X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;ds=sidebyside;f=lustre%2Fost%2Fost_handler.c;h=9c200089ac29df25052a79bb02d7931cb48f422a;hb=d031e92fe730792e3a4dba2f1e8ae90a085c96c5;hp=8b2b7fb3065a9df6fb2a52cef109e1ac0800c869;hpb=cefa8cda2ba2d288ccaa4ec077a6c627592503ea;p=fs%2Flustre-release.git diff --git a/lustre/ost/ost_handler.c b/lustre/ost/ost_handler.c index 8b2b7fb..9c20008 100644 --- a/lustre/ost/ost_handler.c +++ b/lustre/ost/ost_handler.c @@ -106,70 +106,140 @@ static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req, struct obd_trans_info *oti) { struct ost_body *body, *repbody; - __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) }; struct lustre_capa *capa = NULL; int rc; ENTRY; - body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), - lustre_swab_ost_body); + /* Get the request body */ + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); if (body == NULL) RETURN(-EFAULT); if (body->oa.o_id == 0) RETURN(-EPROTO); - if (lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1)) { + /* If there's a DLM request, cancel the locks mentioned in it*/ + if (req_capsule_field_present(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT)) { struct ldlm_request *dlm; - dlm = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*dlm), - lustre_swab_ldlm_request); + + dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ); if (dlm == NULL) 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); + /* If there's a capability, get it */ + if (body->oa.o_valid & OBD_MD_FLOSSCAPA) { + capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1); + if (capa == NULL) { + CERROR("Missing capability for OST DESTROY"); + RETURN (-EFAULT); + } + } - rc = lustre_pack_reply(req, 2, size, NULL); + /* Prepare the reply */ + rc = req_capsule_server_pack(&req->rq_pill); if (rc) RETURN(rc); + /* Get the log cancellation cookie */ if (body->oa.o_valid & OBD_MD_FLCOOKIE) oti->oti_logcookies = &body->oa.o_lcookie; - repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, - sizeof(*repbody)); + + /* Finish the reply */ + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); + + /* Do the destroy and set the reply status accordingly */ req->rq_status = obd_destroy(exp, &body->oa, NULL, oti, NULL, capa); RETURN(0); } +/** + * Helper function for getting server side [start, start+count] DLM lock + * if asked by client. + */ +static int ost_lock_get(struct obd_export *exp, struct obdo *oa, + __u64 start, __u64 count, struct lustre_handle *lh, + int mode, int flags) +{ + struct ldlm_res_id res_id = { .name = { oa->o_id, 0, oa->o_gr, 0} }; + ldlm_policy_data_t policy; + __u64 end = start + count; + + ENTRY; + + LASSERT(!lustre_handle_is_used(lh)); + LASSERT((oa->o_valid & (OBD_MD_FLID | OBD_MD_FLGROUP)) == + (OBD_MD_FLID | OBD_MD_FLGROUP)); + + if (!(oa->o_valid & OBD_MD_FLFLAGS) || + !(oa->o_flags & OBD_FL_SRVLOCK)) + RETURN(0); + + CDEBUG(D_INODE, "OST-side extent lock.\n"); + + policy.l_extent.start = start & CFS_PAGE_MASK; + + /* If ->o_blocks is EOF it means "lock till the end of the + * file". Otherwise, it's size of a hole being punched (in bytes) */ + if (count == OBD_OBJECT_EOF || end < start) + policy.l_extent.end = OBD_OBJECT_EOF; + else + policy.l_extent.end = end | ~CFS_PAGE_MASK; + + 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)); +} + +/* Helper function: release lock, if any. */ +static void ost_lock_put(struct obd_export *exp, + struct lustre_handle *lh, int mode) +{ + ENTRY; + if (lustre_handle_is_used(lh)) + ldlm_lock_decref(lh, mode); + EXIT; +} + static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req) { struct ost_body *body, *repbody; struct obd_info oinfo = { { { 0 } } }; - __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) }; + struct lustre_handle lh = { 0 }; int rc; ENTRY; - body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), - lustre_swab_ost_body); + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); if (body == NULL) RETURN(-EFAULT); - rc = lustre_pack_reply(req, 2, size, NULL); + rc = req_capsule_server_pack(&req->rq_pill); if (rc) RETURN(rc); - repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, - sizeof(*repbody)); + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); repbody->oa = body->oa; + rc = ost_lock_get(exp, &body->oa, 0, OBD_OBJECT_EOF, &lh, LCK_PR, 0); + if (rc) + RETURN(rc); + oinfo.oi_oa = &repbody->oa; - if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) - oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg, - REQ_REC_OFF + 1); + if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) { + oinfo.oi_capa = req_capsule_client_get(&req->rq_pill, + &RMF_CAPA1); + if (oinfo.oi_capa == NULL) { + CERROR("Missing capability for OST GETATTR"); + RETURN (-EFAULT); + } + } + req->rq_status = obd_getattr(exp, &oinfo); + ost_lock_put(exp, &lh, LCK_PR); + ost_drop_id(exp, &repbody->oa); RETURN(0); } @@ -177,20 +247,17 @@ static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req) static int ost_statfs(struct ptlrpc_request *req) { struct obd_statfs *osfs; - __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*osfs) }; int rc; ENTRY; - rc = lustre_pack_reply(req, 2, size, NULL); + rc = req_capsule_server_pack(&req->rq_pill); if (rc) RETURN(rc); - osfs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*osfs)); + osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS); req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs, - cfs_time_current_64() - HZ, 0); - if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) - osfs->os_bfree = osfs->os_bavail = 64; + cfs_time_current_64() - CFS_HZ, 0); if (req->rq_status != 0) CERROR("ost: statfs failed: rc %d\n", req->rq_status); @@ -201,107 +268,39 @@ static int ost_create(struct obd_export *exp, struct ptlrpc_request *req, struct obd_trans_info *oti) { struct ost_body *body, *repbody; - __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) }; int rc; ENTRY; - body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), - lustre_swab_ost_body); + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); if (body == NULL) RETURN(-EFAULT); - rc = lustre_pack_reply(req, 2, size, NULL); + rc = req_capsule_server_pack(&req->rq_pill); if (rc) RETURN(rc); - repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, - sizeof(*repbody)); + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); memcpy(&repbody->oa, &body->oa, sizeof(body->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); } -/* - * Helper function for ost_punch(): if asked by client, acquire [size, EOF] - * lock on the file being truncated. - */ -static int ost_punch_lock_get(struct obd_export *exp, struct obdo *oa, - struct lustre_handle *lh) -{ - int flags; - struct ldlm_res_id res_id; - ldlm_policy_data_t policy; - __u64 start; - __u64 finis; - - ENTRY; - - osc_build_res_name(oa->o_id, oa->o_gr, &res_id); - LASSERT(!lustre_handle_is_used(lh)); - - if (!(oa->o_valid & OBD_MD_FLFLAGS) || - !(oa->o_flags & OBD_FL_TRUNCLOCK)) - RETURN(0); - - CDEBUG(D_INODE, "OST-side truncate lock.\n"); - - start = oa->o_size; - finis = start + oa->o_blocks; - - /* - * standard truncate optimization: if file body is completely - * destroyed, don't send data back to the server. - */ - flags = (start == 0) ? LDLM_AST_DISCARD_DATA : 0; - - policy.l_extent.start = start & CFS_PAGE_MASK; - - /* - * If ->o_blocks is EOF it means "lock till the end of the - * file". Otherwise, it's size of a hole being punched (in bytes) - */ - if (oa->o_blocks == OBD_OBJECT_EOF || finis < start) - policy.l_extent.end = OBD_OBJECT_EOF; - else - policy.l_extent.end = finis | ~CFS_PAGE_MASK; - - 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, - NULL, lh)); -} - -/* - * Helper function for ost_punch(): release lock acquired by - * ost_punch_lock_get(), if any. - */ -static void ost_punch_lock_put(struct obd_export *exp, struct obdo *oa, - struct lustre_handle *lh) -{ - ENTRY; - if (lustre_handle_is_used(lh)) - ldlm_lock_decref(lh, LCK_PW); - EXIT; -} - static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req, struct obd_trans_info *oti) { struct obd_info oinfo = { { { 0 } } }; struct ost_body *body, *repbody; - __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) }; - int rc; + int rc, flags = 0; struct lustre_handle lh = {0,}; ENTRY; /* 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); + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); if (body == NULL) RETURN(-EFAULT); @@ -313,28 +312,38 @@ static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req, (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) RETURN(-EPROTO); - rc = lustre_pack_reply(req, 2, size, NULL); + rc = req_capsule_server_pack(&req->rq_pill); if (rc) RETURN(rc); - repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, - sizeof(*repbody)); - rc = ost_punch_lock_get(exp, oinfo.oi_oa, &lh); + /* standard truncate optimization: if file body is completely + * destroyed, don't send data back to the server. */ + if (oinfo.oi_oa->o_size == 0) + flags |= LDLM_AST_DISCARD_DATA; + + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); + rc = ost_lock_get(exp, oinfo.oi_oa, oinfo.oi_oa->o_size, + oinfo.oi_oa->o_blocks, &lh, LCK_PW, flags); if (rc == 0) { if (oinfo.oi_oa->o_valid & OBD_MD_FLFLAGS && - oinfo.oi_oa->o_flags == OBD_FL_TRUNCLOCK) + oinfo.oi_oa->o_flags == OBD_FL_SRVLOCK) /* - * If OBD_FL_TRUNCLOCK is the only bit set in + * If OBD_FL_SRVLOCK is the only bit set in * ->o_flags, clear OBD_MD_FLFLAGS to avoid falling * through filter_setattr() to filter_iocontrol(). */ oinfo.oi_oa->o_valid &= ~OBD_MD_FLFLAGS; - if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) - oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg, - REQ_REC_OFF + 1); + if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) { + oinfo.oi_capa = req_capsule_client_get(&req->rq_pill, + &RMF_CAPA1); + if (oinfo.oi_capa == NULL) { + CERROR("Missing capability for OST PUNCH"); + RETURN (-EFAULT); + } + } req->rq_status = obd_punch(exp, &oinfo, oti, NULL); - ost_punch_lock_put(exp, oinfo.oi_oa, &lh); + ost_lock_put(exp, &lh, LCK_PW); } repbody->oa = *oinfo.oi_oa; ost_drop_id(exp, &repbody->oa); @@ -345,24 +354,26 @@ static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req) { struct ost_body *body, *repbody; struct lustre_capa *capa = NULL; - __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) }; int rc; ENTRY; - body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), - lustre_swab_ost_body); + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); if (body == NULL) RETURN(-EFAULT); - if (body->oa.o_valid & OBD_MD_FLOSSCAPA) - capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 1); + if (body->oa.o_valid & OBD_MD_FLOSSCAPA) { + capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1); + if (capa == NULL) { + CERROR("Missing capability for OST SYNC"); + RETURN (-EFAULT); + } + } - rc = lustre_pack_reply(req, 2, size, NULL); + rc = req_capsule_server_pack(&req->rq_pill); if (rc) RETURN(rc); - repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, - sizeof(*repbody)); + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); 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); @@ -374,28 +385,30 @@ static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req, struct obd_trans_info *oti) { struct ost_body *body, *repbody; - __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) }; int rc; struct obd_info oinfo = { { { 0 } } }; ENTRY; - body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), - lustre_swab_ost_body); + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); if (body == NULL) RETURN(-EFAULT); - rc = lustre_pack_reply(req, 2, size, NULL); + rc = req_capsule_server_pack(&req->rq_pill); if (rc) RETURN(rc); - repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, - sizeof(*repbody)); + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); repbody->oa = body->oa; oinfo.oi_oa = &repbody->oa; - if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) - oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg, - REQ_REC_OFF + 1); + if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) { + oinfo.oi_capa = req_capsule_client_get(&req->rq_pill, + &RMF_CAPA1); + if (oinfo.oi_capa == NULL) { + CERROR("Missing capability for OST SETATTR"); + RETURN (-EFAULT); + } + } req->rq_status = obd_setattr(exp, &oinfo, oti); ost_drop_id(exp, &repbody->oa); RETURN(0); @@ -461,10 +474,10 @@ static int ost_brw_lock_get(int mode, struct obd_export *exp, if (nrbufs == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK)) RETURN(0); - /* EXPENSIVE ASSERTION */ for (i = 1; i < nrbufs; i ++) - LASSERT((nb[0].flags & OBD_BRW_SRVLOCK) == - (nb[i].flags & OBD_BRW_SRVLOCK)); + if ((nb[0].flags & OBD_BRW_SRVLOCK) != + (nb[i].flags & OBD_BRW_SRVLOCK)) + RETURN(-EFAULT); policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK; policy.l_extent.end = (nb[nrbufs - 1].offset + @@ -473,8 +486,7 @@ 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, - NULL, lh)); + ldlm_glimpse_ast, NULL, 0, NULL, lh)); } static void ost_brw_lock_put(int mode, @@ -493,7 +505,10 @@ 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) @@ -523,36 +538,55 @@ 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, struct obdo *oa, - 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); @@ -562,16 +596,28 @@ static void ost_prolong_locks(struct obd_export *exp, struct obd_ioobj *obj, 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); - EXIT; - return; } } - ldlm_resource_iterate(exp->exp_obd->obd_namespace, &res_id, + opd.opd_oa = oa; + ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id, ost_prolong_locks_iter, &opd); - - EXIT; + RETURN(opd.opd_lock_match); } static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) @@ -585,8 +631,7 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) struct lustre_capa *capa = NULL; struct l_wait_info lwi; struct lustre_handle lockh = { 0 }; - __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) }; - int objcount, niocount, npages, nob = 0, rc, i; + int niocount, npages, nob = 0, rc, i; int no_reply = 0; ENTRY; @@ -599,63 +644,44 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) /* Check if there is eviction in progress, and if so, wait for it to * finish */ - if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) { + if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) { lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq, - !atomic_read(&exp->exp_obd->obd_evict_inprogress), + !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress), &lwi); } 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); - } - - 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); + /* ost_body, ioobj & noibuf_remote are verified and swabbed in + * ost_rw_hpreq_check(). */ + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); + if (body == NULL) GOTO(out, rc = -EFAULT); - } - ioo = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*ioo), - lustre_swab_obd_ioobj); - if (ioo == NULL) { - CERROR("Missing/short ioobj\n"); + /* + * A req_capsule_X_get_array(pill, field, ptr_to_element_count) function + * would be useful here and wherever we get &RMF_OBD_IOOBJ and + * &RMF_NIOBUF_REMOTE. + */ + ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ); + if (ioo == NULL) GOTO(out, rc = -EFAULT); - } niocount = ioo->ioo_bufcnt; - if (niocount > PTLRPC_MAX_BRW_PAGES) { - DEBUG_REQ(D_ERROR, req, "bulk has too many pages (%d)", - niocount); + remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE); + if (remote_nb == NULL) 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]); + if (body->oa.o_valid & OBD_MD_FLOSSCAPA) { + capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1); + if (capa == NULL) { + CERROR("Missing capability for OST BRW READ"); + GOTO(out, rc = -EFAULT); + } } - if (body->oa.o_valid & OBD_MD_FLOSSCAPA) - capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3); - - rc = lustre_pack_reply(req, 2, size, NULL); + rc = req_capsule_server_pack(&req->rq_pill); if (rc) GOTO(out, rc); @@ -692,10 +718,14 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) 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); + if (desc == NULL) + GOTO(out_lock, rc = -ENOMEM); - ost_prolong_locks(exp, ioo, remote_nb, &body->oa, LCK_PW | LCK_PR); + if (!lustre_handle_is_used(&lockh)) + /* no needs to try to prolong lock if server is asked + * to handle locking (= OBD_BRW_SRVLOCK) */ + ost_rw_prolong_locks(req, ioo, remote_nb, &body->oa, + LCK_PW | LCK_PR); nob = 0; for (i = 0; i < npages; i++) { @@ -741,21 +771,22 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) if (rc == 0) { /* Check if there is eviction in progress, and if so, wait for * it to finish */ - if (unlikely(atomic_read(&exp->exp_obd-> - obd_evict_inprogress))) { + if (unlikely(cfs_atomic_read(&exp->exp_obd-> + obd_evict_inprogress))) { lwi = LWI_INTR(NULL, NULL); rc = l_wait_event(exp->exp_obd-> - obd_evict_inprogress_waitq, - !atomic_read(&exp->exp_obd-> - obd_evict_inprogress), + obd_evict_inprogress_waitq, + !cfs_atomic_read(&exp->exp_obd-> + obd_evict_inprogress), &lwi); } - if (exp->exp_failed) + /* Check if client was evicted or tried to reconnect already */ + if (exp->exp_failed || exp->exp_abort_active_req) 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) { @@ -763,15 +794,17 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) do { long timeoutl = req->rq_deadline - cfs_time_current_sec(); - cfs_duration_t timeout = (timeoutl <= 0 || rc) ? + cfs_duration_t timeout = timeoutl <= 0 ? CFS_TICK : cfs_time_seconds(timeoutl); lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1), 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) && @@ -788,6 +821,11 @@ 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)", @@ -809,8 +847,7 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) remote_nb, npages, local_nb, oti, rc); if (rc == 0) { - repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, - sizeof(*repbody)); + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa)); ost_drop_id(exp, &repbody->oa); } @@ -858,9 +895,8 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) struct lustre_handle lockh = {0}; struct lustre_capa *capa = NULL; __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; @@ -879,76 +915,58 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) /* Check if there is eviction in progress, and if so, wait for it to * finish */ - if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) { + if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) { lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq, - !atomic_read(&exp->exp_obd->obd_evict_inprogress), + !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress), &lwi); } 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"); + /* ost_body, ioobj & noibuf_remote are verified and swabbed in + * ost_rw_hpreq_check(). */ + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); + if (body == NULL) GOTO(out, rc = -EFAULT); - } - 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(); - 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); - } + objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ, + RCL_CLIENT) / sizeof(*ioo); + ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ); + if (ioo == NULL) + GOTO(out, rc = -EFAULT); + 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); + /* + * It'd be nice to have a capsule function to indicate how many elements + * there were in a buffer for an RMF that's declared to be an array. + * It's easy enough to compute the number of elements here though. + */ + remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE); + if (remote_nb == NULL || niocount != (req_capsule_get_size(&req->rq_pill, + &RMF_NIOBUF_REMOTE, RCL_CLIENT) / sizeof(*remote_nb))) 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]); + if (body->oa.o_valid & OBD_MD_FLOSSCAPA) { + capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1); + if (capa == NULL) { + CERROR("Missing capability for OST BRW WRITE"); + GOTO(out, rc = -EFAULT); + } } - if (body->oa.o_valid & OBD_MD_FLOSSCAPA) - capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3); - - size[REPLY_REC_OFF + 1] = niocount * sizeof(*rcs); - rc = lustre_pack_reply(req, 3, size, NULL); + req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER, + niocount * sizeof(*rcs)); + rc = req_capsule_server_pack(&req->rq_pill); if (rc != 0) GOTO(out, rc); OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, obd_fail_val); - rcs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, - niocount * sizeof(*rcs)); + rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS); /* * Per-thread array of struct niobuf_{local,remote}'s was allocated by @@ -975,7 +993,10 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) GOTO(out_lock, rc = -ETIMEDOUT); } - ost_prolong_locks(exp, ioo, remote_nb,&body->oa, LCK_PW); + if (!lustre_handle_is_used(&lockh)) + /* no needs to try to prolong lock if server is asked + * to handle locking (= OBD_BRW_SRVLOCK) */ + 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) { @@ -1005,7 +1026,7 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) desc = ptlrpc_prep_bulk_exp(req, npages, BULK_GET_SINK, OST_BULK_PORTAL); if (desc == NULL) - GOTO(out, rc = -ENOMEM); + GOTO(out_lock, rc = -ENOMEM); /* NB Having prepped, we must commit... */ @@ -1014,24 +1035,31 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) local_nb[i].offset & ~CFS_PAGE_MASK, local_nb[i].len); - /* Check if client was evicted while we were doing i/o before touching - network */ - if (desc->bd_export->exp_failed) + rc = sptlrpc_svc_prep_bulk(req, desc); + if (rc != 0) + GOTO(out_lock, rc); + + /* Check if client was evicted or tried to reconnect while we + * were doing i/o before touching network */ + if (desc->bd_export->exp_failed || + desc->bd_export->exp_abort_active_req) 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 { long timeoutl = req->rq_deadline - cfs_time_current_sec(); - cfs_duration_t timeout = (timeoutl <= 0 || rc) ? + cfs_duration_t timeout = timeoutl <= 0 ? CFS_TICK : cfs_time_seconds(timeoutl); 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) && @@ -1048,25 +1076,24 @@ 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)); + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa)); if (unlikely(client_cksum != 0 && rc == 0)) { @@ -1091,6 +1118,12 @@ 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, remote_nb, npages, local_nb, oti, rc); + if (rc == -ENOTCONN) + /* quota acquire process has been given up because + * either the client has been evicted or the client + * has timed out the request already */ + no_reply = 1; + if (exp_connect_rmtclient(exp)) { repbody->oa.o_uid = o_uid; repbody->oa.o_gid = o_gid; @@ -1184,37 +1217,82 @@ out: exp->exp_connection->c_remote_uuid.uuid, libcfs_id2str(req->rq_peer)); } + libcfs_memory_pressure_clr(); RETURN(rc); } +/** + * Implementation of OST_SET_INFO. + * + * OST_SET_INFO is like ioctl(): heavily overloaded. Specifically, it takes a + * "key" and a value RPC buffers as arguments, with the value's contents + * interpreted according to the key. + * + * Value types that need swabbing have swabbing done explicitly, either here or + * in functions called from here. This should be corrected: all swabbing should + * be done in the capsule abstraction, as that will then allow us to move + * swabbing exclusively to the client without having to modify server code + * outside the capsule abstraction's implementation itself. To correct this + * will require minor changes to the capsule abstraction; see the comments for + * req_capsule_extend() in layout.c. + */ static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req) { + struct ost_body *body = NULL, *repbody; char *key, *val = NULL; int keylen, vallen, rc = 0; + int is_grant_shrink = 0; ENTRY; - key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1); + key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY); if (key == NULL) { DEBUG_REQ(D_HA, req, "no set_info key"); RETURN(-EFAULT); } - keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF); + keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY, + RCL_CLIENT); + + vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL, + RCL_CLIENT); + + if ((is_grant_shrink = KEY_IS(KEY_GRANT_SHRINK))) + /* In this case the value is actually an RMF_OST_BODY, so we + * transmutate the type of this PTLRPC */ + req_capsule_extend(&req->rq_pill, &RQF_OST_SET_GRANT_INFO); - rc = lustre_pack_reply(req, 1, NULL, NULL); + rc = req_capsule_server_pack(&req->rq_pill); 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 (is_grant_shrink) { + body = req_capsule_client_get(&req->rq_pill, + &RMF_OST_BODY); + if (!body) + RETURN(-EFAULT); + + repbody = req_capsule_server_get(&req->rq_pill, + &RMF_OST_BODY); + memcpy(repbody, body, sizeof(*body)); + val = (char*)repbody; + } else { + val = req_capsule_client_get(&req->rq_pill, + &RMF_SETINFO_VAL); + } + } 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) && ptlrpc_req_need_swab(req)) { + if (vallen < sizeof(__u32)) + RETURN(-EFAULT); + __swab32s((__u32 *)val); } + /* OBD will also check if KEY_IS(KEY_GRANT_SHRINK), and will cast val to + * a struct ost_body * value */ rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL); out: lustre_msg_set_status(req->rq_repmsg, 0); @@ -1228,8 +1306,6 @@ static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req) struct req_capsule *pill = &req->rq_pill; ENTRY; - req_capsule_set(&req->rq_pill, &RQF_OST_GET_INFO_GENERIC); - /* this common part for get_info rpc */ key = req_capsule_client_get(pill, &RMF_SETINFO_KEY); if (key == NULL) { @@ -1333,7 +1409,7 @@ static int ost_llog_handle_connect(struct obd_export *exp, int rc; ENTRY; - body = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*body)); + body = req_capsule_client_get(&req->rq_pill, &RMF_LLOGD_CONN_BODY); rc = obd_llog_connect(exp, body); RETURN(rc); } @@ -1343,9 +1419,9 @@ do { \ reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | \ OBD_CONNECT_RMT_CLIENT_FORCE | \ OBD_CONNECT_OSS_CAPA); \ - spin_lock(&exp->exp_lock); \ + cfs_spin_lock(&exp->exp_lock); \ exp->exp_connect_flags = reply->ocd_connect_flags; \ - spin_unlock(&exp->exp_lock); \ + cfs_spin_unlock(&exp->exp_lock); \ } while (0) static int ost_init_sec_level(struct ptlrpc_request *req) @@ -1442,9 +1518,9 @@ static int ost_init_sec_level(struct ptlrpc_request *req) if (!filter->fo_fl_oss_capa) reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA; - spin_lock(&exp->exp_lock); + cfs_spin_lock(&exp->exp_lock); exp->exp_connect_flags = reply->ocd_connect_flags; - spin_unlock(&exp->exp_lock); + cfs_spin_unlock(&exp->exp_lock); } break; default: @@ -1454,41 +1530,61 @@ static int ost_init_sec_level(struct ptlrpc_request *req) RETURN(rc); } -static int filter_export_check_flavor(struct filter_obd *filter, - struct obd_export *exp, - struct ptlrpc_request *req) +/* + * 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) { - 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) + 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; - - 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; } - spin_unlock(&exp->exp_lock); + if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) { + cfs_read_lock(&filter->fo_sptlrpc_lock); + sptlrpc_target_choose_flavor(&filter->fo_sptlrpc_rset, + req->rq_sp_from, + req->rq_peer.nid, + &flvr); + cfs_read_unlock(&filter->fo_sptlrpc_lock); + + cfs_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; + } + + cfs_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; } @@ -1590,6 +1686,292 @@ int ost_msg_check_version(struct lustre_msg *msg) return rc; } +/** + * Returns 1 if the given PTLRPC matches the given LDLM locks, or 0 if it does + * not. + */ +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 = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); + if (body == NULL) + RETURN(0); + + objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ, + RCL_CLIENT) / sizeof(*ioo); + ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ); + if (ioo == NULL) + RETURN(0); + + for (niocount = i = 0; i < objcount; i++) + niocount += ioo[i].ioo_bufcnt; + + nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE); + if (nb == NULL || + niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE, + RCL_CLIENT) / sizeof(*nb))) + RETURN(0); + + 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); +} + +/** + * High-priority queue request check for whether the given PTLRPC request (\a + * req) is blocking an LDLM lock cancel. + * + * Returns 1 if the given given PTLRPC request (\a req) is blocking an LDLM lock + * cancel, 0 if it is not, and -EFAULT if the request is malformed. + * + * Only OST_READs, OST_WRITEs and OST_PUNCHes go on the h-p RPC queue. This + * function looks only at OST_READs and OST_WRITEs. + */ +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 = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); + if (body == NULL) + RETURN(-EFAULT); + + objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ, + RCL_CLIENT) / sizeof(*ioo); + ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ); + if (ioo == NULL) + RETURN(-EFAULT); + + for (niocount = i = 0; i < objcount; i++) + niocount += ioo[i].ioo_bufcnt; + nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE); + if (nb == NULL || + niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE, + RCL_CLIENT) / sizeof(*nb))) + RETURN(-EFAULT); + if (niocount != 0 && (nb[0].flags & OBD_BRW_SRVLOCK)) + RETURN(-EFAULT); + + 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); +} + +/** + * Like ost_rw_hpreq_lock_match(), but for OST_PUNCH RPCs. + */ +static int ost_punch_hpreq_lock_match(struct ptlrpc_request *req, + struct ldlm_lock *lock) +{ + struct ost_body *body; + ENTRY; + + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); + if (body == NULL) + RETURN(0); /* can't return -EFAULT here */ + + if (body->oa.o_valid & OBD_MD_FLHANDLE && + body->oa.o_handle.cookie == lock->l_handle.h_cookie) + RETURN(1); + RETURN(0); +} + +/** + * Like ost_rw_hpreq_check(), but for OST_PUNCH RPCs. + */ +static int ost_punch_hpreq_check(struct ptlrpc_request *req) +{ + struct ost_body *body; + + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); + if (body == NULL) + RETURN(-EFAULT); + + LASSERT(!(body->oa.o_valid & OBD_MD_FLFLAGS) || + !(body->oa.o_flags & OBD_FL_SRVLOCK)); + + 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 i; + + /* RPCs on the H-P queue can be inspected before + * ost_handler() initializes their pills, so we + * initialize that here. Capsule initialization is + * idempotent, as is setting the pill's format (provided + * it doesn't change). + */ + req_capsule_init(&req->rq_pill, req, RCL_SERVER); + if (opc == OST_READ) + req_capsule_set(&req->rq_pill, + &RQF_OST_BRW_READ); + else + req_capsule_set(&req->rq_pill, + &RQF_OST_BRW_WRITE); + + body = req_capsule_client_get(&req->rq_pill, + &RMF_OST_BODY); + if (body == NULL) { + CERROR("Missing/short ost_body\n"); + RETURN(-EFAULT); + } + objcount = req_capsule_get_size(&req->rq_pill, + &RMF_OBD_IOOBJ, + RCL_CLIENT) / + sizeof(*ioo); + if (objcount == 0) { + CERROR("Missing/short ioobj\n"); + RETURN(-EFAULT); + } + if (objcount > 1) { + CERROR("too many ioobjs (%d)\n", objcount); + RETURN(-EFAULT); + } + + ioo = req_capsule_client_get(&req->rq_pill, + &RMF_OBD_IOOBJ); + if (ioo == NULL) { + CERROR("Missing/short ioobj\n"); + RETURN(-EFAULT); + } + + for (niocount = i = 0; i < objcount; 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_RPCTRACE, req, + "bulk has too many pages (%d)", + niocount); + RETURN(-EFAULT); + } + + nb = req_capsule_client_get(&req->rq_pill, + &RMF_NIOBUF_REMOTE); + if (nb == NULL) { + CERROR("Missing/short niobuf\n"); + RETURN(-EFAULT); + } + + if (niocount == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK)) + req->rq_ops = &ost_hpreq_rw; + } else if (opc == OST_PUNCH) { + req_capsule_init(&req->rq_pill, req, RCL_SERVER); + req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH); + + body = req_capsule_client_get(&req->rq_pill, + &RMF_OST_BODY); + if (body == NULL) { + CERROR("Missing/short ost_body\n"); + RETURN(-EFAULT); + } + + if (!(body->oa.o_valid & OBD_MD_FLFLAGS) || + !(body->oa.o_flags & OBD_FL_SRVLOCK)) + 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) { @@ -1611,11 +1993,10 @@ int ost_handle(struct ptlrpc_request *req) req_capsule_init(&req->rq_pill, req, RCL_SERVER); - /* XXX identical to MDS */ if (lustre_msg_get_opc(req->rq_reqmsg) != OST_CONNECT) { int recovering; - if (req->rq_export == NULL) { + if (!class_connected_export(req->rq_export)) { CDEBUG(D_HA,"operation %d on unconnected OST from %s\n", lustre_msg_get_opc(req->rq_reqmsg), libcfs_id2str(req->rq_peer)); @@ -1626,9 +2007,9 @@ int ost_handle(struct ptlrpc_request *req) obd = req->rq_export->exp_obd; /* Check for aborted recovery. */ - spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock_bh(&obd->obd_processing_task_lock); recovering = obd->obd_recovering; - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock_bh(&obd->obd_processing_task_lock); if (recovering) { rc = ost_filter_recovery_request(req, obd, &should_process); @@ -1659,13 +2040,8 @@ int ost_handle(struct ptlrpc_request *req) RETURN(0); if (!rc) { rc = ost_init_sec_level(req); - if (!rc) { - struct obd_export *exp = req->rq_export; - - obd = exp->exp_obd; - rc = filter_export_check_flavor(&obd->u.filter, - exp, req); - } + if (!rc) + rc = ost_connect_check_sptlrpc(req); } break; } @@ -1678,16 +2054,16 @@ int ost_handle(struct ptlrpc_request *req) break; case OST_CREATE: CDEBUG(D_INODE, "create\n"); + req_capsule_set(&req->rq_pill, &RQF_OST_CREATE); if (OBD_FAIL_CHECK(OBD_FAIL_OST_CREATE_NET)) RETURN(0); - if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) - GOTO(out, rc = -ENOSPC); if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS)) GOTO(out, rc = -EROFS); rc = ost_create(req->rq_export, req, oti); break; case OST_DESTROY: CDEBUG(D_INODE, "destroy\n"); + req_capsule_set(&req->rq_pill, &RQF_OST_DESTROY); if (OBD_FAIL_CHECK(OBD_FAIL_OST_DESTROY_NET)) RETURN(0); if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS)) @@ -1696,17 +2072,20 @@ int ost_handle(struct ptlrpc_request *req) break; case OST_GETATTR: CDEBUG(D_INODE, "getattr\n"); + req_capsule_set(&req->rq_pill, &RQF_OST_GETATTR); if (OBD_FAIL_CHECK(OBD_FAIL_OST_GETATTR_NET)) RETURN(0); rc = ost_getattr(req->rq_export, req); break; case OST_SETATTR: CDEBUG(D_INODE, "setattr\n"); + req_capsule_set(&req->rq_pill, &RQF_OST_SETATTR); if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_NET)) RETURN(0); rc = ost_setattr(req->rq_export, req, oti); break; case OST_WRITE: + req_capsule_set(&req->rq_pill, &RQF_OST_BRW_WRITE); CDEBUG(D_INODE, "write\n"); /* req->rq_request_portal would be nice, if it was set */ if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){ @@ -1727,6 +2106,7 @@ int ost_handle(struct ptlrpc_request *req) /* ost_brw_write sends its own replies */ RETURN(rc); case OST_READ: + req_capsule_set(&req->rq_pill, &RQF_OST_BRW_READ); CDEBUG(D_INODE, "read\n"); /* req->rq_request_portal would be nice, if it was set */ if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){ @@ -1744,6 +2124,7 @@ int ost_handle(struct ptlrpc_request *req) RETURN(rc); case OST_PUNCH: CDEBUG(D_INODE, "punch\n"); + req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH); if (OBD_FAIL_CHECK(OBD_FAIL_OST_PUNCH_NET)) RETURN(0); if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS)) @@ -1752,23 +2133,26 @@ int ost_handle(struct ptlrpc_request *req) break; case OST_STATFS: CDEBUG(D_INODE, "statfs\n"); + req_capsule_set(&req->rq_pill, &RQF_OST_STATFS); if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_NET)) RETURN(0); rc = ost_statfs(req); break; case OST_SYNC: CDEBUG(D_INODE, "sync\n"); + req_capsule_set(&req->rq_pill, &RQF_OST_SYNC); if (OBD_FAIL_CHECK(OBD_FAIL_OST_SYNC_NET)) RETURN(0); rc = ost_sync(req->rq_export, req); break; case OST_SET_INFO: DEBUG_REQ(D_INODE, req, "set_info"); - req_capsule_set(&req->rq_pill, &RQF_OST_SET_INFO); + req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO); rc = ost_set_info(req->rq_export, req); break; case OST_GET_INFO: DEBUG_REQ(D_INODE, req, "get_info"); + req_capsule_set(&req->rq_pill, &RQF_OST_GET_INFO_GENERIC); rc = ost_get_info(req->rq_export, req); break; #ifdef HAVE_QUOTA_SUPPORT @@ -1930,14 +2314,14 @@ static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg) int rc; ENTRY; - rc = cleanup_group_info(); + rc = cfs_cleanup_group_info(); if (rc) RETURN(rc); lprocfs_ost_init_vars(&lvars); lprocfs_obd_setup(obd, lvars.obd_vars); - sema_init(&ost->ost_health_sem, 1); + cfs_sema_init(&ost->ost_health_sem, 1); if (oss_num_threads) { /* If oss_num_threads is set, it is the min and the max. */ @@ -1948,14 +2332,15 @@ static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg) 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 = + cfs_num_possible_cpus() * CFS_NUM_CACHEPAGES >> (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) 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 = @@ -1965,7 +2350,7 @@ 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); @@ -1994,7 +2379,7 @@ 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); @@ -2011,7 +2396,7 @@ 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); @@ -2050,20 +2435,20 @@ static int ost_cleanup(struct obd_device *obd) ping_evictor_stop(); - spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock_bh(&obd->obd_processing_task_lock); if (obd->obd_recovering) { target_cancel_recovery_timer(obd); obd->obd_recovering = 0; } - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock_bh(&obd->obd_processing_task_lock); - down(&ost->ost_health_sem); + cfs_down(&ost->ost_health_sem); ptlrpc_unregister_service(ost->ost_service); ptlrpc_unregister_service(ost->ost_create_service); ptlrpc_unregister_service(ost->ost_io_service); ost->ost_service = NULL; ost->ost_create_service = NULL; - up(&ost->ost_health_sem); + cfs_up(&ost->ost_health_sem); lprocfs_obd_cleanup(obd); @@ -2075,11 +2460,11 @@ static int ost_health_check(struct obd_device *obd) struct ost_obd *ost = &obd->u.ost; int rc = 0; - down(&ost->ost_health_sem); + cfs_down(&ost->ost_health_sem); rc |= ptlrpc_service_health_check(ost->ost_service); rc |= ptlrpc_service_health_check(ost->ost_create_service); rc |= ptlrpc_service_health_check(ost->ost_io_service); - up(&ost->ost_health_sem); + cfs_up(&ost->ost_health_sem); /* * health_check to return 0 on healthy