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 if (!(body->oa.o_valid & OBD_MD_FLSIZE) ||
170 !(body->oa.o_valid & OBD_MD_FLBLOCKS))
173 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
177 repbody = lustre_msg_buf(req->rq_repmsg, 0);
178 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
179 req->rq_status = obd_punch(conn, &repbody->oa, NULL,
180 repbody->oa.o_blocks, repbody->oa.o_size);
184 static int ost_setattr(struct ptlrpc_request *req)
186 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
187 struct ost_body *body, *repbody;
188 int rc, size = sizeof(*body);
191 body = lustre_msg_buf(req->rq_reqmsg, 0);
193 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
197 repbody = lustre_msg_buf(req->rq_repmsg, 0);
198 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
199 req->rq_status = obd_setattr(conn, &repbody->oa, NULL);
203 static int ost_bulk_timeout(void *data)
205 struct ptlrpc_bulk_desc *desc = data;
208 CERROR("(not yet) starting recovery of client %p\n", desc->bd_client);
212 static int ost_brw_read(struct ptlrpc_request *req)
214 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
215 struct ptlrpc_bulk_desc *desc;
216 void *tmp1, *tmp2, *end2;
217 struct niobuf_remote *remote_nb;
218 struct niobuf_local *local_nb = NULL;
219 struct obd_ioobj *ioo;
220 struct ost_body *body;
221 struct l_wait_info lwi;
222 void *desc_priv = NULL;
223 int rc, cmd, i, j, objcount, niocount, size = sizeof(*body);
226 body = lustre_msg_buf(req->rq_reqmsg, 0);
227 tmp1 = lustre_msg_buf(req->rq_reqmsg, 1);
228 tmp2 = lustre_msg_buf(req->rq_reqmsg, 2);
229 end2 = (char *)tmp2 + req->rq_reqmsg->buflens[2];
230 objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
231 niocount = req->rq_reqmsg->buflens[2] / sizeof(*remote_nb);
234 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
237 for (i = 0; i < objcount; i++) {
238 ost_unpack_ioo(&tmp1, &ioo);
239 if (tmp2 + ioo->ioo_bufcnt > end2) {
241 GOTO(out, rc = -EFAULT);
243 for (j = 0; j < ioo->ioo_bufcnt; j++)
244 ost_unpack_niobuf(&tmp2, &remote_nb);
247 OBD_ALLOC(local_nb, sizeof(*local_nb) * niocount);
248 if (local_nb == NULL)
249 GOTO(out, rc = -ENOMEM);
251 /* The unpackers move tmp1 and tmp2, so reset them before using */
252 ioo = lustre_msg_buf(req->rq_reqmsg, 1);
253 remote_nb = lustre_msg_buf(req->rq_reqmsg, 2);
254 req->rq_status = obd_preprw(cmd, conn, objcount, ioo, niocount,
255 remote_nb, local_nb, &desc_priv);
260 desc = ptlrpc_prep_bulk(req->rq_connection);
262 GOTO(out_local, rc = -ENOMEM);
263 desc->bd_portal = OST_BULK_PORTAL;
265 for (i = 0; i < niocount; i++) {
266 struct ptlrpc_bulk_page *bulk = ptlrpc_prep_bulk_page(desc);
269 GOTO(out_bulk, rc = -ENOMEM);
270 bulk->bp_xid = remote_nb[i].xid;
271 bulk->bp_buf = local_nb[i].addr;
272 bulk->bp_buflen = remote_nb[i].len;
275 rc = ptlrpc_send_bulk(desc);
279 lwi = LWI_TIMEOUT(obd_timeout * HZ, ost_bulk_timeout, desc);
280 rc = l_wait_event(desc->bd_waitq, desc->bd_flags &PTL_BULK_FL_SENT, &lwi);
282 LASSERT(rc == -ETIMEDOUT);
286 req->rq_status = obd_commitrw(cmd, conn, objcount, ioo, niocount,
287 local_nb, desc_priv);
289 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
292 ptlrpc_free_bulk(desc);
294 OBD_FREE(local_nb, sizeof(*local_nb) * niocount);
297 /* It's a lot of work to delay allocating the reply, and a lot
298 * less work to just free it here. */
299 OBD_FREE(req->rq_repmsg, req->rq_replen);
300 req->rq_repmsg = NULL;
301 ptlrpc_error(req->rq_svc, req);
303 ptlrpc_reply(req->rq_svc, req);
307 static int ost_brw_write(struct ptlrpc_request *req)
309 struct lustre_handle *conn = (struct lustre_handle *)req->rq_reqmsg;
310 struct ptlrpc_bulk_desc *desc;
311 struct niobuf_remote *remote_nb;
312 struct niobuf_local *local_nb, *lnb;
313 struct obd_ioobj *ioo;
314 struct ost_body *body;
315 int cmd, rc, i, j, objcount, niocount, size[2] = {sizeof(*body)};
316 void *tmp1, *tmp2, *end2;
317 void *desc_priv = NULL;
319 struct ptlrpc_service *srv;
320 struct l_wait_info lwi;
324 body = lustre_msg_buf(req->rq_reqmsg, 0);
325 tmp1 = lustre_msg_buf(req->rq_reqmsg, 1);
326 tmp2 = lustre_msg_buf(req->rq_reqmsg, 2);
327 end2 = (char *)tmp2 + req->rq_reqmsg->buflens[2];
328 objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
329 niocount = req->rq_reqmsg->buflens[2] / sizeof(*remote_nb);
332 for (i = 0; i < objcount; i++) {
333 ost_unpack_ioo((void *)&tmp1, &ioo);
334 if (tmp2 + ioo->ioo_bufcnt > end2) {
338 for (j = 0; j < ioo->ioo_bufcnt; j++)
339 ost_unpack_niobuf((void *)&tmp2, &remote_nb);
342 size[1] = niocount * sizeof(*remote_nb);
343 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
346 remote_nb = lustre_msg_buf(req->rq_repmsg, 1);
348 OBD_ALLOC(local_nb, niocount * sizeof(*local_nb));
349 if (local_nb == NULL)
350 GOTO(out, rc = -ENOMEM);
352 /* The unpackers move tmp1 and tmp2, so reset them before using */
353 tmp1 = lustre_msg_buf(req->rq_reqmsg, 1);
354 tmp2 = lustre_msg_buf(req->rq_reqmsg, 2);
355 req->rq_status = obd_preprw(cmd, conn, objcount, tmp1, niocount, tmp2,
356 local_nb, &desc_priv);
358 GOTO(out_free, rc = 0); /* XXX is this correct? */
360 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
361 GOTO(fail_preprw, rc = 0);
363 desc = ptlrpc_prep_bulk(req->rq_connection);
365 GOTO(fail_preprw, rc = -ENOMEM);
367 desc->bd_portal = OSC_BULK_PORTAL;
368 desc->bd_desc_private = desc_priv;
369 memcpy(&(desc->bd_conn), &conn, sizeof(conn));
371 srv = req->rq_obd->u.ost.ost_service;
372 spin_lock(&srv->srv_lock);
373 xid = srv->srv_xid++; /* single xid for all pages */
374 spin_unlock(&srv->srv_lock);
376 for (i = 0, lnb = local_nb; i < niocount; i++, lnb++) {
377 struct ptlrpc_bulk_page *bulk;
379 bulk = ptlrpc_prep_bulk_page(desc);
381 GOTO(fail_bulk, rc = -ENOMEM);
383 bulk->bp_xid = xid; /* single xid for all pages */
385 bulk->bp_buf = lnb->addr;
386 bulk->bp_page = lnb->page;
387 bulk->bp_flags = lnb->flags;
388 bulk->bp_dentry = lnb->dentry;
389 bulk->bp_buflen = lnb->len;
392 /* this advances remote_nb */
393 ost_pack_niobuf((void **)&remote_nb, lnb->offset, lnb->len, 0,
397 rc = ptlrpc_register_bulk(desc);
402 ptlrpc_reply(req->rq_svc, req);
404 lwi = LWI_TIMEOUT(obd_timeout * HZ, ost_bulk_timeout, desc);
405 rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_RCVD,
408 if (rc != -ETIMEDOUT)
413 rc = obd_commitrw(cmd, conn, objcount, tmp1, niocount, local_nb,
414 desc->bd_desc_private);
415 ptlrpc_free_bulk(desc);
418 OBD_FREE(local_nb, niocount * sizeof(*local_nb));
422 OBD_FREE(req->rq_repmsg, req->rq_replen);
423 req->rq_repmsg = NULL;
424 ptlrpc_error(req->rq_svc, req);
426 ptlrpc_reply(req->rq_svc, req);
431 ptlrpc_free_bulk(desc);
433 /* FIXME: how do we undo the preprw? */
437 static int ost_handle(struct ptlrpc_request *req)
442 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
443 if (rc || OBD_FAIL_CHECK(OBD_FAIL_OST_HANDLE_UNPACK)) {
444 CERROR("lustre_ost: Invalid request\n");
448 if (req->rq_reqmsg->opc != OST_CONNECT &&
449 req->rq_export == NULL) {
450 CERROR("lustre_ost: operation %d on unconnected OST\n",
451 req->rq_reqmsg->opc);
452 GOTO(out, rc = -ENOTCONN);
455 if (strcmp(req->rq_obd->obd_type->typ_name, "ost") != 0)
456 GOTO(out, rc = -EINVAL);
458 switch (req->rq_reqmsg->opc) {
460 CDEBUG(D_INODE, "connect\n");
461 OBD_FAIL_RETURN(OBD_FAIL_OST_CONNECT_NET, 0);
462 rc = target_handle_connect(req);
465 CDEBUG(D_INODE, "disconnect\n");
466 OBD_FAIL_RETURN(OBD_FAIL_OST_DISCONNECT_NET, 0);
467 rc = target_handle_disconnect(req);
470 CDEBUG(D_INODE, "create\n");
471 OBD_FAIL_RETURN(OBD_FAIL_OST_CREATE_NET, 0);
472 rc = ost_create(req);
475 CDEBUG(D_INODE, "destroy\n");
476 OBD_FAIL_RETURN(OBD_FAIL_OST_DESTROY_NET, 0);
477 rc = ost_destroy(req);
480 CDEBUG(D_INODE, "getattr\n");
481 OBD_FAIL_RETURN(OBD_FAIL_OST_GETATTR_NET, 0);
482 rc = ost_getattr(req);
485 CDEBUG(D_INODE, "setattr\n");
486 OBD_FAIL_RETURN(OBD_FAIL_OST_SETATTR_NET, 0);
487 rc = ost_setattr(req);
490 CDEBUG(D_INODE, "open\n");
491 OBD_FAIL_RETURN(OBD_FAIL_OST_OPEN_NET, 0);
495 CDEBUG(D_INODE, "close\n");
496 OBD_FAIL_RETURN(OBD_FAIL_OST_CLOSE_NET, 0);
500 CDEBUG(D_INODE, "write\n");
501 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
502 rc = ost_brw_write(req);
503 /* ost_brw sends its own replies */
506 CDEBUG(D_INODE, "read\n");
507 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
508 rc = ost_brw_read(req);
509 /* ost_brw sends its own replies */
512 CDEBUG(D_INODE, "punch\n");
513 OBD_FAIL_RETURN(OBD_FAIL_OST_PUNCH_NET, 0);
517 CDEBUG(D_INODE, "statfs\n");
518 OBD_FAIL_RETURN(OBD_FAIL_OST_STATFS_NET, 0);
519 rc = ost_statfs(req);
522 CDEBUG(D_INODE, "enqueue\n");
523 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
524 rc = ldlm_handle_enqueue(req);
529 CDEBUG(D_INODE, "convert\n");
530 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
531 rc = ldlm_handle_convert(req);
536 CDEBUG(D_INODE, "cancel\n");
537 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
538 rc = ldlm_handle_cancel(req);
542 case LDLM_BL_CALLBACK:
543 case LDLM_CP_CALLBACK:
544 CDEBUG(D_INODE, "callback\n");
545 CERROR("callbacks should not happen on OST\n");
547 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
550 req->rq_status = -ENOTSUPP;
551 rc = ptlrpc_error(req->rq_svc, req);
557 //req->rq_status = rc;
559 CERROR("ost: processing error (opcode=%d): %d\n",
560 req->rq_reqmsg->opc, rc);
561 ptlrpc_error(req->rq_svc, req);
563 CDEBUG(D_INODE, "sending reply\n");
564 if (req->rq_repmsg == NULL)
565 CERROR("handler for opcode %d returned rc=0 without "
566 "creating rq_repmsg; needs to return rc != "
567 "0!\n", req->rq_reqmsg->opc);
568 ptlrpc_reply(req->rq_svc, req);
574 #define OST_NUM_THREADS 6
576 /* mount the file system (secretly) */
577 static int ost_setup(struct obd_device *obddev, obd_count len, void *buf)
579 struct obd_ioctl_data* data = buf;
580 struct ost_obd *ost = &obddev->u.ost;
581 struct obd_device *tgt;
586 if (data->ioc_inllen1 < 1) {
587 CERROR("requires a TARGET OBD UUID\n");
590 if (data->ioc_inllen1 > 37) {
591 CERROR("OBD UUID must be less than 38 characters\n");
596 tgt = class_uuid2obd(data->ioc_inlbuf1);
597 if (!tgt || !(tgt->obd_flags & OBD_ATTACHED) ||
598 !(tgt->obd_flags & OBD_SET_UP)) {
599 CERROR("device not attached or not set up (%d)\n",
601 GOTO(error_dec, err = -EINVAL);
604 err = obd_connect(&ost->ost_conn, tgt, NULL, NULL, NULL);
606 CERROR("fail to connect to device %d\n", data->ioc_dev);
607 GOTO(error_dec, err = -EINVAL);
610 ost->ost_service = ptlrpc_init_svc(OST_NEVENTS, OST_NBUFS,
611 OST_BUFSIZE, OST_MAXREQSIZE,
612 OST_REQUEST_PORTAL, OSC_REPLY_PORTAL,
613 "self", ost_handle, "ost");
614 if (!ost->ost_service) {
615 CERROR("failed to start service\n");
616 GOTO(error_disc, err = -EINVAL);
619 for (i = 0; i < OST_NUM_THREADS; i++) {
621 sprintf(name, "lustre_ost_%02d", i);
622 err = ptlrpc_start_thread(obddev, ost->ost_service, name);
624 CERROR("error starting thread #%d: rc %d\n", i, err);
625 GOTO(error_disc, err = -EINVAL);
632 obd_disconnect(&ost->ost_conn);
638 static int ost_cleanup(struct obd_device * obddev)
640 struct ost_obd *ost = &obddev->u.ost;
645 if ( !list_empty(&obddev->obd_exports) ) {
646 CERROR("still has clients!\n");
650 ptlrpc_stop_all_threads(ost->ost_service);
651 ptlrpc_unregister_service(ost->ost_service);
653 err = obd_disconnect(&ost->ost_conn);
655 CERROR("lustre ost: fail to disconnect device\n");
663 /* use obd ops to offer management infrastructure */
664 static struct obd_ops ost_obd_ops = {
666 o_cleanup: ost_cleanup,
669 static int __init ost_init(void)
671 class_register_type(&ost_obd_ops, LUSTRE_OST_NAME);
675 static void __exit ost_exit(void)
677 class_unregister_type(LUSTRE_OST_NAME);
680 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
681 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
682 MODULE_LICENSE("GPL");
684 module_init(ost_init);
685 module_exit(ost_exit);