1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * Copyright (C) 2002 Cluster File Systems, Inc.
7 * Author: Phil Schwan <phil@off.net>
8 * Peter Braam <braam@clusterfs.com>
10 * This code is issued under the GNU General Public License.
11 * See the file COPYING in this distribution
15 #define DEBUG_SUBSYSTEM S_LOV
17 #include <linux/slab.h>
18 #include <linux/module.h>
19 #include <linux/obd_support.h>
20 #include <linux/lustre_lib.h>
21 #include <linux/lustre_net.h>
22 #include <linux/lustre_idl.h>
23 #include <linux/lustre_mds.h>
24 #include <linux/obd_class.h>
25 #include <linux/obd_lov.h>
26 #include <linux/init.h>
28 extern struct obd_device obd_dev[MAX_OBD_DEVICES];
31 static int lov_connect(struct lustre_handle *conn, struct obd_device *obd,
34 struct ptlrpc_request *req;
35 struct lov_obd *lov = &obd->u.lov;
36 struct lustre_handle mdc_conn;
42 rc = class_connect(conn, obd, cluuid);
48 /* retrieve LOV metadata from MDS */
49 rc = obd_connect(&mdc_conn, lov->mdcobd, NULL);
51 CERROR("cannot connect to mdc: rc = %d\n", rc);
52 GOTO(out, rc = -EINVAL);
55 rc = mdc_getlovinfo(obd, &mdc_conn, &uuidarray, &req);
56 rc2 = obd_disconnect(&mdc_conn);
58 CERROR("cannot get lov info or disconnect %d/%d\n", rc, rc2);
59 GOTO(out, (rc) ? rc : rc2 );
63 if (strcmp(obd->obd_uuid, lov->desc.ld_uuid)) {
64 CERROR("lov uuid %s not on mds device (%s)\n",
65 obd->obd_uuid, lov->desc.ld_uuid);
66 GOTO(out, rc = -EINVAL);
68 if (lov->desc.ld_tgt_count > 1000) {
69 CERROR("configuration error: target count > 1000 (%d)\n",
70 lov->desc.ld_tgt_count);
71 GOTO(out, rc = -EINVAL);
73 if (req->rq_repmsg->bufcount < 2 || req->rq_repmsg->buflens[1] <
74 sizeof(uuid_t) * lov->desc.ld_tgt_count) {
75 CERROR("invalid uuid array returned\n");
76 GOTO(out, rc = -EINVAL);
79 lov->bufsize = sizeof(struct lov_tgt_desc) * lov->desc.ld_tgt_count;
80 OBD_ALLOC(lov->tgts, lov->bufsize);
82 CERROR("Out of memory\n");
83 GOTO(out, rc = -ENOMEM);
86 uuidarray = lustre_msg_buf(req->rq_repmsg, 1);
87 for (i = 0 ; i < lov->desc.ld_tgt_count; i++)
88 memcpy(lov->tgts[i].uuid, uuidarray[i], sizeof(uuid_t));
90 for (i = 0 ; i < lov->desc.ld_tgt_count; i++) {
91 struct obd_device *tgt = class_uuid2obd(uuidarray[i]);
93 CERROR("Target %s not attached\n", uuidarray[i]);
94 GOTO(out_mem, rc = -EINVAL);
96 if (!(tgt->obd_flags & OBD_SET_UP)) {
97 CERROR("Target %s not set up\n", uuidarray[i]);
98 GOTO(out_mem, rc = -EINVAL);
100 rc = obd_connect(&lov->tgts[i].conn, tgt, NULL);
102 CERROR("Target %s connect error %d\n",
110 for (i = 0 ; i < lov->desc.ld_tgt_count; i++) {
111 rc2 = obd_disconnect(&lov->tgts[i].conn);
113 CERROR("BAD: Target %s disconnect error %d\n",
116 OBD_FREE(lov->tgts, lov->bufsize);
120 class_disconnect(conn);
121 ptlrpc_free_req(req);
125 static int lov_disconnect(struct lustre_handle *conn)
127 struct obd_device *obd = class_conn2obd(conn);
128 struct lov_obd *lov = &obd->u.lov;
135 for (i = 0 ; i < lov->desc.ld_tgt_count; i++) {
136 rc = obd_disconnect(&lov->tgts[i].conn);
138 CERROR("Target %s disconnect error %d\n",
139 lov->tgts[i].uuid, rc);
143 OBD_FREE(lov->tgts, lov->bufsize);
148 rc = class_disconnect(conn);
154 static int lov_setup(struct obd_device *obd, obd_count len, void *buf)
156 struct obd_ioctl_data* data = buf;
157 struct lov_obd *lov = &obd->u.lov;
161 if (data->ioc_inllen1 < 1) {
162 CERROR("osc setup requires an MDC UUID\n");
166 if (data->ioc_inllen1 > 37) {
167 CERROR("mdc UUID must be less than 38 characters\n");
171 lov->mdcobd = class_uuid2obd(data->ioc_inlbuf1);
173 CERROR("LOV %s cannot locate MDC %s\n", obd->obd_uuid,
181 static inline int lov_stripe_md_size(struct obd_device *obd)
183 struct lov_obd *lov = &obd->u.lov;
186 size = sizeof(struct lov_stripe_md) +
187 lov->desc.ld_tgt_count * sizeof(struct lov_object_id);
191 /* the LOV counts on oa->o_id to be set as the LOV object id */
192 static int lov_create(struct lustre_handle *conn, struct obdo *oa, struct lov_stripe_md **ea)
196 struct obd_export *export = class_conn2export(conn);
198 struct lov_stripe_md *md;
202 CERROR("lov_create needs EA for striping information\n");
207 lov = &export->exp_obd->u.lov;
209 oa->o_easize = lov_stripe_md_size(export->exp_obd);
211 OBD_ALLOC(*ea, oa->o_easize);
217 md->lmd_easize = oa->o_easize;
218 md->lmd_object_id = oa->o_id;
219 if (!md->lmd_stripe_count) {
220 md->lmd_stripe_count = lov->desc.ld_default_stripe_count;
223 if (!md->lmd_stripe_size)
224 md->lmd_stripe_size = lov->desc.ld_default_stripe_size;
228 for (i = 0; i < md->lmd_stripe_count; i++) {
229 struct lov_stripe_md obj_md;
230 struct lov_stripe_md *obj_mdp = &obj_md;
231 /* create data objects with "parent" OA */
232 memcpy(&tmp, oa, sizeof(tmp));
233 tmp.o_easize = sizeof(struct lov_stripe_md);
234 rc = obd_create(&lov->tgts[i].conn, &tmp, &obj_mdp);
236 GOTO(out_cleanup, rc);
237 md->lmd_objects[i].l_object_id = tmp.o_id;
243 for (i2 = 0 ; i2 < i ; i2++) {
244 /* destroy already created objects here */
245 tmp.o_id = md->lmd_objects[i].l_object_id;
246 rc2 = obd_destroy(&lov->tgts[i].conn, &tmp, NULL);
248 CERROR("Failed to remove object from target %d\n",
256 static int lov_destroy(struct lustre_handle *conn, struct obdo *oa,
257 struct lov_stripe_md *md)
261 struct obd_export *export = class_conn2export(conn);
266 CERROR("LOV requires striping ea for desctruction\n");
270 if (!export || !export->exp_obd)
273 lov = &export->exp_obd->u.lov;
274 for (i = 0; i < md->lmd_stripe_count; i++) {
275 /* create data objects with "parent" OA */
276 memcpy(&tmp, oa, sizeof(tmp));
277 tmp.o_id = md->lmd_objects[i].l_object_id;
278 rc = obd_destroy(&lov->tgts[i].conn, &tmp, NULL);
280 CERROR("Error destroying object %Ld on %d\n",
287 static int lov_getattr(struct lustre_handle *conn, struct obdo *oa,
288 struct lov_stripe_md *md)
292 struct obd_export *export = class_conn2export(conn);
297 CERROR("LOV requires striping ea for desctruction\n");
301 if (!export || !export->exp_obd)
304 lov = &export->exp_obd->u.lov;
305 for (i = 0; i < md->lmd_stripe_count; i++) {
306 /* create data objects with "parent" OA */
307 memcpy(&tmp, oa, sizeof(tmp));
308 oa->o_id = md->lmd_objects[i].l_object_id;
310 rc = obd_getattr(&lov->tgts[i].conn, &tmp, NULL);
312 CERROR("Error getattr object %Ld on %d\n",
315 /* XXX can do something more sophisticated here... */
317 obd_id id = oa->o_id;
318 memcpy(oa, &tmp, sizeof(tmp));
325 static int lov_setattr(struct lustre_handle *conn, struct obdo *oa,
326 struct lov_stripe_md *md)
330 struct obd_export *export = class_conn2export(conn);
335 CERROR("LOV requires striping ea for desctruction\n");
339 if (!export || !export->exp_obd)
342 lov = &export->exp_obd->u.lov;
343 for (i = 0; i < md->lmd_stripe_count; i++) {
344 /* create data objects with "parent" OA */
345 memcpy(&tmp, oa, sizeof(tmp));
346 oa->o_id = md->lmd_objects[i].l_object_id;
348 rc = obd_setattr(&lov->tgts[i].conn, &tmp, NULL);
350 CERROR("Error setattr object %Ld on %d\n",
357 static int lov_open(struct lustre_handle *conn, struct obdo *oa,
358 struct lov_stripe_md *md)
360 int rc = 0, rc2 = 0, i;
362 struct obd_export *export = class_conn2export(conn);
367 CERROR("LOV requires striping ea for opening\n");
371 if (!export || !export->exp_obd)
374 lov = &export->exp_obd->u.lov;
375 for (i = 0; i < md->lmd_stripe_count; i++) {
376 /* create data objects with "parent" OA */
377 memcpy(&tmp, oa, sizeof(tmp));
378 oa->o_id = md->lmd_objects[i].l_object_id;
380 rc = obd_open(&lov->tgts[i].conn, &tmp, NULL);
383 CERROR("Error getattr object %Ld on %d\n",
391 static int lov_close(struct lustre_handle *conn, struct obdo *oa,
392 struct lov_stripe_md *md)
396 struct obd_export *export = class_conn2export(conn);
401 CERROR("LOV requires striping ea for desctruction\n");
405 if (!export || !export->exp_obd)
408 lov = &export->exp_obd->u.lov;
409 for (i = 0; i < md->lmd_stripe_count; i++) {
410 /* create data objects with "parent" OA */
411 memcpy(&tmp, oa, sizeof(tmp));
412 oa->o_id = md->lmd_objects[i].l_object_id;
414 rc = obd_close(&lov->tgts[i].conn, &tmp, NULL);
416 CERROR("Error getattr object %Ld on %d\n",
424 #define log2(n) ffz(~(n))
427 /* compute offset in stripe i corresponds to offset "in" */
428 __u64 lov_offset(struct lov_stripe_md *md, __u64 in, int i)
430 __u32 ssz = md->lmd_stripe_size;
431 /* full stripes across all * stripe size */
432 __u32 out = ( ((__u32)in) / (md->lmd_stripe_count * ssz)) * ssz;
433 __u32 off = (__u32)in % (md->lmd_stripe_count * ssz);
435 if ( in == 0xffffffffffffffff ) {
436 return 0xffffffffffffffff;
439 if ( (i+1) * ssz < off )
441 else if ( i * ssz > off )
444 out += (off - (i * ssz)) % ssz;
449 /* compute offset in stripe i corresponds to offset "in" */
450 __u64 lov_stripe(struct lov_stripe_md *md, __u64 in, int *j)
452 __u32 ssz = md->lmd_stripe_size;
454 /* full stripes across all * stripe size */
455 *j = (((__u32) in)/ssz) % md->lmd_stripe_count;
456 off = (__u32)in % (md->lmd_stripe_count * ssz);
457 out = ( ((__u32)in) / (md->lmd_stripe_count * ssz)) * ssz +
458 (off - ((*j) * ssz)) % ssz;;
463 int lov_stripe_which(struct lov_stripe_md *md, __u64 in)
465 __u32 ssz = md->lmd_stripe_size;
467 j = (((__u32) in)/ssz) % md->lmd_stripe_count;
472 /* FIXME: maybe we'll just make one node the authoritative attribute node, then
473 * we can send this 'punch' to just the authoritative node and the nodes
474 * that the punch will affect. */
475 static int lov_punch(struct lustre_handle *conn, struct obdo *oa,
476 struct lov_stripe_md *md,
477 obd_off start, obd_off end)
481 struct obd_export *export = class_conn2export(conn);
486 CERROR("LOV requires striping ea for desctruction\n");
490 if (!export || !export->exp_obd)
493 lov = &export->exp_obd->u.lov;
494 for (i = 0; i < md->lmd_stripe_count; i++) {
495 __u64 starti = lov_offset(md, start, i);
496 __u64 endi = lov_offset(md, end, i);
500 /* create data objects with "parent" OA */
501 memcpy(&tmp, oa, sizeof(tmp));
502 oa->o_id = md->lmd_objects[i].l_object_id;
504 rc = obd_punch(&lov->tgts[i].conn, &tmp, NULL,
507 CERROR("Error punch object %Ld on %d\n",
514 int lov_osc_brw_callback(struct io_cb_data *cbd, int err, int phase)
519 if (phase == CB_PHASE_START)
522 if (phase == CB_PHASE_FINISH) {
525 if (atomic_dec_and_test(&cbd->refcount))
526 ret = cbd->cb(cbd, cbd->err, phase);
534 static inline int lov_brw(int cmd, struct lustre_handle *conn,
535 struct lov_stripe_md *md,
537 struct brw_page *pga,
538 brw_callback_t callback, struct io_cb_data *cbd)
540 int stripe_count = md->lmd_stripe_count;
541 struct obd_export *export = class_conn2export(conn);
547 struct lov_stripe_md md;
549 struct brw_page *ioarr;
553 lov = &export->exp_obd->u.lov;
556 OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
560 OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
562 OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
566 for (i=0 ; i < oa_bufs ; i++ ) {
568 which = lov_stripe_which(md, pga[i].pg->index * PAGE_SIZE);
569 stripeinfo[which].bufct++;
572 for (i=0 ; i < stripe_count ; i++) {
574 stripeinfo[i].index =
575 stripeinfo[i-1].index + stripeinfo[i-1].bufct;
576 stripeinfo[i].md.lmd_object_id =
577 md->lmd_objects[i].l_object_id;
580 for (i=0 ; i < oa_bufs ; i++ ) {
582 which = lov_stripe_which(md, pga[i].pg->index * PAGE_SIZE);
584 shift = stripeinfo[which].index;
585 ioarr[shift + stripeinfo[which].subcount] = pga[i];
586 pga[i].off = lov_offset(md, pga[i].pg->index * PAGE_SIZE, which);
587 stripeinfo[which].subcount++;
591 atomic_set(&cbd->refcount, oa_bufs);
592 for (i=0 ; i < stripe_count ; i++) {
593 int shift = stripeinfo[i].index;
594 if (stripeinfo[i].bufct)
595 obd_brw(cmd, &lov->tgts[i].conn, &stripeinfo[i].md,
596 stripeinfo[i].bufct, &ioarr[shift],
597 lov_osc_brw_callback, cbd);
600 rc = callback(cbd, 0, CB_PHASE_START);
605 static int lov_enqueue(struct lustre_handle *conn, struct lov_stripe_md *md,
606 struct lustre_handle *parent_lock,
607 __u32 type, void *cookie, int cookielen, __u32 mode,
608 int *flags, void *cb, void *data, int datalen,
609 struct lustre_handle *lockhs)
612 struct obd_export *export = class_conn2export(conn);
617 CERROR("LOV requires striping ea for desctruction\n");
621 if (!export || !export->exp_obd)
624 lov = &export->exp_obd->u.lov;
625 for (i = 0; i < md->lmd_stripe_count; i++) {
626 struct ldlm_extent *extent = (struct ldlm_extent *)cookie;
627 struct ldlm_extent sub_ext;
628 struct lov_stripe_md submd;
630 sub_ext.start = lov_offset(md, extent->start, i);
631 sub_ext.end = lov_offset(md, extent->end, i);
632 if ( sub_ext.start == sub_ext.end )
635 submd.lmd_object_id = md->lmd_objects[i].l_object_id;
636 submd.lmd_easize = sizeof(submd);
637 rc = obd_enqueue(&(lov->tgts[i].conn), &submd, parent_lock, type,
638 &sub_ext, sizeof(sub_ext), mode, flags, cb, data, datalen, &(lockhs[i]));
639 // XXX add a lock debug statement here
641 CERROR("Error obd_enqueu object %Ld subobj %Ld\n", md->lmd_object_id,
642 md->lmd_objects[i].l_object_id);
648 static int lov_cancel(struct lustre_handle *conn, struct lov_stripe_md *md, __u32 mode,
649 struct lustre_handle *lockhs)
652 struct obd_export *export = class_conn2export(conn);
657 CERROR("LOV requires striping ea for lock cancellation\n");
661 if (!export || !export->exp_obd)
664 lov = &export->exp_obd->u.lov;
665 for (i = 0; i < md->lmd_stripe_count; i++) {
666 struct lov_stripe_md submd;
668 if ( lockhs[i].addr == 0 )
671 submd.lmd_object_id = md->lmd_objects[i].l_object_id;
672 submd.lmd_easize = sizeof(submd);
673 rc = obd_cancel(&lov->tgts[i].conn, &submd, mode, &lockhs[i]);
675 CERROR("Error punch object %Ld subobj %Ld\n", md->lmd_object_id,
676 md->lmd_objects[i].l_object_id);
685 struct obd_ops lov_obd_ops = {
687 o_connect: lov_connect,
688 o_disconnect: lov_disconnect,
689 o_create: lov_create,
690 o_destroy: lov_destroy,
691 o_getattr: lov_getattr,
692 o_setattr: lov_setattr,
697 o_enqueue: lov_enqueue,
702 #define LOV_VERSION "v0.1"
704 static int __init lov_init(void)
706 printk(KERN_INFO "Lustre Logical Object Volume driver " LOV_VERSION
707 ", info@clusterfs.com\n");
708 return class_register_type(&lov_obd_ops, OBD_LOV_DEVICENAME);
711 static void __exit lov_exit(void)
713 class_unregister_type(OBD_LOV_DEVICENAME);
716 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
717 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver v0.1");
718 MODULE_LICENSE("GPL");
720 module_init(lov_init);
721 module_exit(lov_exit);