1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002, 2003, 2004, 2005, 2006 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.lustre.org.
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 # define EXPORT_SYMTAB
25 #define DEBUG_SUBSYSTEM S_LMV
27 #include <linux/slab.h>
28 #include <linux/module.h>
29 #include <linux/init.h>
30 #include <linux/slab.h>
31 #include <linux/pagemap.h>
33 #include <asm/div64.h>
34 #include <linux/seq_file.h>
35 #include <linux/namei.h>
37 #include <liblustre.h>
40 #include <linux/ext2_fs.h>
42 #include <lustre/lustre_idl.h>
43 #include <lustre_log.h>
44 #include <obd_support.h>
45 #include <lustre_lib.h>
46 #include <lustre_net.h>
47 #include <obd_class.h>
48 #include <lprocfs_status.h>
49 #include <lustre_lite.h>
50 #include <lustre_fid.h>
51 #include "lmv_internal.h"
53 /* not defined for liblustre building */
54 #if !defined(ATOMIC_INIT)
55 #define ATOMIC_INIT(val) { (val) }
59 cfs_mem_cache_t *obj_cache;
60 atomic_t obj_cache_count = ATOMIC_INIT(0);
62 static void lmv_activate_target(struct lmv_obd *lmv,
63 struct lmv_tgt_desc *tgt,
66 if (tgt->ltd_active == activate)
69 tgt->ltd_active = activate;
70 lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
75 * -EINVAL : UUID can't be found in the LMV's target list
76 * -ENOTCONN: The UUID is found, but the target connection is bad (!)
77 * -EBADF : The UUID is found, but the OBD of the wrong type (!)
79 static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid,
82 struct lmv_tgt_desc *tgt;
83 struct obd_device *obd;
87 CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
88 lmv, uuid->uuid, activate);
90 spin_lock(&lmv->lmv_lock);
91 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
92 if (tgt->ltd_exp == NULL)
95 CDEBUG(D_INFO, "lmv idx %d is %s conn "LPX64"\n",
96 i, tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
98 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
102 if (i == lmv->desc.ld_tgt_count)
103 GOTO(out_lmv_lock, rc = -EINVAL);
105 obd = class_exp2obd(tgt->ltd_exp);
107 GOTO(out_lmv_lock, rc = -ENOTCONN);
109 CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
110 obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
111 obd->obd_type->typ_name, i);
112 LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
114 if (tgt->ltd_active == activate) {
115 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
116 activate ? "" : "in");
117 GOTO(out_lmv_lock, rc);
120 CDEBUG(D_INFO, "Marking OBD %p %sactive\n",
121 obd, activate ? "" : "in");
123 lmv_activate_target(lmv, tgt, activate);
128 spin_unlock(&lmv->lmv_lock);
132 static int lmv_set_mdc_data(struct lmv_obd *lmv, struct obd_uuid *uuid,
133 struct obd_connect_data *data)
135 struct lmv_tgt_desc *tgt;
139 LASSERT(data != NULL);
141 spin_lock(&lmv->lmv_lock);
142 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
143 if (tgt->ltd_exp == NULL)
146 if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
147 lmv->datas[tgt->ltd_idx] = *data;
151 spin_unlock(&lmv->lmv_lock);
155 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
156 enum obd_notify_event ev, void *data)
158 struct lmv_obd *lmv = &obd->u.lmv;
159 struct obd_uuid *uuid;
163 if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
164 CERROR("unexpected notification of %s %s!\n",
165 watched->obd_type->typ_name,
170 uuid = &watched->u.cli.cl_target_uuid;
171 if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
173 * Set MDC as active before notifying the observer, so the
174 * observer can use the MDC normally.
176 rc = lmv_set_mdc_active(lmv, uuid,
177 ev == OBD_NOTIFY_ACTIVE);
179 CERROR("%sactivation of %s failed: %d\n",
180 ev == OBD_NOTIFY_ACTIVE ? "" : "de",
184 } else if (ev == OBD_NOTIFY_OCD) {
185 struct obd_connect_data *conn_data =
186 &watched->u.cli.cl_import->imp_connect_data;
188 /* Set connect data to desired target, update
189 * exp_connect_flags. */
190 rc = lmv_set_mdc_data(lmv, uuid, conn_data);
192 CERROR("can't set connect data to target %s, rc %d\n",
198 * XXX: Make sure that ocd_connect_flags from all targets are
199 * the same. Otherwise one of MDTs runs wrong version or
200 * something like this. --umka
202 obd->obd_self_export->exp_connect_flags =
203 conn_data->ocd_connect_flags;
206 else if (ev == OBD_NOTIFY_DISCON) {
207 /* For disconnect event, flush fld cache for failout MDS case. */
208 fld_client_flush(&lmv->lmv_fld);
211 /* Pass the notification up the chain. */
212 if (obd->obd_observer)
213 rc = obd_notify(obd->obd_observer, watched, ev, data);
218 /* this is fake connect function. Its purpose is to initialize lmv and say
219 * caller that everything is okay. Real connection will be performed later. */
220 static int lmv_connect(const struct lu_env *env,
221 struct lustre_handle *conn, struct obd_device *obd,
222 struct obd_uuid *cluuid, struct obd_connect_data *data,
226 struct proc_dir_entry *lmv_proc_dir;
228 struct lmv_obd *lmv = &obd->u.lmv;
229 struct obd_export *exp;
233 rc = class_connect(conn, obd, cluuid);
235 CERROR("class_connection() returned %d\n", rc);
239 exp = class_conn2export(conn);
241 /* we don't want to actually do the underlying connections more than
242 * once, so keep track. */
244 if (lmv->refcount > 1) {
245 class_export_put(exp);
251 lmv->cluuid = *cluuid;
254 lmv->conn_data = *data;
257 lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
259 if (IS_ERR(lmv_proc_dir)) {
260 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
261 obd->obd_type->typ_name, obd->obd_name);
266 /* all real clients should perform actual connection right away, because
267 * it is possible, that LMV will not have opportunity to connect targets
268 * and MDC stuff will be called directly, for instance while reading
269 * ../mdc/../kbytesfree procfs file, etc. */
270 if (data->ocd_connect_flags & OBD_CONNECT_REAL)
271 rc = lmv_check_connect(obd);
276 lprocfs_remove(&lmv_proc_dir);
283 static void lmv_set_timeouts(struct obd_device *obd)
285 struct lmv_tgt_desc *tgts;
290 if (lmv->server_timeout == 0)
293 if (lmv->connected == 0)
296 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
297 if (tgts->ltd_exp == NULL)
300 obd_set_info_async(tgts->ltd_exp, strlen("inter_mds"),
301 "inter_mds", 0, NULL, NULL);
305 static int lmv_init_ea_size(struct obd_export *exp, int easize,
306 int def_easize, int cookiesize)
308 struct obd_device *obd = exp->exp_obd;
309 struct lmv_obd *lmv = &obd->u.lmv;
310 int i, rc = 0, change = 0;
313 if (lmv->max_easize < easize) {
314 lmv->max_easize = easize;
317 if (lmv->max_def_easize < def_easize) {
318 lmv->max_def_easize = def_easize;
321 if (lmv->max_cookiesize < cookiesize) {
322 lmv->max_cookiesize = cookiesize;
328 if (lmv->connected == 0)
331 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
332 if (lmv->tgts[i].ltd_exp == NULL) {
333 CWARN("%s: NULL export for %d\n", obd->obd_name, i);
337 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
340 CERROR("obd_init_ea_size() failed on MDT target %d, "
341 "error %d.\n", i, rc);
348 #define MAX_STRING_SIZE 128
350 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
352 struct lmv_obd *lmv = &obd->u.lmv;
353 struct obd_uuid *cluuid = &lmv->cluuid;
354 struct obd_connect_data *mdc_data = NULL;
355 struct obd_uuid lmv_mdc_uuid = { "LMV_MDC_UUID" };
356 struct lustre_handle conn = {0, };
357 struct obd_device *mdc_obd;
358 struct obd_export *mdc_exp;
359 struct lu_fld_target target;
362 struct proc_dir_entry *lmv_proc_dir;
366 /* for MDS: don't connect to yourself */
367 if (obd_uuid_equals(&tgt->ltd_uuid, cluuid)) {
368 CDEBUG(D_CONFIG, "don't connect back to %s\n", cluuid->uuid);
369 /* XXX - the old code didn't increment active tgt count.
374 mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
377 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
381 CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
382 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
383 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
386 if (!mdc_obd->obd_set_up) {
387 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
391 rc = obd_connect(NULL, &conn, mdc_obd, &lmv_mdc_uuid,
392 &lmv->conn_data, NULL);
394 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
398 mdc_exp = class_conn2export(&conn);
400 /* Init fid sequence client for this mdc. */
401 rc = obd_fid_init(mdc_exp);
405 /* Add new FLD target. */
406 target.ft_srv = NULL;
407 target.ft_exp = mdc_exp;
408 target.ft_idx = tgt->ltd_idx;
410 fld_client_add_target(&lmv->lmv_fld, &target);
412 mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
414 rc = obd_register_observer(mdc_obd, obd);
416 obd_disconnect(mdc_exp);
417 CERROR("target %s register_observer error %d\n",
418 tgt->ltd_uuid.uuid, rc);
422 if (obd->obd_observer) {
423 /* tell the mds_lmv about the new target */
424 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
425 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
427 obd_disconnect(mdc_exp);
433 tgt->ltd_exp = mdc_exp;
434 lmv->desc.ld_active_tgt_count++;
436 /* copy connect data, it may be used later */
437 lmv->datas[tgt->ltd_idx] = *mdc_data;
439 md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
440 lmv->max_def_easize, lmv->max_cookiesize);
442 CDEBUG(D_CONFIG, "connected to %s(%s) successfully (%d)\n",
443 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
444 atomic_read(&obd->obd_refcount));
447 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
449 struct proc_dir_entry *mdc_symlink;
450 char name[MAX_STRING_SIZE + 1];
452 LASSERT(mdc_obd->obd_type != NULL);
453 LASSERT(mdc_obd->obd_type->typ_name != NULL);
454 name[MAX_STRING_SIZE] = '\0';
455 snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
456 mdc_obd->obd_type->typ_name,
458 mdc_symlink = proc_symlink(mdc_obd->obd_name,
460 if (mdc_symlink == NULL) {
461 CERROR("could not register LMV target "
462 "/proc/fs/lustre/%s/%s/target_obds/%s.",
463 obd->obd_type->typ_name, obd->obd_name,
465 lprocfs_remove(&lmv_proc_dir);
473 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
475 struct lmv_obd *lmv = &obd->u.lmv;
476 struct lmv_tgt_desc *tgt;
480 CDEBUG(D_CONFIG, "tgt_uuid: %s.\n", tgt_uuid->uuid);
484 if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
485 lmv_init_unlock(lmv);
486 CERROR("can't add %s, LMV module compiled for %d MDCs. "
487 "That many MDCs already configured.\n",
488 tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
491 if (lmv->desc.ld_tgt_count == 0) {
492 struct obd_device *mdc_obd;
494 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
497 lmv_init_unlock(lmv);
498 CERROR("Target %s not attached\n", tgt_uuid->uuid);
502 rc = obd_llog_init(obd, OBD_LLOG_GROUP, mdc_obd, 0, NULL, tgt_uuid);
504 lmv_init_unlock(lmv);
505 CERROR("lmv failed to setup llogging subsystems\n");
508 spin_lock(&lmv->lmv_lock);
509 tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
510 tgt->ltd_uuid = *tgt_uuid;
511 spin_unlock(&lmv->lmv_lock);
513 if (lmv->connected) {
514 rc = lmv_connect_mdc(obd, tgt);
516 spin_lock(&lmv->lmv_lock);
517 lmv->desc.ld_tgt_count--;
518 memset(tgt, 0, sizeof(*tgt));
519 spin_unlock(&lmv->lmv_lock);
521 int easize = sizeof(struct lmv_stripe_md) +
522 lmv->desc.ld_tgt_count *
523 sizeof(struct lu_fid);
524 lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
528 lmv_init_unlock(lmv);
532 /* performs a check if passed obd is connected. If no - connect it. */
533 int lmv_check_connect(struct obd_device *obd)
535 struct lmv_obd *lmv = &obd->u.lmv;
536 struct lmv_tgt_desc *tgt;
544 if (lmv->connected) {
545 lmv_init_unlock(lmv);
549 if (lmv->desc.ld_tgt_count == 0) {
550 CERROR("%s: no targets configured.\n", obd->obd_name);
554 CDEBUG(D_CONFIG, "time to connect %s to %s\n",
555 lmv->cluuid.uuid, obd->obd_name);
557 LASSERT(lmv->tgts != NULL);
559 for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
560 rc = lmv_connect_mdc(obd, tgt);
565 lmv_set_timeouts(obd);
566 class_export_put(lmv->exp);
568 easize = lmv_get_easize(lmv);
569 lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
570 lmv_init_unlock(lmv);
579 --lmv->desc.ld_active_tgt_count;
580 rc2 = obd_disconnect(tgt->ltd_exp);
582 CERROR("error: LMV target %s disconnect on "
583 "MDC idx %d: error %d\n",
584 tgt->ltd_uuid.uuid, i, rc2);
588 class_disconnect(lmv->exp);
589 lmv_init_unlock(lmv);
593 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
596 struct proc_dir_entry *lmv_proc_dir;
598 struct lmv_obd *lmv = &obd->u.lmv;
599 struct obd_device *mdc_obd;
603 LASSERT(tgt != NULL);
604 LASSERT(obd != NULL);
606 mdc_obd = class_exp2obd(tgt->ltd_exp);
609 mdc_obd->obd_no_recov = obd->obd_no_recov;
612 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
614 struct proc_dir_entry *mdc_symlink;
616 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
618 lprocfs_remove(&mdc_symlink);
620 CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
621 obd->obd_type->typ_name, obd->obd_name,
626 rc = obd_fid_fini(tgt->ltd_exp);
628 CERROR("Can't finanize fids factory\n");
630 CDEBUG(D_OTHER, "Disconnected from %s(%s) successfully\n",
631 tgt->ltd_exp->exp_obd->obd_name,
632 tgt->ltd_exp->exp_obd->obd_uuid.uuid);
634 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
635 rc = obd_disconnect(tgt->ltd_exp);
637 if (tgt->ltd_active) {
638 CERROR("Target %s disconnect error %d\n",
639 tgt->ltd_uuid.uuid, rc);
643 lmv_activate_target(lmv, tgt, 0);
648 static int lmv_disconnect(struct obd_export *exp)
650 struct obd_device *obd = class_exp2obd(exp);
652 struct proc_dir_entry *lmv_proc_dir;
654 struct lmv_obd *lmv = &obd->u.lmv;
661 /* Only disconnect the underlying layers on the final disconnect. */
663 if (lmv->refcount != 0)
666 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
667 if (lmv->tgts[i].ltd_exp == NULL)
669 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
673 lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
675 lprocfs_remove(&lmv_proc_dir);
677 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
678 obd->obd_type->typ_name, obd->obd_name);
684 * This is the case when no real connection is established by
685 * lmv_check_connect().
688 class_export_put(exp);
689 rc = class_disconnect(exp);
690 if (lmv->refcount == 0)
695 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
696 int len, void *karg, void *uarg)
698 struct obd_device *obddev = class_exp2obd(exp);
699 struct lmv_obd *lmv = &obddev->u.lmv;
700 int i, rc = 0, set = 0;
703 if (lmv->desc.ld_tgt_count == 0)
707 case IOC_OBD_STATFS: {
708 struct obd_ioctl_data *data = karg;
709 struct obd_device *mdc_obd;
710 struct obd_statfs stat_buf = {0};
713 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
714 LASSERT(data->ioc_plen1 == sizeof(struct obd_statfs));
716 if ((index >= lmv->desc.ld_tgt_count))
719 if (!lmv->tgts[index].ltd_active)
722 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
726 /* got statfs data */
727 rc = obd_statfs(mdc_obd, &stat_buf,
728 cfs_time_current_64() - HZ, 0);
731 if (copy_to_user(data->ioc_pbuf1, &stat_buf, data->ioc_plen1))
734 rc = copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
739 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
742 if (lmv->tgts[i].ltd_exp == NULL)
745 err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
748 if (lmv->tgts[i].ltd_active) {
749 CERROR("error: iocontrol MDC %s on MDT"
750 "idx %d cmd %x: err = %d\n",
751 lmv->tgts[i].ltd_uuid.uuid,
771 static int lmv_all_chars_policy(int count, const char *name,
782 static int lmv_nid_policy(struct lmv_obd *lmv)
784 struct obd_import *imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
787 * XXX Hack: to get nid we assume that underlying obd device is mdc.
789 id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
790 return id % lmv->desc.ld_tgt_count;
793 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
798 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
800 op_data->op_namelen);
802 return lmv_nid_policy(lmv);
808 CERROR("unsupport type %d \n", type);
812 /* This is _inode_ placement policy function (not name). */
813 static int lmv_placement_policy(struct obd_device *obd,
814 struct md_op_data *op_data,
817 struct lmv_obd *lmv = &obd->u.lmv;
822 LASSERT(mds != NULL);
825 * Allocate new fid on target according to operation type and parent
828 obj = lmv_obj_grab(obd, &op_data->op_fid1);
829 if (obj != NULL || op_data->op_name == NULL ||
830 op_data->op_opc != LUSTRE_OPC_MKDIR) {
832 * Allocate fid for non-dir or for null name or for case parent
839 * If we have this flag turned on, and we see that
840 * parent dir is split, this means, that caller did not
841 * notice split yet. This is race and we would like to
842 * let caller know that.
844 if (op_data->op_bias & MDS_CHECK_SPLIT)
849 * Allocate new fid on same mds where parent fid is located and
850 * where operation will be sent. In case of split dir, ->op_fid1
851 * and ->op_mds here will contain fid and mds of slave directory
852 * object (assigned by caller).
854 *mds = op_data->op_mds;
858 /* XXX: This should be removed later wehn we sure it is not
860 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, mds);
866 * Parent directory is not split and we want to create a
867 * directory in it. Let's calculate where to place it according
870 *mds = lmv_choose_mds(lmv, op_data, NID_TYPE);
878 CERROR("Can't choose MDS, err = %d\n", rc);
880 LASSERT(*mds < lmv->desc.ld_tgt_count);
886 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
889 struct lmv_tgt_desc *tgt = &lmv->tgts[mds];
893 /* New seq alloc and FLD setup should be atomic. */
894 down(&tgt->ltd_fid_sem);
896 /* Asking underlaying tgt layer to allocate new fid. */
897 rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
899 LASSERT(fid_is_sane(fid));
901 /* Client switches to new sequence, setup FLD. */
902 rc = fld_client_create(&lmv->lmv_fld, fid_seq(fid),
905 CERROR("Can't create fld entry, rc %d\n", rc);
906 /* Delete just allocated fid sequence */
907 obd_fid_delete(tgt->ltd_exp, NULL);
910 up(&tgt->ltd_fid_sem);
914 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
915 struct md_op_data *op_data)
917 struct obd_device *obd = class_exp2obd(exp);
918 struct lmv_obd *lmv = &obd->u.lmv;
923 LASSERT(op_data != NULL);
924 LASSERT(fid != NULL);
926 rc = lmv_placement_policy(obd, op_data, &mds);
928 CERROR("Can't get target for allocating fid, "
933 rc = __lmv_fid_alloc(lmv, fid, mds);
935 CERROR("Can't alloc new fid, rc %d\n", rc);
942 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
947 if (lmv_obj_delete(exp, fid)) {
948 CDEBUG(D_OTHER, "lmv object "DFID" is destroyed.\n",
954 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
956 struct lmv_obd *lmv = &obd->u.lmv;
957 struct lprocfs_static_vars lvars;
958 struct lmv_desc *desc;
962 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
963 CERROR("LMV setup requires a descriptor\n");
967 desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
968 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
969 CERROR("descriptor size wrong: %d > %d\n",
970 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
974 lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
976 OBD_ALLOC(lmv->tgts, lmv->tgts_size);
977 if (lmv->tgts == NULL)
980 for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
981 sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
982 lmv->tgts[i].ltd_idx = i;
985 lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
987 OBD_ALLOC(lmv->datas, lmv->datas_size);
988 if (lmv->datas == NULL)
989 GOTO(out_free_tgts, rc = -ENOMEM);
991 obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
992 lmv->desc.ld_tgt_count = 0;
993 lmv->desc.ld_active_tgt_count = 0;
994 lmv->max_cookiesize = 0;
995 lmv->max_def_easize = 0;
998 spin_lock_init(&lmv->lmv_lock);
999 sema_init(&lmv->init_sem, 1);
1001 rc = lmv_obj_setup(obd);
1003 CERROR("Can't setup LMV object manager, "
1005 GOTO(out_free_datas, rc);
1008 lprocfs_lmv_init_vars(&lvars);
1009 lprocfs_obd_setup(obd, lvars.obd_vars);
1012 struct proc_dir_entry *entry;
1014 entry = create_proc_entry("target_obd_status", 0444,
1015 obd->obd_proc_entry);
1016 if (entry != NULL) {
1017 entry->proc_fops = &lmv_proc_target_fops;
1022 rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1023 LUSTRE_CLI_FLD_HASH_DHT);
1025 CERROR("can't init FLD, err %d\n",
1027 GOTO(out_free_datas, rc);
1033 OBD_FREE(lmv->datas, lmv->datas_size);
1036 OBD_FREE(lmv->tgts, lmv->tgts_size);
1041 static int lmv_cleanup(struct obd_device *obd)
1043 struct lmv_obd *lmv = &obd->u.lmv;
1046 fld_client_fini(&lmv->lmv_fld);
1047 lprocfs_obd_cleanup(obd);
1048 lmv_obj_cleanup(obd);
1049 OBD_FREE(lmv->datas, lmv->datas_size);
1050 OBD_FREE(lmv->tgts, lmv->tgts_size);
1055 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1057 struct lustre_cfg *lcfg = buf;
1058 struct obd_uuid tgt_uuid;
1062 switch(lcfg->lcfg_command) {
1064 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1065 GOTO(out, rc = -EINVAL);
1067 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1068 rc = lmv_add_target(obd, &tgt_uuid);
1071 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1072 GOTO(out, rc = -EINVAL);
1079 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1080 __u64 max_age, __u32 flags)
1082 struct lmv_obd *lmv = &obd->u.lmv;
1083 struct obd_statfs *temp;
1087 rc = lmv_check_connect(obd);
1091 OBD_ALLOC(temp, sizeof(*temp));
1095 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1096 if (lmv->tgts[i].ltd_exp == NULL)
1099 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1102 CERROR("can't stat MDS #%d (%s), error %d\n", i,
1103 lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1105 GOTO(out_free_temp, rc);
1110 osfs->os_bavail += temp->os_bavail;
1111 osfs->os_blocks += temp->os_blocks;
1112 osfs->os_ffree += temp->os_ffree;
1113 osfs->os_files += temp->os_files;
1119 OBD_FREE(temp, sizeof(*temp));
1123 static int lmv_getstatus(struct obd_export *exp,
1125 struct obd_capa **pc)
1127 struct obd_device *obd = exp->exp_obd;
1128 struct lmv_obd *lmv = &obd->u.lmv;
1132 rc = lmv_check_connect(obd);
1136 rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1141 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1142 struct obd_capa *oc, obd_valid valid, const char *name,
1143 const char *input, int input_size, int output_size,
1144 int flags, struct ptlrpc_request **request)
1146 struct obd_device *obd = exp->exp_obd;
1147 struct lmv_obd *lmv = &obd->u.lmv;
1148 struct obd_export *tgt_exp;
1152 rc = lmv_check_connect(obd);
1156 tgt_exp = lmv_find_export(lmv, fid);
1157 if (IS_ERR(tgt_exp))
1158 RETURN(PTR_ERR(tgt_exp));
1160 rc = md_getxattr(tgt_exp, fid, oc, valid, name, input, input_size,
1161 output_size, flags, request);
1166 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1167 struct obd_capa *oc, obd_valid valid, const char *name,
1168 const char *input, int input_size, int output_size,
1169 int flags, __u32 suppgid,
1170 struct ptlrpc_request **request)
1172 struct obd_device *obd = exp->exp_obd;
1173 struct lmv_obd *lmv = &obd->u.lmv;
1174 struct obd_export *tgt_exp;
1178 rc = lmv_check_connect(obd);
1182 tgt_exp = lmv_find_export(lmv, fid);
1183 if (IS_ERR(tgt_exp))
1184 RETURN(PTR_ERR(tgt_exp));
1186 rc = md_setxattr(tgt_exp, fid, oc, valid, name,
1187 input, input_size, output_size, flags, suppgid,
1193 static int lmv_getattr(struct obd_export *exp, const struct lu_fid *fid,
1194 struct obd_capa *oc, obd_valid valid, int ea_size,
1195 struct ptlrpc_request **request)
1197 struct obd_device *obd = exp->exp_obd;
1198 struct lmv_obd *lmv = &obd->u.lmv;
1199 struct obd_export *tgt_exp;
1200 struct lmv_obj *obj;
1204 rc = lmv_check_connect(obd);
1208 tgt_exp = lmv_find_export(lmv, fid);
1209 if (IS_ERR(tgt_exp))
1210 RETURN(PTR_ERR(tgt_exp));
1212 rc = md_getattr(tgt_exp, fid, oc, valid, ea_size, request);
1216 obj = lmv_obj_grab(obd, fid);
1218 CDEBUG(D_OTHER, "GETATTR for "DFID" %s\n", PFID(fid),
1219 obj ? "(split)" : "");
1222 * If object is split, then we loop over all the slaves and gather size
1223 * attribute. In ideal world we would have to gather also mds field from
1224 * all slaves, as object is spread over the cluster and this is
1225 * definitely interesting information and it is not good to loss it,
1229 struct mdt_body *body;
1231 if (*request == NULL) {
1236 body = req_capsule_server_get(&(*request)->rq_pill,
1238 LASSERT(body != NULL);
1242 for (i = 0; i < obj->lo_objcount; i++) {
1243 if (lmv->tgts[i].ltd_exp == NULL) {
1244 CWARN("%s: NULL export for %d\n",
1249 /* skip master obj. */
1250 if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid))
1253 lmv_update_body(body, &obj->lo_inodes[i]);
1256 lmv_obj_unlock(obj);
1263 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1264 ldlm_iterator_t it, void *data)
1266 struct obd_device *obd = exp->exp_obd;
1267 struct lmv_obd *lmv = &obd->u.lmv;
1271 rc = lmv_check_connect(obd);
1275 CDEBUG(D_OTHER, "CBDATA for "DFID"\n", PFID(fid));
1278 * With CMD every object can have two locks in different namespaces:
1279 * lookup lock in space of mds storing direntry and update/open lock in
1280 * space of mds storing inode.
1282 for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1283 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1288 static int lmv_close(struct obd_export *exp,
1289 struct md_op_data *op_data,
1290 struct md_open_data *mod,
1291 struct ptlrpc_request **request)
1293 struct obd_device *obd = exp->exp_obd;
1294 struct lmv_obd *lmv = &obd->u.lmv;
1295 struct obd_export *tgt_exp;
1299 rc = lmv_check_connect(obd);
1303 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1304 if (IS_ERR(tgt_exp))
1305 RETURN(PTR_ERR(tgt_exp));
1307 CDEBUG(D_OTHER, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1308 rc = md_close(tgt_exp, op_data, mod, request);
1313 * Called in the case MDS returns -ERESTART on create on open, what means that
1314 * directory is split and its LMV presentation object has to be updated.
1316 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1318 struct obd_device *obd = exp->exp_obd;
1319 struct lmv_obd *lmv = &obd->u.lmv;
1320 struct ptlrpc_request *req = NULL;
1321 struct obd_export *tgt_exp;
1322 struct lmv_obj *obj;
1323 struct lustre_md md;
1329 mealen = lmv_get_easize(lmv);
1331 valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1333 tgt_exp = lmv_find_export(lmv, fid);
1334 if (IS_ERR(tgt_exp))
1335 RETURN(PTR_ERR(tgt_exp));
1337 /* time to update mea of parent fid */
1338 rc = md_getattr(tgt_exp, fid, NULL, valid, mealen, &req);
1340 CERROR("md_getattr() failed, error %d\n", rc);
1344 rc = md_get_lustre_md(tgt_exp, req, NULL, exp, &md);
1346 CERROR("mdc_get_lustre_md() failed, error %d\n", rc);
1351 GOTO(cleanup, rc = -ENODATA);
1353 obj = lmv_obj_create(exp, fid, md.mea);
1359 obd_free_memmd(exp, (struct lov_stripe_md **)&md.mea);
1364 ptlrpc_req_finished(req);
1368 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1369 const void *data, int datalen, int mode, __u32 uid,
1370 __u32 gid, __u32 cap_effective, __u64 rdev,
1371 struct ptlrpc_request **request)
1373 struct obd_device *obd = exp->exp_obd;
1374 struct lmv_obd *lmv = &obd->u.lmv;
1375 struct obd_export *tgt_exp;
1376 struct lmv_obj *obj;
1380 rc = lmv_check_connect(obd);
1384 if (!lmv->desc.ld_active_tgt_count)
1389 obj = lmv_obj_grab(obd, &op_data->op_fid1);
1393 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1394 op_data->op_name, op_data->op_namelen);
1395 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1396 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1397 op_data->op_mds = obj->lo_inodes[mea_idx].li_mds;
1398 tgt_exp = lmv_get_export(lmv, op_data->op_mds);
1401 struct lmv_tgt_desc *tgt;
1403 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1404 op_data->op_bias |= MDS_CHECK_SPLIT;
1405 op_data->op_mds = tgt->ltd_idx;
1406 tgt_exp = tgt->ltd_exp;
1409 if (IS_ERR(tgt_exp))
1410 RETURN(PTR_ERR(tgt_exp));
1412 rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1413 if (rc == -ERESTART)
1418 CDEBUG(D_OTHER, "CREATE '%*s' on "DFID"\n", op_data->op_namelen,
1419 op_data->op_name, PFID(&op_data->op_fid1));
1421 op_data->op_flags |= MF_MDC_CANCEL_FID1;
1422 rc = md_create(tgt_exp, op_data, data, datalen, mode, uid, gid,
1423 cap_effective, rdev, request);
1425 if (*request == NULL)
1427 CDEBUG(D_OTHER, "created - "DFID"\n", PFID(&op_data->op_fid1));
1428 } else if (rc == -ERESTART) {
1429 LASSERT(*request != NULL);
1430 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1431 "Got -ERESTART during create!\n");
1432 ptlrpc_req_finished(*request);
1436 * Directory got split. Time to update local object and repeat
1437 * the request with proper MDS.
1439 rc = lmv_handle_split(exp, &op_data->op_fid1);
1441 rc = lmv_alloc_slave_fids(obd, &op_data->op_fid1,
1442 op_data, &op_data->op_fid2);
1451 static int lmv_done_writing(struct obd_export *exp,
1452 struct md_op_data *op_data,
1453 struct md_open_data *mod)
1455 struct obd_device *obd = exp->exp_obd;
1456 struct lmv_obd *lmv = &obd->u.lmv;
1457 struct obd_export *tgt_exp;
1461 rc = lmv_check_connect(obd);
1465 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1466 if (IS_ERR(tgt_exp))
1467 RETURN(PTR_ERR(tgt_exp));
1469 rc = md_done_writing(tgt_exp, op_data, mod);
1474 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1475 struct lookup_intent *it, struct md_op_data *op_data,
1476 struct lustre_handle *lockh, void *lmm, int lmmsize)
1478 struct obd_device *obd = exp->exp_obd;
1479 struct lmv_obd *lmv = &obd->u.lmv;
1480 struct lmv_stripe_md *mea = op_data->op_mea1;
1481 struct md_op_data *op_data2;
1482 struct obd_export *tgt_exp;
1486 OBD_ALLOC_PTR(op_data2);
1487 if (op_data2 == NULL)
1490 LASSERT(mea != NULL);
1491 for (i = 0; i < mea->mea_count; i++) {
1492 memset(op_data2, 0, sizeof(*op_data2));
1493 op_data2->op_fid1 = mea->mea_ids[i];
1494 op_data2->op_bias = 0;
1496 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
1497 if (IS_ERR(tgt_exp))
1498 GOTO(cleanup, rc = PTR_ERR(tgt_exp));
1500 if (tgt_exp == NULL)
1503 rc = md_enqueue(tgt_exp, einfo, it, op_data2,
1504 lockh + i, lmm, lmmsize, 0);
1506 CDEBUG(D_OTHER, "take lock on slave "DFID" -> %d/%d\n",
1507 PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1512 if (it->d.lustre.it_data) {
1513 struct ptlrpc_request *req;
1514 req = (struct ptlrpc_request *)it->d.lustre.it_data;
1515 ptlrpc_req_finished(req);
1518 if (it->d.lustre.it_status)
1519 GOTO(cleanup, rc = it->d.lustre.it_status);
1524 OBD_FREE_PTR(op_data2);
1527 /* drop all taken locks */
1529 if (lockh[i].cookie)
1530 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1531 lockh[i].cookie = 0;
1538 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1539 struct lookup_intent *it, struct md_op_data *op_data,
1540 struct lustre_handle *lockh, void *lmm, int lmmsize,
1541 int extra_lock_flags)
1543 struct ptlrpc_request *req = it->d.lustre.it_data;
1544 struct obd_device *obd = exp->exp_obd;
1545 struct lmv_obd *lmv = &obd->u.lmv;
1546 struct lustre_handle plock;
1547 struct obd_export *tgt_exp;
1548 struct md_op_data *rdata;
1549 struct lu_fid fid_copy;
1550 struct mdt_body *body;
1554 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1555 LASSERT(body != NULL);
1557 if (!(body->valid & OBD_MD_MDS))
1560 CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID" -> "DFID"\n",
1561 LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1563 /* We got LOOKUP lock, but we really need attrs */
1564 pmode = it->d.lustre.it_lock_mode;
1565 LASSERT(pmode != 0);
1566 memcpy(&plock, lockh, sizeof(plock));
1567 it->d.lustre.it_lock_mode = 0;
1568 it->d.lustre.it_data = NULL;
1569 fid_copy = body->fid1;
1571 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1572 ptlrpc_req_finished(req);
1574 tgt_exp = lmv_find_export(lmv, &fid_copy);
1575 if (IS_ERR(tgt_exp))
1576 GOTO(out, rc = PTR_ERR(tgt_exp));
1578 OBD_ALLOC_PTR(rdata);
1580 GOTO(out, rc = -ENOMEM);
1582 rdata->op_fid1 = fid_copy;
1583 rdata->op_bias = MDS_CROSS_REF;
1585 rc = md_enqueue(tgt_exp, einfo, it, rdata, lockh,
1586 lmm, lmmsize, extra_lock_flags);
1587 OBD_FREE_PTR(rdata);
1590 ldlm_lock_decref(&plock, pmode);
1595 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1596 struct lookup_intent *it, struct md_op_data *op_data,
1597 struct lustre_handle *lockh, void *lmm, int lmmsize,
1598 int extra_lock_flags)
1600 struct obd_device *obd = exp->exp_obd;
1601 struct lmv_obd *lmv = &obd->u.lmv;
1602 struct obd_export *tgt_exp = NULL;
1603 struct lmv_obj *obj;
1607 rc = lmv_check_connect(obd);
1611 if (op_data->op_mea1 && it->it_op == IT_UNLINK) {
1612 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1613 lockh, lmm, lmmsize);
1617 if (op_data->op_namelen) {
1618 obj = lmv_obj_grab(obd, &op_data->op_fid1);
1622 /* directory is split. look for right mds for this
1624 mea_idx = raw_name2idx(obj->lo_hashtype,
1626 (char *)op_data->op_name,
1627 op_data->op_namelen);
1628 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1629 tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1634 if (tgt_exp == NULL)
1635 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1636 if (IS_ERR(tgt_exp))
1637 RETURN(PTR_ERR(tgt_exp));
1639 CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID"\n", LL_IT2STR(it),
1640 PFID(&op_data->op_fid1));
1642 rc = md_enqueue(tgt_exp, einfo, it, op_data, lockh,
1643 lmm, lmmsize, extra_lock_flags);
1645 if (rc == 0 && it->it_op == IT_OPEN)
1646 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1647 lmm, lmmsize, extra_lock_flags);
1652 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1653 struct obd_capa *oc, const char *filename, int namelen,
1654 obd_valid valid, int ea_size, __u32 suppgid,
1655 struct ptlrpc_request **request)
1657 struct obd_device *obd = exp->exp_obd;
1658 struct lmv_obd *lmv = &obd->u.lmv;
1659 struct lu_fid rid = *fid;
1660 struct obd_export *tgt_exp;
1661 struct mdt_body *body;
1662 struct lmv_obj *obj;
1666 rc = lmv_check_connect(obd);
1673 obj = lmv_obj_grab(obd, &rid);
1677 /* Directory is split. Look for right mds for this name */
1678 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1679 filename, namelen - 1);
1680 rid = obj->lo_inodes[mea_idx].li_fid;
1681 tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1683 valid &= ~OBD_MD_FLCKSPLIT;
1685 tgt_exp = lmv_find_export(lmv, &rid);
1686 valid |= OBD_MD_FLCKSPLIT;
1688 if (IS_ERR(tgt_exp))
1689 RETURN(PTR_ERR(tgt_exp));
1691 CDEBUG(D_OTHER, "getattr_name for %*s on "DFID" -> "DFID"\n",
1692 namelen, filename, PFID(fid), PFID(&rid));
1694 rc = md_getattr_name(tgt_exp, &rid, oc, filename, namelen, valid,
1695 ea_size, suppgid, request);
1697 body = req_capsule_server_get(&(*request)->rq_pill,
1699 LASSERT(body != NULL);
1701 if (body->valid & OBD_MD_MDS) {
1702 struct ptlrpc_request *req = NULL;
1705 CDEBUG(D_OTHER, "request attrs for "DFID"\n",
1708 tgt_exp = lmv_find_export(lmv, &rid);
1709 if (IS_ERR(tgt_exp)) {
1710 ptlrpc_req_finished(*request);
1711 RETURN(PTR_ERR(tgt_exp));
1714 rc = md_getattr_name(tgt_exp, &rid, NULL, NULL, 1,
1715 valid | OBD_MD_FLCROSSREF,
1716 ea_size, suppgid, &req);
1717 ptlrpc_req_finished(*request);
1720 } else if (rc == -ERESTART) {
1721 LASSERT(*request != NULL);
1722 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1723 "Got -ERESTART during getattr!\n");
1724 ptlrpc_req_finished(*request);
1728 * Directory got split. Time to update local object and repeat
1729 * the request with proper MDS.
1731 rc = lmv_handle_split(exp, &rid);
1738 #define md_op_data_fid(op_data, fl) \
1739 (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1740 fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1741 fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1742 fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1745 /* @tgt_exp is the export the metadata request is sent.
1746 * @fid_exp is the export the cancel should be sent for the current fid.
1747 * if @fid_exp is NULL, the export is found for the current fid.
1748 * @op_data keeps the current fid, which is pointed through @flag.
1749 * @mode, @bits -- lock match parameters. */
1750 static int lmv_early_cancel(struct lmv_obd *lmv, struct obd_export *tgt_exp,
1751 struct obd_export *fid_exp,
1752 struct md_op_data *op_data,
1753 ldlm_mode_t mode, int bits, int flag)
1755 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1756 ldlm_policy_data_t policy = {{0}};
1760 if (!fid_is_sane(fid))
1763 if (fid_exp == NULL)
1764 fid_exp = lmv_find_export(lmv, fid);
1766 if (tgt_exp == fid_exp) {
1767 /* The export is the same as on the target server, cancel
1768 * will be sent along with the main metadata operation. */
1769 op_data->op_flags |= flag;
1773 policy.l_inodebits.bits = bits;
1774 rc = md_cancel_unused(fid_exp, fid, &policy, mode, LDLM_FL_ASYNC, NULL);
1778 #ifdef EARLY_CANCEL_FOR_STRIPED_DIR_IS_READY
1779 /* Check if the fid in @op_data pointed to by flag is of the same export(s)
1780 * as @tgt_exp. Early cancels will be sent later by mdc code, otherwise, call
1781 * md_cancel_unused for child export(s). */
1782 static int lmv_early_cancel_stripes(struct obd_export *exp,
1783 struct obd_export *tgt_exp,
1784 struct md_op_data *op_data,
1785 ldlm_mode_t mode, int bits, int flag)
1787 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1788 struct obd_device *obd = exp->exp_obd;
1789 struct lmv_obd *lmv = &obd->u.lmv;
1790 struct obd_export *st_exp;
1791 struct lmv_obj *obj;
1795 if (!fid_is_sane(fid))
1798 obj = lmv_obj_grab(obd, fid);
1800 ldlm_policy_data_t policy = {{0}};
1801 struct lu_fid *st_fid;
1804 policy.l_inodebits.bits = bits;
1805 for (i = 0; i < obj->lo_objcount; i++) {
1806 st_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
1807 st_fid = &obj->lo_inodes[i].li_fid;
1808 if (tgt_exp != st_exp) {
1809 rc = md_cancel_unused(st_exp, st_fid, &policy,
1810 mode, LDLM_FL_ASYNC,
1815 /* Some export matches to @tgt_exp, do cancel
1816 * for its fid in mdc */
1818 op_data->op_flags |= flag;
1823 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data,
1831 * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1834 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1835 struct ptlrpc_request **request)
1837 struct obd_device *obd = exp->exp_obd;
1838 struct lmv_obd *lmv = &obd->u.lmv;
1839 struct obd_export *tgt_exp;
1840 struct lmv_obj *obj;
1845 rc = lmv_check_connect(obd);
1852 if (op_data->op_namelen != 0) {
1855 /* Usual link request */
1856 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1858 mea_idx = raw_name2idx(obj->lo_hashtype,
1861 op_data->op_namelen);
1862 op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1863 mds = obj->lo_inodes[mea_idx].li_mds;
1866 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1871 CDEBUG(D_OTHER,"link "DFID":%*s to "DFID"\n",
1872 PFID(&op_data->op_fid2), op_data->op_namelen,
1873 op_data->op_name, PFID(&op_data->op_fid1));
1875 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds);
1879 /* request from MDS to acquire i_links for inode by fid1 */
1880 CDEBUG(D_OTHER, "inc i_nlinks for "DFID"\n",
1881 PFID(&op_data->op_fid1));
1884 CDEBUG(D_OTHER, "forward to MDS #"LPU64" ("DFID")\n",
1885 mds, PFID(&op_data->op_fid1));
1887 op_data->op_fsuid = current->fsuid;
1888 op_data->op_fsgid = current->fsgid;
1889 op_data->op_cap = current->cap_effective;
1891 tgt_exp = lmv->tgts[mds].ltd_exp;
1892 if (op_data->op_namelen) {
1893 op_data->op_flags |= MF_MDC_CANCEL_FID2;
1894 /* Cancel UPDATE lock on child (fid1). */
1895 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
1896 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1899 rc = md_link(tgt_exp, op_data, request);
1900 if (rc == -ERESTART) {
1901 LASSERT(*request != NULL);
1902 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1903 "Got -ERESTART during link!\n");
1904 ptlrpc_req_finished(*request);
1908 * Directory got split. Time to update local object and repeat
1909 * the request with proper MDS.
1911 rc = lmv_handle_split(exp, &op_data->op_fid2);
1919 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
1920 const char *old, int oldlen, const char *new, int newlen,
1921 struct ptlrpc_request **request)
1923 struct obd_export *tgt_exp = NULL, *src_exp;
1924 struct obd_device *obd = exp->exp_obd;
1925 struct lmv_obd *lmv = &obd->u.lmv;
1926 int rc, mea_idx, loop = 0;
1927 struct lmv_obj *obj;
1931 CDEBUG(D_OTHER, "rename %*s in "DFID" to %*s in "DFID"\n",
1932 oldlen, old, PFID(&op_data->op_fid1),
1933 newlen, new, PFID(&op_data->op_fid2));
1935 rc = lmv_check_connect(obd);
1941 * MDS with old dir entry is asking another MDS to create name
1945 "create %*s(%d/%d) in "DFID" pointing "
1946 "to "DFID"\n", newlen, new, oldlen, newlen,
1947 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1));
1949 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds1);
1954 * Target directory can be split, sowe should forward request to
1957 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1959 mea_idx = raw_name2idx(obj->lo_hashtype,
1961 (char *)new, newlen);
1962 op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1963 CDEBUG(D_OTHER, "Parent obj "DFID"\n",
1964 PFID(&op_data->op_fid2));
1973 obj = lmv_obj_grab(obd, &op_data->op_fid1);
1976 * directory is already split, so we have to forward request to
1979 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1980 (char *)old, oldlen);
1981 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1982 mds1 = obj->lo_inodes[mea_idx].li_mds;
1983 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
1986 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
1991 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1994 * Directory is already split, so we have to forward request to
1997 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1998 (char *)new, newlen);
2000 mds2 = obj->lo_inodes[mea_idx].li_mds;
2001 op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
2002 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2005 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2011 op_data->op_fsuid = current->fsuid;
2012 op_data->op_fsgid = current->fsgid;
2013 op_data->op_cap = current->cap_effective;
2015 src_exp = lmv_get_export(lmv, mds1);
2016 tgt_exp = lmv_get_export(lmv, mds2);
2018 /* LOOKUP lock on src child (fid3) should also be cancelled for
2019 * src_exp in mdc_rename. */
2020 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2022 /* Cancel UPDATE locks on tgt parent (fid2), tgt_exp is its
2024 rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data, LCK_EX,
2025 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2027 /* Cancel LOOKUP locks on tgt child (fid4) for parent tgt_exp.*/
2029 rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data,
2030 LCK_EX, MDS_INODELOCK_LOOKUP,
2031 MF_MDC_CANCEL_FID4);
2033 /* XXX: the case when child is a striped dir is not supported.
2034 * Only the master stripe has all locks cancelled early. */
2035 /* Cancel all the locks on tgt child (fid4). */
2037 rc = lmv_early_cancel(lmv, src_exp, NULL, op_data,
2038 LCK_EX, MDS_INODELOCK_FULL,
2039 MF_MDC_CANCEL_FID4);
2043 rc = md_rename(src_exp, op_data, old, oldlen,
2044 new, newlen, request);
2045 if (rc == -ERESTART) {
2046 LASSERT(*request != NULL);
2047 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2048 "Got -ERESTART during rename!\n");
2049 ptlrpc_req_finished(*request);
2053 * Directory got split. Time to update local object and repeat
2054 * the request with proper MDS.
2056 rc = lmv_handle_split(exp, &op_data->op_fid1);
2063 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2064 void *ea, int ealen, void *ea2, int ea2len,
2065 struct ptlrpc_request **request,
2066 struct md_open_data **mod)
2068 struct obd_device *obd = exp->exp_obd;
2069 struct lmv_obd *lmv = &obd->u.lmv;
2070 struct ptlrpc_request *req;
2071 struct obd_export *tgt_exp;
2072 struct lmv_obj *obj;
2076 rc = lmv_check_connect(obd);
2080 obj = lmv_obj_grab(obd, &op_data->op_fid1);
2082 CDEBUG(D_OTHER, "SETATTR for "DFID", valid 0x%x%s\n",
2083 PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2084 obj ? ", split" : "");
2086 op_data->op_flags |= MF_MDC_CANCEL_FID1;
2088 for (i = 0; i < obj->lo_objcount; i++) {
2089 op_data->op_fid1 = obj->lo_inodes[i].li_fid;
2091 tgt_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
2092 if (IS_ERR(tgt_exp)) {
2093 rc = PTR_ERR(tgt_exp);
2097 rc = md_setattr(tgt_exp, op_data, ea, ealen,
2098 ea2, ea2len, &req, mod);
2100 if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid)) {
2102 * this is master object and this request should
2103 * be returned back to llite.
2107 ptlrpc_req_finished(req);
2115 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2116 if (IS_ERR(tgt_exp))
2117 RETURN(PTR_ERR(tgt_exp));
2119 rc = md_setattr(tgt_exp, op_data, ea, ealen, ea2,
2120 ea2len, request, mod);
2125 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2126 struct obd_capa *oc, struct ptlrpc_request **request)
2128 struct obd_device *obd = exp->exp_obd;
2129 struct lmv_obd *lmv = &obd->u.lmv;
2130 struct obd_export *tgt_exp;
2134 rc = lmv_check_connect(obd);
2138 tgt_exp = lmv_find_export(lmv, fid);
2139 if (IS_ERR(tgt_exp))
2140 RETURN(PTR_ERR(tgt_exp));
2142 rc = md_sync(tgt_exp, fid, oc, request);
2146 /* main purpose of LMV blocking ast is to remove split directory LMV
2147 * presentation object (struct lmv_obj) attached to the lock being revoked. */
2148 int lmv_blocking_ast(struct ldlm_lock *lock,
2149 struct ldlm_lock_desc *desc,
2150 void *data, int flag)
2152 struct lustre_handle lockh;
2153 struct lmv_obj *obj;
2158 case LDLM_CB_BLOCKING:
2159 ldlm_lock2handle(lock, &lockh);
2160 rc = ldlm_cli_cancel(&lockh);
2162 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2166 case LDLM_CB_CANCELING:
2167 /* time to drop cached attrs for dirobj */
2168 obj = lock->l_ast_data;
2170 CDEBUG(D_OTHER, "cancel %s on "LPU64"/"LPU64
2171 ", master "DFID"\n",
2172 lock->l_resource->lr_name.name[3] == 1 ?
2173 "LOOKUP" : "UPDATE",
2174 lock->l_resource->lr_name.name[0],
2175 lock->l_resource->lr_name.name[1],
2176 PFID(&obj->lo_fid));
2186 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2190 val = le64_to_cpu(*hash);
2192 val += MAX_HASH_SIZE;
2193 if (val != DIR_END_OFF)
2194 *hash = cpu_to_le64(val - hash_adj);
2197 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2200 struct obd_import *imp;
2203 * XXX Hack: to get nid we assume that underlying obd device is mdc.
2205 imp = class_exp2cliimp(exp);
2206 id = imp->imp_connection->c_self + fid_flatten(fid);
2208 CDEBUG(D_INFO, "node rank: %llx "DFID" %llx %llx\n",
2209 imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2211 return id ^ (id >> 32);
2214 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2215 struct obd_capa *oc, __u64 offset64, struct page *page,
2216 struct ptlrpc_request **request)
2218 struct obd_device *obd = exp->exp_obd;
2219 struct lmv_obd *lmv = &obd->u.lmv;
2220 struct obd_export *tgt_exp;
2221 struct lu_fid rid = *fid;
2222 struct lmv_obj *obj;
2236 rc = lmv_check_connect(obd);
2240 CDEBUG(D_INFO, "READPAGE at %llx from "DFID"\n", offset, PFID(&rid));
2242 obj = lmv_obj_grab(obd, fid);
2246 * This case handle directory lookup in clustered metadata case (i.e.
2247 * split directory is located on multiple md servers.)
2248 * each server keeps directory entries for certain range of hashes.
2249 * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2250 * first server will keep records with hashes [ 0 ... MAX_HASH / N - 1],
2251 * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2253 * readdir can simply start reading entries from 0 - N server in
2254 * order but that will not scale well as all client will request dir in
2255 * to server in same order.
2256 * Following algorithm does optimization:
2257 * Instead of doing readdir in 1, 2, ...., N order, client with a
2258 * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2259 * (every client has rank R)
2260 * But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2261 * since client ask dir from MDS{R} client has pages with offsets
2262 * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2263 * on hash values that we get.
2266 struct lmv_inode *loi;
2270 nr = obj->lo_objcount;
2272 seg_size = MAX_HASH_SIZE;
2273 do_div(seg_size, nr);
2274 loi = obj->lo_inodes;
2275 rank = lmv_node_rank(lmv_get_export(lmv, loi[0].li_mds),
2278 do_div(tgt_tmp, seg_size);
2279 tgt0 = do_div(tgt_tmp, nr);
2280 tgt = (tgt0 + rank) % nr;
2286 * Last segment has unusual length due to division
2289 hash_adj = MAX_HASH_SIZE - seg_size * nr;
2293 hash_adj += rank * seg_size;
2295 CDEBUG(D_INFO, "hash_adj: %x %llx %llx/%x -> %llx/%x\n",
2296 rank, hash_adj, offset, tgt0, offset + hash_adj, tgt);
2298 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2299 rid = obj->lo_inodes[tgt].li_fid;
2300 tgt_exp = lmv_get_export(lmv, loi[tgt].li_mds);
2302 CDEBUG(D_INFO, "forward to "DFID" with offset %lu i %d\n",
2303 PFID(&rid), (unsigned long)offset, tgt);
2305 tgt_exp = lmv_find_export(lmv, &rid);
2307 if (IS_ERR(tgt_exp))
2308 GOTO(cleanup, rc = PTR_ERR(tgt_exp));
2310 rc = md_readpage(tgt_exp, &rid, oc, offset, page, request);
2314 struct lu_dirpage *dp;
2315 struct lu_dirent *ent;
2317 dp = cfs_kmap(page);
2319 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2320 lmv_hash_adjust(&dp->ldp_hash_end, hash_adj);
2321 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2323 for (ent = lu_dirent_start(dp); ent != NULL;
2324 ent = lu_dirent_next(ent))
2325 lmv_hash_adjust(&ent->lde_hash, hash_adj);
2327 if (tgt0 != nr - 1) {
2330 end = le64_to_cpu(dp->ldp_hash_end);
2331 if (end == DIR_END_OFF) {
2332 dp->ldp_hash_end = cpu_to_le32(seg_size *
2334 CDEBUG(D_INFO, ""DFID" reset end %llx tgt %d\n",
2336 le64_to_cpu(dp->ldp_hash_end), tgt);
2342 * Here we could remove "." and ".." from all pages which at not from
2343 * master. But MDS has only "." and ".." for master dir.
2348 lmv_obj_unlock(obj);
2354 static int lmv_unlink_slaves(struct obd_export *exp,
2355 struct md_op_data *op_data,
2356 struct ptlrpc_request **req)
2358 struct obd_device *obd = exp->exp_obd;
2359 struct lmv_obd *lmv = &obd->u.lmv;
2360 struct lmv_stripe_md *mea = op_data->op_mea1;
2361 struct md_op_data *op_data2;
2362 struct obd_export *tgt_exp;
2366 OBD_ALLOC_PTR(op_data2);
2367 if (op_data2 == NULL)
2370 op_data2->op_mode = S_IFDIR;
2371 op_data2->op_fsuid = current->fsuid;
2372 op_data2->op_fsgid = current->fsgid;
2373 op_data2->op_bias = 0;
2375 LASSERT(mea != NULL);
2376 for (i = 0; i < mea->mea_count; i++) {
2377 memset(op_data2, 0, sizeof(*op_data2));
2378 op_data2->op_fid1 = mea->mea_ids[i];
2379 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
2380 if (IS_ERR(tgt_exp))
2381 GOTO(out_free_op_data2, rc = PTR_ERR(tgt_exp));
2383 if (tgt_exp == NULL)
2386 rc = md_unlink(tgt_exp, op_data2, req);
2388 CDEBUG(D_OTHER, "unlink slave "DFID" -> %d\n",
2389 PFID(&mea->mea_ids[i]), rc);
2392 ptlrpc_req_finished(*req);
2396 GOTO(out_free_op_data2, rc);
2401 OBD_FREE_PTR(op_data2);
2405 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2406 struct ptlrpc_request **request)
2408 struct obd_device *obd = exp->exp_obd;
2409 struct lmv_obd *lmv = &obd->u.lmv;
2410 struct obd_export *tgt_exp = NULL;
2411 struct lmv_obj *obj;
2415 rc = lmv_check_connect(obd);
2419 if (op_data->op_namelen == 0 && op_data->op_mea1 != NULL) {
2420 /* mds asks to remove slave objects */
2421 rc = lmv_unlink_slaves(exp, op_data, request);
2428 if (op_data->op_namelen != 0) {
2431 obj = lmv_obj_grab(obd, &op_data->op_fid1);
2433 mea_idx = raw_name2idx(obj->lo_hashtype,
2436 op_data->op_namelen);
2437 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2438 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
2439 tgt_exp = lmv_get_export(lmv,
2440 obj->lo_inodes[mea_idx].li_mds);
2442 CDEBUG(D_OTHER, "unlink '%*s' in "DFID" -> %u\n",
2443 op_data->op_namelen, op_data->op_name,
2444 PFID(&op_data->op_fid1), mea_idx);
2447 CDEBUG(D_OTHER, "drop i_nlink on "DFID"\n",
2448 PFID(&op_data->op_fid1));
2450 if (tgt_exp == NULL) {
2451 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2452 if (IS_ERR(tgt_exp))
2453 RETURN(PTR_ERR(tgt_exp));
2454 op_data->op_bias |= MDS_CHECK_SPLIT;
2457 op_data->op_fsuid = current->fsuid;
2458 op_data->op_fsgid = current->fsgid;
2459 op_data->op_cap = current->cap_effective;
2461 /* If child's fid is given, cancel unused locks for it if it is from
2462 * another export than parent. */
2463 if (op_data->op_namelen) {
2464 /* LOOKUP lock for child (fid3) should also be cancelled on
2465 * parent tgt_exp in mdc_unlink(). */
2466 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2468 /* XXX: the case when child is a striped dir is not supported.
2469 * Only the master stripe has all locks cancelled early. */
2470 /* Cancel FULL locks on child (fid3). */
2471 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
2472 MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2475 rc = md_unlink(tgt_exp, op_data, request);
2476 if (rc == -ERESTART) {
2477 LASSERT(*request != NULL);
2478 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2479 "Got -ERESTART during unlink!\n");
2480 ptlrpc_req_finished(*request);
2484 * Directory got split. Time to update local object and repeat
2485 * the request with proper MDS.
2487 rc = lmv_handle_split(exp, &op_data->op_fid1);
2494 static int lmv_llog_init(struct obd_device *obd, int group,
2495 struct obd_device *tgt, int count,
2496 struct llog_catid *logid, struct obd_uuid *uuid)
2499 struct llog_ctxt *ctxt;
2503 LASSERT(group == OBD_LLOG_GROUP);
2504 rc = llog_setup(obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
2507 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_CONFIG_REPL_CTXT);
2508 llog_initiator_connect(ctxt, tgt);
2509 llog_ctxt_put(ctxt);
2517 static int lmv_llog_finish(struct obd_device *obd, int count)
2519 struct llog_ctxt *ctxt;
2523 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
2525 rc = llog_cleanup(ctxt);
2530 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2535 case OBD_CLEANUP_EARLY:
2536 /* XXX: here should be calling obd_precleanup() down to
2539 case OBD_CLEANUP_SELF_EXP:
2540 rc = obd_llog_finish(obd, 0);
2542 CERROR("failed to cleanup llogging subsystems\n");
2550 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2551 void *key, __u32 *vallen, void *val)
2553 struct obd_device *obd;
2554 struct lmv_obd *lmv;
2558 obd = class_exp2obd(exp);
2560 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2561 exp->exp_handle.h_cookie);
2566 if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2567 struct lmv_tgt_desc *tgts;
2570 rc = lmv_check_connect(obd);
2574 LASSERT(*vallen == sizeof(__u32));
2575 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2578 /* all tgts should be connected when this get called. */
2579 if (!tgts || !tgts->ltd_exp) {
2580 CERROR("target not setup?\n");
2584 if (!obd_get_info(tgts->ltd_exp, keylen, key,
2589 } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2590 rc = lmv_check_connect(obd);
2594 /* forwarding this request to first MDS, it should know LOV
2596 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2598 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2599 exp->exp_connect_flags =
2600 ((struct obd_connect_data *)val)->ocd_connect_flags;
2605 CDEBUG(D_IOCTL, "invalid key\n");
2609 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2610 void *key, obd_count vallen, void *val,
2611 struct ptlrpc_request_set *set)
2613 struct lmv_tgt_desc *tgt;
2614 struct obd_device *obd;
2615 struct lmv_obd *lmv;
2619 obd = class_exp2obd(exp);
2621 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2622 exp->exp_handle.h_cookie);
2627 if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2628 KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2631 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2632 tgt = &lmv->tgts[i];
2637 err = obd_set_info_async(tgt->ltd_exp,
2638 keylen, key, vallen, val, set);
2649 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2650 struct lov_stripe_md *lsm)
2652 struct obd_device *obd = class_exp2obd(exp);
2653 struct lmv_obd *lmv = &obd->u.lmv;
2654 struct lmv_stripe_md *meap, *lsmp;
2658 mea_size = lmv_get_easize(lmv);
2662 if (*lmmp && !lsm) {
2663 OBD_FREE(*lmmp, mea_size);
2668 if (*lmmp == NULL) {
2669 OBD_ALLOC(*lmmp, mea_size);
2677 lsmp = (struct lmv_stripe_md *)lsm;
2678 meap = (struct lmv_stripe_md *)*lmmp;
2680 if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2681 lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2684 meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2685 meap->mea_count = cpu_to_le32(lsmp->mea_count);
2686 meap->mea_master = cpu_to_le32(lsmp->mea_master);
2688 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2689 meap->mea_ids[i] = meap->mea_ids[i];
2690 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2696 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2697 struct lov_mds_md *lmm, int lmm_size)
2699 struct obd_device *obd = class_exp2obd(exp);
2700 struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2701 struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2702 struct lmv_obd *lmv = &obd->u.lmv;
2707 mea_size = lmv_get_easize(lmv);
2711 if (*lsmp != NULL && lmm == NULL) {
2712 OBD_FREE(*tmea, mea_size);
2717 LASSERT(mea_size == lmm_size);
2719 OBD_ALLOC(*tmea, mea_size);
2726 if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2727 mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2728 mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2730 magic = le32_to_cpu(mea->mea_magic);
2732 /* old mea is not handled here */
2736 (*tmea)->mea_magic = magic;
2737 (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2738 (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2740 for (i = 0; i < (*tmea)->mea_count; i++) {
2741 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2742 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2747 static int lmv_cancel_unused(struct obd_export *exp,
2748 const struct lu_fid *fid,
2749 ldlm_policy_data_t *policy,
2750 ldlm_mode_t mode, int flags, void *opaque)
2752 struct obd_device *obd = exp->exp_obd;
2753 struct lmv_obd *lmv = &obd->u.lmv;
2757 LASSERT(fid != NULL);
2759 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2760 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2763 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2764 policy, mode, flags, opaque);
2771 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data)
2773 struct obd_device *obd = exp->exp_obd;
2774 struct lmv_obd *lmv = &obd->u.lmv;
2777 RETURN(md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data));
2780 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2781 const struct lu_fid *fid, ldlm_type_t type,
2782 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2783 struct lustre_handle *lockh)
2785 struct obd_device *obd = exp->exp_obd;
2786 struct lmv_obd *lmv = &obd->u.lmv;
2791 CDEBUG(D_OTHER, "lock match for "DFID"\n", PFID(fid));
2793 /* with CMD every object can have two locks in different namespaces:
2794 * lookup lock in space of mds storing direntry and update/open lock in
2795 * space of mds storing inode. Thus we check all targets, not only that
2796 * one fid was created in. */
2797 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2798 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2799 type, policy, mode, lockh);
2807 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2808 struct obd_export *dt_exp, struct obd_export *md_exp,
2809 struct lustre_md *md)
2811 struct obd_device *obd = exp->exp_obd;
2812 struct lmv_obd *lmv = &obd->u.lmv;
2816 rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2820 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2822 struct obd_device *obd = exp->exp_obd;
2823 struct lmv_obd *lmv = &obd->u.lmv;
2827 obd_free_memmd(exp, (struct lov_stripe_md**)&md->mea);
2828 RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2831 int lmv_set_open_replay_data(struct obd_export *exp,
2832 struct obd_client_handle *och,
2833 struct ptlrpc_request *open_req)
2835 struct obd_device *obd = exp->exp_obd;
2836 struct lmv_obd *lmv = &obd->u.lmv;
2837 struct obd_export *tgt_exp;
2841 tgt_exp = lmv_find_export(lmv, &och->och_fid);
2842 if (IS_ERR(tgt_exp))
2843 RETURN(PTR_ERR(tgt_exp));
2845 RETURN(md_set_open_replay_data(tgt_exp, och, open_req));
2848 int lmv_clear_open_replay_data(struct obd_export *exp,
2849 struct obd_client_handle *och)
2851 struct obd_device *obd = exp->exp_obd;
2852 struct lmv_obd *lmv = &obd->u.lmv;
2853 struct obd_export *tgt_exp;
2856 tgt_exp = lmv_find_export(lmv, &och->och_fid);
2857 if (IS_ERR(tgt_exp))
2858 RETURN(PTR_ERR(tgt_exp));
2860 RETURN(md_clear_open_replay_data(tgt_exp, och));
2863 static int lmv_get_remote_perm(struct obd_export *exp,
2864 const struct lu_fid *fid,
2865 struct obd_capa *oc, __u32 suppgid,
2866 struct ptlrpc_request **request)
2868 struct obd_device *obd = exp->exp_obd;
2869 struct lmv_obd *lmv = &obd->u.lmv;
2870 struct obd_export *tgt_exp;
2875 rc = lmv_check_connect(obd);
2879 tgt_exp = lmv_find_export(lmv, fid);
2880 if (IS_ERR(tgt_exp))
2881 RETURN(PTR_ERR(tgt_exp));
2883 rc = md_get_remote_perm(tgt_exp, fid, oc, suppgid, request);
2888 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2891 struct obd_device *obd = exp->exp_obd;
2892 struct lmv_obd *lmv = &obd->u.lmv;
2893 struct obd_export *tgt_exp;
2897 rc = lmv_check_connect(obd);
2901 tgt_exp = lmv_find_export(lmv, &oc->c_capa.lc_fid);
2902 if (IS_ERR(tgt_exp))
2903 RETURN(PTR_ERR(tgt_exp));
2905 rc = md_renew_capa(tgt_exp, oc, cb);
2909 struct obd_ops lmv_obd_ops = {
2910 .o_owner = THIS_MODULE,
2911 .o_setup = lmv_setup,
2912 .o_cleanup = lmv_cleanup,
2913 .o_precleanup = lmv_precleanup,
2914 .o_process_config = lmv_process_config,
2915 .o_connect = lmv_connect,
2916 .o_disconnect = lmv_disconnect,
2917 .o_statfs = lmv_statfs,
2918 .o_llog_init = lmv_llog_init,
2919 .o_llog_finish = lmv_llog_finish,
2920 .o_get_info = lmv_get_info,
2921 .o_set_info_async = lmv_set_info_async,
2922 .o_packmd = lmv_packmd,
2923 .o_unpackmd = lmv_unpackmd,
2924 .o_notify = lmv_notify,
2925 .o_iocontrol = lmv_iocontrol,
2926 .o_fid_delete = lmv_fid_delete
2929 struct md_ops lmv_md_ops = {
2930 .m_getstatus = lmv_getstatus,
2931 .m_change_cbdata = lmv_change_cbdata,
2932 .m_close = lmv_close,
2933 .m_create = lmv_create,
2934 .m_done_writing = lmv_done_writing,
2935 .m_enqueue = lmv_enqueue,
2936 .m_getattr = lmv_getattr,
2937 .m_getxattr = lmv_getxattr,
2938 .m_getattr_name = lmv_getattr_name,
2939 .m_intent_lock = lmv_intent_lock,
2941 .m_rename = lmv_rename,
2942 .m_setattr = lmv_setattr,
2943 .m_setxattr = lmv_setxattr,
2945 .m_readpage = lmv_readpage,
2946 .m_unlink = lmv_unlink,
2947 .m_init_ea_size = lmv_init_ea_size,
2948 .m_cancel_unused = lmv_cancel_unused,
2949 .m_set_lock_data = lmv_set_lock_data,
2950 .m_lock_match = lmv_lock_match,
2951 .m_get_lustre_md = lmv_get_lustre_md,
2952 .m_free_lustre_md = lmv_free_lustre_md,
2953 .m_set_open_replay_data = lmv_set_open_replay_data,
2954 .m_clear_open_replay_data = lmv_clear_open_replay_data,
2955 .m_get_remote_perm = lmv_get_remote_perm,
2956 .m_renew_capa = lmv_renew_capa
2959 int __init lmv_init(void)
2961 struct lprocfs_static_vars lvars;
2964 obj_cache = cfs_mem_cache_create("lmv_objects",
2965 sizeof(struct lmv_obj),
2968 CERROR("error allocating lmv objects cache\n");
2972 lprocfs_lmv_init_vars(&lvars);
2973 rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
2974 lvars.module_vars, LUSTRE_LMV_NAME, NULL);
2976 cfs_mem_cache_destroy(obj_cache);
2982 static void lmv_exit(void)
2986 class_unregister_type(LUSTRE_LMV_NAME);
2988 rc = cfs_mem_cache_destroy(obj_cache);
2990 "can't free lmv objects cache, %d object(s)"
2991 "still in use\n", atomic_read(&obj_cache_count));
2994 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2995 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
2996 MODULE_LICENSE("GPL");
2998 module_init(lmv_init);
2999 module_exit(lmv_exit);