X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fost%2Fost_handler.c;h=e72d6d1610e940f6aa7fa3d8ebc877591d92eb71;hp=a945bda2f0760a94edb3a5dce54b20a7e68c5d3e;hb=313f87df046b0dfe31d56b902bb66d9bf98d8709;hpb=3912ebcd3faedb3557a945ede8b7357632eeb3e2 diff --git a/lustre/ost/ost_handler.c b/lustre/ost/ost_handler.c index a945bda..e72d6d1 100644 --- a/lustre/ost/ost_handler.c +++ b/lustre/ost/ost_handler.c @@ -5,20 +5,23 @@ * Author: Peter J. Braam * Author: Phil Schwan * - * This file is part of Lustre, http://www.lustre.org. + * This file is part of the Lustre file system, http://www.lustre.org + * Lustre is a trademark of Cluster File Systems, Inc. * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * You may have signed or agreed to another license before downloading + * this software. If so, you are bound by the terms and conditions + * of that agreement, and the following does not apply to you. See the + * LICENSE file included with this distribution for more information. * - * Lustre is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * If you did not agree to a different license, then this copy of Lustre + * is open source software; you can redistribute it and/or modify it + * under the terms of version 2 of the GNU General Public License as + * published by the Free Software Foundation. * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * In either case, Lustre is distributed in the hope that it will be + * useful, but WITHOUT ANY WARRANTY; without even the implied warranty + * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * license text for more details. * * Storage Target Handling functions * Lustre Object Server Module (OST) @@ -36,26 +39,29 @@ #define DEBUG_SUBSYSTEM S_OST #include -#include -#include -#include -#include +#include +#include +#include +#include +#include #include -#include -#include +#include +#include #include -#include -#include +#include +#include "ost_internal.h" -void oti_init(struct obd_trans_info *oti, struct ptlrpc_request *req) -{ - if (oti == NULL) - return; - memset(oti, 0, sizeof *oti); - oti->oti_nid = req->rq_peer.peer_id.nid; - if (req->rq_repmsg && req->rq_reqmsg != 0) - oti->oti_transno = req->rq_repmsg->transno; -} +static int oss_num_threads; +CFS_MODULE_PARM(oss_num_threads, "i", int, 0444, + "number of OSS service threads to start"); + +static int ost_num_threads; +CFS_MODULE_PARM(ost_num_threads, "i", int, 0444, + "number of OST service threads to start (deprecated)"); + +static int oss_num_create_threads; +CFS_MODULE_PARM(oss_num_create_threads, "i", int, 0444, + "number of OSS create threads to start"); void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req) { @@ -66,7 +72,8 @@ void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req) return; if (req->rq_repmsg) - req->rq_repmsg->transno = oti->oti_transno; + lustre_msg_set_transno(req->rq_repmsg, oti->oti_transno); + req->rq_transno = oti->oti_transno; /* XXX 4 == entries in oti_ack_locks??? */ for (ack_lock = oti->oti_ack_locks, i = 0; i < 4; i++, ack_lock++) { @@ -77,62 +84,84 @@ void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req) } } -static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req, +static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req, struct obd_trans_info *oti) { struct ost_body *body, *repbody; - int rc, size = sizeof(*body); + int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) }; ENTRY; - body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body); + body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), + lustre_swab_ost_body); if (body == NULL) RETURN(-EFAULT); - rc = lustre_pack_reply(req, 1, &size, NULL); + if (lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1)) { + struct ldlm_request *dlm; + dlm = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*dlm), + lustre_swab_ldlm_request); + if (dlm == NULL) + RETURN (-EFAULT); + ldlm_request_cancel(req, dlm, 0); + } + + 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); - repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody)); + 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); + req->rq_status = obd_destroy(exp, &body->oa, NULL, oti, NULL); RETURN(0); } static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req) { struct ost_body *body, *repbody; - int rc, size = sizeof(*body); + struct obd_info oinfo = { { { 0 } } }; + int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) }; ENTRY; - body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body); + body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), + lustre_swab_ost_body); if (body == NULL) RETURN(-EFAULT); - rc = lustre_pack_reply(req, 1, &size, NULL); + rc = lustre_pack_reply(req, 2, size, NULL); if (rc) RETURN(rc); - repbody = lustre_msg_buf (req->rq_repmsg, 0, sizeof(*repbody)); - memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); - req->rq_status = obd_getattr(exp, &repbody->oa, NULL); + repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, + sizeof(*repbody)); + 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); + req->rq_status = obd_getattr(exp, &oinfo); RETURN(0); } static int ost_statfs(struct ptlrpc_request *req) { struct obd_statfs *osfs; - int rc, size = sizeof(*osfs); + int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*osfs) }; ENTRY; - rc = lustre_pack_reply(req, 1, &size, NULL); + rc = lustre_pack_reply(req, 2, size, NULL); if (rc) RETURN(rc); - osfs = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*osfs)); + osfs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*osfs)); - req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs, jiffies-HZ); + req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs, + cfs_time_current_64() - HZ); + if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_ENOSPC)) + osfs->os_bfree = osfs->os_bavail = 64; if (req->rq_status != 0) CERROR("ost: statfs failed: rc %d\n", req->rq_status); @@ -143,101 +172,194 @@ static int ost_create(struct obd_export *exp, struct ptlrpc_request *req, struct obd_trans_info *oti) { struct ost_body *body, *repbody; - int rc, size = sizeof(*repbody); + int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) }; ENTRY; - body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body); + body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), + lustre_swab_ost_body); if (body == NULL) RETURN(-EFAULT); - rc = lustre_pack_reply(req, 1, &size, NULL); + rc = lustre_pack_reply(req, 2, size, NULL); if (rc) RETURN(rc); - repbody = lustre_msg_buf (req->rq_repmsg, 0, sizeof(*repbody)); + 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); - req->rq_status = obd_create(exp, &repbody->oa, NULL, 0, NULL, oti); + req->rq_status = obd_create(exp, &repbody->oa, NULL, oti); //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0); RETURN(0); } -static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req, +/* + * 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 = { .name = { oa->o_id, 0, oa->o_gr, 0} }; + ldlm_policy_data_t policy; + __u64 start; + __u64 finis; + + ENTRY; + + 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, 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; - struct lustre_capa *capa = NULL; - int rc, size = sizeof(*repbody); + int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) }; + struct lustre_handle lh = {0,}; ENTRY; - body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body); + /* 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); - if ((body->oa.o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) != + oinfo.oi_oa = &body->oa; + oinfo.oi_policy.l_extent.start = oinfo.oi_oa->o_size; + oinfo.oi_policy.l_extent.end = oinfo.oi_oa->o_blocks; + + if ((oinfo.oi_oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) != (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) RETURN(-EINVAL); - if (body->oa.o_valid & OBD_MD_CAPA) { - capa = lustre_swab_reqbuf(req, 1, sizeof(*capa), - lustre_swab_lustre_capa); - if (capa == NULL) { - CERROR("Missing/short capa\n"); - RETURN(-EFAULT); - } - } - - rc = lustre_pack_reply(req, 1, &size, NULL); + rc = lustre_pack_reply(req, 2, size, NULL); if (rc) RETURN(rc); - repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody)); - memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); - req->rq_status = obd_punch(exp, &repbody->oa, NULL, repbody->oa.o_size, - repbody->oa.o_blocks, oti, capa); - RETURN(0); + repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, + sizeof(*repbody)); + rc = ost_punch_lock_get(exp, oinfo.oi_oa, &lh); + if (rc == 0) { + if (oinfo.oi_oa->o_valid & OBD_MD_FLFLAGS && + oinfo.oi_oa->o_flags == OBD_FL_TRUNCLOCK) + /* + * If OBD_FL_TRUNCLOCK 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); + req->rq_status = obd_punch(exp, &oinfo, oti, NULL); + ost_punch_lock_put(exp, oinfo.oi_oa, &lh); + } + repbody->oa = *oinfo.oi_oa; + RETURN(rc); } static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req) { struct ost_body *body, *repbody; - int rc, size = sizeof(*repbody); + struct lustre_capa *capa = NULL; + int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) }; ENTRY; - body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body); + body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), + lustre_swab_ost_body); if (body == NULL) RETURN(-EFAULT); - rc = lustre_pack_reply(req, 1, &size, NULL); + if (body->oa.o_valid & OBD_MD_FLOSSCAPA) + capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 1); + + rc = lustre_pack_reply(req, 2, size, NULL); if (rc) RETURN(rc); - repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody)); + repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, + sizeof(*repbody)); 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); + repbody->oa.o_blocks, capa); RETURN(0); } -static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req, +static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req, struct obd_trans_info *oti) { struct ost_body *body, *repbody; - int rc, size = sizeof(*repbody); + int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) }; + struct obd_info oinfo = { { { 0 } } }; ENTRY; - body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body); + body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body), + lustre_swab_ost_body); if (body == NULL) RETURN(-EFAULT); - rc = lustre_pack_reply(req, 1, &size, NULL); + rc = lustre_pack_reply(req, 2, size, NULL); if (rc) RETURN(rc); - repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody)); - memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); + repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, + sizeof(*repbody)); + repbody->oa = body->oa; - req->rq_status = obd_setattr(exp, &repbody->oa, NULL, oti); + 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); + req->rq_status = obd_setattr(exp, &oinfo, oti); RETURN(0); } @@ -263,12 +385,19 @@ static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo, 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 >> PAGE_SHIFT; - obd_off pn = (offset + rnb[rnbidx].len - 1)>>PAGE_SHIFT; + obd_off p0 = offset >> CFS_PAGE_SHIFT; + obd_off pn = (offset + rnb[rnbidx].len - 1) >> + CFS_PAGE_SHIFT; LASSERT(rnbidx < nrnb); @@ -296,9 +425,7 @@ static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo, return npages; } - OBD_ALLOC(pp_rnb, sizeof(*pp_rnb) * npages); - if (pp_rnb == NULL) - return -ENOMEM; + pp_rnb = *pp_rnbp; /* now do the actual split */ page = rnbidx = 0; @@ -311,8 +438,8 @@ static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo, LASSERT(rnbidx < nrnb); do { - obd_off poff = off & (PAGE_SIZE - 1); - int pnob = (poff + nob > PAGE_SIZE) ? + obd_off poff = off & ~CFS_PAGE_MASK; + int pnob = (poff + nob > CFS_PAGE_SIZE) ? PAGE_SIZE - poff : nob; LASSERT(page < npages); @@ -337,112 +464,245 @@ static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo, } LASSERT(page == npages); - *pp_rnbp = pp_rnb; return npages; } -static void free_per_page_niobufs (int npages, struct niobuf_remote *pp_rnb, - struct niobuf_remote *rnb) +static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc) { - if (pp_rnb == rnb) /* didn't allocate above */ - return; + __u32 cksum = ~0; + int i; + + for (i = 0; i < desc->bd_iov_count; i++) { + struct page *page = desc->bd_iov[i].kiov_page; + int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK; + char *ptr = kmap(page) + off; + int len = desc->bd_iov[i].kiov_len; + + /* corrupt the data before we compute the checksum, to + * simulate a client->OST data error */ + if (i == 0 && + OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_CHECKSUM_RECEIVE)) + memcpy(ptr, "bad3", min(4, len)); + cksum = crc32_le(cksum, ptr, len); + /* corrupt the data after we compute the checksum, to + * simulate an OST->client data error */ + if (i == 0 && OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_CHECKSUM_SEND)) + memcpy(ptr, "bad4", min(4, len)); + kunmap(page); + } - OBD_FREE(pp_rnb, sizeof(*pp_rnb) * npages); + return cksum; } -#if CHECKSUM_BULK -obd_count ost_checksum_bulk(struct ptlrpc_bulk_desc *desc) +/* + * 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) { - obd_count cksum = 0; - struct ptlrpc_bulk_page *bp; + int i; + struct ost_thread_local_cache *tls; - list_for_each_entry(bp, &desc->bd_page_list, bp_link) { - ost_checksum(&cksum, kmap(bp->bp_page) + bp->bp_pageoffset, - bp->bp_buflen); - kunmap(bp->bp_page); + 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; +} - return cksum; +/* + * 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; } -#endif -static void ost_stime_record(struct ptlrpc_request *req, struct timeval *start, - unsigned rw, unsigned phase) +static int ost_brw_lock_get(int mode, struct obd_export *exp, + struct obd_ioobj *obj, struct niobuf_remote *nb, + struct lustre_handle *lh) { - struct obd_device *obd = req->rq_svc->srv_obddev; - struct timeval stop; - int ind = rw *3 + phase; - - if (obd && obd->obd_type && obd->obd_type->typ_name) { - if (!strcmp(obd->obd_type->typ_name, OBD_OST_DEVICENAME)) { - struct ost_obd *ost = NULL; - - ost = &obd->u.ost; - if (ind >= (sizeof(ost->ost_stimes) / - sizeof(ost->ost_stimes[0]))) - return; - do_gettimeofday(&stop); - - spin_lock(&ost->ost_lock); - lprocfs_stime_record(&ost->ost_stimes[ind],&stop,start); - spin_unlock(&ost->ost_lock); - memcpy(start, &stop, sizeof(*start)); - } - } + int flags = 0; + int nrbufs = obj->ioo_bufcnt; + struct ldlm_res_id res_id = { .name = { obj->ioo_id, 0, + obj->ioo_gr, 0} }; + ldlm_policy_data_t policy; + int i; + + ENTRY; + + LASSERT(mode == LCK_PR || mode == LCK_PW); + LASSERT(!lustre_handle_is_used(lh)); + + 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)); + + policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK; + policy.l_extent.end = (nb[nrbufs - 1].offset + + nb[nrbufs - 1].len - 1) | ~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)); +} + +static void ost_brw_lock_put(int mode, + struct obd_ioobj *obj, struct niobuf_remote *niob, + struct lustre_handle *lh) +{ + ENTRY; + LASSERT(mode == LCK_PR || mode == LCK_PW); + LASSERT((obj->ioo_bufcnt > 0 && (niob[0].flags & OBD_BRW_SRVLOCK)) == + lustre_handle_is_used(lh)); + if (lustre_handle_is_used(lh)) + ldlm_lock_decref(lh, mode); + EXIT; } -static int ost_brw_read(struct ptlrpc_request *req) +struct ost_prolong_data { + struct obd_export *opd_exp; + ldlm_policy_data_t opd_policy; + ldlm_mode_t opd_mode; +}; + +static int ost_prolong_locks_iter(struct ldlm_lock *lock, void *data) +{ + struct ost_prolong_data *opd = data; + + LASSERT(lock->l_resource->lr_type == LDLM_EXTENT); + + if (lock->l_req_mode != lock->l_granted_mode) { + /* scan granted locks only */ + return LDLM_ITER_STOP; + } + + if (lock->l_export != opd->opd_exp) { + /* prolong locks only for given client */ + return LDLM_ITER_CONTINUE; + } + + if (!(lock->l_granted_mode & opd->opd_mode)) { + /* we aren't interesting in all type of locks */ + return LDLM_ITER_CONTINUE; + } + + if (lock->l_policy_data.l_extent.end < opd->opd_policy.l_extent.start || + lock->l_policy_data.l_extent.start > opd->opd_policy.l_extent.end) { + /* the request doesn't cross the lock, skip it */ + return LDLM_ITER_CONTINUE; + } + + if (!(lock->l_flags & LDLM_FL_AST_SENT)) { + /* ignore locks not being cancelled */ + return LDLM_ITER_CONTINUE; + } + + /* OK. this is a possible lock the user holds doing I/O + * let's refresh eviction timer for it */ + ldlm_refresh_waiting_lock(lock); + + 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) +{ + struct ldlm_res_id res_id = { .name = { obj->ioo_id, 0, + obj->ioo_gr, 0} }; + int nrbufs = obj->ioo_bufcnt; + struct ost_prolong_data opd; + + ENTRY; + + opd.opd_mode = mode; + opd.opd_exp = exp; + 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", + 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, + ost_prolong_locks_iter, &opd); +} + +static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) { struct ptlrpc_bulk_desc *desc; - struct niobuf_remote *remote_nb; - struct niobuf_remote *pp_rnb; - struct niobuf_local *local_nb; - struct obd_ioobj *ioo; - struct ost_body *body, *repbody; - struct lustre_capa *capa = NULL; - struct l_wait_info lwi; - struct obd_trans_info oti = { 0 }; - int size[1] = { sizeof(*body) }; - int comms_error = 0; - int niocount; - int npages; - int nob = 0; - int rc; - int i, bufcnt = 0; - struct timeval start; + 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; + struct lustre_capa *capa = NULL; + struct l_wait_info lwi; + struct lustre_handle lockh = { 0 }; + int size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) }; + int comms_error = 0, niocount, npages, nob = 0, rc, i; + int no_reply = 0; ENTRY; + req->rq_bulk_read = 1; + if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK)) GOTO(out, rc = -EIO); OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK | OBD_FAIL_ONCE, (obd_timeout + 1) / 4); - body = lustre_swab_reqbuf(req, bufcnt++, sizeof(*body), + 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); } - ioo = lustre_swab_reqbuf(req, bufcnt++, sizeof(*ioo), + 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); } - if (body->oa.o_valid & OBD_MD_CAPA) { - capa = lustre_swab_reqbuf(req, bufcnt++, sizeof(*capa), - lustre_swab_lustre_capa); - if (capa == NULL) { - CERROR("Missing/short capa\n"); - 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); + GOTO(out, rc = -EFAULT); } - niocount = ioo->ioo_bufcnt; - remote_nb = lustre_swab_reqbuf(req, bufcnt++, + remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2, niocount * sizeof(*remote_nb), lustre_swab_niobuf_remote); if (remote_nb == NULL) { @@ -454,34 +714,59 @@ static int ost_brw_read(struct ptlrpc_request *req) lustre_swab_niobuf_remote (&remote_nb[i]); } - rc = lustre_pack_reply(req, 1, size, NULL); + 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); if (rc) GOTO(out, rc); + /* + * Per-thread array of struct niobuf_{local,remote}'s was allocated by + * 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); - OBD_ALLOC(local_nb, sizeof(*local_nb) * npages); - if (local_nb == NULL) - GOTO(out_pp_rnb, rc = -ENOMEM); + LASSERT(npages <= OST_THREAD_POOL_SIZE); - desc = ptlrpc_prep_bulk_exp (req, npages, + 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_local, rc = -ENOMEM); + GOTO(out, rc = -ENOMEM); - do_gettimeofday(&start); - rc = obd_preprw(OBD_BRW_READ, req->rq_export, &body->oa, 1, - ioo, npages, pp_rnb, local_nb, &oti, capa); - ost_stime_record(req, &start, 0, 0); + rc = ost_brw_lock_get(LCK_PR, req->rq_export, ioo, pp_rnb, &lockh); if (rc != 0) GOTO(out_bulk, rc); - /* We're finishing using body->oa as an input variable */ - body->oa.o_valid = 0; + /* + * If getting the lock took more time than + * client was willing to wait, drop it. b=11330 + */ + if (cfs_time_current_sec() > req->rq_arrival_time.tv_sec + obd_timeout || + OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) { + no_reply = 1; + CERROR("Dropping timed-out read from %s because locking" + "object "LPX64" took %ld seconds.\n", + libcfs_id2str(req->rq_peer), ioo->ioo_id, + cfs_time_current_sec() - req->rq_arrival_time.tv_sec); + goto out_lock; + } + + rc = obd_preprw(OBD_BRW_READ, req->rq_export, &body->oa, 1, + ioo, npages, pp_rnb, local_nb, oti, capa); + if (rc != 0) + GOTO(out_lock, rc); + + ost_prolong_locks(req->rq_export, ioo, pp_rnb, LCK_PW | LCK_PR); nob = 0; for (i = 0; i < npages; i++) { @@ -492,12 +777,13 @@ static int ost_brw_read(struct ptlrpc_request *req) break; } - LASSERT(page_rc <= pp_rnb[i].len); + 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 & (PAGE_SIZE-1), + pp_rnb[i].offset & ~CFS_PAGE_MASK, page_rc); } @@ -509,27 +795,39 @@ static int ost_brw_read(struct ptlrpc_request *req) } } + if (unlikely(body->oa.o_valid & OBD_MD_FLCKSUM)) { + body->oa.o_cksum = ost_checksum_bulk(desc); + body->oa.o_valid = OBD_MD_FLCKSUM; + CDEBUG(D_PAGE,"checksum at read origin: %x\n",body->oa.o_cksum); + } else { + body->oa.o_valid = 0; + } + /* We're finishing using body->oa as an input variable */ + + /* Check if client was evicted while we were doing i/o before touching + network */ if (rc == 0) { - rc = ptlrpc_start_bulk_transfer(desc); + if (desc->bd_export->exp_failed) + rc = -ENOTCONN; + else { + sptlrpc_svc_wrap_bulk(req, desc); + + rc = ptlrpc_start_bulk_transfer(desc); + } + if (rc == 0) { - struct timeval tstart, now; - do_gettimeofday(&tstart); - lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, - ost_bulk_timeout, desc); + lwi = LWI_TIMEOUT_INTERVAL(obd_timeout * HZ / 4, HZ, + ost_bulk_timeout, desc); rc = l_wait_event(desc->bd_waitq, - !ptlrpc_bulk_active(desc), &lwi); + !ptlrpc_bulk_active(desc) || + desc->bd_export->exp_failed, &lwi); LASSERT(rc == 0 || rc == -ETIMEDOUT); - do_gettimeofday(&now); if (rc == -ETIMEDOUT) { - DEBUG_REQ(D_ERROR, req, "timeout on bulk PUT" - ", expt_conn_cnt = %u, real wait %us" - ", arrived %u.%u, served %u.%u", - req->rq_export->exp_conn_cnt, - (unsigned) (now.tv_sec - tstart.tv_sec), - (unsigned) req->rq_arrival_time.tv_sec, - (unsigned) req->rq_arrival_time.tv_usec, - (unsigned) req->rq_rpcd_start.tv_sec, - (unsigned) req->rq_rpcd_start.tv_usec); + DEBUG_REQ(D_ERROR, req, "timeout on bulk PUT"); + ptlrpc_abort_bulk(desc); + } else if (desc->bd_export->exp_failed) { + DEBUG_REQ(D_ERROR, req, "Eviction on bulk PUT"); + rc = -ENOTCONN; ptlrpc_abort_bulk(desc); } else if (!desc->bd_success || desc->bd_nob_transferred != desc->bd_nob) { @@ -542,84 +840,78 @@ static int ost_brw_read(struct ptlrpc_request *req) rc = -ETIMEDOUT; } } else { - DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d\n", rc); + DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d", rc); } comms_error = rc != 0; } - ost_stime_record(req, &start, 0, 1); /* Must commit after prep above in all cases */ rc = obd_commitrw(OBD_BRW_READ, req->rq_export, &body->oa, 1, - ioo, npages, local_nb, &oti, rc); - ost_stime_record(req, &start, 0, 2); + ioo, npages, local_nb, oti, rc); + + ost_nio_pages_put(req, local_nb, npages); if (rc == 0) { - repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody)); + repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, + sizeof(*repbody)); memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa)); - -#if CHECKSUM_BULK - repbody->oa.o_cksum = ost_checksum_bulk(desc); - repbody->oa.o_valid |= OBD_MD_FLCKSUM; -#endif } - out_bulk: +out_lock: + ost_brw_lock_put(LCK_PR, ioo, pp_rnb, &lockh); +out_bulk: ptlrpc_free_bulk(desc); - out_local: - OBD_FREE(local_nb, sizeof(*local_nb) * npages); - out_pp_rnb: - free_per_page_niobufs(npages, pp_rnb, remote_nb); - out: + if (no_reply) + RETURN(rc); +out: LASSERT(rc <= 0); if (rc == 0) { req->rq_status = nob; + target_committed_to_req(req); ptlrpc_reply(req); } else if (!comms_error) { - /* only reply if comms OK */ + /* Only reply if there was no comms problem with bulk */ + target_committed_to_req(req); req->rq_status = rc; ptlrpc_error(req); } else { if (req->rq_reply_state != NULL) { /* reply out callback would free */ - lustre_free_reply_state (req->rq_reply_state); - } - if (req->rq_reqmsg->conn_cnt == req->rq_export->exp_conn_cnt) { - CERROR("bulk IO comms error: " - "evicting %s@%s id %s\n", - req->rq_export->exp_client_uuid.uuid, - req->rq_export->exp_connection->c_remote_uuid.uuid, - req->rq_peerstr); - ptlrpc_fail_export(req->rq_export); - } else { - CERROR("ignoring bulk IO comms error: " - "client reconnected %s@%s id %s\n", - req->rq_export->exp_client_uuid.uuid, - req->rq_export->exp_connection->c_remote_uuid.uuid, - req->rq_peerstr); + ptlrpc_rs_decref(req->rq_reply_state); + req->rq_reply_state = NULL; } + CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - " + "client will retry\n", + req->rq_export->exp_obd->obd_name, + req->rq_export->exp_client_uuid.uuid, + req->rq_export->exp_connection->c_remote_uuid.uuid, + libcfs_id2str(req->rq_peer)); } RETURN(rc); } -int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) +static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) { struct ptlrpc_bulk_desc *desc; struct niobuf_remote *remote_nb; struct niobuf_remote *pp_rnb; struct niobuf_local *local_nb; struct obd_ioobj *ioo; - struct lustre_capa *capa = NULL; struct ost_body *body, *repbody; struct l_wait_info lwi; + struct lustre_handle lockh = {0}; + struct lustre_capa *capa = NULL; __u32 *rcs; - int size[2] = { sizeof(*body) }; - int objcount, niocount, npages; - int comms_error = 0; - int rc, swab, i, j, bufcnt = 0; - struct timeval start; + int size[3] = { sizeof(struct ptlrpc_body), sizeof(*body) }; + int objcount, niocount, npages, comms_error = 0; + int rc, swab, i, j; + obd_count client_cksum, server_cksum = 0; + int no_reply = 0; ENTRY; + req->rq_bulk_write = 1; + if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK)) GOTO(out, rc = -EIO); @@ -628,25 +920,31 @@ int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) (obd_timeout + 1) / 4); swab = lustre_msg_swabbed(req->rq_reqmsg); - body = lustre_swab_reqbuf(req, bufcnt++, sizeof(*body), + 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); } - LASSERT_REQSWAB(req, 1); - objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo); + LASSERT_REQSWAB(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); } - ioo = lustre_msg_buf(req->rq_reqmsg, bufcnt++, + if (objcount > 1) { + CERROR("too many ioobjs (%d)\n", objcount); + GOTO(out, rc = -EFAULT); + } + + 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]); + lustre_swab_obd_ioobj(&ioo[i]); if (ioo[i].ioo_bufcnt == 0) { CERROR("ioo[%d] has zero bufcnt\n", i); GOTO(out, rc = -EFAULT); @@ -654,16 +952,13 @@ int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) niocount += ioo[i].ioo_bufcnt; } - if (body->oa.o_valid & OBD_MD_CAPA) { - capa = lustre_swab_reqbuf(req, bufcnt++, sizeof(*capa), - lustre_swab_lustre_capa); - if (capa == NULL) { - CERROR("Missing/short capa\n"); - GOTO(out, rc = -EFAULT); - } + 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, bufcnt++, + remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2, niocount * sizeof(*remote_nb), lustre_swab_niobuf_remote); if (remote_nb == NULL) { @@ -675,18 +970,22 @@ int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) lustre_swab_niobuf_remote (&remote_nb[i]); } - size[1] = niocount * sizeof(*rcs); - rc = lustre_pack_reply(req, 2, size, NULL); + 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); if (rc != 0) GOTO(out, rc); - rcs = lustre_msg_buf(req->rq_repmsg, 1, niocount * sizeof(*rcs)); + rcs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, + niocount * sizeof(*rcs)); -#if 0 - /* Do snap options here*/ - rc = obd_do_cow(req->rq_export, ioo, objcount, remote_nb); - if (rc) - GOTO(out, rc); -#endif + /* + * Per-thread array of struct niobuf_{local,remote}'s was allocated by + * 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 */ @@ -694,87 +993,162 @@ int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) if (npages < 0) GOTO(out, rc = npages); - OBD_ALLOC(local_nb, sizeof(*local_nb) * npages); - if (local_nb == NULL) - GOTO(out_pp_rnb, rc = -ENOMEM); + LASSERT(npages <= OST_THREAD_POOL_SIZE); + + ost_nio_pages_get(req, local_nb, npages); - desc = ptlrpc_prep_bulk_exp (req, npages, + desc = ptlrpc_prep_bulk_exp(req, npages, BULK_GET_SINK, OST_BULK_PORTAL); if (desc == NULL) - GOTO(out_local, rc = -ENOMEM); + GOTO(out, rc = -ENOMEM); + + rc = ost_brw_lock_get(LCK_PW, req->rq_export, ioo, pp_rnb, &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 + */ + if (cfs_time_current_sec() > req->rq_arrival_time.tv_sec + obd_timeout || + OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) { + no_reply = 1; + CERROR("Dropping timed-out write from %s because locking" + "object "LPX64" took %ld seconds.\n", + libcfs_id2str(req->rq_peer), ioo->ioo_id, + cfs_time_current_sec() - req->rq_arrival_time.tv_sec); + goto out_lock; + } + + ost_prolong_locks(req->rq_export, ioo, pp_rnb, LCK_PW); + + /* obd_preprw clobbers oa->valid, so save what we need */ + client_cksum = body->oa.o_valid & OBD_MD_FLCKSUM ? body->oa.o_cksum : 0; + + /* 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 */ + if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) { + DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info\n"); + body->oa.o_valid &= ~OBD_MD_FLGRANT; + } - do_gettimeofday(&start); rc = obd_preprw(OBD_BRW_WRITE, req->rq_export, &body->oa, objcount, ioo, npages, pp_rnb, local_nb, oti, capa); - ost_stime_record(req, &start, 1, 0); if (rc != 0) - GOTO(out_bulk, rc); + GOTO(out_lock, rc); /* 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 & (PAGE_SIZE - 1), + ptlrpc_prep_bulk_page(desc, local_nb[i].page, + pp_rnb[i].offset & ~CFS_PAGE_MASK, pp_rnb[i].len); - rc = ptlrpc_start_bulk_transfer (desc); + /* 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); if (rc == 0) { - lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, - ost_bulk_timeout, desc); - rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), - &lwi); + lwi = LWI_TIMEOUT_INTERVAL(obd_timeout * HZ / 2, HZ, + ost_bulk_timeout, desc); + rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc) || + desc->bd_export->exp_failed, &lwi); LASSERT(rc == 0 || rc == -ETIMEDOUT); if (rc == -ETIMEDOUT) { DEBUG_REQ(D_ERROR, req, "timeout on bulk GET"); ptlrpc_abort_bulk(desc); + } else if (desc->bd_export->exp_failed) { + 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 ? + desc->bd_success ? "truncated" : "network error on", desc->bd_nob_transferred, desc->bd_nob); /* XXX should this be a different errno? */ rc = -ETIMEDOUT; } } else { - DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d\n", rc); + DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d", rc); } comms_error = rc != 0; - repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody)); + repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, + sizeof(*repbody)); memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa)); -#if CHECKSUM_BULK - if (rc == 0 && (body->oa.o_valid & OBD_MD_FLCKSUM) != 0) { + if (unlikely(client_cksum != 0 && rc == 0)) { static int cksum_counter; - obd_count client_cksum = body->oa.o_cksum; - obd_count cksum = ost_checksum_bulk(desc); - - if (client_cksum != cksum) { - CERROR("Bad checksum: client %x, server %x id %s\n", - client_cksum, cksum, - req->rq_peerstr); - cksum_counter = 1; - repbody->oa.o_cksum = cksum; - } else { - cksum_counter++; - if ((cksum_counter & (-cksum_counter)) == cksum_counter) - CWARN("Checksum %u from NID %s: %x OK\n", - cksum_counter, req->rq_peerstr, cksum); + server_cksum = ost_checksum_bulk(desc); + repbody->oa.o_valid |= OBD_MD_FLCKSUM; + repbody->oa.o_cksum = server_cksum; + cksum_counter++; + if (unlikely(client_cksum != server_cksum)) { + CERROR("client csum %x, server csum %x\n", + client_cksum, server_cksum); + cksum_counter = 0; + } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){ + CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n", + cksum_counter, libcfs_id2str(req->rq_peer), + server_cksum); } } -#endif - ost_stime_record(req, &start, 1, 1); + + sptlrpc_svc_unwrap_bulk(req, desc); + /* Must commit after prep above in all cases */ rc = obd_commitrw(OBD_BRW_WRITE, req->rq_export, &repbody->oa, - objcount, ioo, npages, local_nb, oti, rc); + objcount, ioo, npages, local_nb, oti, rc); + + if (unlikely(client_cksum != server_cksum && rc == 0)) { + int new_cksum = ost_checksum_bulk(desc); + char *msg; + char *via; + char *router; + + if (new_cksum == server_cksum) + msg = "changed in transit before arrival at OST"; + else if (new_cksum == client_cksum) + msg = "initial checksum before message complete"; + else + msg = "changed in transit AND after initial checksum"; + + if (req->rq_peer.nid == desc->bd_sender) { + via = router = ""; + } else { + 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", + req->rq_export->exp_obd->obd_name, msg, + libcfs_id2str(req->rq_peer), + via, router, + body->oa.o_valid & OBD_MD_FLFID ? + body->oa.o_fid : (__u64)0, + body->oa.o_valid & OBD_MD_FLFID ? + body->oa.o_generation :(__u64)0, + 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 ); + 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); - ost_stime_record(req, &start, 1, 2); if (rc == 0) { -#if CHECKSUM_BULK - repbody->oa.o_cksum = ost_checksum_bulk(desc); - repbody->oa.o_valid |= OBD_MD_FLCKSUM; -#endif /* set per-requested niobuf return codes */ for (i = j = 0; i < niocount; i++) { int nob = remote_nb[i].len; @@ -791,208 +1165,162 @@ int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) } LASSERT(j == npages); } - /*XXX This write extents only for write-back cache extents*/ - rc = obd_write_extents(req->rq_export, ioo, objcount, niocount, - local_nb, rc); - out_bulk: + +out_lock: + ost_brw_lock_put(LCK_PW, ioo, pp_rnb, &lockh); +out_bulk: ptlrpc_free_bulk(desc); - out_local: - OBD_FREE(local_nb, sizeof(*local_nb) * npages); - out_pp_rnb: - free_per_page_niobufs(npages, pp_rnb, remote_nb); - out: + if (no_reply) + RETURN(rc); +out: if (rc == 0) { oti_to_request(oti, req); + target_committed_to_req(req); rc = ptlrpc_reply(req); } else if (!comms_error) { /* Only reply if there was no comms problem with bulk */ + target_committed_to_req(req); req->rq_status = rc; ptlrpc_error(req); } else { if (req->rq_reply_state != NULL) { /* reply out callback would free */ - lustre_free_reply_state (req->rq_reply_state); - } - if (req->rq_reqmsg->conn_cnt == req->rq_export->exp_conn_cnt) { - CERROR("%s: bulk IO comm error evicting %s@%s id %s\n", - req->rq_export->exp_obd->obd_name, - req->rq_export->exp_client_uuid.uuid, - req->rq_export->exp_connection->c_remote_uuid.uuid, - req->rq_peerstr); - ptlrpc_fail_export(req->rq_export); - } else { - CERROR("ignoring bulk IO comms error: " - "client reconnected %s@%s id %s\n", - req->rq_export->exp_client_uuid.uuid, - req->rq_export->exp_connection->c_remote_uuid.uuid, - req->rq_peerstr); + ptlrpc_rs_decref(req->rq_reply_state); + req->rq_reply_state = NULL; } + CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - " + "client will retry\n", + req->rq_export->exp_obd->obd_name, + req->rq_export->exp_client_uuid.uuid, + req->rq_export->exp_connection->c_remote_uuid.uuid, + libcfs_id2str(req->rq_peer)); } RETURN(rc); } -EXPORT_SYMBOL(ost_brw_write); - -static int ost_san_brw(struct ptlrpc_request *req, int cmd) -{ - struct niobuf_remote *remote_nb, *res_nb, *pp_rnb; - struct obd_ioobj *ioo; - struct ost_body *body, *repbody; - int rc, i, objcount, niocount, size[2] = {sizeof(*body)}, npages; - int swab; - ENTRY; - - /* XXX not set to use latest protocol */ - - swab = lustre_msg_swabbed(req->rq_reqmsg); - body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body); - if (body == NULL) { - CERROR("Missing/short ost_body\n"); - GOTO(out, rc = -EFAULT); - } - - ioo = lustre_swab_reqbuf(req, 1, sizeof(*ioo), lustre_swab_obd_ioobj); - if (ioo == NULL) { - CERROR("Missing/short ioobj\n"); - GOTO(out, rc = -EFAULT); - } - objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo); - niocount = ioo[0].ioo_bufcnt; - for (i = 1; i < objcount; i++) { - if (swab) - lustre_swab_obd_ioobj (&ioo[i]); - niocount += ioo[i].ioo_bufcnt; - } - - remote_nb = lustre_swab_reqbuf(req, 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]); - } - - /* 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); - - size[1] = npages * sizeof(*pp_rnb); - rc = lustre_pack_reply(req, 2, size, NULL); - if (rc) - GOTO(out_pp_rnb, rc); - - req->rq_status = obd_san_preprw(cmd, req->rq_export, &body->oa, - objcount, ioo, npages, pp_rnb); - - if (req->rq_status) - GOTO(out_pp_rnb, rc = 0); - - repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody)); - memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); - - res_nb = lustre_msg_buf(req->rq_repmsg, 1, size[1]); - memcpy(res_nb, remote_nb, size[1]); - rc = 0; -out_pp_rnb: - free_per_page_niobufs(npages, pp_rnb, remote_nb); -out: - if (rc) { - req->rq_status = rc; - ptlrpc_error(req); - } else - ptlrpc_reply(req); - - return rc; -} static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req) { - char *key, *val; - int keylen, rc = 0; + char *key, *val = NULL; + int keylen, vallen, rc = 0; ENTRY; - key = lustre_msg_buf(req->rq_reqmsg, 0, 1); + key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1); if (key == NULL) { DEBUG_REQ(D_HA, req, "no set_info key"); RETURN(-EFAULT); } - keylen = req->rq_reqmsg->buflens[0]; + keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF); - rc = lustre_pack_reply(req, 0, NULL, NULL); + rc = lustre_pack_reply(req, 1, NULL, NULL); if (rc) RETURN(rc); - val = lustre_msg_buf(req->rq_reqmsg, 1, 0); - if (val == NULL) { - CERROR("val for setinfo can't be NULL\n"); - RETURN(-EFAULT); - } - - if (keylen == 8 && memcmp(key, "auditlog", 8) == 0) { - lustre_swab_reqbuf(req, 1, sizeof(struct audit_msg), - lustre_swab_audit_msg); - } else if (keylen == 5 && strcmp(key, "audit") == 0) { - lustre_swab_reqbuf(req, 1, sizeof(struct audit_attr_msg), - lustre_swab_audit_attr); - } else if (keylen == 9 && strcmp(key, "audit_obj") == 0) { - lustre_swab_reqbuf(req, 1, sizeof(struct obdo), - lustre_swab_obdo); - } else if (keylen == 8 && strcmp(key, "capa_key") == 0) { - lustre_swab_reqbuf(req, 1, sizeof(struct lustre_capa_key), - lustre_swab_lustre_capa_key); + 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 (KEY_IS("evict_by_nid")) { + if (val && vallen) + obd_export_evict_by_nid(exp->exp_obd, val); + + GOTO(out, rc = 0); } - rc = obd_set_info(exp, keylen, key, req->rq_reqmsg->buflens[1], val); - req->rq_repmsg->status = 0; + rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL); +out: + lustre_msg_set_status(req->rq_repmsg, 0); RETURN(rc); } static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req) { char *key; - int keylen, rc = 0, size = sizeof(obd_id); + int keylen, rc = 0; + int size[2] = { sizeof(struct ptlrpc_body), sizeof(obd_id) }; obd_id *reply; ENTRY; - key = lustre_msg_buf(req->rq_reqmsg, 0, 1); + key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1); if (key == NULL) { DEBUG_REQ(D_HA, req, "no get_info key"); RETURN(-EFAULT); } - keylen = req->rq_reqmsg->buflens[0]; + keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF); if (keylen < strlen("last_id") || memcmp(key, "last_id", 7) != 0) RETURN(-EPROTO); - rc = lustre_pack_reply(req, 1, &size, NULL); + rc = lustre_pack_reply(req, 2, size, NULL); if (rc) RETURN(rc); - reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply)); - rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply); - req->rq_repmsg->status = 0; + reply = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply)); + rc = obd_get_info(exp, keylen, key, size, reply); + lustre_msg_set_status(req->rq_repmsg, 0); + RETURN(rc); +} + +static int ost_handle_quotactl(struct ptlrpc_request *req) +{ + struct obd_quotactl *oqctl, *repoqc; + int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) }; + ENTRY; + + oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl), + lustre_swab_obd_quotactl); + if (oqctl == NULL) + GOTO(out, rc = -EPROTO); + + rc = lustre_pack_reply(req, 2, size, NULL); + if (rc) + GOTO(out, rc); + + repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc)); + + req->rq_status = obd_quotactl(req->rq_export, oqctl); + *repoqc = *oqctl; +out: RETURN(rc); } +static int ost_handle_quotacheck(struct ptlrpc_request *req) +{ + struct obd_quotactl *oqctl; + int rc; + ENTRY; + + oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl), + lustre_swab_obd_quotactl); + if (oqctl == NULL) + RETURN(-EPROTO); + + rc = lustre_pack_reply(req, 1, NULL, NULL); + if (rc) { + CERROR("ost: out of memory while packing quotacheck reply\n"); + RETURN(-ENOMEM); + } + + req->rq_status = obd_quotacheck(req->rq_export, oqctl); + RETURN(0); +} + static int ost_llog_handle_connect(struct obd_export *exp, - struct ptlrpc_request *req) + struct ptlrpc_request *req) { struct llogd_conn_body *body; int rc; ENTRY; - body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body)); + body = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*body)); rc = obd_llog_connect(exp, body); RETURN(rc); } + static int ost_filter_recovery_request(struct ptlrpc_request *req, struct obd_device *obd, int *process) { - switch (req->rq_reqmsg->opc) { + switch (lustre_msg_get_opc(req->rq_reqmsg)) { case OST_CONNECT: /* This will never get here, but for completeness. */ case OST_DISCONNECT: *process = 1; @@ -1012,10 +1340,8 @@ static int ost_filter_recovery_request(struct ptlrpc_request *req, default: DEBUG_REQ(D_ERROR, req, "not permitted during recovery"); - *process = 0; - /* XXX what should we set rq_status to here? */ - req->rq_status = -EAGAIN; - RETURN(ptlrpc_error(req)); + *process = -EAGAIN; + RETURN(0); } } @@ -1023,27 +1349,39 @@ int ost_msg_check_version(struct lustre_msg *msg) { int rc; - switch(msg->opc) { + switch(lustre_msg_get_opc(msg)) { case OST_CONNECT: case OST_DISCONNECT: case OBD_PING: + case SEC_CTX_INIT: + case SEC_CTX_INIT_CONT: + case SEC_CTX_FINI: + rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION); + if (rc) + CERROR("bad opc %u version %08x, expecting %08x\n", + lustre_msg_get_opc(msg), + lustre_msg_get_version(msg), + LUSTRE_OBD_VERSION); + break; case OST_CREATE: case OST_DESTROY: case OST_GETATTR: case OST_SETATTR: case OST_WRITE: case OST_READ: - case OST_SAN_READ: - case OST_SAN_WRITE: case OST_PUNCH: case OST_STATFS: case OST_SYNC: case OST_SET_INFO: case OST_GET_INFO: - rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION); + case OST_QUOTACHECK: + case OST_QUOTACTL: + rc = lustre_msg_check_version(msg, LUSTRE_OST_VERSION); if (rc) CERROR("bad opc %u version %08x, expecting %08x\n", - msg->opc, msg->version, LUSTRE_OBD_VERSION); + lustre_msg_get_opc(msg), + lustre_msg_get_version(msg), + LUSTRE_OST_VERSION); break; case LDLM_ENQUEUE: case LDLM_CONVERT: @@ -1053,60 +1391,54 @@ int ost_msg_check_version(struct lustre_msg *msg) rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION); if (rc) CERROR("bad opc %u version %08x, expecting %08x\n", - msg->opc, msg->version, LUSTRE_DLM_VERSION); + lustre_msg_get_opc(msg), + lustre_msg_get_version(msg), + LUSTRE_DLM_VERSION); break; - case OBD_LOG_CANCEL: case LLOG_ORIGIN_CONNECT: + case OBD_LOG_CANCEL: rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION); if (rc) CERROR("bad opc %u version %08x, expecting %08x\n", - msg->opc, msg->version, LUSTRE_LOG_VERSION); - break; - case SEC_INIT: - case SEC_INIT_CONTINUE: - case SEC_FINI: - rc = 0; + lustre_msg_get_opc(msg), + lustre_msg_get_version(msg), + LUSTRE_LOG_VERSION); break; default: - CERROR("OST unexpected opcode %d\n", msg->opc); + CERROR("Unexpected opcode %d\n", lustre_msg_get_opc(msg)); rc = -ENOTSUPP; - break; } return rc; } int ost_handle(struct ptlrpc_request *req) { + struct obd_trans_info trans_info = { 0, }; + struct obd_trans_info *oti = &trans_info; int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0; - struct obd_trans_info *oti = NULL; struct obd_device *obd = NULL; ENTRY; LASSERT(current->journal_info == NULL); - rc = ost_msg_check_version(req->rq_reqmsg); - if (rc) { - CERROR("OST drop mal-formed request\n"); - RETURN(rc); - } - - /* Security opc should NOT trigger any recovery events */ - if (req->rq_reqmsg->opc == SEC_INIT || - req->rq_reqmsg->opc == SEC_INIT_CONTINUE || - req->rq_reqmsg->opc == SEC_FINI) { - GOTO(out_check_req, rc = 0); + /* primordial rpcs don't affect server recovery */ + switch (lustre_msg_get_opc(req->rq_reqmsg)) { + case SEC_CTX_INIT: + case SEC_CTX_INIT_CONT: + case SEC_CTX_FINI: + GOTO(out, rc = 0); } /* XXX identical to MDS */ - if (req->rq_reqmsg->opc != OST_CONNECT) { + if (lustre_msg_get_opc(req->rq_reqmsg) != OST_CONNECT) { int recovering; if (req->rq_export == NULL) { CDEBUG(D_HA,"operation %d on unconnected OST from %s\n", - req->rq_reqmsg->opc, - req->rq_peerstr); + lustre_msg_get_opc(req->rq_reqmsg), + libcfs_id2str(req->rq_peer)); req->rq_status = -ENOTCONN; - GOTO(out_check_req, rc = -ENOTCONN); + GOTO(out, rc = -ENOTCONN); } obd = req->rq_export->exp_obd; @@ -1115,12 +1447,12 @@ int ost_handle(struct ptlrpc_request *req) spin_lock_bh(&obd->obd_processing_task_lock); recovering = obd->obd_recovering; spin_unlock_bh(&obd->obd_processing_task_lock); - if (recovering) { + if (recovering) { rc = ost_filter_recovery_request(req, obd, &should_process); if (rc || !should_process) RETURN(rc); - if (should_process < 0) { + else if (should_process < 0) { req->rq_status = should_process; rc = ptlrpc_error(req); RETURN(rc); @@ -1128,16 +1460,16 @@ int ost_handle(struct ptlrpc_request *req) } } - OBD_ALLOC(oti, sizeof(*oti)); - if (oti == NULL) - RETURN(-ENOMEM); - oti_init(oti, req); + + rc = ost_msg_check_version(req->rq_reqmsg); + if (rc) + RETURN(rc); - switch (req->rq_reqmsg->opc) { + switch (lustre_msg_get_opc(req->rq_reqmsg)) { case OST_CONNECT: { CDEBUG(D_INODE, "connect\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_CONNECT_NET, out_free_oti, rc = 0); + OBD_FAIL_RETURN(OBD_FAIL_OST_CONNECT_NET, 0); rc = target_handle_connect(req); if (!rc) obd = req->rq_export->exp_obd; @@ -1145,74 +1477,84 @@ int ost_handle(struct ptlrpc_request *req) } case OST_DISCONNECT: CDEBUG(D_INODE, "disconnect\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_DISCONNECT_NET, out_free_oti, rc = 0); + OBD_FAIL_RETURN(OBD_FAIL_OST_DISCONNECT_NET, 0); rc = target_handle_disconnect(req); break; case OST_CREATE: CDEBUG(D_INODE, "create\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_ENOSPC, out_check_req, rc = -ENOSPC); - OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS); - OBD_FAIL_GOTO(OBD_FAIL_OST_CREATE_NET, out_free_oti, rc = 0); + OBD_FAIL_RETURN(OBD_FAIL_OST_CREATE_NET, 0); + if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_ENOSPC)) + GOTO(out, rc = -ENOSPC); + if (OBD_FAIL_CHECK_ONCE(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"); - OBD_FAIL_GOTO(OBD_FAIL_OST_DESTROY_NET, out_free_oti, rc = 0); - OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS); + OBD_FAIL_RETURN(OBD_FAIL_OST_DESTROY_NET, 0); + if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_EROFS)) + GOTO(out, rc = -EROFS); rc = ost_destroy(req->rq_export, req, oti); break; case OST_GETATTR: CDEBUG(D_INODE, "getattr\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_GETATTR_NET, out_free_oti, rc = 0); + OBD_FAIL_RETURN(OBD_FAIL_OST_GETATTR_NET, 0); rc = ost_getattr(req->rq_export, req); break; case OST_SETATTR: CDEBUG(D_INODE, "setattr\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_SETATTR_NET, out_free_oti, rc = 0); + OBD_FAIL_RETURN(OBD_FAIL_OST_SETATTR_NET, 0); rc = ost_setattr(req->rq_export, req, oti); break; case OST_WRITE: CDEBUG(D_INODE, "write\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0); - OBD_FAIL_GOTO(OBD_FAIL_OST_ENOSPC, out_check_req, rc = -ENOSPC); - OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS); + /* req->rq_request_portal would be nice, if it was set */ + if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){ + CERROR("%s: deny write request from %s to portal %u\n", + req->rq_export->exp_obd->obd_name, + obd_export_nid2str(req->rq_export), + req->rq_rqbd->rqbd_service->srv_req_portal); + GOTO(out, rc = -EPROTO); + } + OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0); + if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_ENOSPC)) + GOTO(out, rc = -ENOSPC); + if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_EROFS)) + GOTO(out, rc = -EROFS); rc = ost_brw_write(req, oti); LASSERT(current->journal_info == NULL); - /* ost_brw sends its own replies */ - GOTO(out_free_oti, rc); + /* ost_brw_write sends its own replies */ + RETURN(rc); case OST_READ: CDEBUG(D_INODE, "read\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0); - rc = ost_brw_read(req); + /* req->rq_request_portal would be nice, if it was set */ + if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){ + CERROR("%s: deny read request from %s to portal %u\n", + req->rq_export->exp_obd->obd_name, + obd_export_nid2str(req->rq_export), + req->rq_rqbd->rqbd_service->srv_req_portal); + GOTO(out, rc = -EPROTO); + } + OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0); + rc = ost_brw_read(req, oti); LASSERT(current->journal_info == NULL); - /* ost_brw sends its own replies */ - GOTO(out_free_oti, rc); - case OST_SAN_READ: - CDEBUG(D_INODE, "san read\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0); - rc = ost_san_brw(req, OBD_BRW_READ); - /* ost_san_brw sends its own replies */ - GOTO(out_free_oti, rc); - case OST_SAN_WRITE: - CDEBUG(D_INODE, "san write\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0); - rc = ost_san_brw(req, OBD_BRW_WRITE); - /* ost_san_brw sends its own replies */ - GOTO(out_free_oti, rc); + /* ost_brw_read sends its own replies */ + RETURN(rc); case OST_PUNCH: CDEBUG(D_INODE, "punch\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_PUNCH_NET, out_free_oti, rc = 0); - OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS); + OBD_FAIL_RETURN(OBD_FAIL_OST_PUNCH_NET, 0); + if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_EROFS)) + GOTO(out, rc = -EROFS); rc = ost_punch(req->rq_export, req, oti); break; case OST_STATFS: CDEBUG(D_INODE, "statfs\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_STATFS_NET, out_free_oti, rc = 0); + OBD_FAIL_RETURN(OBD_FAIL_OST_STATFS_NET, 0); rc = ost_statfs(req); break; case OST_SYNC: CDEBUG(D_INODE, "sync\n"); - OBD_FAIL_GOTO(OBD_FAIL_OST_SYNC_NET, out_free_oti, rc = 0); + OBD_FAIL_RETURN(OBD_FAIL_OST_SYNC_NET, 0); rc = ost_sync(req->rq_export, req); break; case OST_SET_INFO: @@ -1223,31 +1565,41 @@ int ost_handle(struct ptlrpc_request *req) DEBUG_REQ(D_INODE, req, "get_info"); rc = ost_get_info(req->rq_export, req); break; + case OST_QUOTACHECK: + CDEBUG(D_INODE, "quotacheck\n"); + OBD_FAIL_RETURN(OBD_FAIL_OST_QUOTACHECK_NET, 0); + rc = ost_handle_quotacheck(req); + break; + case OST_QUOTACTL: + CDEBUG(D_INODE, "quotactl\n"); + OBD_FAIL_RETURN(OBD_FAIL_OST_QUOTACTL_NET, 0); + rc = ost_handle_quotactl(req); + break; case OBD_PING: DEBUG_REQ(D_INODE, req, "ping"); rc = target_handle_ping(req); break; /* FIXME - just reply status */ case LLOG_ORIGIN_CONNECT: - DEBUG_REQ(D_INODE, req, "log connect\n"); - rc = ost_llog_handle_connect(req->rq_export, req); + DEBUG_REQ(D_INODE, req, "log connect"); + rc = ost_llog_handle_connect(req->rq_export, req); req->rq_status = rc; - rc = lustre_pack_reply(req, 0, NULL, NULL); + rc = lustre_pack_reply(req, 1, NULL, NULL); if (rc) - GOTO(out_free_oti, rc); - GOTO(out_free_oti, rc = ptlrpc_reply(req)); + RETURN(rc); + RETURN(ptlrpc_reply(req)); case OBD_LOG_CANCEL: CDEBUG(D_INODE, "log cancel\n"); - OBD_FAIL_GOTO(OBD_FAIL_OBD_LOG_CANCEL_NET, out_free_oti, rc = 0); + OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0); rc = llog_origin_handle_cancel(req); req->rq_status = rc; - rc = lustre_pack_reply(req, 0, NULL, NULL); + rc = lustre_pack_reply(req, 1, NULL, NULL); if (rc) - GOTO(out_free_oti, rc); - GOTO(out_free_oti, rc = ptlrpc_reply(req)); + RETURN(rc); + RETURN(ptlrpc_reply(req)); case LDLM_ENQUEUE: CDEBUG(D_INODE, "enqueue\n"); - OBD_FAIL_GOTO(OBD_FAIL_LDLM_ENQUEUE, out_free_oti, rc = 0); + OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0); rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast, ldlm_server_blocking_ast, ldlm_server_glimpse_ast); @@ -1255,12 +1607,12 @@ int ost_handle(struct ptlrpc_request *req) break; case LDLM_CONVERT: CDEBUG(D_INODE, "convert\n"); - OBD_FAIL_GOTO(OBD_FAIL_LDLM_CONVERT, out_free_oti, rc = 0); + OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0); rc = ldlm_handle_convert(req); break; case LDLM_CANCEL: CDEBUG(D_INODE, "cancel\n"); - OBD_FAIL_GOTO(OBD_FAIL_LDLM_CANCEL, out_free_oti, rc = 0); + OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0); rc = ldlm_handle_cancel(req); break; case LDLM_BL_CALLBACK: @@ -1269,60 +1621,103 @@ int ost_handle(struct ptlrpc_request *req) CERROR("callbacks should not happen on OST\n"); /* fall through */ default: - CERROR("Unexpected opcode %d\n", req->rq_reqmsg->opc); + CERROR("Unexpected opcode %d\n", + lustre_msg_get_opc(req->rq_reqmsg)); req->rq_status = -ENOTSUPP; rc = ptlrpc_error(req); - GOTO(out_free_oti, rc); + RETURN(rc); } LASSERT(current->journal_info == NULL); EXIT; /* If we're DISCONNECTing, the export_data is already freed */ - if (!rc && req->rq_reqmsg->opc != OST_DISCONNECT) { - if (!obd->obd_no_transno) { - req->rq_repmsg->last_committed = - obd->obd_last_committed; - } else { - DEBUG_REQ(D_IOCTL, req, - "not sending last_committed update"); - } - CDEBUG(D_INFO, "last_committed "LPU64", xid "LPX64"\n", - obd->obd_last_committed, req->rq_xid); - } - -out_check_req: + if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != OST_DISCONNECT) + target_committed_to_req(req); +out: if (!rc) oti_to_request(oti, req); + target_send_reply(req, rc, fail); - rc = 0; - -out_free_oti: - if (oti) - OBD_FREE(oti, sizeof(*oti)); - return rc; + return 0; } EXPORT_SYMBOL(ost_handle); - -int ost_attach(struct obd_device *dev, obd_count len, void *data) +/* + * free per-thread pool created by ost_thread_init(). + */ +static void ost_thread_done(struct ptlrpc_thread *thread) { - struct lprocfs_static_vars lvars; + int i; + struct ost_thread_local_cache *tls; /* TLS stands for Thread-Local + * Storage */ - lprocfs_init_vars(ost,&lvars); - return lprocfs_obd_attach(dev, lvars.obd_vars); + ENTRY; + + LASSERT(thread != NULL); + + /* + * be prepared to handle partially-initialized pools (because this is + * called from ost_thread_init() for cleanup. + */ + tls = thread->t_data; + if (tls != NULL) { + for (i = 0; i < OST_THREAD_POOL_SIZE; ++ i) { + if (tls->page[i] != NULL) + __cfs_free_page(tls->page[i]); + } + OBD_FREE_PTR(tls); + thread->t_data = NULL; + } + EXIT; } -int ost_detach(struct obd_device *dev) +/* + * initialize per-thread page pool (bug 5137). + */ +static int ost_thread_init(struct ptlrpc_thread *thread) { - return lprocfs_obd_detach(dev); + int result; + int i; + struct ost_thread_local_cache *tls; + + ENTRY; + + LASSERT(thread != NULL); + LASSERT(thread->t_data == NULL); + 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) { + tls->page[i] = cfs_alloc_page(OST_THREAD_POOL_GFP); + if (tls->page[i] == NULL) { + ost_thread_done(thread); + result = -ENOMEM; + break; + } + } + } else + result = -ENOMEM; + RETURN(result); } -extern struct file_operations ost_stimes_fops; +#define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000) -static int ost_setup(struct obd_device *obd, obd_count len, void *buf) +/* Sigh - really, this is an OSS, the _server_, not the _target_ */ +static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg) { struct ost_obd *ost = &obd->u.ost; + struct lprocfs_static_vars lvars; + int oss_min_threads; + int oss_max_threads; + int oss_min_create_threads; + int oss_max_create_threads; int rc; ENTRY; @@ -1334,76 +1729,124 @@ static int ost_setup(struct obd_device *obd, obd_count len, void *buf) if (rc < 0) RETURN(rc); - lprocfs_obd_seq_create(obd, "service_times", 0444, &ost_stimes_fops, - obd); + lprocfs_init_vars(ost, &lvars); + lprocfs_obd_setup(obd, lvars.obd_vars); + + sema_init(&ost->ost_health_sem, 1); + + if (oss_num_threads) { + /* If oss_num_threads is set, it is the min and the 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 = smp_num_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) + oss_min_threads = OSS_THREADS_MAX / 4; + oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4); + } ost->ost_service = ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE, - OST_REQUEST_PORTAL, OSC_REPLY_PORTAL, 30000, - ost_handle, "ost", - obd->obd_proc_entry); + OST_MAXREPSIZE, OST_REQUEST_PORTAL, + OSC_REPLY_PORTAL, + OST_WATCHDOG_TIMEOUT, ost_handle, + LUSTRE_OSS_NAME, obd->obd_proc_entry, + ost_print_req, oss_min_threads, + oss_max_threads, "ll_ost", + LCT_DT_THREAD); if (ost->ost_service == NULL) { CERROR("failed to start service\n"); - RETURN(-ENOMEM); + GOTO(out_lprocfs, rc = -ENOMEM); } - rc = ptlrpc_start_n_threads(obd, ost->ost_service, OST_NUM_THREADS, - "ll_ost"); + rc = ptlrpc_start_threads(obd, ost->ost_service); if (rc) GOTO(out_service, rc = -EINVAL); + if (oss_num_create_threads) { + if (oss_num_create_threads > OSS_MAX_CREATE_THREADS) + oss_num_create_threads = OSS_MAX_CREATE_THREADS; + if (oss_num_create_threads < OSS_DEF_CREATE_THREADS) + oss_num_create_threads = OSS_DEF_CREATE_THREADS; + oss_min_create_threads = oss_max_create_threads = + oss_num_create_threads; + } else { + oss_min_create_threads = OSS_DEF_CREATE_THREADS; + oss_max_create_threads = OSS_MAX_CREATE_THREADS; + } + ost->ost_create_service = ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE, - OST_CREATE_PORTAL, OSC_REPLY_PORTAL, 30000, - ost_handle, "ost_create", - obd->obd_proc_entry); + OST_MAXREPSIZE, OST_CREATE_PORTAL, + OSC_REPLY_PORTAL, + OST_WATCHDOG_TIMEOUT, ost_handle, "ost_create", + obd->obd_proc_entry, ost_print_req, + oss_min_create_threads, + oss_max_create_threads, + "ll_ost_creat", LCT_DT_THREAD); if (ost->ost_create_service == NULL) { CERROR("failed to start OST create service\n"); GOTO(out_service, rc = -ENOMEM); } - - spin_lock_init(&ost->ost_lock); - ost->ost_service->srv_obddev = obd; - - rc = ptlrpc_start_n_threads(obd, ost->ost_create_service, 1, - "ll_ost_creat"); + rc = ptlrpc_start_threads(obd, ost->ost_create_service); if (rc) GOTO(out_create, rc = -EINVAL); - ost->ost_destroy_service = + ost->ost_io_service = ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE, - OST_DESTROY_PORTAL, OSC_REPLY_PORTAL, 30000, - ost_handle, "ost_destroy", - obd->obd_proc_entry); - if (ost->ost_destroy_service == NULL) { - CERROR("failed to start service\n"); + OST_MAXREPSIZE, OST_IO_PORTAL, + OSC_REPLY_PORTAL, + OST_WATCHDOG_TIMEOUT, ost_handle, "ost_io", + obd->obd_proc_entry, ost_print_req, + oss_min_threads, oss_max_threads, + "ll_ost_io", LCT_DT_THREAD); + if (ost->ost_io_service == NULL) { + CERROR("failed to start OST I/O service\n"); GOTO(out_create, rc = -ENOMEM); } - rc = ptlrpc_start_n_threads(obd, ost->ost_destroy_service, - OST_NUM_THREADS, "ll_dstr_ost"); + ost->ost_io_service->srv_init = ost_thread_init; + ost->ost_io_service->srv_done = ost_thread_done; + ost->ost_io_service->srv_cpu_affinity = 1; + rc = ptlrpc_start_threads(obd, ost->ost_io_service); if (rc) - GOTO(out_destroy, rc = -EINVAL); + GOTO(out_io, rc = -EINVAL); + + ping_evictor_start(); RETURN(0); -out_destroy: - ptlrpc_unregister_service(ost->ost_destroy_service); +out_io: + ptlrpc_unregister_service(ost->ost_io_service); + ost->ost_io_service = NULL; out_create: ptlrpc_unregister_service(ost->ost_create_service); + ost->ost_create_service = NULL; out_service: ptlrpc_unregister_service(ost->ost_service); + ost->ost_service = NULL; +out_lprocfs: + lprocfs_obd_cleanup(obd); RETURN(rc); } -extern void lgss_svc_cache_purge_all(void); -static int ost_cleanup(struct obd_device *obd, int flags) +static int ost_cleanup(struct obd_device *obd) { struct ost_obd *ost = &obd->u.ost; int err = 0; ENTRY; + ping_evictor_stop(); + spin_lock_bh(&obd->obd_processing_task_lock); if (obd->obd_recovering) { target_cancel_recovery_timer(obd); @@ -1411,44 +1854,77 @@ static int ost_cleanup(struct obd_device *obd, int flags) } spin_unlock_bh(&obd->obd_processing_task_lock); - ptlrpc_stop_all_threads(ost->ost_service); + down(&ost->ost_health_sem); ptlrpc_unregister_service(ost->ost_service); - - ptlrpc_stop_all_threads(ost->ost_create_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); - ptlrpc_stop_all_threads(ost->ost_destroy_service); - ptlrpc_unregister_service(ost->ost_destroy_service); + lprocfs_obd_cleanup(obd); -#ifdef ENABLE_GSS - /* XXX */ - lgss_svc_cache_purge_all(); -#endif RETURN(err); } +static int ost_health_check(struct obd_device *obd) +{ + struct ost_obd *ost = &obd->u.ost; + int rc = 0; + + 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); + + /* + * health_check to return 0 on healthy + * and 1 on unhealthy. + */ + if( rc != 0) + rc = 1; + + return rc; +} + +struct ost_thread_local_cache *ost_tls(struct ptlrpc_request *r) +{ + return (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data); +} + /* use obd ops to offer management infrastructure */ static struct obd_ops ost_obd_ops = { .o_owner = THIS_MODULE, - .o_attach = ost_attach, - .o_detach = ost_detach, .o_setup = ost_setup, .o_cleanup = ost_cleanup, + .o_health_check = ost_health_check, }; + static int __init ost_init(void) { struct lprocfs_static_vars lvars; + int rc; ENTRY; - lprocfs_init_vars(ost,&lvars); - RETURN(class_register_type(&ost_obd_ops, NULL, lvars.module_vars, - OBD_OST_DEVICENAME)); + lprocfs_init_vars(ost, &lvars); + rc = class_register_type(&ost_obd_ops, NULL, lvars.module_vars, + LUSTRE_OSS_NAME, NULL); + + if (ost_num_threads != 0 && oss_num_threads == 0) { + LCONSOLE_INFO("ost_num_threads module parameter is deprecated, " + "use oss_num_threads instead or unset both for " + "dynamic thread startup\n"); + oss_num_threads = ost_num_threads; + } + + RETURN(rc); } static void /*__exit*/ ost_exit(void) { - class_unregister_type(OBD_OST_DEVICENAME); + class_unregister_type(LUSTRE_OSS_NAME); } MODULE_AUTHOR("Cluster File Systems, Inc. ");