits; + LDLM_LOCK_PUT(lock); + } RETURN(rc); } +int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o, + const struct lu_fid *fid, struct lustre_handle *lh, + enum ldlm_mode mode, __u64 ibits, bool cache) +{ + return mdt_remote_object_lock_try(mti, o, fid, lh, mode, &ibits, 0, + cache); +} + static int mdt_object_local_lock(struct mdt_thread_info *info, struct mdt_object *o, - struct mdt_lock_handle *lh, __u64 ibits, - bool nonblock, bool cos_incompat) + struct mdt_lock_handle *lh, __u64 *ibits, + __u64 trybits, bool cos_incompat) { struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace; union ldlm_policy_data *policy = &info->mti_policy; @@ -2628,8 +2824,9 @@ static int mdt_object_local_lock(struct mdt_thread_info *info, } /* Only enqueue LOOKUP lock for remote object */ - if (mdt_object_remote(o)) - LASSERT(ibits == MDS_INODELOCK_LOOKUP); + if (mdt_object_remote(o)) { + LASSERT(*ibits == MDS_INODELOCK_LOOKUP); + } if (lh->mlh_type == MDT_PDO_LOCK) { /* check for exists after object is locked */ @@ -2640,30 +2837,28 @@ static int mdt_object_local_lock(struct mdt_thread_info *info, /* Non-dir object shouldn't have PDO lock */ if (!S_ISDIR(lu_object_attr(&o->mot_obj))) RETURN(-ENOTDIR); - } - } + } + } - memset(policy, 0, sizeof(*policy)); - fid_build_reg_res_name(mdt_object_fid(o), res_id); + fid_build_reg_res_name(mdt_object_fid(o), res_id); dlmflags |= LDLM_FL_ATOMIC_CB; - if (nonblock) - dlmflags |= LDLM_FL_BLOCK_NOWAIT; - /* - * Take PDO lock on whole directory and build correct @res_id for lock - * on part of directory. - */ - if (lh->mlh_pdo_hash != 0) { - LASSERT(lh->mlh_type == MDT_PDO_LOCK); - mdt_lock_pdo_mode(info, o, lh); - if (lh->mlh_pdo_mode != LCK_NL) { - /* - * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it - * is never going to be sent to client and we do not - * want it slowed down due to possible cancels. - */ - policy->l_inodebits.bits = MDS_INODELOCK_UPDATE; + /* + * Take PDO lock on whole directory and build correct @res_id for lock + * on part of directory. + */ + if (lh->mlh_pdo_hash != 0) { + LASSERT(lh->mlh_type == MDT_PDO_LOCK); + mdt_lock_pdo_mode(info, o, lh); + if (lh->mlh_pdo_mode != LCK_NL) { + /* + * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it + * is never going to be sent to client and we do not + * want it slowed down due to possible cancels. + */ + policy->l_inodebits.bits = MDS_INODELOCK_UPDATE; + policy->l_inodebits.try_bits = 0; rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, policy, res_id, dlmflags, info->mti_exp == NULL ? NULL : @@ -2679,7 +2874,8 @@ static int mdt_object_local_lock(struct mdt_thread_info *info, res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash; } - policy->l_inodebits.bits = ibits; + policy->l_inodebits.bits = *ibits; + policy->l_inodebits.try_bits = trybits; /* * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is @@ -2698,39 +2894,49 @@ out_unlock: (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX)) OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15); + /* Return successfully acquired bits to a caller */ + if (rc == 0) { + struct ldlm_lock *lock = ldlm_handle2lock(&lh->mlh_reg_lh); + + LASSERT(lock); + *ibits = lock->l_policy_data.l_inodebits.bits; + LDLM_LOCK_PUT(lock); + } RETURN(rc); } static int mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o, - struct mdt_lock_handle *lh, __u64 ibits, bool nonblock, - bool cos_incompat) + struct mdt_lock_handle *lh, __u64 *ibits, + __u64 trybits, bool cos_incompat) { struct mdt_lock_handle *local_lh = NULL; int rc; ENTRY; if (!mdt_object_remote(o)) { - rc = mdt_object_local_lock(info, o, lh, ibits, nonblock, + rc = mdt_object_local_lock(info, o, lh, ibits, trybits, cos_incompat); RETURN(rc); } /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */ - ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT | - MDS_INODELOCK_XATTR); + *ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT | + MDS_INODELOCK_XATTR); /* Only enqueue LOOKUP lock for remote object */ - if (ibits & MDS_INODELOCK_LOOKUP) { - rc = mdt_object_local_lock(info, o, lh, MDS_INODELOCK_LOOKUP, - nonblock, cos_incompat); + if (*ibits & MDS_INODELOCK_LOOKUP) { + __u64 local = MDS_INODELOCK_LOOKUP; + + rc = mdt_object_local_lock(info, o, lh, &local, 0, + cos_incompat); if (rc != ELDLM_OK) RETURN(rc); local_lh = lh; } - if (ibits & MDS_INODELOCK_UPDATE) { + if ((*ibits | trybits) & MDS_INODELOCK_UPDATE) { /* Sigh, PDO needs to enqueue 2 locks right now, but * enqueue RPC can only request 1 lock, to avoid extra * RPC, so it will instead enqueue EX lock for remote @@ -2744,11 +2950,11 @@ mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o, lh->mlh_rreg_mode = LCK_EX; lh->mlh_type = MDT_REG_LOCK; } - rc = mdt_remote_object_lock(info, o, mdt_object_fid(o), - &lh->mlh_rreg_lh, - lh->mlh_rreg_mode, - MDS_INODELOCK_UPDATE, nonblock, - false); + + rc = mdt_remote_object_lock_try(info, o, mdt_object_fid(o), + &lh->mlh_rreg_lh, + lh->mlh_rreg_mode, + ibits, trybits, false); if (rc != ELDLM_OK) { if (local_lh != NULL) mdt_object_unlock(info, o, local_lh, rc); @@ -2762,7 +2968,7 @@ mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o, int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o, struct mdt_lock_handle *lh, __u64 ibits) { - return mdt_object_lock_internal(info, o, lh, ibits, false, false); + return mdt_object_lock_internal(info, o, lh, &ibits, 0, false); } int mdt_reint_object_lock(struct mdt_thread_info *info, struct mdt_object *o, @@ -2770,36 +2976,16 @@ int mdt_reint_object_lock(struct mdt_thread_info *info, struct mdt_object *o, bool cos_incompat) { LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX); - return mdt_object_lock_internal(info, o, lh, ibits, false, + return mdt_object_lock_internal(info, o, lh, &ibits, 0, cos_incompat); } int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o, - struct mdt_lock_handle *lh, __u64 ibits) -{ - struct mdt_lock_handle tmp = *lh; - int rc; - - rc = mdt_object_lock_internal(info, o, &tmp, ibits, true, false); - if (rc == 0) - *lh = tmp; - - return rc == 0; -} - -int mdt_reint_object_lock_try(struct mdt_thread_info *info, - struct mdt_object *o, struct mdt_lock_handle *lh, - __u64 ibits, bool cos_incompat) + struct mdt_lock_handle *lh, __u64 *ibits, + __u64 trybits, bool cos_incompat) { - struct mdt_lock_handle tmp = *lh; - int rc; - - LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX); - rc = mdt_object_lock_internal(info, o, &tmp, ibits, true, cos_incompat); - if (rc == 0) - *lh = tmp; - - return rc == 0; + return mdt_object_lock_internal(info, o, lh, ibits, trybits, + cos_incompat); } /** @@ -2815,8 +3001,8 @@ int mdt_reint_object_lock_try(struct mdt_thread_info *info, * \param mode lock mode * \param decref force immediate lock releasing */ -static void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h, - enum ldlm_mode mode, int decref) +void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h, + enum ldlm_mode mode, int decref) { ENTRY; @@ -2844,8 +3030,11 @@ static void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h, ldlm_lock_downgrade(lock, LCK_COS); mode = LCK_COS; } - ptlrpc_save_lock(req, h, mode, cos, - convert_lock); + if (req->rq_export->exp_disconnected) + mdt_fid_unlock(h, mode); + else + ptlrpc_save_lock(req, h, mode, cos, + convert_lock); } else { mdt_fid_unlock(h, mode); } @@ -3033,6 +3222,13 @@ static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags) req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0); + /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD + * by default. If the target object has more ACL entries, then + * enlarge the buffer when necessary. */ + if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER)) + req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER, + LUSTRE_POSIX_ACL_MAX_SIZE_OLD); + rc = req_capsule_server_pack(pill); } RETURN(rc); @@ -3089,6 +3285,8 @@ void mdt_thread_info_init(struct ptlrpc_request *req, info->mti_cross_ref = 0; info->mti_opdata = 0; info->mti_big_lmm_used = 0; + info->mti_big_acl_used = 0; + info->mti_som_valid = 0; info->mti_spec.no_create = 0; info->mti_spec.sp_rm_entry = 0; @@ -3160,13 +3358,14 @@ enum mdt_it_code { MDT_IT_GETXATTR, MDT_IT_LAYOUT, MDT_IT_QUOTA, - MDT_IT_NR + MDT_IT_GLIMPSE, + MDT_IT_BRW, + MDT_IT_NR }; static int mdt_intent_getattr(enum mdt_it_code opcode, - struct mdt_thread_info *info, - struct ldlm_lock **, - __u64); + struct mdt_thread_info *info, + struct ldlm_lock **, __u64); static int mdt_intent_getxattr(enum mdt_it_code opcode, struct mdt_thread_info *info, @@ -3181,6 +3380,20 @@ static int mdt_intent_reint(enum mdt_it_code opcode, struct mdt_thread_info *info, struct ldlm_lock **, __u64); +static int mdt_intent_glimpse(enum mdt_it_code opcode, + struct mdt_thread_info *info, + struct ldlm_lock **lockp, __u64 flags) +{ + return mdt_glimpse_enqueue(info, info->mti_mdt->mdt_namespace, + lockp, flags); +} +static int mdt_intent_brw(enum mdt_it_code opcode, + struct mdt_thread_info *info, + struct ldlm_lock **lockp, __u64 flags) +{ + return mdt_brw_enqueue(info, info->mti_mdt->mdt_namespace, + lockp, flags); +} static struct mdt_it_flavor { const struct req_format *it_fmt; @@ -3252,14 +3465,24 @@ static struct mdt_it_flavor { .it_fmt = &RQF_LDLM_INTENT_LAYOUT, .it_flags = 0, .it_act = mdt_intent_layout - } + }, + [MDT_IT_GLIMPSE] = { + .it_fmt = &RQF_LDLM_INTENT, + .it_flags = 0, + .it_act = mdt_intent_glimpse, + }, + [MDT_IT_BRW] = { + .it_fmt = &RQF_LDLM_INTENT, + .it_flags = 0, + .it_act = mdt_intent_brw, + }, + }; -static int -mdt_intent_lock_replace(struct mdt_thread_info *info, - struct ldlm_lock **lockp, - struct mdt_lock_handle *lh, - __u64 flags, int result) +int mdt_intent_lock_replace(struct mdt_thread_info *info, + struct ldlm_lock **lockp, + struct mdt_lock_handle *lh, + __u64 flags, int result) { struct ptlrpc_request *req = mdt_info_req(info); struct ldlm_lock *lock = *lockp; @@ -3335,6 +3558,8 @@ mdt_intent_lock_replace(struct mdt_thread_info *info, new_lock->l_export = class_export_lock_get(req->rq_export, new_lock); new_lock->l_blocking_ast = lock->l_blocking_ast; new_lock->l_completion_ast = lock->l_completion_ast; + if (ldlm_has_dom(new_lock)) + new_lock->l_glimpse_ast = ldlm_server_glimpse_ast; new_lock->l_remote_handle = lock->l_remote_handle; new_lock->l_flags &= ~LDLM_FL_LOCAL; @@ -3350,10 +3575,9 @@ mdt_intent_lock_replace(struct mdt_thread_info *info, RETURN(ELDLM_LOCK_REPLACED); } -static void mdt_intent_fixup_resent(struct mdt_thread_info *info, - struct ldlm_lock *new_lock, - struct mdt_lock_handle *lh, - __u64 flags) +void mdt_intent_fixup_resent(struct mdt_thread_info *info, + struct ldlm_lock *new_lock, + struct mdt_lock_handle *lh, __u64 flags) { struct ptlrpc_request *req = mdt_info_req(info); struct ldlm_request *dlmreq; @@ -3519,10 +3743,10 @@ static int mdt_intent_layout(enum mdt_it_code opcode, __u64 flags) { struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT]; - struct layout_intent *layout; + struct md_layout_change layout = { .mlc_opc = MD_LAYOUT_NOP }; + struct layout_intent *intent; struct lu_fid *fid; struct mdt_object *obj = NULL; - bool layout_change = false; int layout_size = 0; int rc = 0; ENTRY; @@ -3533,14 +3757,15 @@ static int mdt_intent_layout(enum mdt_it_code opcode, RETURN(-EINVAL); } - layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT); - if (layout == NULL) + intent = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT); + if (intent == NULL) RETURN(-EPROTO); - switch (layout->li_opc) { + switch (intent->li_opc) { case LAYOUT_INTENT_TRUNC: case LAYOUT_INTENT_WRITE: - layout_change = true; + layout.mlc_opc = MD_LAYOUT_WRITE; + layout.mlc_intent = intent; break; case LAYOUT_INTENT_ACCESS: break; @@ -3549,12 +3774,12 @@ static int mdt_intent_layout(enum mdt_it_code opcode, case LAYOUT_INTENT_RELEASE: case LAYOUT_INTENT_RESTORE: CERROR("%s: Unsupported layout intent opc %d\n", - mdt_obd_name(info->mti_mdt), layout->li_opc); + mdt_obd_name(info->mti_mdt), intent->li_opc); rc = -ENOTSUPP; break; default: CERROR("%s: Unknown layout intent opc %d\n", - mdt_obd_name(info->mti_mdt), layout->li_opc); + mdt_obd_name(info->mti_mdt), intent->li_opc); rc = -EINVAL; break; } @@ -3592,8 +3817,8 @@ static int mdt_intent_layout(enum mdt_it_code opcode, GOTO(out_obj, rc); - if (layout_change) { - struct lu_buf *buf = &info->mti_buf; + if (layout.mlc_opc != MD_LAYOUT_NOP) { + struct lu_buf *buf = &layout.mlc_buf; /** * mdt_layout_change is a reint operation, when the request @@ -3637,7 +3862,7 @@ static int mdt_intent_layout(enum mdt_it_code opcode, * lovea, then it's a replay of the layout intent write * RPC. */ - rc = mdt_layout_change(info, obj, layout, buf); + rc = mdt_layout_change(info, obj, &layout); if (rc) GOTO(out_obj, rc); } @@ -3768,6 +3993,12 @@ static int mdt_intent_code(enum ldlm_intent_flags itcode) case IT_QUOTA_CONN: rc = MDT_IT_QUOTA; break; + case IT_GLIMPSE: + rc = MDT_IT_GLIMPSE; + break; + case IT_BRW: + rc = MDT_IT_BRW; + break; default: CERROR("Unknown intent opcode: 0x%08x\n", itcode); rc = -EINVAL; @@ -3839,6 +4070,18 @@ static int mdt_intent_opc(enum ldlm_intent_flags itopc, RETURN(rc); } +static void mdt_ptlrpc_stats_update(struct ptlrpc_request *req, + enum ldlm_intent_flags it_opc) +{ + struct lprocfs_stats *srv_stats = ptlrpc_req2svc(req)->srv_stats; + + /* update stats when IT code is known */ + if (srv_stats != NULL) + lprocfs_counter_incr(srv_stats, + PTLRPC_LAST_CNTR + (it_opc == IT_GLIMPSE ? + LDLM_GLIMPSE_ENQUEUE : LDLM_IBITS_ENQUEUE)); +} + static int mdt_intent_policy(struct ldlm_namespace *ns, struct ldlm_lock **lockp, void *req_cookie, enum ldlm_mode mode, __u64 flags, void *data) @@ -3848,6 +4091,7 @@ static int mdt_intent_policy(struct ldlm_namespace *ns, struct ptlrpc_request *req = req_cookie; struct ldlm_intent *it; struct req_capsule *pill; + const struct ldlm_lock_desc *ldesc; int rc; ENTRY; @@ -3857,37 +4101,37 @@ static int mdt_intent_policy(struct ldlm_namespace *ns, tsi = tgt_ses_info(req->rq_svc_thread->t_env); info = tsi2mdt_info(tsi); - LASSERT(info != NULL); - pill = info->mti_pill; - LASSERT(pill->rc_req == req); + LASSERT(info != NULL); + pill = info->mti_pill; + LASSERT(pill->rc_req == req); + ldesc = &info->mti_dlm_req->lock_desc; - if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) { + if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) { req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC); - it = req_capsule_client_get(pill, &RMF_LDLM_INTENT); - if (it != NULL) { - rc = mdt_intent_opc(it->opc, info, lockp, flags); - if (rc == 0) - rc = ELDLM_OK; - - /* Lock without inodebits makes no sense and will oops - * later in ldlm. Let's check it now to see if we have - * ibits corrupted somewhere in mdt_intent_opc(). - * The case for client miss to set ibits has been - * processed by others. */ - LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\ - lr_type == LDLM_IBITS, - info->mti_dlm_req->lock_desc.\ - l_policy_data.l_inodebits.bits != 0)); - } else - rc = err_serious(-EFAULT); - } else { - /* No intent was provided */ - LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE); + it = req_capsule_client_get(pill, &RMF_LDLM_INTENT); + if (it != NULL) { + mdt_ptlrpc_stats_update(req, it->opc); + rc = mdt_intent_opc(it->opc, info, lockp, flags); + if (rc == 0) + rc = ELDLM_OK; + + /* Lock without inodebits makes no sense and will oops + * later in ldlm. Let's check it now to see if we have + * ibits corrupted somewhere in mdt_intent_opc(). + * The case for client miss to set ibits has been + * processed by others. */ + LASSERT(ergo(ldesc->l_resource.lr_type == LDLM_IBITS, + ldesc->l_policy_data.l_inodebits.bits != 0)); + } else { + rc = err_serious(-EFAULT); + } + } else { + /* No intent was provided */ req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0); - rc = req_capsule_server_pack(pill); - if (rc) - rc = err_serious(rc); - } + rc = req_capsule_server_pack(pill); + if (rc) + rc = err_serious(rc); + } mdt_thread_info_fini(info); RETURN(rc); } @@ -4163,12 +4407,13 @@ static void mdt_stack_pre_fini(const struct lu_env *env, * by osd only doesn't have mdt/mdd slices -bzzz */ lustre_cfg_bufs_reset(bufs, mdt_obd_name(m)); lustre_cfg_bufs_set_string(bufs, 1, NULL); - lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs); - if (lcfg == NULL) + OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen)); + if (!lcfg) RETURN_EXIT; + lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, bufs); top->ld_ops->ldo_process_config(env, top, lcfg); - lustre_cfg_free(lcfg); + OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens)); EXIT; } @@ -4198,13 +4443,14 @@ static void mdt_stack_fini(const struct lu_env *env, if (obd->obd_fail) strcat(flags, "A"); lustre_cfg_bufs_set_string(bufs, 1, flags); - lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs); - if (lcfg == NULL) + OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen)); + if (!lcfg) RETURN_EXIT; + lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs); LASSERT(top); top->ld_ops->ldo_process_config(env, top, lcfg); - lustre_cfg_free(lcfg); + OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens)); lu_site_purge(env, top->ld_site, -1); @@ -4324,9 +4570,10 @@ static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt, lustre_cfg_bufs_set_string(bufs, 2, uuid); lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt); - lcfg = lustre_cfg_new(LCFG_ATTACH, bufs); - if (lcfg == NULL) + OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen)); + if (!lcfg) GOTO(put_profile, rc = -ENOMEM); + lustre_cfg_init(lcfg, LCFG_ATTACH, bufs); rc = class_attach(lcfg); if (rc) @@ -4339,16 +4586,17 @@ static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt, GOTO(lcfg_cleanup, rc = -EINVAL); } - lustre_cfg_free(lcfg); + OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens)); lustre_cfg_bufs_reset(bufs, name); lustre_cfg_bufs_set_string(bufs, 1, uuid); lustre_cfg_bufs_set_string(bufs, 2, dev); lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt); - lcfg = lustre_cfg_new(LCFG_SETUP, bufs); - if (lcfg == NULL) + OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen)); + if (!lcfg) GOTO(class_detach, rc = -ENOMEM); + lustre_cfg_init(lcfg, LCFG_SETUP, bufs); rc = class_setup(obd, lcfg); if (rc) @@ -4385,7 +4633,7 @@ class_detach: if (rc) class_detach(obd, lcfg); lcfg_cleanup: - lustre_cfg_free(lcfg); + OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens)); put_profile: class_put_profile(lprof); free_bufs: @@ -4456,9 +4704,10 @@ static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt, lustre_cfg_bufs_set_string(bufs, 2, uuid); lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt); - lcfg = lustre_cfg_new(LCFG_ATTACH, bufs); - if (lcfg == NULL) + OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen)); + if (!lcfg) GOTO(put_profile, rc = -ENOMEM); + lustre_cfg_init(lcfg, LCFG_ATTACH, bufs); rc = class_attach(lcfg); if (rc) @@ -4471,7 +4720,7 @@ static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt, GOTO(lcfg_cleanup, rc = -EINVAL); } - lustre_cfg_free(lcfg); + OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens)); lustre_cfg_bufs_reset(bufs, qmtname); lustre_cfg_bufs_set_string(bufs, 1, uuid); @@ -4481,9 +4730,10 @@ static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt, lustre_cfg_bufs_set_string(bufs, 3, mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name); - lcfg = lustre_cfg_new(LCFG_SETUP, bufs); - if (lcfg == NULL) + OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen)); + if (!lcfg) GOTO(class_detach, rc = -ENOMEM); + lustre_cfg_init(lcfg, LCFG_SETUP, bufs); rc = class_setup(obd, lcfg); if (rc) @@ -4519,7 +4769,7 @@ class_detach: if (rc) class_detach(obd, lcfg); lcfg_cleanup: - lustre_cfg_free(lcfg); + OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens)); put_profile: class_put_profile(lprof); cleanup_mem: @@ -4564,6 +4814,11 @@ static int mdt_tgt_getxattr(struct tgt_session_info *tsi) return rc; } +#define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET +#define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET +#define OST_BRW_READ OST_READ +#define OST_BRW_WRITE OST_WRITE + static struct tgt_handler mdt_tgt_handlers[] = { TGT_RPC_HANDLER(MDS_FIRST_OPC, 0, MDS_CONNECT, mdt_tgt_connect, @@ -4604,6 +4859,14 @@ TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR, mdt_swap_layouts), }; +static struct tgt_handler mdt_io_ops[] = { +TGT_OST_HDL(HABEO_CORPUS | HABEO_REFERO, OST_BRW_READ, tgt_brw_read), +TGT_OST_HDL(HABEO_CORPUS | MUTABOR, OST_BRW_WRITE, tgt_brw_write), +TGT_OST_HDL(HABEO_CORPUS | HABEO_REFERO | MUTABOR, + OST_PUNCH, mdt_punch_hdl), +TGT_OST_HDL(HABEO_CORPUS | HABEO_REFERO, OST_SYNC, mdt_data_sync), +}; + static struct tgt_handler mdt_sec_ctx_ops[] = { TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle), TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle), @@ -4665,7 +4928,11 @@ static struct tgt_opc_slice mdt_common_slice[] = { .tos_opc_end = LFSCK_LAST_OPC, .tos_hs = tgt_lfsck_handlers }, - + { + .tos_opc_start = OST_FIRST_OPC, + .tos_opc_end = OST_LAST_OPC, + .tos_hs = mdt_io_ops + }, { .tos_hs = NULL } @@ -4686,8 +4953,10 @@ static void mdt_fini(const struct lu_env *env, struct mdt_device *m) mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child)); ping_evictor_stop(); - if (m->mdt_opts.mo_coordinator) - mdt_hsm_cdt_stop(m); + /* Remove the HSM /proc entry so the coordinator cannot be + * restarted by a user while it's shutting down. */ + hsm_cdt_procfs_fini(m); + mdt_hsm_cdt_stop(m); mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT); mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT); @@ -4747,71 +5016,71 @@ static void mdt_fini(const struct lu_env *env, struct mdt_device *m) static int mdt_postrecov(const struct lu_env *, struct mdt_device *); static int mdt_init0(const struct lu_env *env, struct mdt_device *m, - struct lu_device_type *ldt, struct lustre_cfg *cfg) -{ - struct mdt_thread_info *info; - struct obd_device *obd; - const char *dev = lustre_cfg_string(cfg, 0); - const char *num = lustre_cfg_string(cfg, 2); - struct lustre_mount_info *lmi = NULL; - struct lustre_sb_info *lsi; - struct lu_site *s; - struct seq_server_site *ss_site; - const char *identity_upcall = "NONE"; - struct md_device *next; - int rc; - long node_id; - mntopt_t mntopts; - ENTRY; + struct lu_device_type *ldt, struct lustre_cfg *cfg) +{ + const struct dt_device_param *dt_conf; + struct mdt_thread_info *info; + struct obd_device *obd; + const char *dev = lustre_cfg_string(cfg, 0); + const char *num = lustre_cfg_string(cfg, 2); + struct tg_grants_data *tgd = &m->mdt_lut.lut_tgd; + struct lustre_mount_info *lmi = NULL; + struct lustre_sb_info *lsi; + struct lu_site *s; + struct seq_server_site *ss_site; + const char *identity_upcall = "NONE"; + struct md_device *next; + int rc; + long node_id; + mntopt_t mntopts; + ENTRY; lu_device_init(&m->mdt_lu_dev, ldt); - /* - * Environment (env) might be missing mdt_thread_key values at that - * point, if device is allocated when mdt_thread_key is in QUIESCENT - * mode. - * - * Usually device allocation path doesn't use module key values, but - * mdt has to do a lot of work here, so allocate key value. - */ - rc = lu_env_refill((struct lu_env *)env); - if (rc != 0) - RETURN(rc); + /* + * Environment (env) might be missing mdt_thread_key values at that + * point, if device is allocated when mdt_thread_key is in QUIESCENT + * mode. + * + * Usually device allocation path doesn't use module key values, but + * mdt has to do a lot of work here, so allocate key value. + */ + rc = lu_env_refill((struct lu_env *)env); + if (rc != 0) + RETURN(rc); - info = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - LASSERT(info != NULL); + info = lu_context_key_get(&env->le_ctx, &mdt_thread_key); + LASSERT(info != NULL); - obd = class_name2obd(dev); - LASSERT(obd != NULL); + obd = class_name2obd(dev); + LASSERT(obd != NULL); - m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */ + m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */ m->mdt_opts.mo_evict_tgt_nids = 1; - m->mdt_opts.mo_cos = MDT_COS_DEFAULT; - - /* default is coordinator off, it is started through conf_param - * or /proc */ - m->mdt_opts.mo_coordinator = 0; + m->mdt_opts.mo_cos = MDT_COS_DEFAULT; lmi = server_get_mount(dev); - if (lmi == NULL) { - CERROR("Cannot get mount info for %s!\n", dev); - RETURN(-EFAULT); - } else { - lsi = s2lsi(lmi->lmi_sb); - /* CMD is supported only in IAM mode */ - LASSERT(num); - node_id = simple_strtol(num, NULL, 10); + if (lmi == NULL) { + CERROR("Cannot get mount info for %s!\n", dev); + RETURN(-EFAULT); + } else { + lsi = s2lsi(lmi->lmi_sb); + /* CMD is supported only in IAM mode */ + LASSERT(num); + node_id = simple_strtol(num, NULL, 10); obd->u.obt.obt_magic = OBT_MAGIC; if (lsi->lsi_lmd != NULL && lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK) m->mdt_skip_lfsck = 1; } + /* DoM files get IO lock at open by default */ + m->mdt_opts.mo_dom_lock = 1; + m->mdt_squash.rsi_uid = 0; m->mdt_squash.rsi_gid = 0; INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids); init_rwsem(&m->mdt_squash.rsi_sem); spin_lock_init(&m->mdt_lock); - m->mdt_osfs_age = cfs_time_shift_64(-1000); m->mdt_enable_remote_dir = 0; m->mdt_enable_remote_dir_gid = 0; @@ -4836,16 +5105,16 @@ static int mdt_init0(const struct lu_env *env, struct mdt_device *m, s->ld_seq_site = ss_site; ss_site->ss_lu = s; - /* set server index */ + /* set server index */ ss_site->ss_node_id = node_id; /* failover is the default * FIXME: we do not failout mds0/mgs, which may cause some problems. * assumed whose ss_node_id == 0 XXX * */ - obd->obd_replayable = 1; - /* No connection accepted until configurations will finish */ - obd->obd_no_conn = 1; + obd->obd_replayable = 1; + /* No connection accepted until configurations will finish */ + obd->obd_no_conn = 1; if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) { char *str = lustre_cfg_string(cfg, 4); @@ -4865,25 +5134,25 @@ static int mdt_init0(const struct lu_env *env, struct mdt_device *m, snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s", LUSTRE_MDT_NAME, obd->obd_uuid.uuid); - m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name, - LDLM_NAMESPACE_SERVER, - LDLM_NAMESPACE_GREEDY, - LDLM_NS_TYPE_MDT); - if (m->mdt_namespace == NULL) - GOTO(err_fini_seq, rc = -ENOMEM); + m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name, + LDLM_NAMESPACE_SERVER, + LDLM_NAMESPACE_GREEDY, + LDLM_NS_TYPE_MDT); + if (m->mdt_namespace == NULL) + GOTO(err_fini_seq, rc = -ENOMEM); m->mdt_namespace->ns_lvbp = m; m->mdt_namespace->ns_lvbo = &mdt_lvbo; - ldlm_register_intent(m->mdt_namespace, mdt_intent_policy); - /* set obd_namespace for compatibility with old code */ - obd->obd_namespace = m->mdt_namespace; + ldlm_register_intent(m->mdt_namespace, mdt_intent_policy); + /* set obd_namespace for compatibility with old code */ + obd->obd_namespace = m->mdt_namespace; rc = mdt_hsm_cdt_init(m); if (rc != 0) { CERROR("%s: error initializing coordinator, rc %d\n", mdt_obd_name(m), rc); - GOTO(err_free_ns, rc); + GOTO(err_free_ns, rc); } rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice, @@ -4892,31 +5161,37 @@ static int mdt_init0(const struct lu_env *env, struct mdt_device *m, if (rc) GOTO(err_free_hsm, rc); + /* Amount of available space excluded from granting and reserved + * for metadata. It is in percentage and 50% is default value. */ + tgd->tgd_reserved_pcnt = 50; + + if (ONE_MB_BRW_SIZE < (1U << tgd->tgd_blockbits)) + m->mdt_brw_size = 1U << tgd->tgd_blockbits; + else + m->mdt_brw_size = ONE_MB_BRW_SIZE; + rc = mdt_fs_setup(env, m, obd, lsi); if (rc) GOTO(err_tgt, rc); tgt_adapt_sptlrpc_conf(&m->mdt_lut); - next = m->mdt_child; - rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0, - &mntopts); - if (rc) - GOTO(err_fs_cleanup, rc); + next = m->mdt_child; + dt_conf = next->md_ops->mdo_dtconf_get(env, next); - if (mntopts & MNTOPT_USERXATTR) - m->mdt_opts.mo_user_xattr = 1; - else - m->mdt_opts.mo_user_xattr = 0; + mntopts = dt_conf->ddp_mntopts; - rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size); - if (rc) - GOTO(err_fs_cleanup, rc); + if (mntopts & MNTOPT_USERXATTR) + m->mdt_opts.mo_user_xattr = 1; + else + m->mdt_opts.mo_user_xattr = 0; - if (mntopts & MNTOPT_ACL) - m->mdt_opts.mo_acl = 1; - else - m->mdt_opts.mo_acl = 0; + m->mdt_max_ea_size = dt_conf->ddp_max_ea_size; + + if (mntopts & MNTOPT_ACL) + m->mdt_opts.mo_acl = 1; + else + m->mdt_opts.mo_acl = 0; /* XXX: to support suppgid for ACL, we enable identity_upcall * by default, otherwise, maybe got unexpected -EACCESS. */ @@ -4932,11 +5207,11 @@ static int mdt_init0(const struct lu_env *env, struct mdt_device *m, GOTO(err_fs_cleanup, rc); } - rc = mdt_procfs_init(m, dev); - if (rc) { - CERROR("Can't init MDT lprocfs, rc %d\n", rc); - GOTO(err_recovery, rc); - } + rc = mdt_procfs_init(m, dev); + if (rc) { + CERROR("Can't init MDT lprocfs, rc %d\n", rc); + GOTO(err_recovery, rc); + } rc = mdt_quota_init(env, m, cfg); if (rc) @@ -4952,13 +5227,13 @@ static int mdt_init0(const struct lu_env *env, struct mdt_device *m, * when the whole stack is complete and ready * to serve the requests */ - /* Reduce the initial timeout on an MDS because it doesn't need such - * a long timeout as an OST does. Adaptive timeouts will adjust this - * value appropriately. */ - if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT) - ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT; + /* Reduce the initial timeout on an MDS because it doesn't need such + * a long timeout as an OST does. Adaptive timeouts will adjust this + * value appropriately. */ + if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT) + ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT; - RETURN(0); + RETURN(0); err_procfs: mdt_procfs_fini(m); err_recovery: @@ -5058,9 +5333,9 @@ static int mdt_process_config(const struct lu_env *env, cfg); } - if (old_cfg != NULL) - lustre_cfg_free(cfg); - + if (old_cfg) + OBD_FREE(cfg, lustre_cfg_len(cfg->lcfg_bufcount, + cfg->lcfg_buflens)); break; } default: @@ -5092,6 +5367,7 @@ static struct lu_object *mdt_object_alloc(const struct lu_env *env, o->lo_ops = &mdt_obj_ops; spin_lock_init(&mo->mot_write_lock); mutex_init(&mo->mot_lov_mutex); + init_rwsem(&mo->mot_dom_sem); init_rwsem(&mo->mot_open_sem); atomic_set(&mo->mot_open_count, 0); RETURN(o); @@ -5260,9 +5536,10 @@ static int mdt_obd_set_info_async(const struct lu_env *env, * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size * \retval -EBADE client and server feature requirements are incompatible */ -static int mdt_connect_internal(struct obd_export *exp, +static int mdt_connect_internal(const struct lu_env *env, + struct obd_export *exp, struct mdt_device *mdt, - struct obd_connect_data *data) + struct obd_connect_data *data, bool reconnect) { LASSERT(data != NULL); @@ -5294,7 +5571,8 @@ static int mdt_connect_internal(struct obd_export *exp, data->ocd_connect_flags &= ~OBD_CONNECT_XATTR; if (OCD_HAS_FLAG(data, BRW_SIZE)) { - data->ocd_brw_size = min(data->ocd_brw_size, MD_MAX_BRW_SIZE); + data->ocd_brw_size = min(data->ocd_brw_size, + mdt->mdt_brw_size); if (data->ocd_brw_size == 0) { CERROR("%s: cli %s/%p ocd_connect_flags: %#llx " "ocd_version: %x ocd_grant: %d ocd_index: %u " @@ -5308,6 +5586,30 @@ static int mdt_connect_internal(struct obd_export *exp, } } + if (OCD_HAS_FLAG(data, GRANT_PARAM)) { + struct dt_device_param *ddp = &mdt->mdt_lut.lut_dt_conf; + + /* client is reporting its page size, for future use */ + exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits; + data->ocd_grant_blkbits = mdt->mdt_lut.lut_tgd.tgd_blockbits; + /* ddp_inodespace may not be power-of-two value, eg. for ldiskfs + * it's LDISKFS_DIR_REC_LEN(20) = 28. */ + data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1); + /* ocd_grant_tax_kb is in 1K byte blocks */ + data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10; + data->ocd_grant_max_blks = ddp->ddp_max_extent_blks; + } + + if (OCD_HAS_FLAG(data, GRANT)) { + /* Save connect_data we have so far because tgt_grant_connect() + * uses it to calculate grant. */ + exp->exp_connect_data = *data; + tgt_grant_connect(env, exp, data, !reconnect); + } + + if (OCD_HAS_FLAG(data, MAXBYTES)) + data->ocd_maxbytes = mdt->mdt_lut.lut_dt_conf.ddp_maxbytes; + /* NB: Disregard the rule against updating * exp_connect_data.ocd_connect_flags in this case, since * tgt_client_new() needs to know if this is a lightweight @@ -5351,6 +5653,32 @@ static int mdt_connect_internal(struct obd_export *exp, spin_unlock(&exp->exp_lock); } + if (OCD_HAS_FLAG(data, CKSUM)) { + __u32 cksum_types = data->ocd_cksum_types; + + /* The client set in ocd_cksum_types the checksum types it + * supports. We have to mask off the algorithms that we don't + * support */ + data->ocd_cksum_types &= cksum_types_supported_server(); + + if (unlikely(data->ocd_cksum_types == 0)) { + CERROR("%s: Connect with checksum support but no " + "ocd_cksum_types is set\n", + exp->exp_obd->obd_name); + RETURN(-EPROTO); + } + + CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return " + "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp), + cksum_types, data->ocd_cksum_types); + } else { + /* This client does not support OBD_CONNECT_CKSUM + * fall back to CRC32 */ + CDEBUG(D_RPCTRACE, "%s: cli %s does not support " + "OBD_CONNECT_CKSUM, CRC32 will be used\n", + exp->exp_obd->obd_name, obd_export_nid2str(exp)); + } + return 0; } @@ -5475,11 +5803,15 @@ static inline void mdt_disable_slc(struct mdt_device *mdt) static int mdt_obd_disconnect(struct obd_export *exp) { - int rc; - ENTRY; + int rc; - LASSERT(exp); - class_export_get(exp); + ENTRY; + + LASSERT(exp); + class_export_get(exp); + + if (!(exp->exp_flags & OBD_OPT_FORCE)) + tgt_grant_sanity_check(exp->exp_obd, __func__); if ((exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) && !(exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)) { @@ -5493,6 +5825,8 @@ static int mdt_obd_disconnect(struct obd_export *exp) if (rc != 0) CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc); + tgt_grant_discard(exp); + rc = mdt_export_cleanup(exp); nodemap_del_member(exp); class_export_put(exp); @@ -5554,7 +5888,7 @@ static int mdt_obd_connect(const struct lu_env *env, if (rc != 0 && rc != -EEXIST) GOTO(out, rc); - rc = mdt_connect_internal(lexp, mdt, data); + rc = mdt_connect_internal(env, lexp, mdt, data, false); if (rc == 0) { struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd; @@ -5600,7 +5934,8 @@ static int mdt_obd_reconnect(const struct lu_env *env, if (rc != 0 && rc != -EEXIST) RETURN(rc); - rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data); + rc = mdt_connect_internal(env, exp, mdt_dev(obd->obd_lu_dev), data, + true); if (rc == 0) mdt_export_stats_init(obd, exp, localdata); else @@ -5662,6 +5997,17 @@ static int mdt_destroy_export(struct obd_export *exp) LASSERT(list_empty(&exp->exp_outstanding_replies)); LASSERT(list_empty(&exp->exp_mdt_data.med_open_head)); + /* + * discard grants once we're sure no more + * interaction with the client is possible + */ + tgt_grant_discard(exp); + if (exp_connect_flags(exp) & OBD_CONNECT_GRANT) + exp->exp_obd->u.obt.obt_lut->lut_tgd.tgd_tot_granted_clients--; + + if (!(exp->exp_flags & OBD_OPT_FORCE)) + tgt_grant_sanity_check(exp->exp_obd, __func__); + RETURN(0); } @@ -6227,6 +6573,9 @@ static struct obd_ops mdt_obd_device_ops = { .o_destroy_export = mdt_destroy_export, .o_iocontrol = mdt_iocontrol, .o_postrecov = mdt_obd_postrecov, + /* Data-on-MDT IO methods */ + .o_preprw = mdt_obd_preprw, + .o_commitrw = mdt_obd_commitrw, }; static struct lu_device* mdt_device_fini(const struct lu_env *env, @@ -6287,6 +6636,13 @@ static void mdt_key_fini(const struct lu_context *ctx, info->mti_big_lmm = NULL; info->mti_big_lmmsize = 0; } + + if (info->mti_big_acl) { + OBD_FREE_LARGE(info->mti_big_acl, info->mti_big_aclsize); + info->mti_big_acl = NULL; + info->mti_big_aclsize = 0; + } + OBD_FREE_PTR(info); }