1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2001-2003 Cluster File Systems, Inc.
5 * Author: Peter J. Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
8 * This file is part of the Lustre file system, http://www.lustre.org
9 * Lustre is a trademark of Cluster File Systems, Inc.
11 * You may have signed or agreed to another license before downloading
12 * this software. If so, you are bound by the terms and conditions
13 * of that agreement, and the following does not apply to you. See the
14 * LICENSE file included with this distribution for more information.
16 * If you did not agree to a different license, then this copy of Lustre
17 * is open source software; you can redistribute it and/or modify it
18 * under the terms of version 2 of the GNU General Public License as
19 * published by the Free Software Foundation.
21 * In either case, Lustre is distributed in the hope that it will be
22 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
23 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
24 * license text for more details.
26 * Storage Target Handling functions
27 * Lustre Object Server Module (OST)
29 * This server is single threaded at present (but can easily be multi
30 * threaded). For testing and management it is treated as an
31 * obd_device, although it does not export a full OBD method table
32 * (the requests are coming in over the wire, so object target
33 * modules do not have a full method table.)
37 # define EXPORT_SYMTAB
39 #define DEBUG_SUBSYSTEM S_OST
41 #include <linux/module.h>
43 #include <lustre_net.h>
44 #include <lustre_dlm.h>
45 #include <lustre_export.h>
46 #include <lustre_debug.h>
47 #include <linux/init.h>
48 #include <lprocfs_status.h>
49 #include <lustre_commit_confd.h>
50 #include <libcfs/list.h>
51 #include <lustre_quota.h>
52 #include "ost_internal.h"
54 static int oss_num_threads;
55 CFS_MODULE_PARM(oss_num_threads, "i", int, 0444,
56 "number of OSS service threads to start");
58 static int ost_num_threads;
59 CFS_MODULE_PARM(ost_num_threads, "i", int, 0444,
60 "number of OST service threads to start (deprecated)");
62 static int oss_num_create_threads;
63 CFS_MODULE_PARM(oss_num_create_threads, "i", int, 0444,
64 "number of OSS create threads to start");
66 void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req)
68 struct oti_req_ack_lock *ack_lock;
75 lustre_msg_set_transno(req->rq_repmsg, oti->oti_transno);
76 req->rq_transno = oti->oti_transno;
78 /* XXX 4 == entries in oti_ack_locks??? */
79 for (ack_lock = oti->oti_ack_locks, i = 0; i < 4; i++, ack_lock++) {
82 /* XXX not even calling target_send_reply in some cases... */
83 ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode);
87 static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req,
88 struct obd_trans_info *oti)
90 struct ost_body *body, *repbody;
91 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
94 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
95 lustre_swab_ost_body);
99 if (lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1)) {
100 struct ldlm_request *dlm;
101 dlm = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*dlm),
102 lustre_swab_ldlm_request);
105 ldlm_request_cancel(req, dlm, 0);
108 rc = lustre_pack_reply(req, 2, size, NULL);
112 if (body->oa.o_valid & OBD_MD_FLCOOKIE)
113 oti->oti_logcookies = obdo_logcookie(&body->oa);
114 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
116 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
117 req->rq_status = obd_destroy(exp, &body->oa, NULL, oti, NULL);
121 static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req)
123 struct ost_body *body, *repbody;
124 struct obd_info oinfo = { { { 0 } } };
125 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
128 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
129 lustre_swab_ost_body);
133 rc = lustre_pack_reply(req, 2, size, NULL);
137 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
139 repbody->oa = body->oa;
141 oinfo.oi_oa = &repbody->oa;
142 if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
143 oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
145 req->rq_status = obd_getattr(exp, &oinfo);
149 static int ost_statfs(struct ptlrpc_request *req)
151 struct obd_statfs *osfs;
152 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*osfs) };
155 rc = lustre_pack_reply(req, 2, size, NULL);
159 osfs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*osfs));
161 req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs,
162 cfs_time_current_64() - HZ);
163 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
164 osfs->os_bfree = osfs->os_bavail = 64;
165 if (req->rq_status != 0)
166 CERROR("ost: statfs failed: rc %d\n", req->rq_status);
171 static int ost_create(struct obd_export *exp, struct ptlrpc_request *req,
172 struct obd_trans_info *oti)
174 struct ost_body *body, *repbody;
175 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
178 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
179 lustre_swab_ost_body);
183 rc = lustre_pack_reply(req, 2, size, NULL);
187 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
189 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
190 oti->oti_logcookies = obdo_logcookie(&repbody->oa);
191 req->rq_status = obd_create(exp, &repbody->oa, NULL, oti);
192 //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0);
197 * Helper function for ost_punch(): if asked by client, acquire [size, EOF]
198 * lock on the file being truncated.
200 static int ost_punch_lock_get(struct obd_export *exp, struct obdo *oa,
201 struct lustre_handle *lh)
204 struct ldlm_res_id res_id = { .name = { oa->o_id, 0, oa->o_gr, 0} };
205 ldlm_policy_data_t policy;
211 LASSERT(!lustre_handle_is_used(lh));
213 if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
214 !(oa->o_flags & OBD_FL_TRUNCLOCK))
217 CDEBUG(D_INODE, "OST-side truncate lock.\n");
220 finis = start + oa->o_blocks;
223 * standard truncate optimization: if file body is completely
224 * destroyed, don't send data back to the server.
226 flags = (start == 0) ? LDLM_AST_DISCARD_DATA : 0;
228 policy.l_extent.start = start & CFS_PAGE_MASK;
231 * If ->o_blocks is EOF it means "lock till the end of the
232 * file". Otherwise, it's size of a hole being punched (in bytes)
234 if (oa->o_blocks == OBD_OBJECT_EOF || finis < start)
235 policy.l_extent.end = OBD_OBJECT_EOF;
237 policy.l_extent.end = finis | ~CFS_PAGE_MASK;
239 RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
240 LDLM_EXTENT, &policy, LCK_PW, &flags,
241 ldlm_blocking_ast, ldlm_completion_ast,
242 ldlm_glimpse_ast, NULL, 0, NULL, lh));
246 * Helper function for ost_punch(): release lock acquired by
247 * ost_punch_lock_get(), if any.
249 static void ost_punch_lock_put(struct obd_export *exp, struct obdo *oa,
250 struct lustre_handle *lh)
253 if (lustre_handle_is_used(lh))
254 ldlm_lock_decref(lh, LCK_PW);
258 static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req,
259 struct obd_trans_info *oti)
261 struct obd_info oinfo = { { { 0 } } };
262 struct ost_body *body, *repbody;
263 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
264 struct lustre_handle lh = {0,};
267 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
268 CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
270 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
271 lustre_swab_ost_body);
275 oinfo.oi_oa = &body->oa;
276 oinfo.oi_policy.l_extent.start = oinfo.oi_oa->o_size;
277 oinfo.oi_policy.l_extent.end = oinfo.oi_oa->o_blocks;
279 if ((oinfo.oi_oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
280 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
283 rc = lustre_pack_reply(req, 2, size, NULL);
287 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
289 rc = ost_punch_lock_get(exp, oinfo.oi_oa, &lh);
291 if (oinfo.oi_oa->o_valid & OBD_MD_FLFLAGS &&
292 oinfo.oi_oa->o_flags == OBD_FL_TRUNCLOCK)
294 * If OBD_FL_TRUNCLOCK is the only bit set in
295 * ->o_flags, clear OBD_MD_FLFLAGS to avoid falling
296 * through filter_setattr() to filter_iocontrol().
298 oinfo.oi_oa->o_valid &= ~OBD_MD_FLFLAGS;
300 if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
301 oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
303 req->rq_status = obd_punch(exp, &oinfo, oti, NULL);
304 ost_punch_lock_put(exp, oinfo.oi_oa, &lh);
306 repbody->oa = *oinfo.oi_oa;
310 static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req)
312 struct ost_body *body, *repbody;
313 struct lustre_capa *capa = NULL;
314 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
317 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
318 lustre_swab_ost_body);
322 if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
323 capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 1);
325 rc = lustre_pack_reply(req, 2, size, NULL);
329 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
331 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
332 req->rq_status = obd_sync(exp, &repbody->oa, NULL, repbody->oa.o_size,
333 repbody->oa.o_blocks, capa);
337 static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req,
338 struct obd_trans_info *oti)
340 struct ost_body *body, *repbody;
341 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
342 struct obd_info oinfo = { { { 0 } } };
345 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
346 lustre_swab_ost_body);
350 rc = lustre_pack_reply(req, 2, size, NULL);
354 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
356 repbody->oa = body->oa;
358 oinfo.oi_oa = &repbody->oa;
359 if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
360 oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
362 req->rq_status = obd_setattr(exp, &oinfo, oti);
366 static int ost_bulk_timeout(void *data)
369 /* We don't fail the connection here, because having the export
370 * killed makes the (vital) call to commitrw very sad.
375 static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo,
376 struct niobuf_remote *rnb, int nrnb,
377 struct niobuf_remote **pp_rnbp)
379 /* Copy a remote niobuf, splitting it into page-sized chunks
380 * and setting ioo[i].ioo_bufcnt accordingly */
381 struct niobuf_remote *pp_rnb;
389 * array of sufficient size already preallocated by caller
391 LASSERT(pp_rnbp != NULL);
392 LASSERT(*pp_rnbp != NULL);
394 /* first count and check the number of pages required */
395 for (i = 0; i < nioo; i++)
396 for (j = 0; j < ioo->ioo_bufcnt; j++, rnbidx++) {
397 obd_off offset = rnb[rnbidx].offset;
398 obd_off p0 = offset >> CFS_PAGE_SHIFT;
399 obd_off pn = (offset + rnb[rnbidx].len - 1) >>
402 LASSERT(rnbidx < nrnb);
404 npages += (pn + 1 - p0);
406 if (rnb[rnbidx].len == 0) {
407 CERROR("zero len BRW: obj %d objid "LPX64
408 " buf %u\n", i, ioo[i].ioo_id, j);
412 rnb[rnbidx].offset <= rnb[rnbidx-1].offset) {
413 CERROR("unordered BRW: obj %d objid "LPX64
414 " buf %u offset "LPX64" <= "LPX64"\n",
415 i, ioo[i].ioo_id, j, rnb[rnbidx].offset,
421 LASSERT(rnbidx == nrnb);
423 if (npages == nrnb) { /* all niobufs are for single pages */
430 /* now do the actual split */
432 for (i = 0; i < nioo; i++) {
435 for (j = 0; j < ioo[i].ioo_bufcnt; j++, rnbidx++) {
436 obd_off off = rnb[rnbidx].offset;
437 int nob = rnb[rnbidx].len;
439 LASSERT(rnbidx < nrnb);
441 obd_off poff = off & ~CFS_PAGE_MASK;
442 int pnob = (poff + nob > CFS_PAGE_SIZE) ?
443 PAGE_SIZE - poff : nob;
445 LASSERT(page < npages);
446 pp_rnb[page].len = pnob;
447 pp_rnb[page].offset = off;
448 pp_rnb[page].flags = rnb[rnbidx].flags;
450 CDEBUG(0, " obj %d id "LPX64
451 "page %d(%d) "LPX64" for %d, flg %x\n",
452 i, ioo[i].ioo_id, obj_pages, page,
453 pp_rnb[page].offset, pp_rnb[page].len,
463 ioo[i].ioo_bufcnt = obj_pages;
465 LASSERT(page == npages);
470 static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc, int opc)
475 for (i = 0; i < desc->bd_iov_count; i++) {
476 struct page *page = desc->bd_iov[i].kiov_page;
477 int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
478 char *ptr = kmap(page) + off;
479 int len = desc->bd_iov[i].kiov_len;
481 /* corrupt the data before we compute the checksum, to
482 * simulate a client->OST data error */
483 if (i == 0 && opc == OST_WRITE &&
484 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE))
485 memcpy(ptr, "bad3", min(4, len));
486 cksum = crc32_le(cksum, ptr, len);
487 /* corrupt the data after we compute the checksum, to
488 * simulate an OST->client data error */
489 if (i == 0 && opc == OST_READ &&
490 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))
491 memcpy(ptr, "bad4", min(4, len));
499 * populate @nio by @nrpages pages from per-thread page pool
501 static void ost_nio_pages_get(struct ptlrpc_request *req,
502 struct niobuf_local *nio, int nrpages)
505 struct ost_thread_local_cache *tls;
509 LASSERT(nrpages <= OST_THREAD_POOL_SIZE);
510 LASSERT(req != NULL);
511 LASSERT(req->rq_svc_thread != NULL);
514 LASSERT(tls != NULL);
516 memset(nio, 0, nrpages * sizeof *nio);
517 for (i = 0; i < nrpages; ++ i) {
521 LASSERT(page != NULL);
522 POISON_PAGE(page, 0xf1);
524 LL_CDEBUG_PAGE(D_INFO, page, "%d\n", i);
530 * Dual for ost_nio_pages_get(). Poison pages in pool for debugging
532 static void ost_nio_pages_put(struct ptlrpc_request *req,
533 struct niobuf_local *nio, int nrpages)
539 LASSERT(nrpages <= OST_THREAD_POOL_SIZE);
541 for (i = 0; i < nrpages; ++ i)
542 POISON_PAGE(nio[i].page, 0xf2);
546 static int ost_brw_lock_get(int mode, struct obd_export *exp,
547 struct obd_ioobj *obj, struct niobuf_remote *nb,
548 struct lustre_handle *lh)
551 int nrbufs = obj->ioo_bufcnt;
552 struct ldlm_res_id res_id = { .name = { obj->ioo_id, 0,
554 ldlm_policy_data_t policy;
559 LASSERT(mode == LCK_PR || mode == LCK_PW);
560 LASSERT(!lustre_handle_is_used(lh));
562 if (nrbufs == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
565 /* EXPENSIVE ASSERTION */
566 for (i = 1; i < nrbufs; i ++)
567 LASSERT((nb[0].flags & OBD_BRW_SRVLOCK) ==
568 (nb[i].flags & OBD_BRW_SRVLOCK));
570 policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
571 policy.l_extent.end = (nb[nrbufs - 1].offset +
572 nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
574 RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
575 LDLM_EXTENT, &policy, mode, &flags,
576 ldlm_blocking_ast, ldlm_completion_ast,
577 ldlm_glimpse_ast, NULL, 0, NULL, lh));
580 static void ost_brw_lock_put(int mode,
581 struct obd_ioobj *obj, struct niobuf_remote *niob,
582 struct lustre_handle *lh)
585 LASSERT(mode == LCK_PR || mode == LCK_PW);
586 LASSERT((obj->ioo_bufcnt > 0 && (niob[0].flags & OBD_BRW_SRVLOCK)) ==
587 lustre_handle_is_used(lh));
588 if (lustre_handle_is_used(lh))
589 ldlm_lock_decref(lh, mode);
593 struct ost_prolong_data {
594 struct obd_export *opd_exp;
595 ldlm_policy_data_t opd_policy;
596 ldlm_mode_t opd_mode;
599 static int ost_prolong_locks_iter(struct ldlm_lock *lock, void *data)
601 struct ost_prolong_data *opd = data;
603 LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
605 if (lock->l_req_mode != lock->l_granted_mode) {
606 /* scan granted locks only */
607 return LDLM_ITER_STOP;
610 if (lock->l_export != opd->opd_exp) {
611 /* prolong locks only for given client */
612 return LDLM_ITER_CONTINUE;
615 if (!(lock->l_granted_mode & opd->opd_mode)) {
616 /* we aren't interesting in all type of locks */
617 return LDLM_ITER_CONTINUE;
620 if (lock->l_policy_data.l_extent.end < opd->opd_policy.l_extent.start ||
621 lock->l_policy_data.l_extent.start > opd->opd_policy.l_extent.end) {
622 /* the request doesn't cross the lock, skip it */
623 return LDLM_ITER_CONTINUE;
626 if (!(lock->l_flags & LDLM_FL_AST_SENT)) {
627 /* ignore locks not being cancelled */
628 return LDLM_ITER_CONTINUE;
631 /* OK. this is a possible lock the user holds doing I/O
632 * let's refresh eviction timer for it */
633 ldlm_refresh_waiting_lock(lock);
635 return LDLM_ITER_CONTINUE;
638 static void ost_prolong_locks(struct obd_export *exp, struct obd_ioobj *obj,
639 struct niobuf_remote *nb, ldlm_mode_t mode)
641 struct ldlm_res_id res_id = { .name = { obj->ioo_id, 0,
643 int nrbufs = obj->ioo_bufcnt;
644 struct ost_prolong_data opd;
650 opd.opd_policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
651 opd.opd_policy.l_extent.end = (nb[nrbufs - 1].offset +
652 nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
654 CDEBUG(D_DLMTRACE,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
655 res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start,
656 opd.opd_policy.l_extent.end);
657 ldlm_resource_iterate(exp->exp_obd->obd_namespace, &res_id,
658 ost_prolong_locks_iter, &opd);
661 static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti)
663 struct ptlrpc_bulk_desc *desc;
664 struct obd_export *exp = req->rq_export;
665 struct niobuf_remote *remote_nb;
666 struct niobuf_remote *pp_rnb = NULL;
667 struct niobuf_local *local_nb;
668 struct obd_ioobj *ioo;
669 struct ost_body *body, *repbody;
670 struct lustre_capa *capa = NULL;
671 struct l_wait_info lwi;
672 struct lustre_handle lockh = { 0 };
673 int size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
674 int niocount, npages, nob = 0, rc, i;
678 req->rq_bulk_read = 1;
680 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
681 GOTO(out, rc = -EIO);
683 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
685 /* Check if there is eviction in progress, and if so, wait for it to
687 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
688 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
689 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
690 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
694 GOTO(out, rc = -ENOTCONN);
696 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
697 lustre_swab_ost_body);
699 CERROR("Missing/short ost_body\n");
700 GOTO(out, rc = -EFAULT);
703 ioo = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*ioo),
704 lustre_swab_obd_ioobj);
706 CERROR("Missing/short ioobj\n");
707 GOTO(out, rc = -EFAULT);
710 niocount = ioo->ioo_bufcnt;
711 if (niocount > PTLRPC_MAX_BRW_PAGES) {
712 DEBUG_REQ(D_ERROR, req, "bulk has too many pages (%d)",
714 GOTO(out, rc = -EFAULT);
717 remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2,
718 niocount * sizeof(*remote_nb),
719 lustre_swab_niobuf_remote);
720 if (remote_nb == NULL) {
721 CERROR("Missing/short niobuf\n");
722 GOTO(out, rc = -EFAULT);
724 if (lustre_msg_swabbed(req->rq_reqmsg)) { /* swab remaining niobufs */
725 for (i = 1; i < niocount; i++)
726 lustre_swab_niobuf_remote (&remote_nb[i]);
729 if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
730 capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3);
732 rc = lustre_pack_reply(req, 2, size, NULL);
737 * Per-thread array of struct niobuf_{local,remote}'s was allocated by
740 local_nb = ost_tls(req)->local;
741 pp_rnb = ost_tls(req)->remote;
743 /* FIXME all niobuf splitting should be done in obdfilter if needed */
744 /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
745 npages = get_per_page_niobufs(ioo, 1, remote_nb, niocount, &pp_rnb);
747 GOTO(out, rc = npages);
749 LASSERT(npages <= OST_THREAD_POOL_SIZE);
751 ost_nio_pages_get(req, local_nb, npages);
753 desc = ptlrpc_prep_bulk_exp(req, npages,
754 BULK_PUT_SOURCE, OST_BULK_PORTAL);
756 GOTO(out, rc = -ENOMEM);
758 rc = ost_brw_lock_get(LCK_PR, exp, ioo, pp_rnb, &lockh);
763 * If getting the lock took more time than
764 * client was willing to wait, drop it. b=11330
766 if (cfs_time_current_sec() > req->rq_arrival_time.tv_sec + obd_timeout ||
767 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
769 CERROR("Dropping timed-out read from %s because locking"
770 "object "LPX64" took %ld seconds.\n",
771 libcfs_id2str(req->rq_peer), ioo->ioo_id,
772 cfs_time_current_sec() - req->rq_arrival_time.tv_sec);
773 GOTO(out_lock, rc = -ETIMEDOUT);
776 rc = obd_preprw(OBD_BRW_READ, exp, &body->oa, 1,
777 ioo, npages, pp_rnb, local_nb, oti, capa);
781 ost_prolong_locks(exp, ioo, pp_rnb, LCK_PW | LCK_PR);
784 for (i = 0; i < npages; i++) {
785 int page_rc = local_nb[i].rc;
787 if (page_rc < 0) { /* error */
792 LASSERTF(page_rc <= pp_rnb[i].len, "page_rc (%d) > "
793 "pp_rnb[%d].len (%d)\n", page_rc, i, pp_rnb[i].len);
795 if (page_rc != 0) { /* some data! */
796 LASSERT (local_nb[i].page != NULL);
797 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
798 pp_rnb[i].offset & ~CFS_PAGE_MASK,
802 if (page_rc != pp_rnb[i].len) { /* short read */
803 /* All subsequent pages should be 0 */
805 LASSERT(local_nb[i].rc == 0);
810 if (unlikely(body->oa.o_valid & OBD_MD_FLCKSUM)) {
811 body->oa.o_cksum = ost_checksum_bulk(desc, OST_READ);
812 body->oa.o_valid = OBD_MD_FLCKSUM;
813 CDEBUG(D_PAGE,"checksum at read origin: %x\n",body->oa.o_cksum);
815 body->oa.o_valid = 0;
817 /* We're finishing using body->oa as an input variable */
819 /* Check if client was evicted while we were doing i/o before touching
822 /* Check if there is eviction in progress, and if so, wait for
824 if (unlikely(atomic_read(&exp->exp_obd->
825 obd_evict_inprogress))) {
826 lwi = LWI_INTR(NULL, NULL);
827 rc = l_wait_event(exp->exp_obd->
828 obd_evict_inprogress_waitq,
829 !atomic_read(&exp->exp_obd->
830 obd_evict_inprogress),
836 sptlrpc_svc_wrap_bulk(req, desc);
838 rc = ptlrpc_start_bulk_transfer(desc);
842 lwi = LWI_TIMEOUT_INTERVAL(obd_timeout * HZ / 4, HZ,
843 ost_bulk_timeout, desc);
844 rc = l_wait_event(desc->bd_waitq,
845 !ptlrpc_bulk_active(desc) ||
846 exp->exp_failed, &lwi);
847 LASSERT(rc == 0 || rc == -ETIMEDOUT);
848 if (rc == -ETIMEDOUT) {
849 DEBUG_REQ(D_ERROR, req, "timeout on bulk PUT");
850 ptlrpc_abort_bulk(desc);
851 } else if (exp->exp_failed) {
852 DEBUG_REQ(D_ERROR, req, "Eviction on bulk PUT");
854 ptlrpc_abort_bulk(desc);
855 } else if (!desc->bd_success ||
856 desc->bd_nob_transferred != desc->bd_nob) {
857 DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)",
859 "truncated" : "network error on",
860 desc->bd_nob_transferred,
862 /* XXX should this be a different errno? */
866 DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d", rc);
871 /* Must commit after prep above in all cases */
872 rc = obd_commitrw(OBD_BRW_READ, exp, &body->oa, 1,
873 ioo, npages, local_nb, oti, rc);
875 ost_nio_pages_put(req, local_nb, npages);
878 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
880 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
884 ost_brw_lock_put(LCK_PR, ioo, pp_rnb, &lockh);
886 ptlrpc_free_bulk(desc);
890 req->rq_status = nob;
891 target_committed_to_req(req);
893 } else if (!no_reply) {
894 /* Only reply if there was no comms problem with bulk */
895 target_committed_to_req(req);
899 if (req->rq_reply_state != NULL) {
900 /* reply out callback would free */
901 ptlrpc_rs_decref(req->rq_reply_state);
902 req->rq_reply_state = NULL;
904 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
905 "client will retry\n",
906 exp->exp_obd->obd_name,
907 exp->exp_client_uuid.uuid,
908 exp->exp_connection->c_remote_uuid.uuid,
909 libcfs_id2str(req->rq_peer));
915 static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
917 struct ptlrpc_bulk_desc *desc;
918 struct obd_export *exp = req->rq_export;
919 struct niobuf_remote *remote_nb;
920 struct niobuf_remote *pp_rnb;
921 struct niobuf_local *local_nb;
922 struct obd_ioobj *ioo;
923 struct ost_body *body, *repbody;
924 struct l_wait_info lwi;
925 struct lustre_handle lockh = {0};
926 struct lustre_capa *capa = NULL;
928 int size[3] = { sizeof(struct ptlrpc_body), sizeof(*body) };
929 int objcount, niocount, npages;
931 obd_count client_cksum, server_cksum = 0;
935 req->rq_bulk_write = 1;
937 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
938 GOTO(out, rc = -EIO);
939 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
940 GOTO(out, rc = -EFAULT);
942 /* pause before transaction has been started */
943 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
945 /* Check if there is eviction in progress, and if so, wait for it to
947 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
948 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
949 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
950 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
954 GOTO(out, rc = -ENOTCONN);
956 swab = lustre_msg_swabbed(req->rq_reqmsg);
957 body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
958 lustre_swab_ost_body);
960 CERROR("Missing/short ost_body\n");
961 GOTO(out, rc = -EFAULT);
964 lustre_set_req_swabbed(req, REQ_REC_OFF + 1);
965 objcount = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1) /
968 CERROR("Missing/short ioobj\n");
969 GOTO(out, rc = -EFAULT);
972 CERROR("too many ioobjs (%d)\n", objcount);
973 GOTO(out, rc = -EFAULT);
976 ioo = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1,
977 objcount * sizeof(*ioo));
978 LASSERT (ioo != NULL);
979 for (niocount = i = 0; i < objcount; i++) {
981 lustre_swab_obd_ioobj(&ioo[i]);
982 if (ioo[i].ioo_bufcnt == 0) {
983 CERROR("ioo[%d] has zero bufcnt\n", i);
984 GOTO(out, rc = -EFAULT);
986 niocount += ioo[i].ioo_bufcnt;
989 if (niocount > PTLRPC_MAX_BRW_PAGES) {
990 DEBUG_REQ(D_ERROR, req, "bulk has too many pages (%d)",
992 GOTO(out, rc = -EFAULT);
995 remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2,
996 niocount * sizeof(*remote_nb),
997 lustre_swab_niobuf_remote);
998 if (remote_nb == NULL) {
999 CERROR("Missing/short niobuf\n");
1000 GOTO(out, rc = -EFAULT);
1002 if (swab) { /* swab the remaining niobufs */
1003 for (i = 1; i < niocount; i++)
1004 lustre_swab_niobuf_remote (&remote_nb[i]);
1007 if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
1008 capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3);
1010 size[REPLY_REC_OFF + 1] = niocount * sizeof(*rcs);
1011 rc = lustre_pack_reply(req, 3, size, NULL);
1014 rcs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1,
1015 niocount * sizeof(*rcs));
1018 * Per-thread array of struct niobuf_{local,remote}'s was allocated by
1019 * ost_thread_init().
1021 local_nb = ost_tls(req)->local;
1022 pp_rnb = ost_tls(req)->remote;
1024 /* FIXME all niobuf splitting should be done in obdfilter if needed */
1025 /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
1026 npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
1028 GOTO(out, rc = npages);
1030 LASSERT(npages <= OST_THREAD_POOL_SIZE);
1032 ost_nio_pages_get(req, local_nb, npages);
1034 desc = ptlrpc_prep_bulk_exp(req, npages,
1035 BULK_GET_SINK, OST_BULK_PORTAL);
1037 GOTO(out, rc = -ENOMEM);
1039 rc = ost_brw_lock_get(LCK_PW, exp, ioo, pp_rnb, &lockh);
1044 * If getting the lock took more time than
1045 * client was willing to wait, drop it. b=11330
1047 if (cfs_time_current_sec() > req->rq_arrival_time.tv_sec + obd_timeout ||
1048 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
1050 CERROR("Dropping timed-out write from %s because locking"
1051 "object "LPX64" took %ld seconds.\n",
1052 libcfs_id2str(req->rq_peer), ioo->ioo_id,
1053 cfs_time_current_sec() - req->rq_arrival_time.tv_sec);
1054 GOTO(out_lock, rc = -ETIMEDOUT);
1057 ost_prolong_locks(exp, ioo, pp_rnb, LCK_PW);
1059 /* obd_preprw clobbers oa->valid, so save what we need */
1060 client_cksum = body->oa.o_valid & OBD_MD_FLCKSUM ? body->oa.o_cksum : 0;
1062 /* Because we already sync grant info with client when reconnect,
1063 * grant info will be cleared for resent req, then fed_grant and
1064 * total_grant will not be modified in following preprw_write */
1065 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
1066 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
1067 body->oa.o_valid &= ~OBD_MD_FLGRANT;
1070 rc = obd_preprw(OBD_BRW_WRITE, exp, &body->oa, objcount,
1071 ioo, npages, pp_rnb, local_nb, oti, capa);
1075 /* NB Having prepped, we must commit... */
1077 for (i = 0; i < npages; i++)
1078 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
1079 pp_rnb[i].offset & ~CFS_PAGE_MASK,
1082 /* Check if client was evicted while we were doing i/o before touching
1084 if (desc->bd_export->exp_failed)
1087 rc = ptlrpc_start_bulk_transfer (desc);
1089 lwi = LWI_TIMEOUT_INTERVAL(obd_timeout * HZ / 2, HZ,
1090 ost_bulk_timeout, desc);
1091 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc) ||
1092 desc->bd_export->exp_failed, &lwi);
1093 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1094 if (rc == -ETIMEDOUT) {
1095 DEBUG_REQ(D_ERROR, req, "timeout on bulk GET");
1096 ptlrpc_abort_bulk(desc);
1097 } else if (desc->bd_export->exp_failed) {
1098 DEBUG_REQ(D_ERROR, req, "Eviction on bulk GET");
1100 ptlrpc_abort_bulk(desc);
1101 } else if (!desc->bd_success ||
1102 desc->bd_nob_transferred != desc->bd_nob) {
1103 DEBUG_REQ(D_ERROR, req, "%s bulk GET %d(%d)",
1105 "truncated" : "network error on",
1106 desc->bd_nob_transferred, desc->bd_nob);
1107 /* XXX should this be a different errno? */
1111 DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d", rc);
1116 /* let client retry if unwrap failed */
1117 rc = sptlrpc_svc_unwrap_bulk(req, desc);
1120 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1122 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
1124 if (unlikely(client_cksum != 0 && rc == 0)) {
1125 static int cksum_counter;
1126 server_cksum = ost_checksum_bulk(desc, OST_WRITE);
1127 repbody->oa.o_valid |= OBD_MD_FLCKSUM;
1128 repbody->oa.o_cksum = server_cksum;
1130 if (unlikely(client_cksum != server_cksum)) {
1131 CERROR("client csum %x, server csum %x\n",
1132 client_cksum, server_cksum);
1134 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
1135 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
1136 cksum_counter, libcfs_id2str(req->rq_peer),
1141 /* Must commit after prep above in all cases */
1142 rc = obd_commitrw(OBD_BRW_WRITE, exp, &repbody->oa,
1143 objcount, ioo, npages, local_nb, oti, rc);
1145 if (unlikely(client_cksum != server_cksum && rc == 0)) {
1146 int new_cksum = ost_checksum_bulk(desc, OST_WRITE);
1151 if (new_cksum == server_cksum)
1152 msg = "changed in transit before arrival at OST";
1153 else if (new_cksum == client_cksum)
1154 msg = "initial checksum before message complete";
1156 msg = "changed in transit AND after initial checksum";
1158 if (req->rq_peer.nid == desc->bd_sender) {
1162 router = libcfs_nid2str(desc->bd_sender);
1165 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: %s from "
1166 "%s%s%s inum "LPU64"/"LPU64" object "
1167 LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1168 exp->exp_obd->obd_name, msg,
1169 libcfs_id2str(req->rq_peer),
1171 body->oa.o_valid & OBD_MD_FLFID ?
1172 body->oa.o_fid : (__u64)0,
1173 body->oa.o_valid & OBD_MD_FLFID ?
1174 body->oa.o_generation :(__u64)0,
1176 body->oa.o_valid & OBD_MD_FLGROUP ?
1177 body->oa.o_gr : (__u64)0,
1179 pp_rnb[npages-1].offset+pp_rnb[npages-1].len
1181 CERROR("client csum %x, original server csum %x, "
1182 "server csum now %x\n",
1183 client_cksum, server_cksum, new_cksum);
1186 ost_nio_pages_put(req, local_nb, npages);
1189 /* set per-requested niobuf return codes */
1190 for (i = j = 0; i < niocount; i++) {
1191 int nob = remote_nb[i].len;
1195 LASSERT(j < npages);
1196 if (local_nb[j].rc < 0)
1197 rcs[i] = local_nb[j].rc;
1198 nob -= pp_rnb[j].len;
1203 LASSERT(j == npages);
1207 ost_brw_lock_put(LCK_PW, ioo, pp_rnb, &lockh);
1209 ptlrpc_free_bulk(desc);
1212 oti_to_request(oti, req);
1213 target_committed_to_req(req);
1214 rc = ptlrpc_reply(req);
1215 } else if (!no_reply) {
1216 /* Only reply if there was no comms problem with bulk */
1217 target_committed_to_req(req);
1218 req->rq_status = rc;
1221 if (req->rq_reply_state != NULL) {
1222 /* reply out callback would free */
1223 ptlrpc_rs_decref(req->rq_reply_state);
1224 req->rq_reply_state = NULL;
1226 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
1227 "client will retry\n",
1228 exp->exp_obd->obd_name,
1229 exp->exp_client_uuid.uuid,
1230 exp->exp_connection->c_remote_uuid.uuid,
1231 libcfs_id2str(req->rq_peer));
1236 static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
1238 char *key, *val = NULL;
1239 int keylen, vallen, rc = 0;
1242 key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1);
1244 DEBUG_REQ(D_HA, req, "no set_info key");
1247 keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF);
1249 rc = lustre_pack_reply(req, 1, NULL, NULL);
1253 vallen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1);
1255 val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, 0);
1257 if (KEY_IS("evict_by_nid")) {
1259 obd_export_evict_by_nid(exp->exp_obd, val);
1264 rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL);
1266 lustre_msg_set_status(req->rq_repmsg, 0);
1270 static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req)
1274 int size[2] = { sizeof(struct ptlrpc_body), 0 };
1277 key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1);
1279 DEBUG_REQ(D_HA, req, "no get_info key");
1282 keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF);
1284 /* call once to get the size to allocate the reply buffer */
1285 rc = obd_get_info(exp, keylen, key, &size[1], NULL);
1289 rc = lustre_pack_reply(req, 2, size, NULL);
1293 reply = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply));
1294 /* call again to fill in the reply buffer */
1295 rc = obd_get_info(exp, keylen, key, size, reply);
1296 lustre_msg_set_status(req->rq_repmsg, 0);
1301 static int ost_handle_quotactl(struct ptlrpc_request *req)
1303 struct obd_quotactl *oqctl, *repoqc;
1304 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) };
1307 oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1308 lustre_swab_obd_quotactl);
1310 GOTO(out, rc = -EPROTO);
1312 rc = lustre_pack_reply(req, 2, size, NULL);
1316 repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc));
1318 req->rq_status = obd_quotactl(req->rq_export, oqctl);
1324 static int ost_handle_quotacheck(struct ptlrpc_request *req)
1326 struct obd_quotactl *oqctl;
1330 oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1331 lustre_swab_obd_quotactl);
1335 rc = lustre_pack_reply(req, 1, NULL, NULL);
1339 req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1343 static int ost_llog_handle_connect(struct obd_export *exp,
1344 struct ptlrpc_request *req)
1346 struct llogd_conn_body *body;
1350 body = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*body));
1351 rc = obd_llog_connect(exp, body);
1356 static int ost_filter_recovery_request(struct ptlrpc_request *req,
1357 struct obd_device *obd, int *process)
1359 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1360 case OST_CONNECT: /* This will never get here, but for completeness. */
1361 case OST_DISCONNECT:
1372 case OBD_LOG_CANCEL:
1374 *process = target_queue_recovery_request(req, obd);
1378 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1384 int ost_msg_check_version(struct lustre_msg *msg)
1388 switch(lustre_msg_get_opc(msg)) {
1390 case OST_DISCONNECT:
1393 case SEC_CTX_INIT_CONT:
1395 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1397 CERROR("bad opc %u version %08x, expecting %08x\n",
1398 lustre_msg_get_opc(msg),
1399 lustre_msg_get_version(msg),
1400 LUSTRE_OBD_VERSION);
1413 case OST_QUOTACHECK:
1415 rc = lustre_msg_check_version(msg, LUSTRE_OST_VERSION);
1417 CERROR("bad opc %u version %08x, expecting %08x\n",
1418 lustre_msg_get_opc(msg),
1419 lustre_msg_get_version(msg),
1420 LUSTRE_OST_VERSION);
1425 case LDLM_BL_CALLBACK:
1426 case LDLM_CP_CALLBACK:
1427 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1429 CERROR("bad opc %u version %08x, expecting %08x\n",
1430 lustre_msg_get_opc(msg),
1431 lustre_msg_get_version(msg),
1432 LUSTRE_DLM_VERSION);
1434 case LLOG_ORIGIN_CONNECT:
1435 case OBD_LOG_CANCEL:
1436 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1438 CERROR("bad opc %u version %08x, expecting %08x\n",
1439 lustre_msg_get_opc(msg),
1440 lustre_msg_get_version(msg),
1441 LUSTRE_LOG_VERSION);
1444 CERROR("Unexpected opcode %d\n", lustre_msg_get_opc(msg));
1450 int ost_handle(struct ptlrpc_request *req)
1452 struct obd_trans_info trans_info = { 0, };
1453 struct obd_trans_info *oti = &trans_info;
1454 int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0;
1455 struct obd_device *obd = NULL;
1458 LASSERT(current->journal_info == NULL);
1460 /* primordial rpcs don't affect server recovery */
1461 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1463 case SEC_CTX_INIT_CONT:
1468 /* XXX identical to MDS */
1469 if (lustre_msg_get_opc(req->rq_reqmsg) != OST_CONNECT) {
1472 if (req->rq_export == NULL) {
1473 CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
1474 lustre_msg_get_opc(req->rq_reqmsg),
1475 libcfs_id2str(req->rq_peer));
1476 req->rq_status = -ENOTCONN;
1477 GOTO(out, rc = -ENOTCONN);
1480 obd = req->rq_export->exp_obd;
1482 /* Check for aborted recovery. */
1483 spin_lock_bh(&obd->obd_processing_task_lock);
1484 recovering = obd->obd_recovering;
1485 spin_unlock_bh(&obd->obd_processing_task_lock);
1487 rc = ost_filter_recovery_request(req, obd,
1489 if (rc || !should_process)
1491 else if (should_process < 0) {
1492 req->rq_status = should_process;
1493 rc = ptlrpc_error(req);
1501 rc = ost_msg_check_version(req->rq_reqmsg);
1505 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1507 CDEBUG(D_INODE, "connect\n");
1508 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET))
1510 rc = target_handle_connect(req);
1512 obd = req->rq_export->exp_obd;
1515 case OST_DISCONNECT:
1516 CDEBUG(D_INODE, "disconnect\n");
1517 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DISCONNECT_NET))
1519 rc = target_handle_disconnect(req);
1522 CDEBUG(D_INODE, "create\n");
1523 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CREATE_NET))
1525 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
1526 GOTO(out, rc = -ENOSPC);
1527 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1528 GOTO(out, rc = -EROFS);
1529 rc = ost_create(req->rq_export, req, oti);
1532 CDEBUG(D_INODE, "destroy\n");
1533 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DESTROY_NET))
1535 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1536 GOTO(out, rc = -EROFS);
1537 rc = ost_destroy(req->rq_export, req, oti);
1540 CDEBUG(D_INODE, "getattr\n");
1541 if (OBD_FAIL_CHECK(OBD_FAIL_OST_GETATTR_NET))
1543 rc = ost_getattr(req->rq_export, req);
1546 CDEBUG(D_INODE, "setattr\n");
1547 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_NET))
1549 rc = ost_setattr(req->rq_export, req, oti);
1552 CDEBUG(D_INODE, "write\n");
1553 /* req->rq_request_portal would be nice, if it was set */
1554 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
1555 CERROR("%s: deny write request from %s to portal %u\n",
1556 req->rq_export->exp_obd->obd_name,
1557 obd_export_nid2str(req->rq_export),
1558 req->rq_rqbd->rqbd_service->srv_req_portal);
1559 GOTO(out, rc = -EPROTO);
1561 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
1563 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
1564 GOTO(out, rc = -ENOSPC);
1565 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1566 GOTO(out, rc = -EROFS);
1567 rc = ost_brw_write(req, oti);
1568 LASSERT(current->journal_info == NULL);
1569 /* ost_brw_write sends its own replies */
1572 CDEBUG(D_INODE, "read\n");
1573 /* req->rq_request_portal would be nice, if it was set */
1574 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
1575 CERROR("%s: deny read request from %s to portal %u\n",
1576 req->rq_export->exp_obd->obd_name,
1577 obd_export_nid2str(req->rq_export),
1578 req->rq_rqbd->rqbd_service->srv_req_portal);
1579 GOTO(out, rc = -EPROTO);
1581 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
1583 rc = ost_brw_read(req, oti);
1584 LASSERT(current->journal_info == NULL);
1585 /* ost_brw_read sends its own replies */
1588 CDEBUG(D_INODE, "punch\n");
1589 if (OBD_FAIL_CHECK(OBD_FAIL_OST_PUNCH_NET))
1591 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1592 GOTO(out, rc = -EROFS);
1593 rc = ost_punch(req->rq_export, req, oti);
1596 CDEBUG(D_INODE, "statfs\n");
1597 if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_NET))
1599 rc = ost_statfs(req);
1602 CDEBUG(D_INODE, "sync\n");
1603 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SYNC_NET))
1605 rc = ost_sync(req->rq_export, req);
1608 DEBUG_REQ(D_INODE, req, "set_info");
1609 rc = ost_set_info(req->rq_export, req);
1612 DEBUG_REQ(D_INODE, req, "get_info");
1613 rc = ost_get_info(req->rq_export, req);
1615 case OST_QUOTACHECK:
1616 CDEBUG(D_INODE, "quotacheck\n");
1617 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACHECK_NET))
1619 rc = ost_handle_quotacheck(req);
1622 CDEBUG(D_INODE, "quotactl\n");
1623 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACTL_NET))
1625 rc = ost_handle_quotactl(req);
1628 DEBUG_REQ(D_INODE, req, "ping");
1629 rc = target_handle_ping(req);
1631 /* FIXME - just reply status */
1632 case LLOG_ORIGIN_CONNECT:
1633 DEBUG_REQ(D_INODE, req, "log connect");
1634 rc = ost_llog_handle_connect(req->rq_export, req);
1635 req->rq_status = rc;
1636 rc = lustre_pack_reply(req, 1, NULL, NULL);
1639 RETURN(ptlrpc_reply(req));
1640 case OBD_LOG_CANCEL:
1641 CDEBUG(D_INODE, "log cancel\n");
1642 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1644 rc = llog_origin_handle_cancel(req);
1645 req->rq_status = rc;
1646 rc = lustre_pack_reply(req, 1, NULL, NULL);
1649 RETURN(ptlrpc_reply(req));
1651 CDEBUG(D_INODE, "enqueue\n");
1652 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE))
1654 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1655 ldlm_server_blocking_ast,
1656 ldlm_server_glimpse_ast);
1657 fail = OBD_FAIL_OST_LDLM_REPLY_NET;
1660 CDEBUG(D_INODE, "convert\n");
1661 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CONVERT))
1663 rc = ldlm_handle_convert(req);
1666 CDEBUG(D_INODE, "cancel\n");
1667 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
1669 rc = ldlm_handle_cancel(req);
1671 case LDLM_BL_CALLBACK:
1672 case LDLM_CP_CALLBACK:
1673 CDEBUG(D_INODE, "callback\n");
1674 CERROR("callbacks should not happen on OST\n");
1677 CERROR("Unexpected opcode %d\n",
1678 lustre_msg_get_opc(req->rq_reqmsg));
1679 req->rq_status = -ENOTSUPP;
1680 rc = ptlrpc_error(req);
1684 LASSERT(current->journal_info == NULL);
1687 /* If we're DISCONNECTing, the export_data is already freed */
1688 if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != OST_DISCONNECT)
1689 target_committed_to_req(req);
1693 oti_to_request(oti, req);
1695 target_send_reply(req, rc, fail);
1698 EXPORT_SYMBOL(ost_handle);
1700 * free per-thread pool created by ost_thread_init().
1702 static void ost_thread_done(struct ptlrpc_thread *thread)
1705 struct ost_thread_local_cache *tls; /* TLS stands for Thread-Local
1710 LASSERT(thread != NULL);
1713 * be prepared to handle partially-initialized pools (because this is
1714 * called from ost_thread_init() for cleanup.
1716 tls = thread->t_data;
1718 for (i = 0; i < OST_THREAD_POOL_SIZE; ++ i) {
1719 if (tls->page[i] != NULL)
1720 OBD_PAGE_FREE(tls->page[i]);
1723 thread->t_data = NULL;
1729 * initialize per-thread page pool (bug 5137).
1731 static int ost_thread_init(struct ptlrpc_thread *thread)
1735 struct ost_thread_local_cache *tls;
1739 LASSERT(thread != NULL);
1740 LASSERT(thread->t_data == NULL);
1741 LASSERTF(thread->t_id <= OSS_THREADS_MAX, "%u\n", thread->t_id);
1746 thread->t_data = tls;
1750 for (i = 0; i < OST_THREAD_POOL_SIZE; ++ i) {
1751 OBD_PAGE_ALLOC(tls->page[i], OST_THREAD_POOL_GFP);
1752 if (tls->page[i] == NULL) {
1753 ost_thread_done(thread);
1763 #define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000)
1765 /* Sigh - really, this is an OSS, the _server_, not the _target_ */
1766 static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
1768 struct ost_obd *ost = &obd->u.ost;
1769 struct lprocfs_static_vars lvars;
1770 int oss_min_threads;
1771 int oss_max_threads;
1772 int oss_min_create_threads;
1773 int oss_max_create_threads;
1777 rc = cleanup_group_info();
1781 rc = llog_start_commit_thread();
1785 lprocfs_init_vars(ost, &lvars);
1786 lprocfs_obd_setup(obd, lvars.obd_vars);
1788 sema_init(&ost->ost_health_sem, 1);
1790 if (oss_num_threads) {
1791 /* If oss_num_threads is set, it is the min and the max. */
1792 if (oss_num_threads > OSS_THREADS_MAX)
1793 oss_num_threads = OSS_THREADS_MAX;
1794 if (oss_num_threads < OSS_THREADS_MIN)
1795 oss_num_threads = OSS_THREADS_MIN;
1796 oss_max_threads = oss_min_threads = oss_num_threads;
1798 /* Base min threads on memory and cpus */
1799 oss_min_threads = num_possible_cpus() * num_physpages >>
1800 (27 - CFS_PAGE_SHIFT);
1801 if (oss_min_threads < OSS_THREADS_MIN)
1802 oss_min_threads = OSS_THREADS_MIN;
1803 /* Insure a 4x range for dynamic threads */
1804 if (oss_min_threads > OSS_THREADS_MAX / 4)
1805 oss_min_threads = OSS_THREADS_MAX / 4;
1806 oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4);
1810 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1811 OST_MAXREPSIZE, OST_REQUEST_PORTAL,
1813 OST_WATCHDOG_TIMEOUT, ost_handle,
1814 LUSTRE_OSS_NAME, obd->obd_proc_entry,
1815 ost_print_req, oss_min_threads,
1816 oss_max_threads, "ll_ost",
1818 if (ost->ost_service == NULL) {
1819 CERROR("failed to start service\n");
1820 GOTO(out_lprocfs, rc = -ENOMEM);
1823 rc = ptlrpc_start_threads(obd, ost->ost_service);
1825 GOTO(out_service, rc = -EINVAL);
1827 if (oss_num_create_threads) {
1828 if (oss_num_create_threads > OSS_MAX_CREATE_THREADS)
1829 oss_num_create_threads = OSS_MAX_CREATE_THREADS;
1830 if (oss_num_create_threads < OSS_DEF_CREATE_THREADS)
1831 oss_num_create_threads = OSS_DEF_CREATE_THREADS;
1832 oss_min_create_threads = oss_max_create_threads =
1833 oss_num_create_threads;
1835 oss_min_create_threads = OSS_DEF_CREATE_THREADS;
1836 oss_max_create_threads = OSS_MAX_CREATE_THREADS;
1839 ost->ost_create_service =
1840 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1841 OST_MAXREPSIZE, OST_CREATE_PORTAL,
1843 OST_WATCHDOG_TIMEOUT, ost_handle, "ost_create",
1844 obd->obd_proc_entry, ost_print_req,
1845 oss_min_create_threads,
1846 oss_max_create_threads,
1847 "ll_ost_creat", LCT_DT_THREAD);
1848 if (ost->ost_create_service == NULL) {
1849 CERROR("failed to start OST create service\n");
1850 GOTO(out_service, rc = -ENOMEM);
1853 rc = ptlrpc_start_threads(obd, ost->ost_create_service);
1855 GOTO(out_create, rc = -EINVAL);
1857 ost->ost_io_service =
1858 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1859 OST_MAXREPSIZE, OST_IO_PORTAL,
1861 OST_WATCHDOG_TIMEOUT, ost_handle, "ost_io",
1862 obd->obd_proc_entry, ost_print_req,
1863 oss_min_threads, oss_max_threads,
1864 "ll_ost_io", LCT_DT_THREAD);
1865 if (ost->ost_io_service == NULL) {
1866 CERROR("failed to start OST I/O service\n");
1867 GOTO(out_create, rc = -ENOMEM);
1870 ost->ost_io_service->srv_init = ost_thread_init;
1871 ost->ost_io_service->srv_done = ost_thread_done;
1872 ost->ost_io_service->srv_cpu_affinity = 1;
1873 rc = ptlrpc_start_threads(obd, ost->ost_io_service);
1875 GOTO(out_io, rc = -EINVAL);
1877 ping_evictor_start();
1882 ptlrpc_unregister_service(ost->ost_io_service);
1883 ost->ost_io_service = NULL;
1885 ptlrpc_unregister_service(ost->ost_create_service);
1886 ost->ost_create_service = NULL;
1888 ptlrpc_unregister_service(ost->ost_service);
1889 ost->ost_service = NULL;
1891 lprocfs_obd_cleanup(obd);
1895 static int ost_cleanup(struct obd_device *obd)
1897 struct ost_obd *ost = &obd->u.ost;
1901 ping_evictor_stop();
1903 spin_lock_bh(&obd->obd_processing_task_lock);
1904 if (obd->obd_recovering) {
1905 target_cancel_recovery_timer(obd);
1906 obd->obd_recovering = 0;
1908 spin_unlock_bh(&obd->obd_processing_task_lock);
1910 down(&ost->ost_health_sem);
1911 ptlrpc_unregister_service(ost->ost_service);
1912 ptlrpc_unregister_service(ost->ost_create_service);
1913 ptlrpc_unregister_service(ost->ost_io_service);
1914 ost->ost_service = NULL;
1915 ost->ost_create_service = NULL;
1916 up(&ost->ost_health_sem);
1918 lprocfs_obd_cleanup(obd);
1923 static int ost_health_check(struct obd_device *obd)
1925 struct ost_obd *ost = &obd->u.ost;
1928 down(&ost->ost_health_sem);
1929 rc |= ptlrpc_service_health_check(ost->ost_service);
1930 rc |= ptlrpc_service_health_check(ost->ost_create_service);
1931 rc |= ptlrpc_service_health_check(ost->ost_io_service);
1932 up(&ost->ost_health_sem);
1935 * health_check to return 0 on healthy
1936 * and 1 on unhealthy.
1944 struct ost_thread_local_cache *ost_tls(struct ptlrpc_request *r)
1946 return (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
1949 /* use obd ops to offer management infrastructure */
1950 static struct obd_ops ost_obd_ops = {
1951 .o_owner = THIS_MODULE,
1952 .o_setup = ost_setup,
1953 .o_cleanup = ost_cleanup,
1954 .o_health_check = ost_health_check,
1958 static int __init ost_init(void)
1960 struct lprocfs_static_vars lvars;
1964 lprocfs_init_vars(ost, &lvars);
1965 rc = class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
1966 LUSTRE_OSS_NAME, NULL);
1968 if (ost_num_threads != 0 && oss_num_threads == 0) {
1969 LCONSOLE_INFO("ost_num_threads module parameter is deprecated, "
1970 "use oss_num_threads instead or unset both for "
1971 "dynamic thread startup\n");
1972 oss_num_threads = ost_num_threads;
1978 static void /*__exit*/ ost_exit(void)
1980 class_unregister_type(LUSTRE_OSS_NAME);
1983 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1984 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
1985 MODULE_LICENSE("GPL");
1987 module_init(ost_init);
1988 module_exit(ost_exit);