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, sizeof(KEY_INTERMDS),
301 KEY_INTERMDS, 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->obd_olg, 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 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1013 0444, &lmv_proc_target_fops, obd);
1015 CWARN("Error adding the target_obd_status file\n");
1018 rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1019 LUSTRE_CLI_FLD_HASH_DHT);
1021 CERROR("can't init FLD, err %d\n",
1023 GOTO(out_free_datas, rc);
1029 OBD_FREE(lmv->datas, lmv->datas_size);
1032 OBD_FREE(lmv->tgts, lmv->tgts_size);
1037 static int lmv_cleanup(struct obd_device *obd)
1039 struct lmv_obd *lmv = &obd->u.lmv;
1042 fld_client_fini(&lmv->lmv_fld);
1043 lprocfs_obd_cleanup(obd);
1044 lmv_obj_cleanup(obd);
1045 OBD_FREE(lmv->datas, lmv->datas_size);
1046 OBD_FREE(lmv->tgts, lmv->tgts_size);
1051 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1053 struct lustre_cfg *lcfg = buf;
1054 struct obd_uuid tgt_uuid;
1058 switch(lcfg->lcfg_command) {
1060 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1061 GOTO(out, rc = -EINVAL);
1063 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1064 rc = lmv_add_target(obd, &tgt_uuid);
1067 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1068 GOTO(out, rc = -EINVAL);
1075 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1076 __u64 max_age, __u32 flags)
1078 struct lmv_obd *lmv = &obd->u.lmv;
1079 struct obd_statfs *temp;
1083 rc = lmv_check_connect(obd);
1087 OBD_ALLOC(temp, sizeof(*temp));
1091 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1092 if (lmv->tgts[i].ltd_exp == NULL)
1095 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1098 CERROR("can't stat MDS #%d (%s), error %d\n", i,
1099 lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1101 GOTO(out_free_temp, rc);
1106 osfs->os_bavail += temp->os_bavail;
1107 osfs->os_blocks += temp->os_blocks;
1108 osfs->os_ffree += temp->os_ffree;
1109 osfs->os_files += temp->os_files;
1115 OBD_FREE(temp, sizeof(*temp));
1119 static int lmv_getstatus(struct obd_export *exp,
1121 struct obd_capa **pc)
1123 struct obd_device *obd = exp->exp_obd;
1124 struct lmv_obd *lmv = &obd->u.lmv;
1128 rc = lmv_check_connect(obd);
1132 rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1137 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1138 struct obd_capa *oc, obd_valid valid, const char *name,
1139 const char *input, int input_size, int output_size,
1140 int flags, struct ptlrpc_request **request)
1142 struct obd_device *obd = exp->exp_obd;
1143 struct lmv_obd *lmv = &obd->u.lmv;
1144 struct obd_export *tgt_exp;
1148 rc = lmv_check_connect(obd);
1152 tgt_exp = lmv_find_export(lmv, fid);
1153 if (IS_ERR(tgt_exp))
1154 RETURN(PTR_ERR(tgt_exp));
1156 rc = md_getxattr(tgt_exp, fid, oc, valid, name, input, input_size,
1157 output_size, flags, request);
1162 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1163 struct obd_capa *oc, obd_valid valid, const char *name,
1164 const char *input, int input_size, int output_size,
1165 int flags, __u32 suppgid,
1166 struct ptlrpc_request **request)
1168 struct obd_device *obd = exp->exp_obd;
1169 struct lmv_obd *lmv = &obd->u.lmv;
1170 struct obd_export *tgt_exp;
1174 rc = lmv_check_connect(obd);
1178 tgt_exp = lmv_find_export(lmv, fid);
1179 if (IS_ERR(tgt_exp))
1180 RETURN(PTR_ERR(tgt_exp));
1182 rc = md_setxattr(tgt_exp, fid, oc, valid, name,
1183 input, input_size, output_size, flags, suppgid,
1189 static int lmv_getattr(struct obd_export *exp, const struct lu_fid *fid,
1190 struct obd_capa *oc, obd_valid valid, int ea_size,
1191 struct ptlrpc_request **request)
1193 struct obd_device *obd = exp->exp_obd;
1194 struct lmv_obd *lmv = &obd->u.lmv;
1195 struct obd_export *tgt_exp;
1196 struct lmv_obj *obj;
1200 rc = lmv_check_connect(obd);
1204 tgt_exp = lmv_find_export(lmv, fid);
1205 if (IS_ERR(tgt_exp))
1206 RETURN(PTR_ERR(tgt_exp));
1208 rc = md_getattr(tgt_exp, fid, oc, valid, ea_size, request);
1212 obj = lmv_obj_grab(obd, fid);
1214 CDEBUG(D_OTHER, "GETATTR for "DFID" %s\n", PFID(fid),
1215 obj ? "(split)" : "");
1218 * If object is split, then we loop over all the slaves and gather size
1219 * attribute. In ideal world we would have to gather also mds field from
1220 * all slaves, as object is spread over the cluster and this is
1221 * definitely interesting information and it is not good to loss it,
1225 struct mdt_body *body;
1227 if (*request == NULL) {
1232 body = req_capsule_server_get(&(*request)->rq_pill,
1234 LASSERT(body != NULL);
1238 for (i = 0; i < obj->lo_objcount; i++) {
1239 if (lmv->tgts[i].ltd_exp == NULL) {
1240 CWARN("%s: NULL export for %d\n",
1245 /* skip master obj. */
1246 if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid))
1249 lmv_update_body(body, &obj->lo_inodes[i]);
1252 lmv_obj_unlock(obj);
1259 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1260 ldlm_iterator_t it, void *data)
1262 struct obd_device *obd = exp->exp_obd;
1263 struct lmv_obd *lmv = &obd->u.lmv;
1267 rc = lmv_check_connect(obd);
1271 CDEBUG(D_OTHER, "CBDATA for "DFID"\n", PFID(fid));
1274 * With CMD every object can have two locks in different namespaces:
1275 * lookup lock in space of mds storing direntry and update/open lock in
1276 * space of mds storing inode.
1278 for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1279 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1284 static int lmv_close(struct obd_export *exp,
1285 struct md_op_data *op_data,
1286 struct md_open_data *mod,
1287 struct ptlrpc_request **request)
1289 struct obd_device *obd = exp->exp_obd;
1290 struct lmv_obd *lmv = &obd->u.lmv;
1291 struct obd_export *tgt_exp;
1295 rc = lmv_check_connect(obd);
1299 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1300 if (IS_ERR(tgt_exp))
1301 RETURN(PTR_ERR(tgt_exp));
1303 CDEBUG(D_OTHER, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1304 rc = md_close(tgt_exp, op_data, mod, request);
1309 * Called in the case MDS returns -ERESTART on create on open, what means that
1310 * directory is split and its LMV presentation object has to be updated.
1312 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1314 struct obd_device *obd = exp->exp_obd;
1315 struct lmv_obd *lmv = &obd->u.lmv;
1316 struct ptlrpc_request *req = NULL;
1317 struct obd_export *tgt_exp;
1318 struct lmv_obj *obj;
1319 struct lustre_md md;
1325 mealen = lmv_get_easize(lmv);
1327 valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1329 tgt_exp = lmv_find_export(lmv, fid);
1330 if (IS_ERR(tgt_exp))
1331 RETURN(PTR_ERR(tgt_exp));
1333 /* time to update mea of parent fid */
1334 rc = md_getattr(tgt_exp, fid, NULL, valid, mealen, &req);
1336 CERROR("md_getattr() failed, error %d\n", rc);
1340 rc = md_get_lustre_md(tgt_exp, req, NULL, exp, &md);
1342 CERROR("mdc_get_lustre_md() failed, error %d\n", rc);
1347 GOTO(cleanup, rc = -ENODATA);
1349 obj = lmv_obj_create(exp, fid, md.mea);
1355 /* XXX LOV STACKING */
1356 obd_free_memmd(exp, (void *)&md.mea);
1361 ptlrpc_req_finished(req);
1365 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1366 const void *data, int datalen, int mode, __u32 uid,
1367 __u32 gid, __u32 cap_effective, __u64 rdev,
1368 struct ptlrpc_request **request)
1370 struct obd_device *obd = exp->exp_obd;
1371 struct lmv_obd *lmv = &obd->u.lmv;
1372 struct obd_export *tgt_exp;
1373 struct lmv_obj *obj;
1377 rc = lmv_check_connect(obd);
1381 if (!lmv->desc.ld_active_tgt_count)
1386 obj = lmv_obj_grab(obd, &op_data->op_fid1);
1390 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1391 op_data->op_name, op_data->op_namelen);
1392 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1393 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1394 op_data->op_mds = obj->lo_inodes[mea_idx].li_mds;
1395 tgt_exp = lmv_get_export(lmv, op_data->op_mds);
1398 struct lmv_tgt_desc *tgt;
1400 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1401 op_data->op_bias |= MDS_CHECK_SPLIT;
1402 op_data->op_mds = tgt->ltd_idx;
1403 tgt_exp = tgt->ltd_exp;
1406 if (IS_ERR(tgt_exp))
1407 RETURN(PTR_ERR(tgt_exp));
1409 rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1410 if (rc == -ERESTART)
1415 CDEBUG(D_OTHER, "CREATE '%*s' on "DFID"\n", op_data->op_namelen,
1416 op_data->op_name, PFID(&op_data->op_fid1));
1418 op_data->op_flags |= MF_MDC_CANCEL_FID1;
1419 rc = md_create(tgt_exp, op_data, data, datalen, mode, uid, gid,
1420 cap_effective, rdev, request);
1422 if (*request == NULL)
1424 CDEBUG(D_OTHER, "created - "DFID"\n", PFID(&op_data->op_fid1));
1425 } else if (rc == -ERESTART) {
1426 LASSERT(*request != NULL);
1427 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1428 "Got -ERESTART during create!\n");
1429 ptlrpc_req_finished(*request);
1433 * Directory got split. Time to update local object and repeat
1434 * the request with proper MDS.
1436 rc = lmv_handle_split(exp, &op_data->op_fid1);
1438 rc = lmv_alloc_slave_fids(obd, &op_data->op_fid1,
1439 op_data, &op_data->op_fid2);
1448 static int lmv_done_writing(struct obd_export *exp,
1449 struct md_op_data *op_data,
1450 struct md_open_data *mod)
1452 struct obd_device *obd = exp->exp_obd;
1453 struct lmv_obd *lmv = &obd->u.lmv;
1454 struct obd_export *tgt_exp;
1458 rc = lmv_check_connect(obd);
1462 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1463 if (IS_ERR(tgt_exp))
1464 RETURN(PTR_ERR(tgt_exp));
1466 rc = md_done_writing(tgt_exp, op_data, mod);
1471 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1472 struct lookup_intent *it, struct md_op_data *op_data,
1473 struct lustre_handle *lockh, void *lmm, int lmmsize)
1475 struct obd_device *obd = exp->exp_obd;
1476 struct lmv_obd *lmv = &obd->u.lmv;
1477 struct lmv_stripe_md *mea = op_data->op_mea1;
1478 struct md_op_data *op_data2;
1479 struct obd_export *tgt_exp;
1483 OBD_ALLOC_PTR(op_data2);
1484 if (op_data2 == NULL)
1487 LASSERT(mea != NULL);
1488 for (i = 0; i < mea->mea_count; i++) {
1489 memset(op_data2, 0, sizeof(*op_data2));
1490 op_data2->op_fid1 = mea->mea_ids[i];
1491 op_data2->op_bias = 0;
1493 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
1494 if (IS_ERR(tgt_exp))
1495 GOTO(cleanup, rc = PTR_ERR(tgt_exp));
1497 if (tgt_exp == NULL)
1500 rc = md_enqueue(tgt_exp, einfo, it, op_data2,
1501 lockh + i, lmm, lmmsize, 0);
1503 CDEBUG(D_OTHER, "take lock on slave "DFID" -> %d/%d\n",
1504 PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1509 if (it->d.lustre.it_data) {
1510 struct ptlrpc_request *req;
1511 req = (struct ptlrpc_request *)it->d.lustre.it_data;
1512 ptlrpc_req_finished(req);
1515 if (it->d.lustre.it_status)
1516 GOTO(cleanup, rc = it->d.lustre.it_status);
1521 OBD_FREE_PTR(op_data2);
1524 /* drop all taken locks */
1526 if (lockh[i].cookie)
1527 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1528 lockh[i].cookie = 0;
1535 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1536 struct lookup_intent *it, struct md_op_data *op_data,
1537 struct lustre_handle *lockh, void *lmm, int lmmsize,
1538 int extra_lock_flags)
1540 struct ptlrpc_request *req = it->d.lustre.it_data;
1541 struct obd_device *obd = exp->exp_obd;
1542 struct lmv_obd *lmv = &obd->u.lmv;
1543 struct lustre_handle plock;
1544 struct obd_export *tgt_exp;
1545 struct md_op_data *rdata;
1546 struct lu_fid fid_copy;
1547 struct mdt_body *body;
1551 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1552 LASSERT(body != NULL);
1554 if (!(body->valid & OBD_MD_MDS))
1557 CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID" -> "DFID"\n",
1558 LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1560 /* We got LOOKUP lock, but we really need attrs */
1561 pmode = it->d.lustre.it_lock_mode;
1562 LASSERT(pmode != 0);
1563 memcpy(&plock, lockh, sizeof(plock));
1564 it->d.lustre.it_lock_mode = 0;
1565 it->d.lustre.it_data = NULL;
1566 fid_copy = body->fid1;
1568 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1569 ptlrpc_req_finished(req);
1571 tgt_exp = lmv_find_export(lmv, &fid_copy);
1572 if (IS_ERR(tgt_exp))
1573 GOTO(out, rc = PTR_ERR(tgt_exp));
1575 OBD_ALLOC_PTR(rdata);
1577 GOTO(out, rc = -ENOMEM);
1579 rdata->op_fid1 = fid_copy;
1580 rdata->op_bias = MDS_CROSS_REF;
1582 rc = md_enqueue(tgt_exp, einfo, it, rdata, lockh,
1583 lmm, lmmsize, extra_lock_flags);
1584 OBD_FREE_PTR(rdata);
1587 ldlm_lock_decref(&plock, pmode);
1592 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1593 struct lookup_intent *it, struct md_op_data *op_data,
1594 struct lustre_handle *lockh, void *lmm, int lmmsize,
1595 int extra_lock_flags)
1597 struct obd_device *obd = exp->exp_obd;
1598 struct lmv_obd *lmv = &obd->u.lmv;
1599 struct obd_export *tgt_exp = NULL;
1600 struct lmv_obj *obj;
1604 rc = lmv_check_connect(obd);
1608 if (op_data->op_mea1 && it->it_op == IT_UNLINK) {
1609 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1610 lockh, lmm, lmmsize);
1614 if (op_data->op_namelen) {
1615 obj = lmv_obj_grab(obd, &op_data->op_fid1);
1619 /* directory is split. look for right mds for this
1621 mea_idx = raw_name2idx(obj->lo_hashtype,
1623 (char *)op_data->op_name,
1624 op_data->op_namelen);
1625 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1626 tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1631 if (tgt_exp == NULL)
1632 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1633 if (IS_ERR(tgt_exp))
1634 RETURN(PTR_ERR(tgt_exp));
1636 CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID"\n", LL_IT2STR(it),
1637 PFID(&op_data->op_fid1));
1639 rc = md_enqueue(tgt_exp, einfo, it, op_data, lockh,
1640 lmm, lmmsize, extra_lock_flags);
1642 if (rc == 0 && it->it_op == IT_OPEN)
1643 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1644 lmm, lmmsize, extra_lock_flags);
1649 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1650 struct obd_capa *oc, const char *filename, int namelen,
1651 obd_valid valid, int ea_size, __u32 suppgid,
1652 struct ptlrpc_request **request)
1654 struct obd_device *obd = exp->exp_obd;
1655 struct lmv_obd *lmv = &obd->u.lmv;
1656 struct lu_fid rid = *fid;
1657 struct obd_export *tgt_exp;
1658 struct mdt_body *body;
1659 struct lmv_obj *obj;
1663 rc = lmv_check_connect(obd);
1670 obj = lmv_obj_grab(obd, &rid);
1674 /* Directory is split. Look for right mds for this name */
1675 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1676 filename, namelen - 1);
1677 rid = obj->lo_inodes[mea_idx].li_fid;
1678 tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1680 valid &= ~OBD_MD_FLCKSPLIT;
1682 tgt_exp = lmv_find_export(lmv, &rid);
1683 valid |= OBD_MD_FLCKSPLIT;
1685 if (IS_ERR(tgt_exp))
1686 RETURN(PTR_ERR(tgt_exp));
1688 CDEBUG(D_OTHER, "getattr_name for %*s on "DFID" -> "DFID"\n",
1689 namelen, filename, PFID(fid), PFID(&rid));
1691 rc = md_getattr_name(tgt_exp, &rid, oc, filename, namelen, valid,
1692 ea_size, suppgid, request);
1694 body = req_capsule_server_get(&(*request)->rq_pill,
1696 LASSERT(body != NULL);
1698 if (body->valid & OBD_MD_MDS) {
1699 struct ptlrpc_request *req = NULL;
1702 CDEBUG(D_OTHER, "request attrs for "DFID"\n",
1705 tgt_exp = lmv_find_export(lmv, &rid);
1706 if (IS_ERR(tgt_exp)) {
1707 ptlrpc_req_finished(*request);
1708 RETURN(PTR_ERR(tgt_exp));
1711 rc = md_getattr_name(tgt_exp, &rid, NULL, NULL, 1,
1712 valid | OBD_MD_FLCROSSREF,
1713 ea_size, suppgid, &req);
1714 ptlrpc_req_finished(*request);
1717 } else if (rc == -ERESTART) {
1718 LASSERT(*request != NULL);
1719 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1720 "Got -ERESTART during getattr!\n");
1721 ptlrpc_req_finished(*request);
1725 * Directory got split. Time to update local object and repeat
1726 * the request with proper MDS.
1728 rc = lmv_handle_split(exp, &rid);
1735 #define md_op_data_fid(op_data, fl) \
1736 (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1737 fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1738 fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1739 fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1742 /* @tgt_exp is the export the metadata request is sent.
1743 * @fid_exp is the export the cancel should be sent for the current fid.
1744 * if @fid_exp is NULL, the export is found for the current fid.
1745 * @op_data keeps the current fid, which is pointed through @flag.
1746 * @mode, @bits -- lock match parameters. */
1747 static int lmv_early_cancel(struct lmv_obd *lmv, struct obd_export *tgt_exp,
1748 struct obd_export *fid_exp,
1749 struct md_op_data *op_data,
1750 ldlm_mode_t mode, int bits, int flag)
1752 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1753 ldlm_policy_data_t policy = {{0}};
1757 if (!fid_is_sane(fid))
1760 if (fid_exp == NULL)
1761 fid_exp = lmv_find_export(lmv, fid);
1763 if (tgt_exp == fid_exp) {
1764 /* The export is the same as on the target server, cancel
1765 * will be sent along with the main metadata operation. */
1766 op_data->op_flags |= flag;
1770 policy.l_inodebits.bits = bits;
1771 rc = md_cancel_unused(fid_exp, fid, &policy, mode, LDLM_FL_ASYNC, NULL);
1775 #ifdef EARLY_CANCEL_FOR_STRIPED_DIR_IS_READY
1776 /* Check if the fid in @op_data pointed to by flag is of the same export(s)
1777 * as @tgt_exp. Early cancels will be sent later by mdc code, otherwise, call
1778 * md_cancel_unused for child export(s). */
1779 static int lmv_early_cancel_stripes(struct obd_export *exp,
1780 struct obd_export *tgt_exp,
1781 struct md_op_data *op_data,
1782 ldlm_mode_t mode, int bits, int flag)
1784 struct lu_fid *fid = md_op_data_fid(op_data, flag);
1785 struct obd_device *obd = exp->exp_obd;
1786 struct lmv_obd *lmv = &obd->u.lmv;
1787 struct obd_export *st_exp;
1788 struct lmv_obj *obj;
1792 if (!fid_is_sane(fid))
1795 obj = lmv_obj_grab(obd, fid);
1797 ldlm_policy_data_t policy = {{0}};
1798 struct lu_fid *st_fid;
1801 policy.l_inodebits.bits = bits;
1802 for (i = 0; i < obj->lo_objcount; i++) {
1803 st_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
1804 st_fid = &obj->lo_inodes[i].li_fid;
1805 if (tgt_exp != st_exp) {
1806 rc = md_cancel_unused(st_exp, st_fid, &policy,
1807 mode, LDLM_FL_ASYNC,
1812 /* Some export matches to @tgt_exp, do cancel
1813 * for its fid in mdc */
1815 op_data->op_flags |= flag;
1820 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data,
1828 * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1831 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1832 struct ptlrpc_request **request)
1834 struct obd_device *obd = exp->exp_obd;
1835 struct lmv_obd *lmv = &obd->u.lmv;
1836 struct obd_export *tgt_exp;
1837 struct lmv_obj *obj;
1842 rc = lmv_check_connect(obd);
1849 if (op_data->op_namelen != 0) {
1852 /* Usual link request */
1853 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1855 mea_idx = raw_name2idx(obj->lo_hashtype,
1858 op_data->op_namelen);
1859 op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1860 mds = obj->lo_inodes[mea_idx].li_mds;
1863 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1868 CDEBUG(D_OTHER,"link "DFID":%*s to "DFID"\n",
1869 PFID(&op_data->op_fid2), op_data->op_namelen,
1870 op_data->op_name, PFID(&op_data->op_fid1));
1872 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds);
1876 /* request from MDS to acquire i_links for inode by fid1 */
1877 CDEBUG(D_OTHER, "inc i_nlinks for "DFID"\n",
1878 PFID(&op_data->op_fid1));
1881 CDEBUG(D_OTHER, "forward to MDS #"LPU64" ("DFID")\n",
1882 mds, PFID(&op_data->op_fid1));
1884 op_data->op_fsuid = current->fsuid;
1885 op_data->op_fsgid = current->fsgid;
1886 op_data->op_cap = current->cap_effective;
1888 tgt_exp = lmv->tgts[mds].ltd_exp;
1889 if (op_data->op_namelen) {
1890 op_data->op_flags |= MF_MDC_CANCEL_FID2;
1891 /* Cancel UPDATE lock on child (fid1). */
1892 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
1893 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1896 rc = md_link(tgt_exp, op_data, request);
1897 if (rc == -ERESTART) {
1898 LASSERT(*request != NULL);
1899 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1900 "Got -ERESTART during link!\n");
1901 ptlrpc_req_finished(*request);
1905 * Directory got split. Time to update local object and repeat
1906 * the request with proper MDS.
1908 rc = lmv_handle_split(exp, &op_data->op_fid2);
1916 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
1917 const char *old, int oldlen, const char *new, int newlen,
1918 struct ptlrpc_request **request)
1920 struct obd_export *tgt_exp = NULL, *src_exp;
1921 struct obd_device *obd = exp->exp_obd;
1922 struct lmv_obd *lmv = &obd->u.lmv;
1923 int rc, mea_idx, loop = 0;
1924 struct lmv_obj *obj;
1928 CDEBUG(D_OTHER, "rename %*s in "DFID" to %*s in "DFID"\n",
1929 oldlen, old, PFID(&op_data->op_fid1),
1930 newlen, new, PFID(&op_data->op_fid2));
1932 rc = lmv_check_connect(obd);
1938 * MDS with old dir entry is asking another MDS to create name
1942 "create %*s(%d/%d) in "DFID" pointing "
1943 "to "DFID"\n", newlen, new, oldlen, newlen,
1944 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1));
1946 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds1);
1951 * Target directory can be split, sowe should forward request to
1954 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1956 mea_idx = raw_name2idx(obj->lo_hashtype,
1958 (char *)new, newlen);
1959 op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1960 CDEBUG(D_OTHER, "Parent obj "DFID"\n",
1961 PFID(&op_data->op_fid2));
1970 obj = lmv_obj_grab(obd, &op_data->op_fid1);
1973 * directory is already split, so we have to forward request to
1976 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1977 (char *)old, oldlen);
1978 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1979 mds1 = obj->lo_inodes[mea_idx].li_mds;
1980 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
1983 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
1988 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1991 * Directory is already split, so we have to forward request to
1994 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1995 (char *)new, newlen);
1997 mds2 = obj->lo_inodes[mea_idx].li_mds;
1998 op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1999 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2002 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2008 op_data->op_fsuid = current->fsuid;
2009 op_data->op_fsgid = current->fsgid;
2010 op_data->op_cap = current->cap_effective;
2012 src_exp = lmv_get_export(lmv, mds1);
2013 tgt_exp = lmv_get_export(lmv, mds2);
2015 /* LOOKUP lock on src child (fid3) should also be cancelled for
2016 * src_exp in mdc_rename. */
2017 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2019 /* Cancel UPDATE locks on tgt parent (fid2), tgt_exp is its
2021 rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data, LCK_EX,
2022 MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2024 /* Cancel LOOKUP locks on tgt child (fid4) for parent tgt_exp.*/
2026 rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data,
2027 LCK_EX, MDS_INODELOCK_LOOKUP,
2028 MF_MDC_CANCEL_FID4);
2030 /* XXX: the case when child is a striped dir is not supported.
2031 * Only the master stripe has all locks cancelled early. */
2032 /* Cancel all the locks on tgt child (fid4). */
2034 rc = lmv_early_cancel(lmv, src_exp, NULL, op_data,
2035 LCK_EX, MDS_INODELOCK_FULL,
2036 MF_MDC_CANCEL_FID4);
2040 rc = md_rename(src_exp, op_data, old, oldlen,
2041 new, newlen, request);
2042 if (rc == -ERESTART) {
2043 LASSERT(*request != NULL);
2044 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2045 "Got -ERESTART during rename!\n");
2046 ptlrpc_req_finished(*request);
2050 * Directory got split. Time to update local object and repeat
2051 * the request with proper MDS.
2053 rc = lmv_handle_split(exp, &op_data->op_fid1);
2060 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2061 void *ea, int ealen, void *ea2, int ea2len,
2062 struct ptlrpc_request **request,
2063 struct md_open_data **mod)
2065 struct obd_device *obd = exp->exp_obd;
2066 struct lmv_obd *lmv = &obd->u.lmv;
2067 struct ptlrpc_request *req;
2068 struct obd_export *tgt_exp;
2069 struct lmv_obj *obj;
2073 rc = lmv_check_connect(obd);
2077 obj = lmv_obj_grab(obd, &op_data->op_fid1);
2079 CDEBUG(D_OTHER, "SETATTR for "DFID", valid 0x%x%s\n",
2080 PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2081 obj ? ", split" : "");
2083 op_data->op_flags |= MF_MDC_CANCEL_FID1;
2085 for (i = 0; i < obj->lo_objcount; i++) {
2086 op_data->op_fid1 = obj->lo_inodes[i].li_fid;
2088 tgt_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
2089 if (IS_ERR(tgt_exp)) {
2090 rc = PTR_ERR(tgt_exp);
2094 rc = md_setattr(tgt_exp, op_data, ea, ealen,
2095 ea2, ea2len, &req, mod);
2097 if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid)) {
2099 * this is master object and this request should
2100 * be returned back to llite.
2104 ptlrpc_req_finished(req);
2112 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2113 if (IS_ERR(tgt_exp))
2114 RETURN(PTR_ERR(tgt_exp));
2116 rc = md_setattr(tgt_exp, op_data, ea, ealen, ea2,
2117 ea2len, request, mod);
2122 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2123 struct obd_capa *oc, struct ptlrpc_request **request)
2125 struct obd_device *obd = exp->exp_obd;
2126 struct lmv_obd *lmv = &obd->u.lmv;
2127 struct obd_export *tgt_exp;
2131 rc = lmv_check_connect(obd);
2135 tgt_exp = lmv_find_export(lmv, fid);
2136 if (IS_ERR(tgt_exp))
2137 RETURN(PTR_ERR(tgt_exp));
2139 rc = md_sync(tgt_exp, fid, oc, request);
2143 /* main purpose of LMV blocking ast is to remove split directory LMV
2144 * presentation object (struct lmv_obj) attached to the lock being revoked. */
2145 int lmv_blocking_ast(struct ldlm_lock *lock,
2146 struct ldlm_lock_desc *desc,
2147 void *data, int flag)
2149 struct lustre_handle lockh;
2150 struct lmv_obj *obj;
2155 case LDLM_CB_BLOCKING:
2156 ldlm_lock2handle(lock, &lockh);
2157 rc = ldlm_cli_cancel(&lockh);
2159 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2163 case LDLM_CB_CANCELING:
2164 /* time to drop cached attrs for dirobj */
2165 obj = lock->l_ast_data;
2167 CDEBUG(D_OTHER, "cancel %s on "LPU64"/"LPU64
2168 ", master "DFID"\n",
2169 lock->l_resource->lr_name.name[3] == 1 ?
2170 "LOOKUP" : "UPDATE",
2171 lock->l_resource->lr_name.name[0],
2172 lock->l_resource->lr_name.name[1],
2173 PFID(&obj->lo_fid));
2183 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2187 val = le64_to_cpu(*hash);
2189 val += MAX_HASH_SIZE;
2190 if (val != DIR_END_OFF)
2191 *hash = cpu_to_le64(val - hash_adj);
2194 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2197 struct obd_import *imp;
2200 * XXX Hack: to get nid we assume that underlying obd device is mdc.
2202 imp = class_exp2cliimp(exp);
2203 id = imp->imp_connection->c_self + fid_flatten(fid);
2205 CDEBUG(D_INFO, "node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2206 imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2208 return id ^ (id >> 32);
2211 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2212 struct obd_capa *oc, __u64 offset64, struct page *page,
2213 struct ptlrpc_request **request)
2215 struct obd_device *obd = exp->exp_obd;
2216 struct lmv_obd *lmv = &obd->u.lmv;
2217 struct obd_export *tgt_exp;
2218 struct lu_fid rid = *fid;
2219 struct lmv_obj *obj;
2233 rc = lmv_check_connect(obd);
2237 CDEBUG(D_INFO, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2239 obj = lmv_obj_grab(obd, fid);
2243 * This case handle directory lookup in clustered metadata case (i.e.
2244 * split directory is located on multiple md servers.)
2245 * each server keeps directory entries for certain range of hashes.
2246 * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2247 * first server will keep records with hashes [ 0 ... MAX_HASH / N - 1],
2248 * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2250 * readdir can simply start reading entries from 0 - N server in
2251 * order but that will not scale well as all client will request dir in
2252 * to server in same order.
2253 * Following algorithm does optimization:
2254 * Instead of doing readdir in 1, 2, ...., N order, client with a
2255 * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2256 * (every client has rank R)
2257 * But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2258 * since client ask dir from MDS{R} client has pages with offsets
2259 * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2260 * on hash values that we get.
2263 struct lmv_inode *loi;
2267 nr = obj->lo_objcount;
2269 seg_size = MAX_HASH_SIZE;
2270 do_div(seg_size, nr);
2271 loi = obj->lo_inodes;
2272 rank = lmv_node_rank(lmv_get_export(lmv, loi[0].li_mds),
2275 do_div(tgt_tmp, seg_size);
2276 tgt0 = do_div(tgt_tmp, nr);
2277 tgt = (tgt0 + rank) % nr;
2283 * Last segment has unusual length due to division
2286 hash_adj = MAX_HASH_SIZE - seg_size * nr;
2290 hash_adj += rank * seg_size;
2292 CDEBUG(D_INFO, "hash_adj: %x "LPX64" "LPX64"/%x -> "LPX64"/%x\n",
2293 rank, hash_adj, offset, tgt0, offset + hash_adj, tgt);
2295 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2296 rid = obj->lo_inodes[tgt].li_fid;
2297 tgt_exp = lmv_get_export(lmv, loi[tgt].li_mds);
2299 CDEBUG(D_INFO, "forward to "DFID" with offset %lu i %d\n",
2300 PFID(&rid), (unsigned long)offset, tgt);
2302 tgt_exp = lmv_find_export(lmv, &rid);
2304 if (IS_ERR(tgt_exp))
2305 GOTO(cleanup, rc = PTR_ERR(tgt_exp));
2307 rc = md_readpage(tgt_exp, &rid, oc, offset, page, request);
2311 struct lu_dirpage *dp;
2312 struct lu_dirent *ent;
2314 dp = cfs_kmap(page);
2316 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2317 lmv_hash_adjust(&dp->ldp_hash_end, hash_adj);
2318 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2320 for (ent = lu_dirent_start(dp); ent != NULL;
2321 ent = lu_dirent_next(ent))
2322 lmv_hash_adjust(&ent->lde_hash, hash_adj);
2324 if (tgt0 != nr - 1) {
2327 end = le64_to_cpu(dp->ldp_hash_end);
2328 if (end == DIR_END_OFF) {
2329 dp->ldp_hash_end = cpu_to_le32(seg_size *
2332 ""DFID" reset end "LPX64" tgt %d\n",
2334 le64_to_cpu(dp->ldp_hash_end), tgt);
2340 * Here we could remove "." and ".." from all pages which at not from
2341 * master. But MDS has only "." and ".." for master dir.
2346 lmv_obj_unlock(obj);
2352 static int lmv_unlink_slaves(struct obd_export *exp,
2353 struct md_op_data *op_data,
2354 struct ptlrpc_request **req)
2356 struct obd_device *obd = exp->exp_obd;
2357 struct lmv_obd *lmv = &obd->u.lmv;
2358 struct lmv_stripe_md *mea = op_data->op_mea1;
2359 struct md_op_data *op_data2;
2360 struct obd_export *tgt_exp;
2364 OBD_ALLOC_PTR(op_data2);
2365 if (op_data2 == NULL)
2368 op_data2->op_mode = S_IFDIR;
2369 op_data2->op_fsuid = current->fsuid;
2370 op_data2->op_fsgid = current->fsgid;
2371 op_data2->op_bias = 0;
2373 LASSERT(mea != NULL);
2374 for (i = 0; i < mea->mea_count; i++) {
2375 memset(op_data2, 0, sizeof(*op_data2));
2376 op_data2->op_fid1 = mea->mea_ids[i];
2377 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
2378 if (IS_ERR(tgt_exp))
2379 GOTO(out_free_op_data2, rc = PTR_ERR(tgt_exp));
2381 if (tgt_exp == NULL)
2384 rc = md_unlink(tgt_exp, op_data2, req);
2386 CDEBUG(D_OTHER, "unlink slave "DFID" -> %d\n",
2387 PFID(&mea->mea_ids[i]), rc);
2390 ptlrpc_req_finished(*req);
2394 GOTO(out_free_op_data2, rc);
2399 OBD_FREE_PTR(op_data2);
2403 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2404 struct ptlrpc_request **request)
2406 struct obd_device *obd = exp->exp_obd;
2407 struct lmv_obd *lmv = &obd->u.lmv;
2408 struct obd_export *tgt_exp = NULL;
2409 struct lmv_obj *obj;
2413 rc = lmv_check_connect(obd);
2417 if (op_data->op_namelen == 0 && op_data->op_mea1 != NULL) {
2418 /* mds asks to remove slave objects */
2419 rc = lmv_unlink_slaves(exp, op_data, request);
2426 if (op_data->op_namelen != 0) {
2429 obj = lmv_obj_grab(obd, &op_data->op_fid1);
2431 mea_idx = raw_name2idx(obj->lo_hashtype,
2434 op_data->op_namelen);
2435 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2436 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
2437 tgt_exp = lmv_get_export(lmv,
2438 obj->lo_inodes[mea_idx].li_mds);
2440 CDEBUG(D_OTHER, "unlink '%*s' in "DFID" -> %u\n",
2441 op_data->op_namelen, op_data->op_name,
2442 PFID(&op_data->op_fid1), mea_idx);
2445 CDEBUG(D_OTHER, "drop i_nlink on "DFID"\n",
2446 PFID(&op_data->op_fid1));
2448 if (tgt_exp == NULL) {
2449 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2450 if (IS_ERR(tgt_exp))
2451 RETURN(PTR_ERR(tgt_exp));
2452 op_data->op_bias |= MDS_CHECK_SPLIT;
2455 op_data->op_fsuid = current->fsuid;
2456 op_data->op_fsgid = current->fsgid;
2457 op_data->op_cap = current->cap_effective;
2459 /* If child's fid is given, cancel unused locks for it if it is from
2460 * another export than parent. */
2461 if (op_data->op_namelen) {
2462 /* LOOKUP lock for child (fid3) should also be cancelled on
2463 * parent tgt_exp in mdc_unlink(). */
2464 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2466 /* XXX: the case when child is a striped dir is not supported.
2467 * Only the master stripe has all locks cancelled early. */
2468 /* Cancel FULL locks on child (fid3). */
2469 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
2470 MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2473 rc = md_unlink(tgt_exp, op_data, request);
2474 if (rc == -ERESTART) {
2475 LASSERT(*request != NULL);
2476 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2477 "Got -ERESTART during unlink!\n");
2478 ptlrpc_req_finished(*request);
2482 * Directory got split. Time to update local object and repeat
2483 * the request with proper MDS.
2485 rc = lmv_handle_split(exp, &op_data->op_fid1);
2492 static int lmv_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2493 struct obd_device *tgt, int count,
2494 struct llog_catid *logid, struct obd_uuid *uuid)
2497 struct llog_ctxt *ctxt;
2501 LASSERT(group == OBD_LLOG_GROUP);
2502 rc = llog_setup(obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
2505 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_CONFIG_REPL_CTXT);
2506 llog_initiator_connect(ctxt, tgt);
2507 llog_ctxt_put(ctxt);
2515 static int lmv_llog_finish(struct obd_device *obd, int count)
2517 struct llog_ctxt *ctxt;
2521 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
2523 rc = llog_cleanup(ctxt);
2528 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2533 case OBD_CLEANUP_EARLY:
2534 /* XXX: here should be calling obd_precleanup() down to
2537 case OBD_CLEANUP_EXPORTS:
2538 rc = obd_llog_finish(obd, 0);
2540 CERROR("failed to cleanup llogging subsystems\n");
2548 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2549 void *key, __u32 *vallen, void *val)
2551 struct obd_device *obd;
2552 struct lmv_obd *lmv;
2556 obd = class_exp2obd(exp);
2558 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2559 exp->exp_handle.h_cookie);
2564 if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2565 struct lmv_tgt_desc *tgts;
2568 rc = lmv_check_connect(obd);
2572 LASSERT(*vallen == sizeof(__u32));
2573 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2576 /* all tgts should be connected when this get called. */
2577 if (!tgts || !tgts->ltd_exp) {
2578 CERROR("target not setup?\n");
2582 if (!obd_get_info(tgts->ltd_exp, keylen, key,
2587 } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2588 rc = lmv_check_connect(obd);
2592 /* forwarding this request to first MDS, it should know LOV
2594 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2596 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2597 exp->exp_connect_flags =
2598 ((struct obd_connect_data *)val)->ocd_connect_flags;
2603 CDEBUG(D_IOCTL, "invalid key\n");
2607 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2608 void *key, obd_count vallen, void *val,
2609 struct ptlrpc_request_set *set)
2611 struct lmv_tgt_desc *tgt;
2612 struct obd_device *obd;
2613 struct lmv_obd *lmv;
2617 obd = class_exp2obd(exp);
2619 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2620 exp->exp_handle.h_cookie);
2625 if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2626 KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2629 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2630 tgt = &lmv->tgts[i];
2635 err = obd_set_info_async(tgt->ltd_exp,
2636 keylen, key, vallen, val, set);
2647 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2648 struct lov_stripe_md *lsm)
2650 struct obd_device *obd = class_exp2obd(exp);
2651 struct lmv_obd *lmv = &obd->u.lmv;
2652 struct lmv_stripe_md *meap, *lsmp;
2656 mea_size = lmv_get_easize(lmv);
2660 if (*lmmp && !lsm) {
2661 OBD_FREE(*lmmp, mea_size);
2666 if (*lmmp == NULL) {
2667 OBD_ALLOC(*lmmp, mea_size);
2675 lsmp = (struct lmv_stripe_md *)lsm;
2676 meap = (struct lmv_stripe_md *)*lmmp;
2678 if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2679 lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2682 meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2683 meap->mea_count = cpu_to_le32(lsmp->mea_count);
2684 meap->mea_master = cpu_to_le32(lsmp->mea_master);
2686 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2687 meap->mea_ids[i] = meap->mea_ids[i];
2688 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2694 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2695 struct lov_mds_md *lmm, int lmm_size)
2697 struct obd_device *obd = class_exp2obd(exp);
2698 struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2699 struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2700 struct lmv_obd *lmv = &obd->u.lmv;
2705 mea_size = lmv_get_easize(lmv);
2709 if (*lsmp != NULL && lmm == NULL) {
2710 OBD_FREE(*tmea, mea_size);
2715 LASSERT(mea_size == lmm_size);
2717 OBD_ALLOC(*tmea, mea_size);
2724 if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2725 mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2726 mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2728 magic = le32_to_cpu(mea->mea_magic);
2730 /* old mea is not handled here */
2734 (*tmea)->mea_magic = magic;
2735 (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2736 (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2738 for (i = 0; i < (*tmea)->mea_count; i++) {
2739 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2740 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2745 static int lmv_cancel_unused(struct obd_export *exp,
2746 const struct lu_fid *fid,
2747 ldlm_policy_data_t *policy,
2748 ldlm_mode_t mode, int flags, void *opaque)
2750 struct obd_device *obd = exp->exp_obd;
2751 struct lmv_obd *lmv = &obd->u.lmv;
2755 LASSERT(fid != NULL);
2757 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2758 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2761 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2762 policy, mode, flags, opaque);
2769 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data)
2771 struct obd_device *obd = exp->exp_obd;
2772 struct lmv_obd *lmv = &obd->u.lmv;
2775 RETURN(md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data));
2778 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2779 const struct lu_fid *fid, ldlm_type_t type,
2780 ldlm_policy_data_t *policy, ldlm_mode_t mode,
2781 struct lustre_handle *lockh)
2783 struct obd_device *obd = exp->exp_obd;
2784 struct lmv_obd *lmv = &obd->u.lmv;
2789 CDEBUG(D_OTHER, "lock match for "DFID"\n", PFID(fid));
2791 /* with CMD every object can have two locks in different namespaces:
2792 * lookup lock in space of mds storing direntry and update/open lock in
2793 * space of mds storing inode. Thus we check all targets, not only that
2794 * one fid was created in. */
2795 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2796 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2797 type, policy, mode, lockh);
2805 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2806 struct obd_export *dt_exp, struct obd_export *md_exp,
2807 struct lustre_md *md)
2809 struct obd_device *obd = exp->exp_obd;
2810 struct lmv_obd *lmv = &obd->u.lmv;
2814 rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2818 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2820 struct obd_device *obd = exp->exp_obd;
2821 struct lmv_obd *lmv = &obd->u.lmv;
2825 /* XXX LOV STACKING */
2827 obd_free_memmd(exp, (void *)&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 int lmv_intent_getattr_async(struct obd_export *exp,
2910 struct md_enqueue_info *minfo,
2911 struct ldlm_enqueue_info *einfo)
2913 struct obd_device *obd = exp->exp_obd;
2914 struct lmv_obd *lmv = &obd->u.lmv;
2915 struct obd_export *tgt_exp;
2919 rc = lmv_check_connect(obd);
2923 if (fid_is_zero(&minfo->mi_data.op_fid2))
2924 tgt_exp = lmv_find_export(lmv, &minfo->mi_data.op_fid1);
2926 tgt_exp = lmv_find_export(lmv, &minfo->mi_data.op_fid2);
2927 if (IS_ERR(tgt_exp))
2928 RETURN(PTR_ERR(tgt_exp));
2930 rc = md_intent_getattr_async(tgt_exp, minfo, einfo);
2934 int lmv_revalidate_lock(struct obd_export *exp,
2935 struct lookup_intent *it,
2938 struct obd_device *obd = exp->exp_obd;
2939 struct lmv_obd *lmv = &obd->u.lmv;
2940 struct obd_export *tgt_exp;
2944 rc = lmv_check_connect(obd);
2948 tgt_exp = lmv_find_export(lmv, fid);
2949 if (IS_ERR(tgt_exp))
2950 RETURN(PTR_ERR(tgt_exp));
2952 rc = md_revalidate_lock(tgt_exp, it, fid);
2957 struct obd_ops lmv_obd_ops = {
2958 .o_owner = THIS_MODULE,
2959 .o_setup = lmv_setup,
2960 .o_cleanup = lmv_cleanup,
2961 .o_precleanup = lmv_precleanup,
2962 .o_process_config = lmv_process_config,
2963 .o_connect = lmv_connect,
2964 .o_disconnect = lmv_disconnect,
2965 .o_statfs = lmv_statfs,
2966 .o_llog_init = lmv_llog_init,
2967 .o_llog_finish = lmv_llog_finish,
2968 .o_get_info = lmv_get_info,
2969 .o_set_info_async = lmv_set_info_async,
2970 .o_packmd = lmv_packmd,
2971 .o_unpackmd = lmv_unpackmd,
2972 .o_notify = lmv_notify,
2973 .o_iocontrol = lmv_iocontrol,
2974 .o_fid_delete = lmv_fid_delete
2977 struct md_ops lmv_md_ops = {
2978 .m_getstatus = lmv_getstatus,
2979 .m_change_cbdata = lmv_change_cbdata,
2980 .m_close = lmv_close,
2981 .m_create = lmv_create,
2982 .m_done_writing = lmv_done_writing,
2983 .m_enqueue = lmv_enqueue,
2984 .m_getattr = lmv_getattr,
2985 .m_getxattr = lmv_getxattr,
2986 .m_getattr_name = lmv_getattr_name,
2987 .m_intent_lock = lmv_intent_lock,
2989 .m_rename = lmv_rename,
2990 .m_setattr = lmv_setattr,
2991 .m_setxattr = lmv_setxattr,
2993 .m_readpage = lmv_readpage,
2994 .m_unlink = lmv_unlink,
2995 .m_init_ea_size = lmv_init_ea_size,
2996 .m_cancel_unused = lmv_cancel_unused,
2997 .m_set_lock_data = lmv_set_lock_data,
2998 .m_lock_match = lmv_lock_match,
2999 .m_get_lustre_md = lmv_get_lustre_md,
3000 .m_free_lustre_md = lmv_free_lustre_md,
3001 .m_set_open_replay_data = lmv_set_open_replay_data,
3002 .m_clear_open_replay_data = lmv_clear_open_replay_data,
3003 .m_renew_capa = lmv_renew_capa,
3004 .m_get_remote_perm = lmv_get_remote_perm,
3005 .m_intent_getattr_async = lmv_intent_getattr_async,
3006 .m_revalidate_lock = lmv_revalidate_lock
3009 int __init lmv_init(void)
3011 struct lprocfs_static_vars lvars;
3014 obj_cache = cfs_mem_cache_create("lmv_objects",
3015 sizeof(struct lmv_obj),
3018 CERROR("error allocating lmv objects cache\n");
3022 lprocfs_lmv_init_vars(&lvars);
3023 rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3024 lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3026 cfs_mem_cache_destroy(obj_cache);
3032 static void lmv_exit(void)
3036 class_unregister_type(LUSTRE_LMV_NAME);
3038 rc = cfs_mem_cache_destroy(obj_cache);
3040 "can't free lmv objects cache, %d object(s)"
3041 "still in use\n", atomic_read(&obj_cache_count));
3044 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3045 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3046 MODULE_LICENSE("GPL");
3048 module_init(lmv_init);
3049 module_exit(lmv_exit);