1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ost/ost_handler.c
38 * Author: Peter J. Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
43 # define EXPORT_SYMTAB
45 #define DEBUG_SUBSYSTEM S_OST
47 #include <linux/module.h>
48 #include <obd_cksum.h>
50 #include <lustre_net.h>
51 #include <lustre_dlm.h>
52 #include <lustre_export.h>
53 #include <lustre_debug.h>
54 #include <linux/init.h>
55 #include <lprocfs_status.h>
56 #include <libcfs/list.h>
57 #include <lustre_quota.h>
58 #include "ost_internal.h"
60 static int oss_num_threads;
61 CFS_MODULE_PARM(oss_num_threads, "i", int, 0444,
62 "number of OSS service threads to start");
64 static int ost_num_threads;
65 CFS_MODULE_PARM(ost_num_threads, "i", int, 0444,
66 "number of OST service threads to start (deprecated)");
68 static int oss_num_create_threads;
69 CFS_MODULE_PARM(oss_num_create_threads, "i", int, 0444,
70 "number of OSS create threads to start");
72 void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req)
74 struct oti_req_ack_lock *ack_lock;
81 lustre_msg_set_transno(req->rq_repmsg, oti->oti_transno);
82 req->rq_transno = oti->oti_transno;
84 /* XXX 4 == entries in oti_ack_locks??? */
85 for (ack_lock = oti->oti_ack_locks, i = 0; i < 4; i++, ack_lock++) {
88 /* XXX not even calling target_send_reply in some cases... */
89 ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode);
93 static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req,
94 struct obd_trans_info *oti)
96 struct ost_body *body, *repbody;
97 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
101 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
102 lustre_swab_ost_body);
106 if (body->oa.o_id == 0)
109 if (lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1)) {
110 struct ldlm_request *dlm;
111 dlm = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*dlm),
112 lustre_swab_ldlm_request);
115 ldlm_request_cancel(req, dlm, 0);
118 rc = lustre_pack_reply(req, 2, size, NULL);
122 if (body->oa.o_valid & OBD_MD_FLCOOKIE)
123 oti->oti_logcookies = &body->oa.o_lcookie;
124 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
126 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
127 req->rq_status = obd_destroy(exp, &body->oa, NULL, oti, NULL);
131 static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req)
133 struct ost_body *body, *repbody;
134 struct obd_info oinfo = { { { 0 } } };
135 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
139 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
140 lustre_swab_ost_body);
144 rc = lustre_pack_reply(req, 2, size, NULL);
148 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
150 repbody->oa = body->oa;
152 oinfo.oi_oa = &repbody->oa;
153 if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
154 oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
156 req->rq_status = obd_getattr(exp, &oinfo);
160 static int ost_statfs(struct ptlrpc_request *req)
162 struct obd_statfs *osfs;
163 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*osfs) };
167 rc = lustre_pack_reply(req, 2, size, NULL);
171 osfs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*osfs));
173 req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs,
174 cfs_time_current_64() - HZ, 0);
175 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
176 osfs->os_bfree = osfs->os_bavail = 64;
177 if (req->rq_status != 0)
178 CERROR("ost: statfs failed: rc %d\n", req->rq_status);
183 static int ost_create(struct obd_export *exp, struct ptlrpc_request *req,
184 struct obd_trans_info *oti)
186 struct ost_body *body, *repbody;
187 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
191 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
192 lustre_swab_ost_body);
196 rc = lustre_pack_reply(req, 2, size, NULL);
200 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
202 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
203 oti->oti_logcookies = &repbody->oa.o_lcookie;
204 req->rq_status = obd_create(exp, &repbody->oa, NULL, oti);
205 //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0);
210 * Helper function for ost_punch(): if asked by client, acquire [size, EOF]
211 * lock on the file being truncated.
213 static int ost_punch_lock_get(struct obd_export *exp, struct obdo *oa,
214 struct lustre_handle *lh)
217 struct ldlm_res_id res_id;
218 ldlm_policy_data_t policy;
224 osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
225 LASSERT(!lustre_handle_is_used(lh));
227 if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
228 !(oa->o_flags & OBD_FL_TRUNCLOCK))
231 CDEBUG(D_INODE, "OST-side truncate lock.\n");
234 finis = start + oa->o_blocks;
237 * standard truncate optimization: if file body is completely
238 * destroyed, don't send data back to the server.
240 flags = (start == 0) ? LDLM_AST_DISCARD_DATA : 0;
242 policy.l_extent.start = start & CFS_PAGE_MASK;
245 * If ->o_blocks is EOF it means "lock till the end of the
246 * file". Otherwise, it's size of a hole being punched (in bytes)
248 if (oa->o_blocks == OBD_OBJECT_EOF || finis < start)
249 policy.l_extent.end = OBD_OBJECT_EOF;
251 policy.l_extent.end = finis | ~CFS_PAGE_MASK;
253 RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
254 LDLM_EXTENT, &policy, LCK_PW, &flags,
255 ldlm_blocking_ast, ldlm_completion_ast,
256 ldlm_glimpse_ast, NULL, 0, NULL, lh));
260 * Helper function for ost_punch(): release lock acquired by
261 * ost_punch_lock_get(), if any.
263 static void ost_punch_lock_put(struct obd_export *exp, struct obdo *oa,
264 struct lustre_handle *lh)
267 if (lustre_handle_is_used(lh))
268 ldlm_lock_decref(lh, LCK_PW);
272 static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req,
273 struct obd_trans_info *oti)
275 struct obd_info oinfo = { { { 0 } } };
276 struct ost_body *body, *repbody;
277 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
279 struct lustre_handle lh = {0,};
282 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
283 CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
285 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
286 lustre_swab_ost_body);
290 oinfo.oi_oa = &body->oa;
291 oinfo.oi_policy.l_extent.start = oinfo.oi_oa->o_size;
292 oinfo.oi_policy.l_extent.end = oinfo.oi_oa->o_blocks;
294 if ((oinfo.oi_oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
295 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
298 rc = lustre_pack_reply(req, 2, size, NULL);
302 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
304 rc = ost_punch_lock_get(exp, oinfo.oi_oa, &lh);
306 if (oinfo.oi_oa->o_valid & OBD_MD_FLFLAGS &&
307 oinfo.oi_oa->o_flags == OBD_FL_TRUNCLOCK)
309 * If OBD_FL_TRUNCLOCK is the only bit set in
310 * ->o_flags, clear OBD_MD_FLFLAGS to avoid falling
311 * through filter_setattr() to filter_iocontrol().
313 oinfo.oi_oa->o_valid &= ~OBD_MD_FLFLAGS;
315 if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
316 oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
318 req->rq_status = obd_punch(exp, &oinfo, oti, NULL);
319 ost_punch_lock_put(exp, oinfo.oi_oa, &lh);
321 repbody->oa = *oinfo.oi_oa;
325 static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req)
327 struct ost_body *body, *repbody;
328 struct lustre_capa *capa = NULL;
329 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
333 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
334 lustre_swab_ost_body);
338 if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
339 capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 1);
341 rc = lustre_pack_reply(req, 2, size, NULL);
345 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
347 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
348 req->rq_status = obd_sync(exp, &repbody->oa, NULL, repbody->oa.o_size,
349 repbody->oa.o_blocks, capa);
353 static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req,
354 struct obd_trans_info *oti)
356 struct ost_body *body, *repbody;
357 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
359 struct obd_info oinfo = { { { 0 } } };
362 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
363 lustre_swab_ost_body);
367 rc = lustre_pack_reply(req, 2, size, NULL);
371 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
373 repbody->oa = body->oa;
375 oinfo.oi_oa = &repbody->oa;
376 if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
377 oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
379 req->rq_status = obd_setattr(exp, &oinfo, oti);
383 static int ost_bulk_timeout(void *data)
386 /* We don't fail the connection here, because having the export
387 * killed makes the (vital) call to commitrw very sad.
392 static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo,
393 struct niobuf_remote *rnb, int nrnb,
394 struct niobuf_remote **pp_rnbp)
396 /* Copy a remote niobuf, splitting it into page-sized chunks
397 * and setting ioo[i].ioo_bufcnt accordingly */
398 struct niobuf_remote *pp_rnb;
406 * array of sufficient size already preallocated by caller
408 LASSERT(pp_rnbp != NULL);
409 LASSERT(*pp_rnbp != NULL);
411 /* first count and check the number of pages required */
412 for (i = 0; i < nioo; i++)
413 for (j = 0; j < ioo->ioo_bufcnt; j++, rnbidx++) {
414 obd_off offset = rnb[rnbidx].offset;
415 obd_off p0 = offset >> CFS_PAGE_SHIFT;
416 obd_off pn = (offset + rnb[rnbidx].len - 1) >>
419 LASSERT(rnbidx < nrnb);
421 npages += (pn + 1 - p0);
423 if (rnb[rnbidx].len == 0) {
424 CERROR("zero len BRW: obj %d objid "LPX64
425 " buf %u\n", i, ioo[i].ioo_id, j);
429 rnb[rnbidx].offset <= rnb[rnbidx-1].offset) {
430 CERROR("unordered BRW: obj %d objid "LPX64
431 " buf %u offset "LPX64" <= "LPX64"\n",
432 i, ioo[i].ioo_id, j, rnb[rnbidx].offset,
438 LASSERT(rnbidx == nrnb);
440 if (npages == nrnb) { /* all niobufs are for single pages */
447 /* now do the actual split */
449 for (i = 0; i < nioo; i++) {
452 for (j = 0; j < ioo[i].ioo_bufcnt; j++, rnbidx++) {
453 obd_off off = rnb[rnbidx].offset;
454 int nob = rnb[rnbidx].len;
456 LASSERT(rnbidx < nrnb);
458 obd_off poff = off & ~CFS_PAGE_MASK;
459 int pnob = (poff + nob > CFS_PAGE_SIZE) ?
460 CFS_PAGE_SIZE - poff : nob;
462 LASSERT(page < npages);
463 pp_rnb[page].len = pnob;
464 pp_rnb[page].offset = off;
465 pp_rnb[page].flags = rnb[rnbidx].flags;
467 CDEBUG(0, " obj %d id "LPX64
468 "page %d(%d) "LPX64" for %d, flg %x\n",
469 i, ioo[i].ioo_id, obj_pages, page,
470 pp_rnb[page].offset, pp_rnb[page].len,
480 ioo[i].ioo_bufcnt = obj_pages;
482 LASSERT(page == npages);
487 static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc, int opc,
488 cksum_type_t cksum_type)
493 cksum = init_checksum(cksum_type);
494 for (i = 0; i < desc->bd_iov_count; i++) {
495 struct page *page = desc->bd_iov[i].kiov_page;
496 int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
497 char *ptr = kmap(page) + off;
498 int len = desc->bd_iov[i].kiov_len;
500 /* corrupt the data before we compute the checksum, to
501 * simulate a client->OST data error */
502 if (i == 0 && opc == OST_WRITE &&
503 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE))
504 memcpy(ptr, "bad3", min(4, len));
505 cksum = compute_checksum(cksum, ptr, len, cksum_type);
506 /* corrupt the data after we compute the checksum, to
507 * simulate an OST->client data error */
508 if (i == 0 && opc == OST_READ &&
509 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))
510 memcpy(ptr, "bad4", min(4, len));
518 * populate @nio by @nrpages pages from per-thread page pool
520 static void ost_nio_pages_get(struct ptlrpc_request *req,
521 struct niobuf_local *nio, int nrpages)
524 struct ost_thread_local_cache *tls;
528 LASSERT(nrpages <= OST_THREAD_POOL_SIZE);
529 LASSERT(req != NULL);
530 LASSERT(req->rq_svc_thread != NULL);
533 LASSERT(tls != NULL);
535 memset(nio, 0, nrpages * sizeof *nio);
536 for (i = 0; i < nrpages; ++ i) {
540 LASSERT(page != NULL);
541 POISON_PAGE(page, 0xf1);
543 LL_CDEBUG_PAGE(D_INFO, page, "%d\n", i);
549 * Dual for ost_nio_pages_get(). Poison pages in pool for debugging
551 static void ost_nio_pages_put(struct ptlrpc_request *req,
552 struct niobuf_local *nio, int nrpages)
558 LASSERT(nrpages <= OST_THREAD_POOL_SIZE);
560 for (i = 0; i < nrpages; ++ i)
561 POISON_PAGE(nio[i].page, 0xf2);
565 static int ost_brw_lock_get(int mode, struct obd_export *exp,
566 struct obd_ioobj *obj, struct niobuf_remote *nb,
567 struct lustre_handle *lh)
570 int nrbufs = obj->ioo_bufcnt;
571 struct ldlm_res_id res_id;
572 ldlm_policy_data_t policy;
576 osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
577 LASSERT(mode == LCK_PR || mode == LCK_PW);
578 LASSERT(!lustre_handle_is_used(lh));
580 if (nrbufs == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
583 /* EXPENSIVE ASSERTION */
584 for (i = 1; i < nrbufs; i ++)
585 LASSERT((nb[0].flags & OBD_BRW_SRVLOCK) ==
586 (nb[i].flags & OBD_BRW_SRVLOCK));
588 policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
589 policy.l_extent.end = (nb[nrbufs - 1].offset +
590 nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
592 RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
593 LDLM_EXTENT, &policy, mode, &flags,
594 ldlm_blocking_ast, ldlm_completion_ast,
595 ldlm_glimpse_ast, NULL, 0, NULL, lh));
598 static void ost_brw_lock_put(int mode,
599 struct obd_ioobj *obj, struct niobuf_remote *niob,
600 struct lustre_handle *lh)
603 LASSERT(mode == LCK_PR || mode == LCK_PW);
604 LASSERT((obj->ioo_bufcnt > 0 && (niob[0].flags & OBD_BRW_SRVLOCK)) ==
605 lustre_handle_is_used(lh));
606 if (lustre_handle_is_used(lh))
607 ldlm_lock_decref(lh, mode);
611 struct ost_prolong_data {
612 struct obd_export *opd_exp;
613 ldlm_policy_data_t opd_policy;
614 ldlm_mode_t opd_mode;
617 static int ost_prolong_locks_iter(struct ldlm_lock *lock, void *data)
619 struct ost_prolong_data *opd = data;
621 LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
623 if (lock->l_req_mode != lock->l_granted_mode) {
624 /* scan granted locks only */
625 return LDLM_ITER_STOP;
628 if (lock->l_export != opd->opd_exp) {
629 /* prolong locks only for given client */
630 return LDLM_ITER_CONTINUE;
633 if (!(lock->l_granted_mode & opd->opd_mode)) {
634 /* we aren't interesting in all type of locks */
635 return LDLM_ITER_CONTINUE;
638 if (lock->l_policy_data.l_extent.end < opd->opd_policy.l_extent.start ||
639 lock->l_policy_data.l_extent.start > opd->opd_policy.l_extent.end) {
640 /* the request doesn't cross the lock, skip it */
641 return LDLM_ITER_CONTINUE;
644 if (!(lock->l_flags & LDLM_FL_AST_SENT)) {
645 /* ignore locks not being cancelled */
646 return LDLM_ITER_CONTINUE;
649 /* OK. this is a possible lock the user holds doing I/O
650 * let's refresh eviction timer for it */
651 ldlm_refresh_waiting_lock(lock);
653 return LDLM_ITER_CONTINUE;
656 static void ost_prolong_locks(struct obd_export *exp, struct obd_ioobj *obj,
657 struct niobuf_remote *nb, struct obdo *oa,
660 struct ldlm_res_id res_id;
661 int nrbufs = obj->ioo_bufcnt;
662 struct ost_prolong_data opd;
665 osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
669 opd.opd_policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
670 opd.opd_policy.l_extent.end = (nb[nrbufs - 1].offset +
671 nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
673 CDEBUG(D_DLMTRACE,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
674 res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start,
675 opd.opd_policy.l_extent.end);
677 if (oa->o_valid & OBD_MD_FLHANDLE) {
678 struct ldlm_lock *lock;
680 lock = ldlm_handle2lock(&oa->o_handle);
682 ost_prolong_locks_iter(lock, &opd);
689 ldlm_resource_iterate(exp->exp_obd->obd_namespace, &res_id,
690 ost_prolong_locks_iter, &opd);
695 static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti)
697 struct ptlrpc_bulk_desc *desc;
698 struct obd_export *exp = req->rq_export;
699 struct niobuf_remote *remote_nb;
700 struct niobuf_remote *pp_rnb = NULL;
701 struct niobuf_local *local_nb;
702 struct obd_ioobj *ioo;
703 struct ost_body *body, *repbody;
704 struct lustre_capa *capa = NULL;
705 struct l_wait_info lwi;
706 struct lustre_handle lockh = { 0 };
707 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
708 int niocount, npages, nob = 0, rc, i;
712 req->rq_bulk_read = 1;
714 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
715 GOTO(out, rc = -EIO);
717 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
719 /* Check if there is eviction in progress, and if so, wait for it to
721 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
722 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
723 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
724 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
728 GOTO(out, rc = -ENOTCONN);
730 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
731 lustre_swab_ost_body);
733 CERROR("Missing/short ost_body\n");
734 GOTO(out, rc = -EFAULT);
737 ioo = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*ioo),
738 lustre_swab_obd_ioobj);
740 CERROR("Missing/short ioobj\n");
741 GOTO(out, rc = -EFAULT);
744 niocount = ioo->ioo_bufcnt;
745 if (niocount > PTLRPC_MAX_BRW_PAGES) {
746 DEBUG_REQ(D_ERROR, req, "bulk has too many pages (%d)",
748 GOTO(out, rc = -EFAULT);
751 remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2,
752 niocount * sizeof(*remote_nb),
753 lustre_swab_niobuf_remote);
754 if (remote_nb == NULL) {
755 CERROR("Missing/short niobuf\n");
756 GOTO(out, rc = -EFAULT);
758 if (lustre_msg_swabbed(req->rq_reqmsg)) { /* swab remaining niobufs */
759 for (i = 1; i < niocount; i++)
760 lustre_swab_niobuf_remote (&remote_nb[i]);
763 if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
764 capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3);
766 rc = lustre_pack_reply(req, 2, size, NULL);
771 * Per-thread array of struct niobuf_{local,remote}'s was allocated by
774 local_nb = ost_tls(req)->local;
775 pp_rnb = ost_tls(req)->remote;
777 /* FIXME all niobuf splitting should be done in obdfilter if needed */
778 /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
779 npages = get_per_page_niobufs(ioo, 1, remote_nb, niocount, &pp_rnb);
781 GOTO(out, rc = npages);
783 LASSERT(npages <= OST_THREAD_POOL_SIZE);
785 ost_nio_pages_get(req, local_nb, npages);
787 desc = ptlrpc_prep_bulk_exp(req, npages,
788 BULK_PUT_SOURCE, OST_BULK_PORTAL);
790 GOTO(out, rc = -ENOMEM);
792 rc = ost_brw_lock_get(LCK_PR, exp, ioo, pp_rnb, &lockh);
797 * If getting the lock took more time than
798 * client was willing to wait, drop it. b=11330
800 if (cfs_time_current_sec() > req->rq_deadline ||
801 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
803 CERROR("Dropping timed-out read from %s because locking"
804 "object "LPX64" took %ld seconds (limit was %ld).\n",
805 libcfs_id2str(req->rq_peer), ioo->ioo_id,
806 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
807 req->rq_deadline - req->rq_arrival_time.tv_sec);
808 GOTO(out_lock, rc = -ETIMEDOUT);
811 rc = obd_preprw(OBD_BRW_READ, exp, &body->oa, 1,
812 ioo, npages, pp_rnb, local_nb, oti, capa);
816 ost_prolong_locks(exp, ioo, pp_rnb, &body->oa, LCK_PW | LCK_PR);
819 for (i = 0; i < npages; i++) {
820 int page_rc = local_nb[i].rc;
822 if (page_rc < 0) { /* error */
827 LASSERTF(page_rc <= pp_rnb[i].len, "page_rc (%d) > "
828 "pp_rnb[%d].len (%d)\n", page_rc, i, pp_rnb[i].len);
830 if (page_rc != 0) { /* some data! */
831 LASSERT (local_nb[i].page != NULL);
832 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
833 pp_rnb[i].offset & ~CFS_PAGE_MASK,
837 if (page_rc != pp_rnb[i].len) { /* short read */
840 /* All subsequent pages should be 0 */
842 LASSERTF(local_nb[i].rc == 0,
843 "page_rc %d, pp_rnb[%u].len=%d, "
844 "local_nb[%u/%u].rc=%d\n",
845 page_rc, j, pp_rnb[j].len,
846 i, npages, local_nb[i].rc);
851 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
852 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
854 if (body->oa.o_valid & OBD_MD_FLFLAGS)
855 cksum_type = cksum_type_unpack(body->oa.o_flags);
856 body->oa.o_flags = cksum_type_pack(cksum_type);
857 body->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
858 body->oa.o_cksum = ost_checksum_bulk(desc, OST_READ, cksum_type);
859 CDEBUG(D_PAGE,"checksum at read origin: %x\n",body->oa.o_cksum);
861 body->oa.o_valid = 0;
863 /* We're finishing using body->oa as an input variable */
865 /* Check if client was evicted while we were doing i/o before touching
868 /* Check if there is eviction in progress, and if so, wait for
870 if (unlikely(atomic_read(&exp->exp_obd->
871 obd_evict_inprogress))) {
872 lwi = LWI_INTR(NULL, NULL);
873 rc = l_wait_event(exp->exp_obd->
874 obd_evict_inprogress_waitq,
875 !atomic_read(&exp->exp_obd->
876 obd_evict_inprogress),
882 sptlrpc_svc_wrap_bulk(req, desc);
884 rc = ptlrpc_start_bulk_transfer(desc);
888 time_t start = cfs_time_current_sec();
890 long timeoutl = req->rq_deadline -
891 cfs_time_current_sec();
892 cfs_duration_t timeout = (timeoutl <= 0 || rc) ?
893 CFS_TICK : cfs_time_seconds(timeoutl);
894 lwi = LWI_TIMEOUT_INTERVAL(timeout,
898 rc = l_wait_event(desc->bd_waitq,
899 !ptlrpc_bulk_active(desc) ||
900 exp->exp_failed, &lwi);
901 LASSERT(rc == 0 || rc == -ETIMEDOUT);
902 /* Wait again if we changed deadline */
903 } while ((rc == -ETIMEDOUT) &&
904 (req->rq_deadline > cfs_time_current_sec()));
906 if (rc == -ETIMEDOUT) {
907 DEBUG_REQ(D_ERROR, req,
908 "timeout on bulk PUT after %ld%+lds",
909 req->rq_deadline - start,
910 cfs_time_current_sec() -
912 ptlrpc_abort_bulk(desc);
913 } else if (exp->exp_failed) {
914 DEBUG_REQ(D_ERROR, req, "Eviction on bulk PUT");
916 ptlrpc_abort_bulk(desc);
917 } else if (!desc->bd_success ||
918 desc->bd_nob_transferred != desc->bd_nob) {
919 DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)",
921 "truncated" : "network error on",
922 desc->bd_nob_transferred,
924 /* XXX should this be a different errno? */
928 DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d", rc);
933 /* Must commit after prep above in all cases */
934 rc = obd_commitrw(OBD_BRW_READ, exp, &body->oa, 1,
935 ioo, npages, local_nb, oti, rc);
937 ost_nio_pages_put(req, local_nb, npages);
940 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
942 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
946 ost_brw_lock_put(LCK_PR, ioo, pp_rnb, &lockh);
948 ptlrpc_free_bulk(desc);
952 req->rq_status = nob;
953 ptlrpc_lprocfs_brw(req, nob);
954 target_committed_to_req(req);
956 } else if (!no_reply) {
957 /* Only reply if there was no comms problem with bulk */
958 target_committed_to_req(req);
962 /* reply out callback would free */
963 ptlrpc_req_drop_rs(req);
964 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
965 "client will retry\n",
966 exp->exp_obd->obd_name,
967 exp->exp_client_uuid.uuid,
968 exp->exp_connection->c_remote_uuid.uuid,
969 libcfs_id2str(req->rq_peer));
975 static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
977 struct ptlrpc_bulk_desc *desc;
978 struct obd_export *exp = req->rq_export;
979 struct niobuf_remote *remote_nb;
980 struct niobuf_remote *pp_rnb;
981 struct niobuf_local *local_nb;
982 struct obd_ioobj *ioo;
983 struct ost_body *body, *repbody;
984 struct l_wait_info lwi;
985 struct lustre_handle lockh = {0};
986 struct lustre_capa *capa = NULL;
988 __u32 size[3] = { sizeof(struct ptlrpc_body), sizeof(*body) };
989 int objcount, niocount, npages;
991 obd_count client_cksum = 0, server_cksum = 0;
992 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
996 req->rq_bulk_write = 1;
998 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
999 GOTO(out, rc = -EIO);
1000 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
1001 GOTO(out, rc = -EFAULT);
1003 /* pause before transaction has been started */
1004 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
1006 /* Check if there is eviction in progress, and if so, wait for it to
1008 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
1009 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
1010 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
1011 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
1014 if (exp->exp_failed)
1015 GOTO(out, rc = -ENOTCONN);
1017 swab = lustre_msg_swabbed(req->rq_reqmsg);
1018 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
1019 lustre_swab_ost_body);
1021 CERROR("Missing/short ost_body\n");
1022 GOTO(out, rc = -EFAULT);
1025 lustre_set_req_swabbed(req, REQ_REC_OFF + 1);
1026 objcount = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1) /
1028 if (objcount == 0) {
1029 CERROR("Missing/short ioobj\n");
1030 GOTO(out, rc = -EFAULT);
1033 CERROR("too many ioobjs (%d)\n", objcount);
1034 GOTO(out, rc = -EFAULT);
1037 ioo = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1,
1038 objcount * sizeof(*ioo));
1039 LASSERT (ioo != NULL);
1040 for (niocount = i = 0; i < objcount; i++) {
1042 lustre_swab_obd_ioobj(&ioo[i]);
1043 if (ioo[i].ioo_bufcnt == 0) {
1044 CERROR("ioo[%d] has zero bufcnt\n", i);
1045 GOTO(out, rc = -EFAULT);
1047 niocount += ioo[i].ioo_bufcnt;
1050 if (niocount > PTLRPC_MAX_BRW_PAGES) {
1051 DEBUG_REQ(D_ERROR, req, "bulk has too many pages (%d)",
1053 GOTO(out, rc = -EFAULT);
1056 remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2,
1057 niocount * sizeof(*remote_nb),
1058 lustre_swab_niobuf_remote);
1059 if (remote_nb == NULL) {
1060 CERROR("Missing/short niobuf\n");
1061 GOTO(out, rc = -EFAULT);
1063 if (swab) { /* swab the remaining niobufs */
1064 for (i = 1; i < niocount; i++)
1065 lustre_swab_niobuf_remote (&remote_nb[i]);
1068 if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
1069 capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3);
1071 size[REPLY_REC_OFF + 1] = niocount * sizeof(*rcs);
1072 rc = lustre_pack_reply(req, 3, size, NULL);
1075 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, obd_fail_val);
1076 rcs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1,
1077 niocount * sizeof(*rcs));
1080 * Per-thread array of struct niobuf_{local,remote}'s was allocated by
1081 * ost_thread_init().
1083 local_nb = ost_tls(req)->local;
1084 pp_rnb = ost_tls(req)->remote;
1086 /* FIXME all niobuf splitting should be done in obdfilter if needed */
1087 /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
1088 npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
1090 GOTO(out, rc = npages);
1092 LASSERT(npages <= OST_THREAD_POOL_SIZE);
1094 ost_nio_pages_get(req, local_nb, npages);
1096 desc = ptlrpc_prep_bulk_exp(req, npages,
1097 BULK_GET_SINK, OST_BULK_PORTAL);
1099 GOTO(out, rc = -ENOMEM);
1101 rc = ost_brw_lock_get(LCK_PW, exp, ioo, pp_rnb, &lockh);
1106 * If getting the lock took more time than
1107 * client was willing to wait, drop it. b=11330
1109 if (cfs_time_current_sec() > req->rq_deadline ||
1110 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
1112 CERROR("Dropping timed-out write from %s because locking "
1113 "object "LPX64" took %ld seconds (limit was %ld).\n",
1114 libcfs_id2str(req->rq_peer), ioo->ioo_id,
1115 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
1116 req->rq_deadline - req->rq_arrival_time.tv_sec);
1117 GOTO(out_lock, rc = -ETIMEDOUT);
1120 ost_prolong_locks(exp, ioo, pp_rnb, &body->oa, LCK_PW);
1122 /* obd_preprw clobbers oa->valid, so save what we need */
1123 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1124 client_cksum = body->oa.o_cksum;
1125 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1126 cksum_type = cksum_type_unpack(body->oa.o_flags);
1129 /* Because we already sync grant info with client when reconnect,
1130 * grant info will be cleared for resent req, then fed_grant and
1131 * total_grant will not be modified in following preprw_write */
1132 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
1133 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
1134 body->oa.o_valid &= ~OBD_MD_FLGRANT;
1137 rc = obd_preprw(OBD_BRW_WRITE, exp, &body->oa, objcount,
1138 ioo, npages, pp_rnb, local_nb, oti, capa);
1142 /* NB Having prepped, we must commit... */
1144 for (i = 0; i < npages; i++)
1145 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
1146 pp_rnb[i].offset & ~CFS_PAGE_MASK,
1149 /* Check if client was evicted while we were doing i/o before touching
1151 if (desc->bd_export->exp_failed)
1154 rc = ptlrpc_start_bulk_transfer (desc);
1156 time_t start = cfs_time_current_sec();
1158 long timeoutl = req->rq_deadline -
1159 cfs_time_current_sec();
1160 cfs_duration_t timeout = (timeoutl <= 0 || rc) ?
1161 CFS_TICK : cfs_time_seconds(timeoutl);
1162 lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
1163 ost_bulk_timeout, desc);
1164 rc = l_wait_event(desc->bd_waitq,
1165 !ptlrpc_bulk_active(desc) ||
1166 desc->bd_export->exp_failed, &lwi);
1167 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1168 /* Wait again if we changed deadline */
1169 } while ((rc == -ETIMEDOUT) &&
1170 (req->rq_deadline > cfs_time_current_sec()));
1172 if (rc == -ETIMEDOUT) {
1173 DEBUG_REQ(D_ERROR, req,
1174 "timeout on bulk GET after %ld%+lds",
1175 req->rq_deadline - start,
1176 cfs_time_current_sec() -
1178 ptlrpc_abort_bulk(desc);
1179 } else if (desc->bd_export->exp_failed) {
1180 DEBUG_REQ(D_ERROR, req, "Eviction on bulk GET");
1182 ptlrpc_abort_bulk(desc);
1183 } else if (!desc->bd_success ||
1184 desc->bd_nob_transferred != desc->bd_nob) {
1185 DEBUG_REQ(D_ERROR, req, "%s bulk GET %d(%d)",
1187 "truncated" : "network error on",
1188 desc->bd_nob_transferred, desc->bd_nob);
1189 /* XXX should this be a different errno? */
1193 DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d", rc);
1198 sptlrpc_svc_unwrap_bulk(req, desc);
1200 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1202 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
1204 if (unlikely(client_cksum != 0 && rc == 0)) {
1205 static int cksum_counter;
1206 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1207 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
1208 repbody->oa.o_flags |= cksum_type_pack(cksum_type);
1209 server_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1210 repbody->oa.o_cksum = server_cksum;
1212 if (unlikely(client_cksum != server_cksum)) {
1213 CERROR("client csum %x, server csum %x\n",
1214 client_cksum, server_cksum);
1216 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
1217 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
1218 cksum_counter, libcfs_id2str(req->rq_peer),
1223 /* Must commit after prep above in all cases */
1224 rc = obd_commitrw(OBD_BRW_WRITE, exp, &repbody->oa,
1225 objcount, ioo, npages, local_nb, oti, rc);
1227 if (unlikely(client_cksum != server_cksum && rc == 0)) {
1228 int new_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1233 if (new_cksum == server_cksum)
1234 msg = "changed in transit before arrival at OST";
1235 else if (new_cksum == client_cksum)
1236 msg = "initial checksum before message complete";
1238 msg = "changed in transit AND after initial checksum";
1240 if (req->rq_peer.nid == desc->bd_sender) {
1244 router = libcfs_nid2str(desc->bd_sender);
1247 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: %s from "
1248 "%s%s%s inum "LPU64"/"LPU64" object "
1249 LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1250 exp->exp_obd->obd_name, msg,
1251 libcfs_id2str(req->rq_peer),
1253 body->oa.o_valid & OBD_MD_FLFID ?
1254 body->oa.o_fid : (__u64)0,
1255 body->oa.o_valid & OBD_MD_FLFID ?
1256 body->oa.o_generation :(__u64)0,
1258 body->oa.o_valid & OBD_MD_FLGROUP ?
1259 body->oa.o_gr : (__u64)0,
1261 pp_rnb[npages-1].offset+pp_rnb[npages-1].len
1263 CERROR("client csum %x, original server csum %x, "
1264 "server csum now %x\n",
1265 client_cksum, server_cksum, new_cksum);
1268 ost_nio_pages_put(req, local_nb, npages);
1273 /* set per-requested niobuf return codes */
1274 for (i = j = 0; i < niocount; i++) {
1275 int len = remote_nb[i].len;
1280 LASSERT(j < npages);
1281 if (local_nb[j].rc < 0)
1282 rcs[i] = local_nb[j].rc;
1283 len -= pp_rnb[j].len;
1288 LASSERT(j == npages);
1289 ptlrpc_lprocfs_brw(req, nob);
1293 ost_brw_lock_put(LCK_PW, ioo, pp_rnb, &lockh);
1295 ptlrpc_free_bulk(desc);
1298 oti_to_request(oti, req);
1299 target_committed_to_req(req);
1300 rc = ptlrpc_reply(req);
1301 } else if (!no_reply) {
1302 /* Only reply if there was no comms problem with bulk */
1303 target_committed_to_req(req);
1304 req->rq_status = rc;
1307 /* reply out callback would free */
1308 ptlrpc_req_drop_rs(req);
1309 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
1310 "client will retry\n",
1311 exp->exp_obd->obd_name,
1312 exp->exp_client_uuid.uuid,
1313 exp->exp_connection->c_remote_uuid.uuid,
1314 libcfs_id2str(req->rq_peer));
1319 static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
1321 char *key, *val = NULL;
1322 int keylen, vallen, rc = 0;
1325 key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1);
1327 DEBUG_REQ(D_HA, req, "no set_info key");
1330 keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF);
1332 rc = lustre_pack_reply(req, 1, NULL, NULL);
1336 vallen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1);
1338 val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, 0);
1340 if (KEY_IS(KEY_EVICT_BY_NID)) {
1342 obd_export_evict_by_nid(exp->exp_obd, val);
1347 rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL);
1349 lustre_msg_set_status(req->rq_repmsg, 0);
1353 static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req)
1356 int keylen, replylen, rc = 0;
1357 struct req_capsule *pill = &req->rq_pill;
1360 req_capsule_set(&req->rq_pill, &RQF_OST_GET_INFO_GENERIC);
1362 /* this common part for get_info rpc */
1363 key = req_capsule_client_get(pill, &RMF_SETINFO_KEY);
1365 DEBUG_REQ(D_HA, req, "no get_info key");
1368 keylen = req_capsule_get_size(pill, &RMF_SETINFO_KEY, RCL_CLIENT);
1370 rc = obd_get_info(exp, keylen, key, &replylen, NULL, NULL);
1374 req_capsule_set_size(pill, &RMF_GENERIC_DATA,
1375 RCL_SERVER, replylen);
1377 rc = req_capsule_server_pack(pill);
1381 reply = req_capsule_server_get(pill, &RMF_GENERIC_DATA);
1385 /* call again to fill in the reply buffer */
1386 rc = obd_get_info(exp, keylen, key, &replylen, reply, NULL);
1388 lustre_msg_set_status(req->rq_repmsg, 0);
1392 static int ost_handle_quotactl(struct ptlrpc_request *req)
1394 struct obd_quotactl *oqctl, *repoqc;
1395 __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) };
1399 oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1400 lustre_swab_obd_quotactl);
1402 GOTO(out, rc = -EPROTO);
1404 rc = lustre_pack_reply(req, 2, size, NULL);
1408 repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc));
1410 req->rq_status = obd_quotactl(req->rq_export, oqctl);
1416 static int ost_handle_quotacheck(struct ptlrpc_request *req)
1418 struct obd_quotactl *oqctl;
1422 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1426 rc = req_capsule_server_pack(&req->rq_pill);
1428 CERROR("ost: out of memory while packing quotacheck reply\n");
1432 req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1436 static int ost_llog_handle_connect(struct obd_export *exp,
1437 struct ptlrpc_request *req)
1439 struct llogd_conn_body *body;
1443 body = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*body));
1444 rc = obd_llog_connect(exp, body);
1448 static int filter_export_check_flavor(struct filter_obd *filter,
1449 struct obd_export *exp,
1450 struct ptlrpc_request *req)
1455 * this should be done in filter_connect()/filter_reconnect(), but
1456 * we can't obtain information like NID, which stored in incoming
1457 * request, thus can't decide what flavor to use. so we do it here.
1459 * This hack should be removed after the OST stack be rewritten, just
1460 * like what we are doing in mdt_obd_connect()/mdt_obd_reconnect().
1462 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_INVALID)
1465 CDEBUG(D_SEC, "from %s\n", sptlrpc_part2name(req->rq_sp_from));
1466 spin_lock(&exp->exp_lock);
1467 exp->exp_sp_peer = req->rq_sp_from;
1469 read_lock(&filter->fo_sptlrpc_lock);
1470 sptlrpc_rule_set_choose(&filter->fo_sptlrpc_rset, exp->exp_sp_peer,
1471 req->rq_peer.nid, &exp->exp_flvr);
1472 read_unlock(&filter->fo_sptlrpc_lock);
1474 if (exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
1475 CERROR("invalid rpc flavor %x, expect %x, from %s\n",
1476 req->rq_flvr.sf_rpc, exp->exp_flvr.sf_rpc,
1477 libcfs_nid2str(req->rq_peer.nid));
1478 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_INVALID;
1482 spin_unlock(&exp->exp_lock);
1487 static int ost_filter_recovery_request(struct ptlrpc_request *req,
1488 struct obd_device *obd, int *process)
1490 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1491 case OST_CONNECT: /* This will never get here, but for completeness. */
1492 case OST_DISCONNECT:
1503 case OBD_LOG_CANCEL:
1505 *process = target_queue_recovery_request(req, obd);
1509 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1515 int ost_msg_check_version(struct lustre_msg *msg)
1519 switch(lustre_msg_get_opc(msg)) {
1521 case OST_DISCONNECT:
1524 case SEC_CTX_INIT_CONT:
1526 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1528 CERROR("bad opc %u version %08x, expecting %08x\n",
1529 lustre_msg_get_opc(msg),
1530 lustre_msg_get_version(msg),
1531 LUSTRE_OBD_VERSION);
1544 case OST_QUOTACHECK:
1546 rc = lustre_msg_check_version(msg, LUSTRE_OST_VERSION);
1548 CERROR("bad opc %u version %08x, expecting %08x\n",
1549 lustre_msg_get_opc(msg),
1550 lustre_msg_get_version(msg),
1551 LUSTRE_OST_VERSION);
1556 case LDLM_BL_CALLBACK:
1557 case LDLM_CP_CALLBACK:
1558 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1560 CERROR("bad opc %u version %08x, expecting %08x\n",
1561 lustre_msg_get_opc(msg),
1562 lustre_msg_get_version(msg),
1563 LUSTRE_DLM_VERSION);
1565 case LLOG_ORIGIN_CONNECT:
1566 case OBD_LOG_CANCEL:
1567 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1569 CERROR("bad opc %u version %08x, expecting %08x\n",
1570 lustre_msg_get_opc(msg),
1571 lustre_msg_get_version(msg),
1572 LUSTRE_LOG_VERSION);
1575 CERROR("Unexpected opcode %d\n", lustre_msg_get_opc(msg));
1581 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1582 int ost_handle(struct ptlrpc_request *req)
1584 struct obd_trans_info trans_info = { 0, };
1585 struct obd_trans_info *oti = &trans_info;
1586 int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0;
1587 struct obd_device *obd = NULL;
1590 LASSERT(current->journal_info == NULL);
1592 /* primordial rpcs don't affect server recovery */
1593 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1595 case SEC_CTX_INIT_CONT:
1600 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1602 /* XXX identical to MDS */
1603 if (lustre_msg_get_opc(req->rq_reqmsg) != OST_CONNECT) {
1606 if (req->rq_export == NULL) {
1607 CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
1608 lustre_msg_get_opc(req->rq_reqmsg),
1609 libcfs_id2str(req->rq_peer));
1610 req->rq_status = -ENOTCONN;
1611 GOTO(out, rc = -ENOTCONN);
1614 obd = req->rq_export->exp_obd;
1616 /* Check for aborted recovery. */
1617 spin_lock_bh(&obd->obd_processing_task_lock);
1618 recovering = obd->obd_recovering;
1619 spin_unlock_bh(&obd->obd_processing_task_lock);
1621 rc = ost_filter_recovery_request(req, obd,
1623 if (rc || !should_process)
1625 else if (should_process < 0) {
1626 req->rq_status = should_process;
1627 rc = ptlrpc_error(req);
1635 rc = ost_msg_check_version(req->rq_reqmsg);
1639 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1641 CDEBUG(D_INODE, "connect\n");
1642 req_capsule_set(&req->rq_pill, &RQF_OST_CONNECT);
1643 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET))
1645 rc = target_handle_connect(req);
1646 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2))
1649 struct obd_export *exp = req->rq_export;
1653 rc = filter_export_check_flavor(&obd->u.filter,
1658 case OST_DISCONNECT:
1659 CDEBUG(D_INODE, "disconnect\n");
1660 req_capsule_set(&req->rq_pill, &RQF_OST_DISCONNECT);
1661 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DISCONNECT_NET))
1663 rc = target_handle_disconnect(req);
1666 CDEBUG(D_INODE, "create\n");
1667 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CREATE_NET))
1669 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
1670 GOTO(out, rc = -ENOSPC);
1671 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1672 GOTO(out, rc = -EROFS);
1673 rc = ost_create(req->rq_export, req, oti);
1676 CDEBUG(D_INODE, "destroy\n");
1677 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DESTROY_NET))
1679 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1680 GOTO(out, rc = -EROFS);
1681 rc = ost_destroy(req->rq_export, req, oti);
1684 CDEBUG(D_INODE, "getattr\n");
1685 if (OBD_FAIL_CHECK(OBD_FAIL_OST_GETATTR_NET))
1687 rc = ost_getattr(req->rq_export, req);
1690 CDEBUG(D_INODE, "setattr\n");
1691 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_NET))
1693 rc = ost_setattr(req->rq_export, req, oti);
1696 CDEBUG(D_INODE, "write\n");
1697 /* req->rq_request_portal would be nice, if it was set */
1698 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
1699 CERROR("%s: deny write request from %s to portal %u\n",
1700 req->rq_export->exp_obd->obd_name,
1701 obd_export_nid2str(req->rq_export),
1702 req->rq_rqbd->rqbd_service->srv_req_portal);
1703 GOTO(out, rc = -EPROTO);
1705 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
1707 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
1708 GOTO(out, rc = -ENOSPC);
1709 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1710 GOTO(out, rc = -EROFS);
1711 rc = ost_brw_write(req, oti);
1712 LASSERT(current->journal_info == NULL);
1713 /* ost_brw_write sends its own replies */
1716 CDEBUG(D_INODE, "read\n");
1717 /* req->rq_request_portal would be nice, if it was set */
1718 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
1719 CERROR("%s: deny read request from %s to portal %u\n",
1720 req->rq_export->exp_obd->obd_name,
1721 obd_export_nid2str(req->rq_export),
1722 req->rq_rqbd->rqbd_service->srv_req_portal);
1723 GOTO(out, rc = -EPROTO);
1725 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
1727 rc = ost_brw_read(req, oti);
1728 LASSERT(current->journal_info == NULL);
1729 /* ost_brw_read sends its own replies */
1732 CDEBUG(D_INODE, "punch\n");
1733 if (OBD_FAIL_CHECK(OBD_FAIL_OST_PUNCH_NET))
1735 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1736 GOTO(out, rc = -EROFS);
1737 rc = ost_punch(req->rq_export, req, oti);
1740 CDEBUG(D_INODE, "statfs\n");
1741 if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_NET))
1743 rc = ost_statfs(req);
1746 CDEBUG(D_INODE, "sync\n");
1747 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SYNC_NET))
1749 rc = ost_sync(req->rq_export, req);
1752 DEBUG_REQ(D_INODE, req, "set_info");
1753 req_capsule_set(&req->rq_pill, &RQF_OST_SET_INFO);
1754 rc = ost_set_info(req->rq_export, req);
1757 DEBUG_REQ(D_INODE, req, "get_info");
1758 rc = ost_get_info(req->rq_export, req);
1760 case OST_QUOTACHECK:
1761 CDEBUG(D_INODE, "quotacheck\n");
1762 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACHECK);
1763 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACHECK_NET))
1765 rc = ost_handle_quotacheck(req);
1768 CDEBUG(D_INODE, "quotactl\n");
1769 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACTL);
1770 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACTL_NET))
1772 rc = ost_handle_quotactl(req);
1775 DEBUG_REQ(D_INODE, req, "ping");
1776 req_capsule_set(&req->rq_pill, &RQF_OBD_PING);
1777 rc = target_handle_ping(req);
1779 /* FIXME - just reply status */
1780 case LLOG_ORIGIN_CONNECT:
1781 DEBUG_REQ(D_INODE, req, "log connect");
1782 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_CONNECT);
1783 rc = ost_llog_handle_connect(req->rq_export, req);
1784 req->rq_status = rc;
1785 rc = req_capsule_server_pack(&req->rq_pill);
1788 RETURN(ptlrpc_reply(req));
1789 case OBD_LOG_CANCEL:
1790 CDEBUG(D_INODE, "log cancel\n");
1791 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1792 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1794 rc = llog_origin_handle_cancel(req);
1795 req->rq_status = rc;
1796 rc = req_capsule_server_pack(&req->rq_pill);
1799 RETURN(ptlrpc_reply(req));
1801 CDEBUG(D_INODE, "enqueue\n");
1802 req_capsule_set(&req->rq_pill, &RQF_LDLM_ENQUEUE);
1803 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE))
1805 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1806 ldlm_server_blocking_ast,
1807 ldlm_server_glimpse_ast);
1808 fail = OBD_FAIL_OST_LDLM_REPLY_NET;
1811 CDEBUG(D_INODE, "convert\n");
1812 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
1813 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CONVERT))
1815 rc = ldlm_handle_convert(req);
1818 CDEBUG(D_INODE, "cancel\n");
1819 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
1820 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
1822 rc = ldlm_handle_cancel(req);
1824 case LDLM_BL_CALLBACK:
1825 case LDLM_CP_CALLBACK:
1826 CDEBUG(D_INODE, "callback\n");
1827 CERROR("callbacks should not happen on OST\n");
1830 CERROR("Unexpected opcode %d\n",
1831 lustre_msg_get_opc(req->rq_reqmsg));
1832 req->rq_status = -ENOTSUPP;
1833 rc = ptlrpc_error(req);
1837 LASSERT(current->journal_info == NULL);
1840 /* If we're DISCONNECTing, the export_data is already freed */
1841 if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != OST_DISCONNECT)
1842 target_committed_to_req(req);
1846 oti_to_request(oti, req);
1848 target_send_reply(req, rc, fail);
1851 EXPORT_SYMBOL(ost_handle);
1853 * free per-thread pool created by ost_thread_init().
1855 static void ost_thread_done(struct ptlrpc_thread *thread)
1858 struct ost_thread_local_cache *tls; /* TLS stands for Thread-Local
1863 LASSERT(thread != NULL);
1866 * be prepared to handle partially-initialized pools (because this is
1867 * called from ost_thread_init() for cleanup.
1869 tls = thread->t_data;
1871 for (i = 0; i < OST_THREAD_POOL_SIZE; ++ i) {
1872 if (tls->page[i] != NULL)
1873 OBD_PAGE_FREE(tls->page[i]);
1876 thread->t_data = NULL;
1882 * initialize per-thread page pool (bug 5137).
1884 static int ost_thread_init(struct ptlrpc_thread *thread)
1888 struct ost_thread_local_cache *tls;
1892 LASSERT(thread != NULL);
1893 LASSERT(thread->t_data == NULL);
1894 LASSERTF(thread->t_id <= OSS_THREADS_MAX, "%u\n", thread->t_id);
1899 thread->t_data = tls;
1903 for (i = 0; i < OST_THREAD_POOL_SIZE; ++ i) {
1904 OBD_PAGE_ALLOC(tls->page[i], OST_THREAD_POOL_GFP);
1905 if (tls->page[i] == NULL) {
1906 ost_thread_done(thread);
1916 #define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000)
1918 /* Sigh - really, this is an OSS, the _server_, not the _target_ */
1919 static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
1921 struct ost_obd *ost = &obd->u.ost;
1922 struct lprocfs_static_vars lvars;
1923 int oss_min_threads;
1924 int oss_max_threads;
1925 int oss_min_create_threads;
1926 int oss_max_create_threads;
1930 rc = cleanup_group_info();
1934 lprocfs_ost_init_vars(&lvars);
1935 lprocfs_obd_setup(obd, lvars.obd_vars);
1937 sema_init(&ost->ost_health_sem, 1);
1939 if (oss_num_threads) {
1940 /* If oss_num_threads is set, it is the min and the max. */
1941 if (oss_num_threads > OSS_THREADS_MAX)
1942 oss_num_threads = OSS_THREADS_MAX;
1943 if (oss_num_threads < OSS_THREADS_MIN)
1944 oss_num_threads = OSS_THREADS_MIN;
1945 oss_max_threads = oss_min_threads = oss_num_threads;
1947 /* Base min threads on memory and cpus */
1948 oss_min_threads = num_possible_cpus() * num_physpages >>
1949 (27 - CFS_PAGE_SHIFT);
1950 if (oss_min_threads < OSS_THREADS_MIN)
1951 oss_min_threads = OSS_THREADS_MIN;
1952 /* Insure a 4x range for dynamic threads */
1953 if (oss_min_threads > OSS_THREADS_MAX / 4)
1954 oss_min_threads = OSS_THREADS_MAX / 4;
1955 oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4);
1959 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1960 OST_MAXREPSIZE, OST_REQUEST_PORTAL,
1961 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
1962 ost_handle, LUSTRE_OSS_NAME,
1963 obd->obd_proc_entry, target_print_req,
1964 oss_min_threads, oss_max_threads,
1965 "ll_ost", LCT_DT_THREAD);
1966 if (ost->ost_service == NULL) {
1967 CERROR("failed to start service\n");
1968 GOTO(out_lprocfs, rc = -ENOMEM);
1971 rc = ptlrpc_start_threads(obd, ost->ost_service);
1973 GOTO(out_service, rc = -EINVAL);
1975 if (oss_num_create_threads) {
1976 if (oss_num_create_threads > OSS_MAX_CREATE_THREADS)
1977 oss_num_create_threads = OSS_MAX_CREATE_THREADS;
1978 if (oss_num_create_threads < OSS_MIN_CREATE_THREADS)
1979 oss_num_create_threads = OSS_MIN_CREATE_THREADS;
1980 oss_min_create_threads = oss_max_create_threads =
1981 oss_num_create_threads;
1983 oss_min_create_threads = OSS_MIN_CREATE_THREADS;
1984 oss_max_create_threads = OSS_MAX_CREATE_THREADS;
1987 ost->ost_create_service =
1988 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1989 OST_MAXREPSIZE, OST_CREATE_PORTAL,
1990 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
1991 ost_handle, "ost_create",
1992 obd->obd_proc_entry, target_print_req,
1993 oss_min_create_threads, oss_max_create_threads,
1994 "ll_ost_creat", LCT_DT_THREAD);
1995 if (ost->ost_create_service == NULL) {
1996 CERROR("failed to start OST create service\n");
1997 GOTO(out_service, rc = -ENOMEM);
2000 rc = ptlrpc_start_threads(obd, ost->ost_create_service);
2002 GOTO(out_create, rc = -EINVAL);
2004 ost->ost_io_service =
2005 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2006 OST_MAXREPSIZE, OST_IO_PORTAL,
2007 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2008 ost_handle, "ost_io",
2009 obd->obd_proc_entry, target_print_req,
2010 oss_min_threads, oss_max_threads,
2011 "ll_ost_io", LCT_DT_THREAD);
2012 if (ost->ost_io_service == NULL) {
2013 CERROR("failed to start OST I/O service\n");
2014 GOTO(out_create, rc = -ENOMEM);
2017 ost->ost_io_service->srv_init = ost_thread_init;
2018 ost->ost_io_service->srv_done = ost_thread_done;
2019 ost->ost_io_service->srv_cpu_affinity = 1;
2020 rc = ptlrpc_start_threads(obd, ost->ost_io_service);
2022 GOTO(out_io, rc = -EINVAL);
2024 ping_evictor_start();
2029 ptlrpc_unregister_service(ost->ost_io_service);
2030 ost->ost_io_service = NULL;
2032 ptlrpc_unregister_service(ost->ost_create_service);
2033 ost->ost_create_service = NULL;
2035 ptlrpc_unregister_service(ost->ost_service);
2036 ost->ost_service = NULL;
2038 lprocfs_obd_cleanup(obd);
2042 static int ost_cleanup(struct obd_device *obd)
2044 struct ost_obd *ost = &obd->u.ost;
2048 ping_evictor_stop();
2050 spin_lock_bh(&obd->obd_processing_task_lock);
2051 if (obd->obd_recovering) {
2052 target_cancel_recovery_timer(obd);
2053 obd->obd_recovering = 0;
2055 spin_unlock_bh(&obd->obd_processing_task_lock);
2057 down(&ost->ost_health_sem);
2058 ptlrpc_unregister_service(ost->ost_service);
2059 ptlrpc_unregister_service(ost->ost_create_service);
2060 ptlrpc_unregister_service(ost->ost_io_service);
2061 ost->ost_service = NULL;
2062 ost->ost_create_service = NULL;
2063 up(&ost->ost_health_sem);
2065 lprocfs_obd_cleanup(obd);
2070 static int ost_health_check(struct obd_device *obd)
2072 struct ost_obd *ost = &obd->u.ost;
2075 down(&ost->ost_health_sem);
2076 rc |= ptlrpc_service_health_check(ost->ost_service);
2077 rc |= ptlrpc_service_health_check(ost->ost_create_service);
2078 rc |= ptlrpc_service_health_check(ost->ost_io_service);
2079 up(&ost->ost_health_sem);
2082 * health_check to return 0 on healthy
2083 * and 1 on unhealthy.
2091 struct ost_thread_local_cache *ost_tls(struct ptlrpc_request *r)
2093 return (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
2096 /* use obd ops to offer management infrastructure */
2097 static struct obd_ops ost_obd_ops = {
2098 .o_owner = THIS_MODULE,
2099 .o_setup = ost_setup,
2100 .o_cleanup = ost_cleanup,
2101 .o_health_check = ost_health_check,
2105 static int __init ost_init(void)
2107 struct lprocfs_static_vars lvars;
2111 lprocfs_ost_init_vars(&lvars);
2112 rc = class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
2113 LUSTRE_OSS_NAME, NULL);
2115 if (ost_num_threads != 0 && oss_num_threads == 0) {
2116 LCONSOLE_INFO("ost_num_threads module parameter is deprecated, "
2117 "use oss_num_threads instead or unset both for "
2118 "dynamic thread startup\n");
2119 oss_num_threads = ost_num_threads;
2125 static void /*__exit*/ ost_exit(void)
2127 class_unregister_type(LUSTRE_OSS_NAME);
2130 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2131 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
2132 MODULE_LICENSE("GPL");
2134 module_init(ost_init);
2135 module_exit(ost_exit);