X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fost%2Fost_handler.c;h=faf551da8cc8df020d8fc14c674257868f1a8657;hp=bc414e64efdf72785011eff50a6243a860608496;hb=3dcf18d393b990e143fbb11fba2ec6137b7b8b43;hpb=112a9a630f3f89b10f77a67df35233cfc4e23686 diff --git a/lustre/ost/ost_handler.c b/lustre/ost/ost_handler.c index bc414e6..faf551d 100644 --- a/lustre/ost/ost_handler.c +++ b/lustre/ost/ost_handler.c @@ -16,8 +16,8 @@ * in the LICENSE file that accompanied this code). * * You should have received a copy of the GNU General Public License - * version 2 along with this program; If not, see [sun.com URL with a - * copy of GPLv2]. + * version 2 along with this program; If not, see + * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf * * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, * CA 95054 USA or visit www.sun.com if you need additional information or @@ -69,6 +69,18 @@ static int oss_num_create_threads; CFS_MODULE_PARM(oss_num_create_threads, "i", int, 0444, "number of OSS create threads to start"); +/** + * Do not return server-side uid/gid to remote client + */ +static void ost_drop_id(struct obd_export *exp, struct obdo *oa) +{ + if (exp_connect_rmtclient(exp)) { + oa->o_uid = -1; + oa->o_gid = -1; + oa->o_valid &= ~(OBD_MD_FLUID | OBD_MD_FLGID); + } +} + void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req) { struct oti_req_ack_lock *ack_lock; @@ -86,7 +98,7 @@ void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req) if (!ack_lock->mode) break; /* XXX not even calling target_send_reply in some cases... */ - ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode); + ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode, 0); } } @@ -95,6 +107,7 @@ static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req, { struct ost_body *body, *repbody; __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) }; + struct lustre_capa *capa = NULL; int rc; ENTRY; @@ -114,17 +127,20 @@ static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req, RETURN (-EFAULT); ldlm_request_cancel(req, dlm, 0); } - + + if (body->oa.o_valid & OBD_MD_FLOSSCAPA) + capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 2); + rc = lustre_pack_reply(req, 2, size, NULL); if (rc) RETURN(rc); if (body->oa.o_valid & OBD_MD_FLCOOKIE) - oti->oti_logcookies = obdo_logcookie(&body->oa); + oti->oti_logcookies = &body->oa.o_lcookie; repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repbody)); memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); - req->rq_status = obd_destroy(exp, &body->oa, NULL, oti, NULL); + req->rq_status = obd_destroy(exp, &body->oa, NULL, oti, NULL, capa); RETURN(0); } @@ -154,6 +170,7 @@ static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req) oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 1); req->rq_status = obd_getattr(exp, &oinfo); + ost_drop_id(exp, &repbody->oa); RETURN(0); } @@ -200,7 +217,8 @@ static int ost_create(struct obd_export *exp, struct ptlrpc_request *req, repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repbody)); memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); - oti->oti_logcookies = obdo_logcookie(&repbody->oa); + oti->oti_logcookies = &repbody->oa.o_lcookie; + req->rq_status = obd_create(exp, &repbody->oa, NULL, oti); //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0); RETURN(0); @@ -253,7 +271,8 @@ static int ost_punch_lock_get(struct obd_export *exp, struct obdo *oa, RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id, LDLM_EXTENT, &policy, LCK_PW, &flags, ldlm_blocking_ast, ldlm_completion_ast, - ldlm_glimpse_ast, NULL, 0, NULL, lh)); + ldlm_glimpse_ast, NULL, 0, NULL, + NULL, lh)); } /* @@ -282,10 +301,9 @@ static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req, /* check that we do support OBD_CONNECT_TRUNCLOCK. */ CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK); - body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), - lustre_swab_ost_body); - if (body == NULL) - RETURN(-EFAULT); + /* ost_body is varified and swabbed in ost_hpreq_handler() */ + body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body)); + LASSERT(body != NULL); oinfo.oi_oa = &body->oa; oinfo.oi_policy.l_extent.start = oinfo.oi_oa->o_size; @@ -293,7 +311,7 @@ static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req, if ((oinfo.oi_oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) != (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) - RETURN(-EINVAL); + RETURN(-EPROTO); rc = lustre_pack_reply(req, 2, size, NULL); if (rc) @@ -319,6 +337,7 @@ static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req, ost_punch_lock_put(exp, oinfo.oi_oa, &lh); } repbody->oa = *oinfo.oi_oa; + ost_drop_id(exp, &repbody->oa); RETURN(rc); } @@ -347,6 +366,7 @@ static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req) memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); req->rq_status = obd_sync(exp, &repbody->oa, NULL, repbody->oa.o_size, repbody->oa.o_blocks, capa); + ost_drop_id(exp, &repbody->oa); RETURN(0); } @@ -377,6 +397,7 @@ static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req, oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 1); req->rq_status = obd_setattr(exp, &oinfo, oti); + ost_drop_id(exp, &repbody->oa); RETURN(0); } @@ -389,101 +410,6 @@ static int ost_bulk_timeout(void *data) RETURN(1); } -static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo, - struct niobuf_remote *rnb, int nrnb, - struct niobuf_remote **pp_rnbp) -{ - /* Copy a remote niobuf, splitting it into page-sized chunks - * and setting ioo[i].ioo_bufcnt accordingly */ - struct niobuf_remote *pp_rnb; - int i; - int j; - int page; - int rnbidx = 0; - int npages = 0; - - /* - * array of sufficient size already preallocated by caller - */ - LASSERT(pp_rnbp != NULL); - LASSERT(*pp_rnbp != NULL); - - /* first count and check the number of pages required */ - for (i = 0; i < nioo; i++) - for (j = 0; j < ioo->ioo_bufcnt; j++, rnbidx++) { - obd_off offset = rnb[rnbidx].offset; - obd_off p0 = offset >> CFS_PAGE_SHIFT; - obd_off pn = (offset + rnb[rnbidx].len - 1) >> - CFS_PAGE_SHIFT; - - LASSERT(rnbidx < nrnb); - - npages += (pn + 1 - p0); - - if (rnb[rnbidx].len == 0) { - CERROR("zero len BRW: obj %d objid "LPX64 - " buf %u\n", i, ioo[i].ioo_id, j); - return -EINVAL; - } - if (j > 0 && - rnb[rnbidx].offset <= rnb[rnbidx-1].offset) { - CERROR("unordered BRW: obj %d objid "LPX64 - " buf %u offset "LPX64" <= "LPX64"\n", - i, ioo[i].ioo_id, j, rnb[rnbidx].offset, - rnb[rnbidx].offset); - return -EINVAL; - } - } - - LASSERT(rnbidx == nrnb); - - if (npages == nrnb) { /* all niobufs are for single pages */ - *pp_rnbp = rnb; - return npages; - } - - pp_rnb = *pp_rnbp; - - /* now do the actual split */ - page = rnbidx = 0; - for (i = 0; i < nioo; i++) { - int obj_pages = 0; - - for (j = 0; j < ioo[i].ioo_bufcnt; j++, rnbidx++) { - obd_off off = rnb[rnbidx].offset; - int nob = rnb[rnbidx].len; - - LASSERT(rnbidx < nrnb); - do { - obd_off poff = off & ~CFS_PAGE_MASK; - int pnob = (poff + nob > CFS_PAGE_SIZE) ? - CFS_PAGE_SIZE - poff : nob; - - LASSERT(page < npages); - pp_rnb[page].len = pnob; - pp_rnb[page].offset = off; - pp_rnb[page].flags = rnb[rnbidx].flags; - - CDEBUG(0, " obj %d id "LPX64 - "page %d(%d) "LPX64" for %d, flg %x\n", - i, ioo[i].ioo_id, obj_pages, page, - pp_rnb[page].offset, pp_rnb[page].len, - pp_rnb[page].flags); - page++; - obj_pages++; - - off += pnob; - nob -= pnob; - } while (nob > 0); - LASSERT(nob == 0); - } - ioo[i].ioo_bufcnt = obj_pages; - } - LASSERT(page == npages); - - return npages; -} - static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc, int opc, cksum_type_t cksum_type) { @@ -506,62 +432,17 @@ static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc, int opc, /* corrupt the data after we compute the checksum, to * simulate an OST->client data error */ if (i == 0 && opc == OST_READ && - OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) + OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) { memcpy(ptr, "bad4", min(4, len)); + /* nobody should use corrupted page again */ + ClearPageUptodate(page); + } kunmap(page); } return cksum; } -/* - * populate @nio by @nrpages pages from per-thread page pool - */ -static void ost_nio_pages_get(struct ptlrpc_request *req, - struct niobuf_local *nio, int nrpages) -{ - int i; - struct ost_thread_local_cache *tls; - - ENTRY; - - LASSERT(nrpages <= OST_THREAD_POOL_SIZE); - LASSERT(req != NULL); - LASSERT(req->rq_svc_thread != NULL); - - tls = ost_tls(req); - LASSERT(tls != NULL); - - memset(nio, 0, nrpages * sizeof *nio); - for (i = 0; i < nrpages; ++ i) { - struct page *page; - - page = tls->page[i]; - LASSERT(page != NULL); - POISON_PAGE(page, 0xf1); - nio[i].page = page; - LL_CDEBUG_PAGE(D_INFO, page, "%d\n", i); - } - EXIT; -} - -/* - * Dual for ost_nio_pages_get(). Poison pages in pool for debugging - */ -static void ost_nio_pages_put(struct ptlrpc_request *req, - struct niobuf_local *nio, int nrpages) -{ - int i; - - ENTRY; - - LASSERT(nrpages <= OST_THREAD_POOL_SIZE); - - for (i = 0; i < nrpages; ++ i) - POISON_PAGE(nio[i].page, 0xf2); - EXIT; -} - static int ost_brw_lock_get(int mode, struct obd_export *exp, struct obd_ioobj *obj, struct niobuf_remote *nb, struct lustre_handle *lh) @@ -592,7 +473,8 @@ static int ost_brw_lock_get(int mode, struct obd_export *exp, RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id, LDLM_EXTENT, &policy, mode, &flags, ldlm_blocking_ast, ldlm_completion_ast, - ldlm_glimpse_ast, NULL, 0, NULL, lh)); + ldlm_glimpse_ast, NULL, 0, NULL, + NULL, lh)); } static void ost_brw_lock_put(int mode, @@ -611,7 +493,10 @@ static void ost_brw_lock_put(int mode, struct ost_prolong_data { struct obd_export *opd_exp; ldlm_policy_data_t opd_policy; + struct obdo *opd_oa; ldlm_mode_t opd_mode; + int opd_lock_match; + int opd_timeout; }; static int ost_prolong_locks_iter(struct ldlm_lock *lock, void *data) @@ -641,47 +526,93 @@ static int ost_prolong_locks_iter(struct ldlm_lock *lock, void *data) return LDLM_ITER_CONTINUE; } + /* Fill the obdo with the matched lock handle. + * XXX: it is possible in some cases the IO RPC is covered by several + * locks, even for the write case, so it may need to be a lock list. */ + if (opd->opd_oa && !(opd->opd_oa->o_valid & OBD_MD_FLHANDLE)) { + opd->opd_oa->o_handle.cookie = lock->l_handle.h_cookie; + opd->opd_oa->o_valid |= OBD_MD_FLHANDLE; + } + if (!(lock->l_flags & LDLM_FL_AST_SENT)) { /* ignore locks not being cancelled */ return LDLM_ITER_CONTINUE; } + CDEBUG(D_DLMTRACE,"refresh lock: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n", + lock->l_resource->lr_name.name[0], + lock->l_resource->lr_name.name[1], + opd->opd_policy.l_extent.start, opd->opd_policy.l_extent.end); /* OK. this is a possible lock the user holds doing I/O * let's refresh eviction timer for it */ - ldlm_refresh_waiting_lock(lock); + ldlm_refresh_waiting_lock(lock, opd->opd_timeout); + opd->opd_lock_match = 1; return LDLM_ITER_CONTINUE; } -static void ost_prolong_locks(struct obd_export *exp, struct obd_ioobj *obj, - struct niobuf_remote *nb, ldlm_mode_t mode) +static int ost_rw_prolong_locks(struct ptlrpc_request *req, struct obd_ioobj *obj, + struct niobuf_remote *nb, struct obdo *oa, + ldlm_mode_t mode) { struct ldlm_res_id res_id; int nrbufs = obj->ioo_bufcnt; - struct ost_prolong_data opd; + struct ost_prolong_data opd = { 0 }; ENTRY; osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id); opd.opd_mode = mode; - opd.opd_exp = exp; + opd.opd_exp = req->rq_export; opd.opd_policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK; opd.opd_policy.l_extent.end = (nb[nrbufs - 1].offset + nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK; - CDEBUG(D_DLMTRACE,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n", + /* prolong locks for the current service time of the corresponding + * portal (= OST_IO_PORTAL) */ + opd.opd_timeout = AT_OFF ? obd_timeout / 2: + max(at_est2timeout(at_get(&req->rq_rqbd-> + rqbd_service->srv_at_estimate)), ldlm_timeout); + + CDEBUG(D_INFO,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n", res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start, opd.opd_policy.l_extent.end); - ldlm_resource_iterate(exp->exp_obd->obd_namespace, &res_id, + + if (oa->o_valid & OBD_MD_FLHANDLE) { + struct ldlm_lock *lock; + + lock = ldlm_handle2lock(&oa->o_handle); + if (lock != NULL) { + ost_prolong_locks_iter(lock, &opd); + if (opd.opd_lock_match) { + LDLM_LOCK_PUT(lock); + RETURN(1); + } + + /* Check if the lock covers the whole IO region, + * otherwise iterate through the resource. */ + if (lock->l_policy_data.l_extent.end >= + opd.opd_policy.l_extent.end && + lock->l_policy_data.l_extent.start <= + opd.opd_policy.l_extent.start) { + LDLM_LOCK_PUT(lock); + RETURN(0); + } + LDLM_LOCK_PUT(lock); + } + } + + opd.opd_oa = oa; + ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id, ost_prolong_locks_iter, &opd); + RETURN(opd.opd_lock_match); } static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) { - struct ptlrpc_bulk_desc *desc; + struct ptlrpc_bulk_desc *desc = NULL; struct obd_export *exp = req->rq_export; struct niobuf_remote *remote_nb; - struct niobuf_remote *pp_rnb = NULL; struct niobuf_local *local_nb; struct obd_ioobj *ioo; struct ost_body *body, *repbody; @@ -711,38 +642,18 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) if (exp->exp_failed) GOTO(out, rc = -ENOTCONN); - body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), - lustre_swab_ost_body); - if (body == NULL) { - CERROR("Missing/short ost_body\n"); - GOTO(out, rc = -EFAULT); - } + /* ost_body, ioobj & noibuf_remote are verified and swabbed in + * ost_rw_hpreq_check(). */ + body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body)); + LASSERT(body != NULL); - ioo = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*ioo), - lustre_swab_obd_ioobj); - if (ioo == NULL) { - CERROR("Missing/short ioobj\n"); - GOTO(out, rc = -EFAULT); - } + ioo = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, sizeof(*ioo)); + LASSERT(ioo != NULL); niocount = ioo->ioo_bufcnt; - if (niocount > PTLRPC_MAX_BRW_PAGES) { - DEBUG_REQ(D_ERROR, req, "bulk has too many pages (%d)", - niocount); - GOTO(out, rc = -EFAULT); - } - - remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2, - niocount * sizeof(*remote_nb), - lustre_swab_niobuf_remote); - if (remote_nb == NULL) { - CERROR("Missing/short niobuf\n"); - GOTO(out, rc = -EFAULT); - } - if (lustre_msg_swabbed(req->rq_reqmsg)) { /* swab remaining niobufs */ - for (i = 1; i < niocount; i++) - lustre_swab_niobuf_remote (&remote_nb[i]); - } + remote_nb = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, + niocount * sizeof(*remote_nb)); + LASSERT(remote_nb != NULL); if (body->oa.o_valid & OBD_MD_FLOSSCAPA) capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3); @@ -756,28 +667,12 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) * ost_thread_init(). */ local_nb = ost_tls(req)->local; - pp_rnb = ost_tls(req)->remote; - - /* FIXME all niobuf splitting should be done in obdfilter if needed */ - /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */ - npages = get_per_page_niobufs(ioo, 1, remote_nb, niocount, &pp_rnb); - if (npages < 0) - GOTO(out, rc = npages); - - LASSERT(npages <= OST_THREAD_POOL_SIZE); - - ost_nio_pages_get(req, local_nb, npages); - - desc = ptlrpc_prep_bulk_exp(req, npages, - BULK_PUT_SOURCE, OST_BULK_PORTAL); - if (desc == NULL) - GOTO(out, rc = -ENOMEM); - rc = ost_brw_lock_get(LCK_PR, exp, ioo, pp_rnb, &lockh); + rc = ost_brw_lock_get(LCK_PR, exp, ioo, remote_nb, &lockh); if (rc != 0) GOTO(out_bulk, rc); - /* + /* * If getting the lock took more time than * client was willing to wait, drop it. b=11330 */ @@ -792,12 +687,18 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) GOTO(out_lock, rc = -ETIMEDOUT); } - rc = obd_preprw(OBD_BRW_READ, exp, &body->oa, 1, - ioo, npages, pp_rnb, local_nb, oti, capa); + npages = OST_THREAD_POOL_SIZE; + rc = obd_preprw(OBD_BRW_READ, exp, &body->oa, 1, ioo, + remote_nb, &npages, local_nb, oti, capa); if (rc != 0) GOTO(out_lock, rc); - ost_prolong_locks(exp, ioo, pp_rnb, LCK_PW | LCK_PR); + desc = ptlrpc_prep_bulk_exp(req, npages, + BULK_PUT_SOURCE, OST_BULK_PORTAL); + if (desc == NULL) /* XXX: check all cleanup stuff */ + GOTO(out, rc = -ENOMEM); + + ost_rw_prolong_locks(req, ioo, remote_nb, &body->oa, LCK_PW | LCK_PR); nob = 0; for (i = 0; i < npages; i++) { @@ -808,17 +709,15 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) break; } - LASSERTF(page_rc <= pp_rnb[i].len, "page_rc (%d) > " - "pp_rnb[%d].len (%d)\n", page_rc, i, pp_rnb[i].len); nob += page_rc; if (page_rc != 0) { /* some data! */ LASSERT (local_nb[i].page != NULL); ptlrpc_prep_bulk_page(desc, local_nb[i].page, - pp_rnb[i].offset & ~CFS_PAGE_MASK, + local_nb[i].offset & ~CFS_PAGE_MASK, page_rc); } - if (page_rc != pp_rnb[i].len) { /* short read */ + if (page_rc != local_nb[i].len) { /* short read */ /* All subsequent pages should be 0 */ while(++i < npages) LASSERT(local_nb[i].rc == 0); @@ -857,9 +756,9 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) if (exp->exp_failed) rc = -ENOTCONN; else { - sptlrpc_svc_wrap_bulk(req, desc); - - rc = ptlrpc_start_bulk_transfer(desc); + rc = sptlrpc_svc_wrap_bulk(req, desc); + if (rc == 0) + rc = ptlrpc_start_bulk_transfer(desc); } if (rc == 0) { @@ -874,8 +773,10 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) ost_bulk_timeout, desc); rc = l_wait_event(desc->bd_waitq, - !ptlrpc_bulk_active(desc) || - exp->exp_failed, &lwi); + !ptlrpc_server_bulk_active(desc) || + exp->exp_failed || + exp->exp_abort_active_req, + &lwi); LASSERT(rc == 0 || rc == -ETIMEDOUT); /* Wait again if we changed deadline */ } while ((rc == -ETIMEDOUT) && @@ -892,6 +793,11 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) DEBUG_REQ(D_ERROR, req, "Eviction on bulk PUT"); rc = -ENOTCONN; ptlrpc_abort_bulk(desc); + } else if (exp->exp_abort_active_req) { + DEBUG_REQ(D_ERROR, req, "Reconnect on bulk PUT"); + /* we don't reply anyway */ + rc = -ETIMEDOUT; + ptlrpc_abort_bulk(desc); } else if (!desc->bd_success || desc->bd_nob_transferred != desc->bd_nob) { DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)", @@ -909,25 +815,26 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) } /* Must commit after prep above in all cases */ - rc = obd_commitrw(OBD_BRW_READ, exp, &body->oa, 1, - ioo, npages, local_nb, oti, rc); - - ost_nio_pages_put(req, local_nb, npages); + rc = obd_commitrw(OBD_BRW_READ, exp, &body->oa, 1, ioo, + remote_nb, npages, local_nb, oti, rc); if (rc == 0) { repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repbody)); memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa)); + ost_drop_id(exp, &repbody->oa); } out_lock: - ost_brw_lock_put(LCK_PR, ioo, pp_rnb, &lockh); + ost_brw_lock_put(LCK_PR, ioo, remote_nb, &lockh); out_bulk: - ptlrpc_free_bulk(desc); + if (desc) + ptlrpc_free_bulk(desc); out: LASSERT(rc <= 0); if (rc == 0) { req->rq_status = nob; + ptlrpc_lprocfs_brw(req, nob); target_committed_to_req(req); ptlrpc_reply(req); } else if (!no_reply) { @@ -951,10 +858,9 @@ out: static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) { - struct ptlrpc_bulk_desc *desc; + struct ptlrpc_bulk_desc *desc = NULL; struct obd_export *exp = req->rq_export; struct niobuf_remote *remote_nb; - struct niobuf_remote *pp_rnb; struct niobuf_local *local_nb; struct obd_ioobj *ioo; struct ost_body *body, *repbody; @@ -964,10 +870,11 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) __u32 *rcs; __u32 size[3] = { sizeof(struct ptlrpc_body), sizeof(*body) }; int objcount, niocount, npages; - int rc, swab, i, j; + int rc, i, j; obd_count client_cksum = 0, server_cksum = 0; cksum_type_t cksum_type = OBD_CKSUM_CRC32; - int no_reply = 0; + int no_reply = 0; + __u32 o_uid = 0, o_gid = 0; ENTRY; req->rq_bulk_write = 1; @@ -991,56 +898,26 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) if (exp->exp_failed) GOTO(out, rc = -ENOTCONN); - swab = lustre_msg_swabbed(req->rq_reqmsg); - body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), - lustre_swab_ost_body); - if (body == NULL) { - CERROR("Missing/short ost_body\n"); - GOTO(out, rc = -EFAULT); - } + /* ost_body, ioobj & noibuf_remote are verified and swabbed in + * ost_rw_hpreq_check(). */ + body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body)); + LASSERT(body != NULL); - lustre_set_req_swabbed(req, REQ_REC_OFF + 1); - objcount = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1) / - sizeof(*ioo); - if (objcount == 0) { - CERROR("Missing/short ioobj\n"); - GOTO(out, rc = -EFAULT); - } - if (objcount > 1) { - CERROR("too many ioobjs (%d)\n", objcount); - GOTO(out, rc = -EFAULT); - } + if ((body->oa.o_flags & OBD_BRW_MEMALLOC) && + (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)) + libcfs_memory_pressure_set(); + objcount = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1) / + sizeof(*ioo); ioo = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, objcount * sizeof(*ioo)); - LASSERT (ioo != NULL); - for (niocount = i = 0; i < objcount; i++) { - if (swab) - lustre_swab_obd_ioobj(&ioo[i]); - if (ioo[i].ioo_bufcnt == 0) { - CERROR("ioo[%d] has zero bufcnt\n", i); - GOTO(out, rc = -EFAULT); - } + LASSERT(ioo != NULL); + for (niocount = i = 0; i < objcount; i++) niocount += ioo[i].ioo_bufcnt; - } - - if (niocount > PTLRPC_MAX_BRW_PAGES) { - DEBUG_REQ(D_ERROR, req, "bulk has too many pages (%d)", - niocount); - GOTO(out, rc = -EFAULT); - } - remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2, - niocount * sizeof(*remote_nb), - lustre_swab_niobuf_remote); - if (remote_nb == NULL) { - CERROR("Missing/short niobuf\n"); - GOTO(out, rc = -EFAULT); - } - if (swab) { /* swab the remaining niobufs */ - for (i = 1; i < niocount; i++) - lustre_swab_niobuf_remote (&remote_nb[i]); - } + remote_nb = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, + niocount * sizeof(*remote_nb)); + LASSERT(remote_nb != NULL); if (body->oa.o_valid & OBD_MD_FLOSSCAPA) capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3); @@ -1058,28 +935,12 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) * ost_thread_init(). */ local_nb = ost_tls(req)->local; - pp_rnb = ost_tls(req)->remote; - - /* FIXME all niobuf splitting should be done in obdfilter if needed */ - /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */ - npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb); - if (npages < 0) - GOTO(out, rc = npages); - - LASSERT(npages <= OST_THREAD_POOL_SIZE); - ost_nio_pages_get(req, local_nb, npages); - - desc = ptlrpc_prep_bulk_exp(req, npages, - BULK_GET_SINK, OST_BULK_PORTAL); - if (desc == NULL) - GOTO(out, rc = -ENOMEM); - - rc = ost_brw_lock_get(LCK_PW, exp, ioo, pp_rnb, &lockh); + rc = ost_brw_lock_get(LCK_PW, exp, ioo, remote_nb, &lockh); if (rc != 0) GOTO(out_bulk, rc); - /* + /* * If getting the lock took more time than * client was willing to wait, drop it. b=11330 */ @@ -1094,7 +955,7 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) GOTO(out_lock, rc = -ETIMEDOUT); } - ost_prolong_locks(exp, ioo, pp_rnb, LCK_PW); + ost_rw_prolong_locks(req, ioo, remote_nb,&body->oa, LCK_PW); /* obd_preprw clobbers oa->valid, so save what we need */ if (body->oa.o_valid & OBD_MD_FLCKSUM) { @@ -1102,33 +963,47 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) if (body->oa.o_valid & OBD_MD_FLFLAGS) cksum_type = cksum_type_unpack(body->oa.o_flags); } - + /* Because we already sync grant info with client when reconnect, - * grant info will be cleared for resent req, then fed_grant and - * total_grant will not be modified in following preprw_write */ + * grant info will be cleared for resent req, then fed_grant and + * total_grant will not be modified in following preprw_write */ if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) { DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info"); body->oa.o_valid &= ~OBD_MD_FLGRANT; } + if (exp_connect_rmtclient(exp)) { + o_uid = body->oa.o_uid; + o_gid = body->oa.o_gid; + } + npages = OST_THREAD_POOL_SIZE; rc = obd_preprw(OBD_BRW_WRITE, exp, &body->oa, objcount, - ioo, npages, pp_rnb, local_nb, oti, capa); + ioo, remote_nb, &npages, local_nb, oti, capa); if (rc != 0) GOTO(out_lock, rc); + desc = ptlrpc_prep_bulk_exp(req, npages, + BULK_GET_SINK, OST_BULK_PORTAL); + if (desc == NULL) + GOTO(out, rc = -ENOMEM); + /* NB Having prepped, we must commit... */ for (i = 0; i < npages; i++) ptlrpc_prep_bulk_page(desc, local_nb[i].page, - pp_rnb[i].offset & ~CFS_PAGE_MASK, - pp_rnb[i].len); + local_nb[i].offset & ~CFS_PAGE_MASK, + local_nb[i].len); + + rc = sptlrpc_svc_prep_bulk(req, desc); + if (rc != 0) + GOTO(out_lock, rc); /* Check if client was evicted while we were doing i/o before touching network */ if (desc->bd_export->exp_failed) rc = -ENOTCONN; else - rc = ptlrpc_start_bulk_transfer (desc); + rc = ptlrpc_start_bulk_transfer(desc); if (rc == 0) { time_t start = cfs_time_current_sec(); do { @@ -1139,8 +1014,10 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1), ost_bulk_timeout, desc); rc = l_wait_event(desc->bd_waitq, - !ptlrpc_bulk_active(desc) || - desc->bd_export->exp_failed, &lwi); + !ptlrpc_server_bulk_active(desc) || + desc->bd_export->exp_failed || + desc->bd_export->exp_abort_active_req, + &lwi); LASSERT(rc == 0 || rc == -ETIMEDOUT); /* Wait again if we changed deadline */ } while ((rc == -ETIMEDOUT) && @@ -1157,23 +1034,23 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) DEBUG_REQ(D_ERROR, req, "Eviction on bulk GET"); rc = -ENOTCONN; ptlrpc_abort_bulk(desc); - } else if (!desc->bd_success || - desc->bd_nob_transferred != desc->bd_nob) { - DEBUG_REQ(D_ERROR, req, "%s bulk GET %d(%d)", - desc->bd_success ? - "truncated" : "network error on", - desc->bd_nob_transferred, desc->bd_nob); + } else if (desc->bd_export->exp_abort_active_req) { + DEBUG_REQ(D_ERROR, req, "Reconnect on bulk GET"); + /* we don't reply anyway */ + rc = -ETIMEDOUT; + ptlrpc_abort_bulk(desc); + } else if (!desc->bd_success) { + DEBUG_REQ(D_ERROR, req, "network error on bulk GET"); /* XXX should this be a different errno? */ rc = -ETIMEDOUT; + } else { + rc = sptlrpc_svc_unwrap_bulk(req, desc); } } else { DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d", rc); } no_reply = rc != 0; - if (rc == 0) - sptlrpc_svc_unwrap_bulk(req, desc); - repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repbody)); memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa)); @@ -1198,8 +1075,12 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) } /* Must commit after prep above in all cases */ - rc = obd_commitrw(OBD_BRW_WRITE, exp, &repbody->oa, - objcount, ioo, npages, local_nb, oti, rc); + rc = obd_commitrw(OBD_BRW_WRITE, exp, &repbody->oa, objcount, ioo, + remote_nb, npages, local_nb, oti, rc); + if (exp_connect_rmtclient(exp)) { + repbody->oa.o_uid = o_uid; + repbody->oa.o_gid = o_gid; + } if (unlikely(client_cksum != server_cksum && rc == 0)) { int new_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type); @@ -1220,7 +1101,7 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) via = " via "; router = libcfs_nid2str(desc->bd_sender); } - + LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: %s from " "%s%s%s inum "LPU64"/"LPU64" object " LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n", @@ -1234,38 +1115,41 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) body->oa.o_id, body->oa.o_valid & OBD_MD_FLGROUP ? body->oa.o_gr : (__u64)0, - pp_rnb[0].offset, - pp_rnb[npages-1].offset+pp_rnb[npages-1].len - - 1 ); + local_nb[0].offset, + local_nb[npages-1].offset + + local_nb[npages-1].len - 1 ); CERROR("client csum %x, original server csum %x, " "server csum now %x\n", client_cksum, server_cksum, new_cksum); } - ost_nio_pages_put(req, local_nb, npages); - if (rc == 0) { + int nob = 0; + /* set per-requested niobuf return codes */ for (i = j = 0; i < niocount; i++) { - int nob = remote_nb[i].len; + int len = remote_nb[i].len; + nob += len; rcs[i] = 0; do { LASSERT(j < npages); if (local_nb[j].rc < 0) rcs[i] = local_nb[j].rc; - nob -= pp_rnb[j].len; + len -= local_nb[j].len; j++; - } while (nob > 0); - LASSERT(nob == 0); + } while (len > 0); + LASSERT(len == 0); } LASSERT(j == npages); + ptlrpc_lprocfs_brw(req, nob); } out_lock: - ost_brw_lock_put(LCK_PW, ioo, pp_rnb, &lockh); + ost_brw_lock_put(LCK_PW, ioo, remote_nb, &lockh); out_bulk: - ptlrpc_free_bulk(desc); + if (desc) + ptlrpc_free_bulk(desc); out: if (rc == 0) { oti_to_request(oti, req); @@ -1286,11 +1170,14 @@ out: exp->exp_connection->c_remote_uuid.uuid, libcfs_id2str(req->rq_peer)); } + libcfs_memory_pressure_clr(); RETURN(rc); } static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req) { + struct ost_body *body = NULL, *repbody; + __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) }; char *key, *val = NULL; int keylen, vallen, rc = 0; ENTRY; @@ -1302,19 +1189,41 @@ static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req) } keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF); - rc = lustre_pack_reply(req, 1, NULL, NULL); - if (rc) - RETURN(rc); + if (KEY_IS(KEY_GRANT_SHRINK)) { + rc = lustre_pack_reply(req, 2, size, NULL); + if (rc) + RETURN(rc); + } else { + rc = lustre_pack_reply(req, 1, NULL, NULL); + if (rc) + RETURN(rc); + } vallen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1); - if (vallen) - val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, 0); + if (vallen) { + if (KEY_IS(KEY_GRANT_SHRINK)) { + body = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, + sizeof(*body), + lustre_swab_ost_body); + if (!body) + RETURN(-EFAULT); + + repbody = lustre_msg_buf(req->rq_repmsg, + REPLY_REC_OFF, + sizeof(*repbody)); + memcpy(repbody, body, sizeof(*body)); + val = (char*)repbody; + } else + val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1,0); + } if (KEY_IS(KEY_EVICT_BY_NID)) { if (val && vallen) obd_export_evict_by_nid(exp->exp_obd, val); - GOTO(out, rc = 0); + } else if (KEY_IS(KEY_MDS_CONN) && lustre_msg_swabbed(req->rq_reqmsg)) { + /* Val's are not swabbed automatically */ + __swab32s((__u32 *)val); } rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL); @@ -1362,26 +1271,25 @@ static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req) RETURN(rc); } +#ifdef HAVE_QUOTA_SUPPORT static int ost_handle_quotactl(struct ptlrpc_request *req) { struct obd_quotactl *oqctl, *repoqc; - __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) }; int rc; ENTRY; - oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl), - lustre_swab_obd_quotactl); + oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL); if (oqctl == NULL) GOTO(out, rc = -EPROTO); - rc = lustre_pack_reply(req, 2, size, NULL); + rc = req_capsule_server_pack(&req->rq_pill); if (rc) GOTO(out, rc); - repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc)); - + repoqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL); req->rq_status = obd_quotactl(req->rq_export, oqctl); *repoqc = *oqctl; + out: RETURN(rc); } @@ -1397,15 +1305,38 @@ static int ost_handle_quotacheck(struct ptlrpc_request *req) RETURN(-EPROTO); rc = req_capsule_server_pack(&req->rq_pill); - if (rc) { - CERROR("ost: out of memory while packing quotacheck reply\n"); + if (rc) RETURN(-ENOMEM); - } req->rq_status = obd_quotacheck(req->rq_export, oqctl); RETURN(0); } +static int ost_handle_quota_adjust_qunit(struct ptlrpc_request *req) +{ + struct quota_adjust_qunit *oqaq, *repoqa; + struct lustre_quota_ctxt *qctxt; + int rc; + ENTRY; + + qctxt = &req->rq_export->exp_obd->u.obt.obt_qctxt; + oqaq = req_capsule_client_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT); + if (oqaq == NULL) + GOTO(out, rc = -EPROTO); + + rc = req_capsule_server_pack(&req->rq_pill); + if (rc) + GOTO(out, rc); + + repoqa = req_capsule_server_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT); + req->rq_status = obd_quota_adjust_qunit(req->rq_export, oqaq, qctxt); + *repoqa = *oqaq; + + out: + RETURN(rc); +} +#endif + static int ost_llog_handle_connect(struct obd_export *exp, struct ptlrpc_request *req) { @@ -1418,41 +1349,177 @@ static int ost_llog_handle_connect(struct obd_export *exp, RETURN(rc); } -static int filter_export_check_flavor(struct filter_obd *filter, - struct obd_export *exp, - struct ptlrpc_request *req) +#define ost_init_sec_none(reply, exp) \ +do { \ + reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | \ + OBD_CONNECT_RMT_CLIENT_FORCE | \ + OBD_CONNECT_OSS_CAPA); \ + spin_lock(&exp->exp_lock); \ + exp->exp_connect_flags = reply->ocd_connect_flags; \ + spin_unlock(&exp->exp_lock); \ +} while (0) + +static int ost_init_sec_level(struct ptlrpc_request *req) { - int rc = 0; - - /* FIXME - * this should be done in filter_connect()/filter_reconnect(), but - * we can't obtain information like NID, which stored in incoming - * request, thus can't decide what flavor to use. so we do it here. - * - * This hack should be removed after the OST stack be rewritten, just - * like what we are doing in mdt_obd_connect()/mdt_obd_reconnect(). - */ - if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_INVALID) - return 0; + struct obd_export *exp = req->rq_export; + struct req_capsule *pill = &req->rq_pill; + struct obd_device *obd = exp->exp_obd; + struct filter_obd *filter = &obd->u.filter; + char *client = libcfs_nid2str(req->rq_peer.nid); + struct obd_connect_data *data, *reply; + int rc = 0, remote; + ENTRY; + + data = req_capsule_client_get(pill, &RMF_CONNECT_DATA); + reply = req_capsule_server_get(pill, &RMF_CONNECT_DATA); + if (data == NULL || reply == NULL) + RETURN(-EFAULT); + + /* connection from MDT is always trusted */ + if (req->rq_auth_usr_mdt) { + ost_init_sec_none(reply, exp); + RETURN(0); + } + + /* no GSS support case */ + if (!req->rq_auth_gss) { + if (filter->fo_sec_level > LUSTRE_SEC_NONE) { + CWARN("client %s -> target %s does not user GSS, " + "can not run under security level %d.\n", + client, obd->obd_name, filter->fo_sec_level); + RETURN(-EACCES); + } else { + ost_init_sec_none(reply, exp); + RETURN(0); + } + } + + /* old version case */ + if (unlikely(!(data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT) || + !(data->ocd_connect_flags & OBD_CONNECT_OSS_CAPA))) { + if (filter->fo_sec_level > LUSTRE_SEC_NONE) { + CWARN("client %s -> target %s uses old version, " + "can not run under security level %d.\n", + client, obd->obd_name, filter->fo_sec_level); + RETURN(-EACCES); + } else { + CWARN("client %s -> target %s uses old version, " + "run under security level %d.\n", + client, obd->obd_name, filter->fo_sec_level); + ost_init_sec_none(reply, exp); + RETURN(0); + } + } + + remote = data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT_FORCE; + if (remote) { + if (!req->rq_auth_remote) + CDEBUG(D_SEC, "client (local realm) %s -> target %s " + "asked to be remote.\n", client, obd->obd_name); + } else if (req->rq_auth_remote) { + remote = 1; + CDEBUG(D_SEC, "client (remote realm) %s -> target %s is set " + "as remote by default.\n", client, obd->obd_name); + } + + if (remote) { + if (!filter->fo_fl_oss_capa) { + CDEBUG(D_SEC, "client %s -> target %s is set as remote," + " but OSS capabilities are not enabled: %d.\n", + client, obd->obd_name, filter->fo_fl_oss_capa); + RETURN(-EACCES); + } + } + + switch (filter->fo_sec_level) { + case LUSTRE_SEC_NONE: + if (!remote) { + ost_init_sec_none(reply, exp); + break; + } else { + CDEBUG(D_SEC, "client %s -> target %s is set as remote, " + "can not run under security level %d.\n", + client, obd->obd_name, filter->fo_sec_level); + RETURN(-EACCES); + } + case LUSTRE_SEC_REMOTE: + if (!remote) + ost_init_sec_none(reply, exp); + break; + case LUSTRE_SEC_ALL: + if (!remote) { + reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | + OBD_CONNECT_RMT_CLIENT_FORCE); + if (!filter->fo_fl_oss_capa) + reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA; + + spin_lock(&exp->exp_lock); + exp->exp_connect_flags = reply->ocd_connect_flags; + spin_unlock(&exp->exp_lock); + } + break; + default: + RETURN(-EINVAL); + } + + RETURN(rc); +} - CDEBUG(D_SEC, "from %s\n", sptlrpc_part2name(req->rq_sp_from)); - spin_lock(&exp->exp_lock); - exp->exp_sp_peer = req->rq_sp_from; - - read_lock(&filter->fo_sptlrpc_lock); - sptlrpc_rule_set_choose(&filter->fo_sptlrpc_rset, exp->exp_sp_peer, - req->rq_peer.nid, &exp->exp_flvr); - read_unlock(&filter->fo_sptlrpc_lock); - - if (exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) { - CERROR("invalid rpc flavor %x, expect %x, from %s\n", - req->rq_flvr.sf_rpc, exp->exp_flvr.sf_rpc, - libcfs_nid2str(req->rq_peer.nid)); - exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_INVALID; - rc = -EACCES; +/* + * FIXME + * this should be done in filter_connect()/filter_reconnect(), but + * we can't obtain information like NID, which stored in incoming + * request, thus can't decide what flavor to use. so we do it here. + * + * This hack should be removed after the OST stack be rewritten, just + * like what we are doing in mdt_obd_connect()/mdt_obd_reconnect(). + */ +static int ost_connect_check_sptlrpc(struct ptlrpc_request *req) +{ + struct obd_export *exp = req->rq_export; + struct filter_obd *filter = &exp->exp_obd->u.filter; + struct sptlrpc_flavor flvr; + int rc = 0; + + if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name, + LUSTRE_ECHO_NAME) == 0)) { + exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY; + return 0; } - spin_unlock(&exp->exp_lock); + if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) { + read_lock(&filter->fo_sptlrpc_lock); + sptlrpc_target_choose_flavor(&filter->fo_sptlrpc_rset, + req->rq_sp_from, + req->rq_peer.nid, + &flvr); + read_unlock(&filter->fo_sptlrpc_lock); + + spin_lock(&exp->exp_lock); + + exp->exp_sp_peer = req->rq_sp_from; + exp->exp_flvr = flvr; + + if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY && + exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) { + CERROR("unauthorized rpc flavor %x from %s, " + "expect %x\n", req->rq_flvr.sf_rpc, + libcfs_nid2str(req->rq_peer.nid), + exp->exp_flvr.sf_rpc); + rc = -EACCES; + } + + spin_unlock(&exp->exp_lock); + } else { + if (exp->exp_sp_peer != req->rq_sp_from) { + CERROR("RPC source %s doesn't match %s\n", + sptlrpc_part2name(req->rq_sp_from), + sptlrpc_part2name(exp->exp_sp_peer)); + rc = -EACCES; + } else { + rc = sptlrpc_target_export_check(exp, req); + } + } return rc; } @@ -1514,8 +1581,11 @@ int ost_msg_check_version(struct lustre_msg *msg) case OST_SYNC: case OST_SET_INFO: case OST_GET_INFO: +#ifdef HAVE_QUOTA_SUPPORT case OST_QUOTACHECK: case OST_QUOTACTL: + case OST_QUOTA_ADJUST_QUNIT: +#endif rc = lustre_msg_check_version(msg, LUSTRE_OST_VERSION); if (rc) CERROR("bad opc %u version %08x, expecting %08x\n", @@ -1551,6 +1621,261 @@ int ost_msg_check_version(struct lustre_msg *msg) return rc; } +static int ost_rw_hpreq_lock_match(struct ptlrpc_request *req, + struct ldlm_lock *lock) +{ + struct niobuf_remote *nb; + struct obd_ioobj *ioo; + struct ost_body *body; + int objcount, niocount; + int mode, opc, i; + __u64 start, end; + ENTRY; + + opc = lustre_msg_get_opc(req->rq_reqmsg); + LASSERT(opc == OST_READ || opc == OST_WRITE); + + /* As the request may be covered by several locks, do not look at + * o_handle, look at the RPC IO region. */ + body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), + lustre_swab_obdo); + objcount = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1) / + sizeof(*ioo); + ioo = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, + objcount * sizeof(*ioo)); + LASSERT(ioo != NULL); + for (niocount = i = 0; i < objcount; i++) + niocount += ioo[i].ioo_bufcnt; + + nb = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, + niocount * sizeof(*nb)); + LASSERT(nb != NULL); + + mode = LCK_PW; + if (opc == OST_READ) + mode |= LCK_PR; + + start = nb[0].offset & CFS_PAGE_MASK; + end = (nb[ioo->ioo_bufcnt - 1].offset + + nb[ioo->ioo_bufcnt - 1].len - 1) | ~CFS_PAGE_MASK; + + LASSERT(lock->l_resource != NULL); + if (!osc_res_name_eq(ioo->ioo_id, ioo->ioo_gr, + &lock->l_resource->lr_name)) + RETURN(0); + + if (!(lock->l_granted_mode & mode)) + RETURN(0); + + if (lock->l_policy_data.l_extent.end < start || + lock->l_policy_data.l_extent.start > end) + RETURN(0); + + RETURN(1); +} + +/** + * Swab buffers needed to call ost_rw_prolong_locks() and call it. + * Return the value from ost_rw_prolong_locks() which is non-zero if + * there is a cancelled lock which is waiting for this IO request. + */ +static int ost_rw_hpreq_check(struct ptlrpc_request *req) +{ + struct niobuf_remote *nb; + struct obd_ioobj *ioo; + struct ost_body *body; + int objcount, niocount; + int mode, opc, i; + ENTRY; + + opc = lustre_msg_get_opc(req->rq_reqmsg); + LASSERT(opc == OST_READ || opc == OST_WRITE); + + body = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*body)); + LASSERT(body != NULL); + + objcount = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1) / + sizeof(*ioo); + ioo = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, + objcount * sizeof(*ioo)); + LASSERT(ioo != NULL); + + for (niocount = i = 0; i < objcount; i++) + niocount += ioo[i].ioo_bufcnt; + nb = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, + niocount * sizeof(*nb)); + LASSERT(nb != NULL); + LASSERT(niocount == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK)); + + mode = LCK_PW; + if (opc == OST_READ) + mode |= LCK_PR; + RETURN(ost_rw_prolong_locks(req, ioo, nb, &body->oa, mode)); +} + +static int ost_punch_prolong_locks(struct ptlrpc_request *req, struct obdo *oa) +{ + struct ldlm_res_id res_id = { .name = { oa->o_id } }; + struct ost_prolong_data opd = { 0 }; + __u64 start, end; + ENTRY; + + start = oa->o_size; + end = start + oa->o_blocks; + + opd.opd_mode = LCK_PW; + opd.opd_exp = req->rq_export; + opd.opd_policy.l_extent.start = start & CFS_PAGE_MASK; + if (oa->o_blocks == OBD_OBJECT_EOF || end < start) + opd.opd_policy.l_extent.end = OBD_OBJECT_EOF; + else + opd.opd_policy.l_extent.end = end | ~CFS_PAGE_MASK; + + /* prolong locks for the current service time of the corresponding + * portal (= OST_IO_PORTAL) */ + opd.opd_timeout = AT_OFF ? obd_timeout / 2: + max(at_est2timeout(at_get(&req->rq_rqbd-> + rqbd_service->srv_at_estimate)), ldlm_timeout); + + CDEBUG(D_DLMTRACE,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n", + res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start, + opd.opd_policy.l_extent.end); + + opd.opd_oa = oa; + ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id, + ost_prolong_locks_iter, &opd); + RETURN(opd.opd_lock_match); +} + +static int ost_punch_hpreq_lock_match(struct ptlrpc_request *req, + struct ldlm_lock *lock) +{ + struct ost_body *body; + ENTRY; + + body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), + lustre_swab_obdo); + LASSERT(body != NULL); + + if (body->oa.o_valid & OBD_MD_FLHANDLE && + body->oa.o_handle.cookie == lock->l_handle.h_cookie) + RETURN(1); + RETURN(0); +} + +static int ost_punch_hpreq_check(struct ptlrpc_request *req) +{ + struct ost_body *body = lustre_msg_buf(req->rq_reqmsg, + REQ_REC_OFF, sizeof(*body)); + LASSERT(body != NULL); + LASSERT(!(body->oa.o_valid & OBD_MD_FLFLAGS) || + !(body->oa.o_flags & OBD_FL_TRUNCLOCK)); + + RETURN(ost_punch_prolong_locks(req, &body->oa)); +} + +struct ptlrpc_hpreq_ops ost_hpreq_rw = { + .hpreq_lock_match = ost_rw_hpreq_lock_match, + .hpreq_check = ost_rw_hpreq_check, +}; + +struct ptlrpc_hpreq_ops ost_hpreq_punch = { + .hpreq_lock_match = ost_punch_hpreq_lock_match, + .hpreq_check = ost_punch_hpreq_check, +}; + +/** Assign high priority operations to the request if needed. */ +static int ost_hpreq_handler(struct ptlrpc_request *req) +{ + ENTRY; + if (req->rq_export) { + int opc = lustre_msg_get_opc(req->rq_reqmsg); + struct ost_body *body; + + if (opc == OST_READ || opc == OST_WRITE) { + struct niobuf_remote *nb; + struct obd_ioobj *ioo; + int objcount, niocount; + int swab, i; + + body = lustre_swab_reqbuf(req, REQ_REC_OFF, + sizeof(*body), + lustre_swab_obdo); + if (!body) { + CERROR("Missing/short ost_body\n"); + RETURN(-EFAULT); + } + objcount = lustre_msg_buflen(req->rq_reqmsg, + REQ_REC_OFF + 1) / + sizeof(*ioo); + if (objcount == 0) { + CERROR("Missing/short ioobj\n"); + RETURN(-EFAULT); + } + if (objcount > 1) { + CERROR("too many ioobjs (%d)\n", objcount); + RETURN(-EFAULT); + } + + swab = !lustre_req_swabbed(req, REQ_REC_OFF + 1) && + lustre_msg_swabbed(req->rq_reqmsg); + ioo = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, + objcount * sizeof(*ioo), + lustre_swab_obd_ioobj); + if (!ioo) { + CERROR("Missing/short ioobj\n"); + RETURN(-EFAULT); + } + for (niocount = i = 0; i < objcount; i++) { + if (i > 0 && swab) + lustre_swab_obd_ioobj(&ioo[i]); + if (ioo[i].ioo_bufcnt == 0) { + CERROR("ioo[%d] has zero bufcnt\n", i); + RETURN(-EFAULT); + } + niocount += ioo[i].ioo_bufcnt; + } + if (niocount > PTLRPC_MAX_BRW_PAGES) { + DEBUG_REQ(D_ERROR, req, "bulk has too many " + "pages (%d)", niocount); + RETURN(-EFAULT); + } + + swab = !lustre_req_swabbed(req, REQ_REC_OFF + 2) && + lustre_msg_swabbed(req->rq_reqmsg); + nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2, + niocount * sizeof(*nb), + lustre_swab_niobuf_remote); + if (!nb) { + CERROR("Missing/short niobuf\n"); + RETURN(-EFAULT); + } + + if (swab) { + /* swab remaining niobufs */ + for (i = 1; i < niocount; i++) + lustre_swab_niobuf_remote(&nb[i]); + } + + if (niocount == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK)) + req->rq_ops = &ost_hpreq_rw; + } else if (opc == OST_PUNCH) { + body = lustre_swab_reqbuf(req, REQ_REC_OFF, + sizeof(*body), + lustre_swab_obdo); + if (!body) { + CERROR("Missing/short ost_body\n"); + RETURN(-EFAULT); + } + + if (!(body->oa.o_valid & OBD_MD_FLFLAGS) || + !(body->oa.o_flags & OBD_FL_TRUNCLOCK)) + req->rq_ops = &ost_hpreq_punch; + } + } + RETURN(0); +} + /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */ int ost_handle(struct ptlrpc_request *req) { @@ -1604,7 +1929,7 @@ int ost_handle(struct ptlrpc_request *req) } oti_init(oti, req); - + rc = ost_msg_check_version(req->rq_reqmsg); if (rc) RETURN(rc); @@ -1619,12 +1944,9 @@ int ost_handle(struct ptlrpc_request *req) if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)) RETURN(0); if (!rc) { - struct obd_export *exp = req->rq_export; - - obd = exp->exp_obd; - - rc = filter_export_check_flavor(&obd->u.filter, - exp, req); + rc = ost_init_sec_level(req); + if (!rc) + rc = ost_connect_check_sptlrpc(req); } break; } @@ -1730,6 +2052,7 @@ int ost_handle(struct ptlrpc_request *req) DEBUG_REQ(D_INODE, req, "get_info"); rc = ost_get_info(req->rq_export, req); break; +#ifdef HAVE_QUOTA_SUPPORT case OST_QUOTACHECK: CDEBUG(D_INODE, "quotacheck\n"); req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACHECK); @@ -1744,6 +2067,12 @@ int ost_handle(struct ptlrpc_request *req) RETURN(0); rc = ost_handle_quotactl(req); break; + case OST_QUOTA_ADJUST_QUNIT: + CDEBUG(D_INODE, "quota_adjust_qunit\n"); + req_capsule_set(&req->rq_pill, &RQF_OST_QUOTA_ADJUST_QUNIT); + rc = ost_handle_quota_adjust_qunit(req); + break; +#endif case OBD_PING: DEBUG_REQ(D_INODE, req, "ping"); req_capsule_set(&req->rq_pill, &RQF_OBD_PING); @@ -1765,6 +2094,8 @@ int ost_handle(struct ptlrpc_request *req) if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET)) RETURN(0); rc = llog_origin_handle_cancel(req); + if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP)) + RETURN(0); req->rq_status = rc; rc = req_capsule_server_pack(&req->rq_pill); if (rc) @@ -1827,7 +2158,6 @@ EXPORT_SYMBOL(ost_handle); */ static void ost_thread_done(struct ptlrpc_thread *thread) { - int i; struct ost_thread_local_cache *tls; /* TLS stands for Thread-Local * Storage */ @@ -1841,10 +2171,6 @@ static void ost_thread_done(struct ptlrpc_thread *thread) */ tls = thread->t_data; if (tls != NULL) { - for (i = 0; i < OST_THREAD_POOL_SIZE; ++ i) { - if (tls->page[i] != NULL) - OBD_PAGE_FREE(tls->page[i]); - } OBD_FREE_PTR(tls); thread->t_data = NULL; } @@ -1856,8 +2182,6 @@ static void ost_thread_done(struct ptlrpc_thread *thread) */ static int ost_thread_init(struct ptlrpc_thread *thread) { - int result; - int i; struct ost_thread_local_cache *tls; ENTRY; @@ -1867,23 +2191,10 @@ static int ost_thread_init(struct ptlrpc_thread *thread) LASSERTF(thread->t_id <= OSS_THREADS_MAX, "%u\n", thread->t_id); OBD_ALLOC_PTR(tls); - if (tls != NULL) { - result = 0; - thread->t_data = tls; - /* - * populate pool - */ - for (i = 0; i < OST_THREAD_POOL_SIZE; ++ i) { - OBD_PAGE_ALLOC(tls->page[i], OST_THREAD_POOL_GFP); - if (tls->page[i] == NULL) { - ost_thread_done(thread); - result = -ENOMEM; - break; - } - } - } else - result = -ENOMEM; - RETURN(result); + if (tls == NULL) + RETURN(-ENOMEM); + thread->t_data = tls; + RETURN(0); } #define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000) @@ -1911,21 +2222,21 @@ static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg) if (oss_num_threads) { /* If oss_num_threads is set, it is the min and the max. */ - if (oss_num_threads > OSS_THREADS_MAX) + if (oss_num_threads > OSS_THREADS_MAX) oss_num_threads = OSS_THREADS_MAX; if (oss_num_threads < OSS_THREADS_MIN) oss_num_threads = OSS_THREADS_MIN; oss_max_threads = oss_min_threads = oss_num_threads; } else { /* Base min threads on memory and cpus */ - oss_min_threads = num_possible_cpus() * num_physpages >> + oss_min_threads = num_possible_cpus() * num_physpages >> (27 - CFS_PAGE_SHIFT); if (oss_min_threads < OSS_THREADS_MIN) oss_min_threads = OSS_THREADS_MIN; /* Insure a 4x range for dynamic threads */ - if (oss_min_threads > OSS_THREADS_MAX / 4) + if (oss_min_threads > OSS_THREADS_MAX / 4) oss_min_threads = OSS_THREADS_MAX / 4; - oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4); + oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4 + 1); } ost->ost_service = @@ -1935,7 +2246,7 @@ static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg) ost_handle, LUSTRE_OSS_NAME, obd->obd_proc_entry, target_print_req, oss_min_threads, oss_max_threads, - "ll_ost", LCT_DT_THREAD); + "ll_ost", LCT_DT_THREAD, NULL); if (ost->ost_service == NULL) { CERROR("failed to start service\n"); GOTO(out_lprocfs, rc = -ENOMEM); @@ -1964,7 +2275,7 @@ static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg) ost_handle, "ost_create", obd->obd_proc_entry, target_print_req, oss_min_create_threads, oss_max_create_threads, - "ll_ost_creat", LCT_DT_THREAD); + "ll_ost_creat", LCT_DT_THREAD, NULL); if (ost->ost_create_service == NULL) { CERROR("failed to start OST create service\n"); GOTO(out_service, rc = -ENOMEM); @@ -1981,7 +2292,7 @@ static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg) ost_handle, "ost_io", obd->obd_proc_entry, target_print_req, oss_min_threads, oss_max_threads, - "ll_ost_io", LCT_DT_THREAD); + "ll_ost_io", LCT_DT_THREAD, ost_hpreq_handler); if (ost->ost_io_service == NULL) { CERROR("failed to start OST I/O service\n"); GOTO(out_create, rc = -ENOMEM);