1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2001, 2002 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 Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 * Storage Target Handling functions
24 * Lustre Object Server Module (OST)
26 * This server is single threaded at present (but can easily be multi
27 * threaded). For testing and management it is treated as an
28 * obd_device, although it does not export a full OBD method table
29 * (the requests are coming in over the wire, so object target
30 * modules do not have a full method table.)
34 #define DEBUG_SUBSYSTEM S_OST
36 #include <linux/module.h>
37 #include <linux/obd_ost.h>
38 #include <linux/lustre_net.h>
39 #include <linux/lustre_dlm.h>
40 #include <linux/init.h>
42 static int ost_destroy(struct ptlrpc_request *req)
44 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
45 struct ost_body *body;
46 int rc, size = sizeof(*body);
49 body = lustre_msg_buf(req->rq_reqmsg, 0);
51 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
55 req->rq_status = obd_destroy(conn, &body->oa, NULL);
59 static int ost_getattr(struct ptlrpc_request *req)
61 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
62 struct ost_body *body, *repbody;
63 int rc, size = sizeof(*body);
66 body = lustre_msg_buf(req->rq_reqmsg, 0);
68 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
72 repbody = lustre_msg_buf(req->rq_repmsg, 0);
73 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
74 req->rq_status = obd_getattr(conn, &repbody->oa, NULL);
78 static int ost_statfs(struct ptlrpc_request *req)
80 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
81 struct obd_statfs *osfs;
82 int rc, size = sizeof(*osfs);
85 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
89 osfs = lustre_msg_buf(req->rq_repmsg, 0);
90 memset(osfs, 0, size);
92 rc = obd_statfs(conn, osfs);
94 CERROR("ost: statfs failed: rc %d\n", rc);
98 obd_statfs_pack(osfs, osfs);
103 static int ost_open(struct ptlrpc_request *req)
105 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
106 struct ost_body *body, *repbody;
107 int rc, size = sizeof(*body);
110 body = lustre_msg_buf(req->rq_reqmsg, 0);
112 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
116 repbody = lustre_msg_buf(req->rq_repmsg, 0);
117 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
118 req->rq_status = obd_open(conn, &repbody->oa, NULL);
122 static int ost_close(struct ptlrpc_request *req)
124 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
125 struct ost_body *body, *repbody;
126 int rc, size = sizeof(*body);
129 body = lustre_msg_buf(req->rq_reqmsg, 0);
131 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
135 repbody = lustre_msg_buf(req->rq_repmsg, 0);
136 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
137 req->rq_status = obd_close(conn, &repbody->oa, NULL);
141 static int ost_create(struct ptlrpc_request *req)
143 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
144 struct ost_body *body, *repbody;
145 int rc, size = sizeof(*body);
148 body = lustre_msg_buf(req->rq_reqmsg, 0);
150 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
154 repbody = lustre_msg_buf(req->rq_repmsg, 0);
155 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
156 req->rq_status = obd_create(conn, &repbody->oa, NULL);
160 static int ost_punch(struct ptlrpc_request *req)
162 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
163 struct ost_body *body, *repbody;
164 int rc, size = sizeof(*body);
167 body = lustre_msg_buf(req->rq_reqmsg, 0);
169 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
173 repbody = lustre_msg_buf(req->rq_repmsg, 0);
174 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
175 req->rq_status = obd_punch(conn, &repbody->oa, NULL,
176 repbody->oa.o_blocks, repbody->oa.o_size);
180 static int ost_setattr(struct ptlrpc_request *req)
182 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
183 struct ost_body *body, *repbody;
184 int rc, size = sizeof(*body);
187 body = lustre_msg_buf(req->rq_reqmsg, 0);
189 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
193 repbody = lustre_msg_buf(req->rq_repmsg, 0);
194 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
195 req->rq_status = obd_setattr(conn, &repbody->oa, NULL);
199 static int ost_bulk_timeout(void *data)
201 struct ptlrpc_bulk_desc *desc = data;
204 CERROR("(not yet) starting recovery of client %p\n", desc->bd_client);
208 static int ost_brw_read(struct ptlrpc_request *req)
210 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
211 struct ptlrpc_bulk_desc *desc;
212 void *tmp1, *tmp2, *end2;
213 struct niobuf_remote *remote_nb;
214 struct niobuf_local *local_nb = NULL;
215 struct obd_ioobj *ioo;
216 struct ost_body *body;
217 struct l_wait_info lwi;
218 void *desc_priv = NULL;
219 int rc, cmd, i, j, objcount, niocount, size = sizeof(*body);
222 body = lustre_msg_buf(req->rq_reqmsg, 0);
223 tmp1 = lustre_msg_buf(req->rq_reqmsg, 1);
224 tmp2 = lustre_msg_buf(req->rq_reqmsg, 2);
225 end2 = (char *)tmp2 + req->rq_reqmsg->buflens[2];
226 objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
227 niocount = req->rq_reqmsg->buflens[2] / sizeof(*remote_nb);
230 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
233 for (i = 0; i < objcount; i++) {
234 ost_unpack_ioo(&tmp1, &ioo);
235 if (tmp2 + ioo->ioo_bufcnt > end2) {
237 GOTO(out, rc = -EFAULT);
239 for (j = 0; j < ioo->ioo_bufcnt; j++)
240 ost_unpack_niobuf(&tmp2, &remote_nb);
243 OBD_ALLOC(local_nb, sizeof(*local_nb) * niocount);
244 if (local_nb == NULL)
245 GOTO(out, rc = -ENOMEM);
247 /* The unpackers move tmp1 and tmp2, so reset them before using */
248 ioo = lustre_msg_buf(req->rq_reqmsg, 1);
249 remote_nb = lustre_msg_buf(req->rq_reqmsg, 2);
250 req->rq_status = obd_preprw(cmd, conn, objcount, ioo, niocount,
251 remote_nb, local_nb, &desc_priv);
256 desc = ptlrpc_prep_bulk(req->rq_connection);
258 GOTO(out_local, rc = -ENOMEM);
259 desc->bd_portal = OST_BULK_PORTAL;
261 for (i = 0; i < niocount; i++) {
262 struct ptlrpc_bulk_page *bulk = ptlrpc_prep_bulk_page(desc);
265 GOTO(out_bulk, rc = -ENOMEM);
266 bulk->bp_xid = remote_nb[i].xid;
267 bulk->bp_buf = local_nb[i].addr;
268 bulk->bp_buflen = remote_nb[i].len;
271 rc = ptlrpc_send_bulk(desc);
275 lwi = LWI_TIMEOUT(obd_timeout * HZ, ost_bulk_timeout, desc);
276 rc = l_wait_event(desc->bd_waitq, desc->bd_flags &PTL_BULK_FL_SENT, &lwi);
278 LASSERT(rc == -ETIMEDOUT);
282 req->rq_status = obd_commitrw(cmd, conn, objcount, ioo, niocount,
283 local_nb, desc_priv);
285 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
288 ptlrpc_free_bulk(desc);
290 OBD_FREE(local_nb, sizeof(*local_nb) * niocount);
293 /* It's a lot of work to delay allocating the reply, and a lot
294 * less work to just free it here. */
295 OBD_FREE(req->rq_repmsg, req->rq_replen);
296 req->rq_repmsg = NULL;
297 ptlrpc_error(req->rq_svc, req);
299 ptlrpc_reply(req->rq_svc, req);
303 static int ost_brw_write(struct ptlrpc_request *req)
305 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
306 struct ptlrpc_bulk_desc *desc;
307 struct niobuf_remote *remote_nb;
308 struct niobuf_local *local_nb, *lnb;
309 struct obd_ioobj *ioo;
310 struct ost_body *body;
311 int cmd, rc, i, j, objcount, niocount, size[2] = {sizeof(*body)};
312 void *tmp1, *tmp2, *end2;
313 void *desc_priv = NULL;
315 struct ptlrpc_service *srv;
316 struct l_wait_info lwi;
320 body = lustre_msg_buf(req->rq_reqmsg, 0);
321 tmp1 = lustre_msg_buf(req->rq_reqmsg, 1);
322 tmp2 = lustre_msg_buf(req->rq_reqmsg, 2);
323 end2 = (char *)tmp2 + req->rq_reqmsg->buflens[2];
324 objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
325 niocount = req->rq_reqmsg->buflens[2] / sizeof(*remote_nb);
328 for (i = 0; i < objcount; i++) {
329 ost_unpack_ioo((void *)&tmp1, &ioo);
330 if (tmp2 + ioo->ioo_bufcnt > end2) {
334 for (j = 0; j < ioo->ioo_bufcnt; j++)
335 ost_unpack_niobuf((void *)&tmp2, &remote_nb);
338 size[1] = niocount * sizeof(*remote_nb);
339 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
342 remote_nb = lustre_msg_buf(req->rq_repmsg, 1);
344 OBD_ALLOC(local_nb, niocount * sizeof(*local_nb));
345 if (local_nb == NULL)
346 GOTO(out, rc = -ENOMEM);
348 /* The unpackers move tmp1 and tmp2, so reset them before using */
349 tmp1 = lustre_msg_buf(req->rq_reqmsg, 1);
350 tmp2 = lustre_msg_buf(req->rq_reqmsg, 2);
351 req->rq_status = obd_preprw(cmd, conn, objcount, tmp1, niocount, tmp2,
352 local_nb, &desc_priv);
354 GOTO(out_free, rc = 0); /* XXX is this correct? */
356 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
357 GOTO(fail_preprw, rc = 0);
359 desc = ptlrpc_prep_bulk(req->rq_connection);
361 GOTO(fail_preprw, rc = -ENOMEM);
363 desc->bd_portal = OSC_BULK_PORTAL;
364 desc->bd_desc_private = desc_priv;
365 memcpy(&(desc->bd_conn), &conn, sizeof(conn));
367 srv = req->rq_obd->u.ost.ost_service;
368 spin_lock(&srv->srv_lock);
369 xid = srv->srv_xid++; /* single xid for all pages */
370 spin_unlock(&srv->srv_lock);
372 for (i = 0, lnb = local_nb; i < niocount; i++, lnb++) {
373 struct ptlrpc_bulk_page *bulk;
375 bulk = ptlrpc_prep_bulk_page(desc);
377 GOTO(fail_bulk, rc = -ENOMEM);
379 bulk->bp_xid = xid; /* single xid for all pages */
381 bulk->bp_buf = lnb->addr;
382 bulk->bp_page = lnb->page;
383 bulk->bp_flags = lnb->flags;
384 bulk->bp_dentry = lnb->dentry;
385 bulk->bp_buflen = lnb->len;
388 /* this advances remote_nb */
389 ost_pack_niobuf((void **)&remote_nb, lnb->offset, lnb->len, 0,
393 rc = ptlrpc_register_bulk(desc);
398 ptlrpc_reply(req->rq_svc, req);
400 lwi = LWI_TIMEOUT(obd_timeout * HZ, ost_bulk_timeout, desc);
401 rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_RCVD,
404 if (rc != -ETIMEDOUT)
409 rc = obd_commitrw(cmd, conn, objcount, tmp1, niocount, local_nb,
410 desc->bd_desc_private);
411 ptlrpc_free_bulk(desc);
414 OBD_FREE(local_nb, niocount * sizeof(*local_nb));
418 OBD_FREE(req->rq_repmsg, req->rq_replen);
419 req->rq_repmsg = NULL;
420 ptlrpc_error(req->rq_svc, req);
422 ptlrpc_reply(req->rq_svc, req);
427 ptlrpc_free_bulk(desc);
429 /* FIXME: how do we undo the preprw? */
433 static int ost_handle(struct ptlrpc_request *req)
438 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
439 if (rc || OBD_FAIL_CHECK(OBD_FAIL_OST_HANDLE_UNPACK)) {
440 CERROR("lustre_ost: Invalid request\n");
444 if (req->rq_reqmsg->opc != OST_CONNECT &&
445 req->rq_export == NULL) {
446 CERROR("lustre_ost: operation %d on unconnected OST\n",
447 req->rq_reqmsg->opc);
448 GOTO(out, rc = -ENOTCONN);
451 if (strcmp(req->rq_obd->obd_type->typ_name, "ost") != 0)
452 GOTO(out, rc = -EINVAL);
454 switch (req->rq_reqmsg->opc) {
456 CDEBUG(D_INODE, "connect\n");
457 OBD_FAIL_RETURN(OBD_FAIL_OST_CONNECT_NET, 0);
458 rc = target_handle_connect(req);
461 CDEBUG(D_INODE, "disconnect\n");
462 OBD_FAIL_RETURN(OBD_FAIL_OST_DISCONNECT_NET, 0);
463 rc = target_handle_disconnect(req);
466 CDEBUG(D_INODE, "create\n");
467 OBD_FAIL_RETURN(OBD_FAIL_OST_CREATE_NET, 0);
468 rc = ost_create(req);
471 CDEBUG(D_INODE, "destroy\n");
472 OBD_FAIL_RETURN(OBD_FAIL_OST_DESTROY_NET, 0);
473 rc = ost_destroy(req);
476 CDEBUG(D_INODE, "getattr\n");
477 OBD_FAIL_RETURN(OBD_FAIL_OST_GETATTR_NET, 0);
478 rc = ost_getattr(req);
481 CDEBUG(D_INODE, "setattr\n");
482 OBD_FAIL_RETURN(OBD_FAIL_OST_SETATTR_NET, 0);
483 rc = ost_setattr(req);
486 CDEBUG(D_INODE, "open\n");
487 OBD_FAIL_RETURN(OBD_FAIL_OST_OPEN_NET, 0);
491 CDEBUG(D_INODE, "close\n");
492 OBD_FAIL_RETURN(OBD_FAIL_OST_CLOSE_NET, 0);
496 CDEBUG(D_INODE, "write\n");
497 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
498 rc = ost_brw_write(req);
499 /* ost_brw sends its own replies */
502 CDEBUG(D_INODE, "read\n");
503 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
504 rc = ost_brw_read(req);
505 /* ost_brw sends its own replies */
508 CDEBUG(D_INODE, "punch\n");
509 OBD_FAIL_RETURN(OBD_FAIL_OST_PUNCH_NET, 0);
513 CDEBUG(D_INODE, "statfs\n");
514 OBD_FAIL_RETURN(OBD_FAIL_OST_STATFS_NET, 0);
515 rc = ost_statfs(req);
518 CDEBUG(D_INODE, "enqueue\n");
519 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
520 rc = ldlm_handle_enqueue(req);
525 CDEBUG(D_INODE, "convert\n");
526 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
527 rc = ldlm_handle_convert(req);
532 CDEBUG(D_INODE, "cancel\n");
533 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
534 rc = ldlm_handle_cancel(req);
538 case LDLM_BL_CALLBACK:
539 case LDLM_CP_CALLBACK:
540 CDEBUG(D_INODE, "callback\n");
541 CERROR("callbacks should not happen on OST\n");
543 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
546 req->rq_status = -ENOTSUPP;
547 rc = ptlrpc_error(req->rq_svc, req);
553 //req->rq_status = rc;
555 CERROR("ost: processing error (opcode=%d): %d\n",
556 req->rq_reqmsg->opc, rc);
557 ptlrpc_error(req->rq_svc, req);
559 CDEBUG(D_INODE, "sending reply\n");
560 if (req->rq_repmsg == NULL)
561 CERROR("handler for opcode %d returned rc=0 without "
562 "creating rq_repmsg; needs to return rc != "
563 "0!\n", req->rq_reqmsg->opc);
564 ptlrpc_reply(req->rq_svc, req);
570 #define OST_NUM_THREADS 6
572 /* mount the file system (secretly) */
573 static int ost_setup(struct obd_device *obddev, obd_count len, void *buf)
575 struct obd_ioctl_data* data = buf;
576 struct ost_obd *ost = &obddev->u.ost;
577 struct obd_device *tgt;
582 if (data->ioc_inllen1 < 1) {
583 CERROR("requires a TARGET OBD UUID\n");
586 if (data->ioc_inllen1 > 37) {
587 CERROR("OBD UUID must be less than 38 characters\n");
592 tgt = class_uuid2obd(data->ioc_inlbuf1);
593 if (!tgt || !(tgt->obd_flags & OBD_ATTACHED) ||
594 !(tgt->obd_flags & OBD_SET_UP)) {
595 CERROR("device not attached or not set up (%d)\n",
597 GOTO(error_dec, err = -EINVAL);
600 err = obd_connect(&ost->ost_conn, tgt, NULL, NULL, NULL);
602 CERROR("fail to connect to device %d\n", data->ioc_dev);
603 GOTO(error_dec, err = -EINVAL);
606 ost->ost_service = ptlrpc_init_svc(OST_NEVENTS, OST_NBUFS,
607 OST_BUFSIZE, OST_MAXREQSIZE,
608 OST_REQUEST_PORTAL, OSC_REPLY_PORTAL,
609 "self", ost_handle, "ost");
610 if (!ost->ost_service) {
611 CERROR("failed to start service\n");
612 GOTO(error_disc, err = -EINVAL);
615 for (i = 0; i < OST_NUM_THREADS; i++) {
617 sprintf(name, "lustre_ost_%02d", i);
618 err = ptlrpc_start_thread(obddev, ost->ost_service, name);
620 CERROR("error starting thread #%d: rc %d\n", i, err);
621 GOTO(error_disc, err = -EINVAL);
628 obd_disconnect(&ost->ost_conn);
634 static int ost_cleanup(struct obd_device * obddev)
636 struct ost_obd *ost = &obddev->u.ost;
641 if ( !list_empty(&obddev->obd_exports) ) {
642 CERROR("still has clients!\n");
646 ptlrpc_stop_all_threads(ost->ost_service);
647 ptlrpc_unregister_service(ost->ost_service);
649 err = obd_disconnect(&ost->ost_conn);
651 CERROR("lustre ost: fail to disconnect device\n");
659 /* use obd ops to offer management infrastructure */
660 static struct obd_ops ost_obd_ops = {
662 o_cleanup: ost_cleanup,
665 static int __init ost_init(void)
667 class_register_type(&ost_obd_ops, LUSTRE_OST_NAME);
671 static void __exit ost_exit(void)
673 class_unregister_type(LUSTRE_OST_NAME);
676 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
677 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
678 MODULE_LICENSE("GPL");
680 module_init(ost_init);
681 module_exit(ost_exit);