--- /dev/null
+/*
+ * GPL HEADER START
+ *
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.gnu.org/licenses/gpl-2.0.html
+ *
+ * GPL HEADER END
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ *
+ * lustre/mdt/mdt_restriper.c
+ *
+ * Lustre directory restripe and auto-split
+ */
+
+#define DEBUG_SUBSYSTEM S_MDS
+
+#include <linux/sched.h>
+#include <linux/kthread.h>
+#include "mdt_internal.h"
+
+/* add directory into splitting list and wake up restripe thread */
+void mdt_auto_split_add(struct mdt_thread_info *info, struct mdt_object *o)
+{
+ struct mdt_device *mdt = info->mti_mdt;
+ struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
+
+ spin_lock(&restriper->mdr_lock);
+ if (mdt->mdt_enable_dir_auto_split && !o->mot_restriping) {
+ o->mot_restriping = 1;
+ mdt_object_get(NULL, o);
+ LASSERT(list_empty(&o->mot_restripe_linkage));
+ list_add_tail(&o->mot_restripe_linkage,
+ &restriper->mdr_auto_splitting);
+
+ CDEBUG(D_INFO, "add "DFID" into auto split list.\n",
+ PFID(mdt_object_fid(o)));
+ }
+ spin_unlock(&restriper->mdr_lock);
+
+ wake_up_process(restriper->mdr_task);
+}
+
+void mdt_restripe_migrate_add(struct mdt_thread_info *info,
+ struct mdt_object *o)
+{
+ struct mdt_device *mdt = info->mti_mdt;
+ struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
+
+ spin_lock(&restriper->mdr_lock);
+ if (!o->mot_restriping) {
+ o->mot_restriping = 1;
+ o->mot_restripe_offset = 0;
+ mdt_object_get(NULL, o);
+ LASSERT(list_empty(&o->mot_restripe_linkage));
+ list_add_tail(&o->mot_restripe_linkage,
+ &restriper->mdr_migrating);
+
+ CDEBUG(D_INFO, "add "DFID" into migrate list.\n",
+ PFID(mdt_object_fid(o)));
+ }
+ spin_unlock(&restriper->mdr_lock);
+
+ wake_up_process(restriper->mdr_task);
+}
+
+void mdt_restripe_update_add(struct mdt_thread_info *info,
+ struct mdt_object *o)
+{
+ struct mdt_device *mdt = info->mti_mdt;
+ struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
+
+ spin_lock(&restriper->mdr_lock);
+ if (!o->mot_restriping) {
+ /* update LMV */
+ o->mot_restriping = 1;
+ mdt_object_get(NULL, o);
+ if (list_empty(&restriper->mdr_updating))
+ restriper->mdr_update_time = ktime_get_real_seconds();
+ LASSERT(list_empty(&o->mot_restripe_linkage));
+ list_add_tail(&o->mot_restripe_linkage,
+ &restriper->mdr_updating);
+
+ CDEBUG(D_INFO, "add "DFID" into update list.\n",
+ PFID(mdt_object_fid(o)));
+ }
+ spin_unlock(&restriper->mdr_lock);
+
+ wake_up_process(restriper->mdr_task);
+}
+
+static inline int mdt_fid_alloc(const struct lu_env *env,
+ struct mdt_device *mdt,
+ struct lu_fid *fid,
+ struct mdt_object *parent,
+ const struct lu_name *name)
+{
+ struct lu_device *next = &mdt->mdt_child->md_lu_dev;
+ struct lu_object *o = lu_object_next(&parent->mot_obj);
+
+ return next->ld_ops->ldo_fid_alloc(env, next, fid, o, name);
+}
+
+static void mdt_auto_split_prep(struct mdt_thread_info *info,
+ struct md_op_spec *spec,
+ struct md_attr *ma,
+ u32 lum_stripe_count)
+{
+ struct lu_attr *attr = &ma->ma_attr;
+ struct lmv_user_md_v1 *lum;
+
+ attr->la_ctime = attr->la_mtime = ktime_get_real_seconds();
+ attr->la_valid = LA_CTIME | LA_MTIME;
+
+ lum = &info->mti_mdt->mdt_restriper.mdr_lmv.lmv_user_md;
+ lum->lum_magic = cpu_to_le32(LMV_USER_MAGIC);
+ lum->lum_stripe_count = cpu_to_le32(lum_stripe_count);
+ lum->lum_stripe_offset = cpu_to_le32(LMV_OFFSET_DEFAULT);
+ lum->lum_hash_type = 0;
+
+ spec->u.sp_ea.eadatalen = sizeof(*lum);
+ spec->u.sp_ea.eadata = lum;
+ spec->sp_cr_flags = MDS_OPEN_HAS_EA;
+ spec->no_create = 0;
+ spec->sp_migrate_close = 0;
+}
+
+/* restripe directory: split or merge stripes */
+int mdt_restripe_internal(struct mdt_thread_info *info,
+ struct mdt_object *parent,
+ struct mdt_object *child,
+ const struct lu_name *lname,
+ struct lu_fid *tfid,
+ struct md_op_spec *spec,
+ struct md_attr *ma)
+{
+ const struct lu_env *env = info->mti_env;
+ struct mdt_device *mdt = info->mti_mdt;
+ struct lmv_user_md *lum = spec->u.sp_ea.eadata;
+ struct lmv_mds_md_v1 *lmv;
+ u32 lmv_stripe_count = 0;
+ int rc;
+
+ ENTRY;
+
+ rc = mdt_stripe_get(info, child, ma, XATTR_NAME_LMV);
+ if (rc)
+ RETURN(rc);
+
+ if (ma->ma_valid & MA_LMV) {
+ lmv = &ma->ma_lmv->lmv_md_v1;
+ if (!lmv_is_sane(lmv))
+ RETURN(-EBADF);
+
+ /* don't allow restripe if dir layout is changing */
+ if (lmv_is_layout_changing(lmv))
+ RETURN(-EBUSY);
+
+ /* check whether stripe count and hash unchanged */
+ if (lum->lum_stripe_count == lmv->lmv_stripe_count &&
+ lum->lum_hash_type == lmv->lmv_hash_type)
+ RETURN(-EALREADY);
+
+ lmv_stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
+ } else if (le32_to_cpu(lum->lum_stripe_count) < 2) {
+ /* stripe count unchanged for plain directory */
+ RETURN(-EALREADY);
+ }
+
+ if (le32_to_cpu(lum->lum_stripe_count) > lmv_stripe_count) {
+ /* split */
+ struct md_layout_change *mlc = &info->mti_mlc;
+ struct mdt_object *tobj = NULL;
+ s64 mtime = ma->ma_attr.la_mtime;
+
+ ma->ma_need = MA_INODE;
+ ma->ma_valid = 0;
+ rc = mdt_attr_get_complex(info, child, ma);
+ if (rc)
+ RETURN(rc);
+
+ if (!(ma->ma_valid & MA_INODE))
+ RETURN(-EBADF);
+
+ /* mtime is from from client or set outside */
+ ma->ma_attr.la_mtime = mtime;
+
+ if (!lmv_stripe_count) {
+ /* if child is plain directory, allocate @tobj as the
+ * master object, and make child the first stripe of
+ * @tobj.
+ */
+ tobj = mdt_object_new(env, mdt, tfid);
+ if (unlikely(IS_ERR(tobj)))
+ RETURN(PTR_ERR(tobj));
+ }
+
+ mlc->mlc_opc = MD_LAYOUT_SPLIT;
+ mlc->mlc_parent = mdt_object_child(parent);
+ mlc->mlc_target = tobj ? mdt_object_child(tobj) : NULL;
+ mlc->mlc_attr = &ma->ma_attr;
+ mlc->mlc_name = lname;
+ mlc->mlc_spec = spec;
+ rc = mo_layout_change(env, mdt_object_child(child), mlc);
+ if (!rc) {
+ /* FID and attr need to be replied to client for manual
+ * restripe.
+ */
+ ma->ma_need = MA_INODE;
+ ma->ma_valid = 0;
+ rc = mdt_attr_get_complex(info,
+ lmv_stripe_count ? child : tobj, ma);
+ }
+ if (tobj)
+ mdt_object_put(env, tobj);
+ else
+ *tfid = *mdt_object_fid(child);
+ } else {
+ /* merge only needs to override LMV */
+ struct lu_buf *buf = &info->mti_buf;
+ __u32 version;
+
+ LASSERT(ma->ma_valid & MA_LMV);
+ lmv = &ma->ma_lmv->lmv_md_v1;
+ version = cpu_to_le32(lmv->lmv_layout_version);
+
+ /* adjust 0 to 1 */
+ if (lum->lum_stripe_count == 0)
+ lum->lum_stripe_count = cpu_to_le32(1);
+
+ lmv->lmv_hash_type |= cpu_to_le32(LMV_HASH_FLAG_MERGE |
+ LMV_HASH_FLAG_MIGRATION);
+ lmv->lmv_merge_offset = lum->lum_stripe_count;
+ lmv->lmv_merge_hash = lum->lum_hash_type;
+ lmv->lmv_layout_version = cpu_to_le32(++version);
+
+ buf->lb_buf = lmv;
+ buf->lb_len = sizeof(*lmv);
+ rc = mo_xattr_set(env, mdt_object_child(child), buf,
+ XATTR_NAME_LMV, LU_XATTR_REPLACE);
+ if (rc)
+ RETURN(rc);
+
+ *tfid = *mdt_object_fid(child);
+ ma->ma_need = MA_INODE;
+ ma->ma_valid = 0;
+ rc = mdt_attr_get_complex(info, child, ma);
+ }
+
+ RETURN(rc);
+}
+
+static int mdt_auto_split(struct mdt_thread_info *info)
+{
+ const struct lu_env *env = info->mti_env;
+ struct mdt_device *mdt = info->mti_mdt;
+ struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
+ struct md_attr *ma = &info->mti_attr;
+ struct md_op_spec *spec = &info->mti_spec;
+ struct lu_name *lname = &info->mti_name;
+ struct lu_fid *fid = &info->mti_tmp_fid2;
+ struct mdt_object *parent = NULL;
+ struct mdt_object *child = NULL;
+ struct mdt_object *stripe = NULL;
+ struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
+ struct mdt_lock_handle *lhp;
+ struct mdt_lock_handle *lhc;
+ u32 lmv_stripe_count = 0;
+ u32 lum_stripe_count = 0;
+ int rc;
+
+ ENTRY;
+
+ if (!atomic_read(&mdt->mdt_mds_mds_conns))
+ RETURN(-EINVAL);
+
+ spin_lock(&restriper->mdr_lock);
+ if (!list_empty(&restriper->mdr_auto_splitting)) {
+ child = list_entry(restriper->mdr_auto_splitting.next,
+ typeof(*child), mot_restripe_linkage);
+ list_del_init(&child->mot_restripe_linkage);
+ }
+ spin_unlock(&restriper->mdr_lock);
+
+ if (!child)
+ RETURN(0);
+
+ LASSERT(child->mot_restriping);
+
+ rc = mdt_stripe_get(info, child, ma, XATTR_NAME_LMV);
+ if (rc)
+ GOTO(out, rc);
+
+ if (ma->ma_valid & MA_LMV) {
+ /* stripe dirent exceeds threshold, find its master object */
+ struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
+
+ /* auto-split won't be done on striped directory master object
+ * directly, because it's triggered when dirent count exceeds
+ * threshold, however dirent count of master object is its
+ * stripe count.
+ */
+ if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_STRIPE)
+ GOTO(out, rc = -EINVAL);
+
+ lmv_stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
+
+ /* save stripe to clear 'restriping' flag in the end to avoid
+ * trigger auto-split multiple times.
+ */
+ stripe = child;
+ child = NULL;
+
+ /* get master object FID from linkea */
+ rc = mdt_attr_get_pfid(info, stripe, &ma->ma_pfid);
+ if (rc)
+ GOTO(out, rc);
+
+ child = mdt_object_find(env, mdt, &ma->ma_pfid);
+ if (IS_ERR(child))
+ GOTO(out, rc = PTR_ERR(child));
+
+ spin_lock(&restriper->mdr_lock);
+ if (child->mot_restriping) {
+ /* race? */
+ spin_unlock(&restriper->mdr_lock);
+ GOTO(out, rc = -EBUSY);
+ }
+ child->mot_restriping = 1;
+ spin_unlock(&restriper->mdr_lock);
+
+ /* skip if master object is remote, let the first stripe
+ * to start splitting because dir split needs to be done
+ * on where master object is.
+ */
+ if (mdt_object_remote(child))
+ GOTO(restriping_clear, rc = -EREMOTE);
+ }
+
+ /* striped directory split adds mdr_auto_split_delta stripes */
+ lum_stripe_count = min_t(unsigned int,
+ lmv_stripe_count +
+ mdt->mdt_restriper.mdr_dir_split_delta,
+ atomic_read(&mdt->mdt_mds_mds_conns) + 1);
+ if (lmv_stripe_count >= lum_stripe_count)
+ GOTO(restriping_clear, rc = -EALREADY);
+
+ /* get dir name and parent FID */
+ rc = mdt_attr_get_pfid_name(info, child, fid, lname);
+ if (rc)
+ GOTO(restriping_clear, rc);
+
+ /* copy name out because mti_linkea will be used later, and name should
+ * end with '\0'
+ */
+ memcpy(info->mti_filename, lname->ln_name, lname->ln_namelen);
+ info->mti_filename[lname->ln_namelen] = '\0';
+ lname->ln_name = info->mti_filename;
+ CDEBUG(D_INFO, "split "DFID"/"DNAME" to count %u (MDT count %d)\n",
+ PFID(fid), PNAME(lname), lum_stripe_count,
+ atomic_read(&mdt->mdt_mds_mds_conns) + 1);
+
+ parent = mdt_object_find(env, mdt, fid);
+ if (IS_ERR(parent))
+ GOTO(restriping_clear, rc = PTR_ERR(parent));
+
+ rc = mdt_fid_alloc(env, mdt, fid, child, NULL);
+ if (rc < 0)
+ GOTO(restriping_clear, rc);
+
+ lhp = &info->mti_lh[MDT_LH_PARENT];
+ mdt_lock_pdo_init(lhp, LCK_PW, lname);
+ rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
+ true);
+ if (rc)
+ GOTO(restriping_clear, rc);
+
+ lhc = &info->mti_lh[MDT_LH_CHILD];
+ mdt_lock_reg_init(lhc, LCK_EX);
+ if (mdt_object_remote(parent)) {
+ /* enqueue object remote LOOKUP lock */
+ rc = mdt_remote_object_lock(info, parent, mdt_object_fid(child),
+ &lhc->mlh_rreg_lh,
+ lhc->mlh_rreg_mode,
+ MDS_INODELOCK_LOOKUP, false);
+ if (rc != ELDLM_OK)
+ GOTO(unlock_parent, rc);
+ }
+
+ rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
+ true);
+ if (rc)
+ GOTO(unlock_child, rc);
+
+ mdt_auto_split_prep(info, spec, ma, lum_stripe_count);
+
+ rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
+ EXIT;
+
+unlock_child:
+ mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
+unlock_parent:
+ mdt_object_unlock(info, parent, lhp, rc);
+restriping_clear:
+ child->mot_restriping = 0;
+ LASSERT(list_empty(&child->mot_restripe_linkage));
+out:
+ /* -EALREADY: dir is split already.
+ * -EBUSY: dir is opened, or is splitting by others.
+ * -EREMOTE: dir is remote.
+ */
+ if (rc && rc != -EALREADY && rc != -EBUSY && rc != -EREMOTE)
+ CERROR("%s: split "DFID"/"DNAME" to count %u failed: rc = %d\n",
+ mdt_obd_name(mdt), PFID(mdt_object_fid(child)),
+ PNAME(lname), lum_stripe_count, rc);
+
+ if (!IS_ERR_OR_NULL(child))
+ mdt_object_put(env, child);
+
+ if (stripe) {
+ LASSERT(stripe->mot_restriping);
+ LASSERT(list_empty(&stripe->mot_restripe_linkage));
+ stripe->mot_restriping = 0;
+ /* lock may not be taken, don't cache stripe LMV */
+ mo_invalidate(env, mdt_object_child(stripe));
+ mdt_object_put(env, stripe);
+ }
+
+ if (!IS_ERR_OR_NULL(parent))
+ mdt_object_put(env, parent);
+
+ return rc;
+}
+
+/* sub-files under one stripe are migrated, clear MIGRATION flag in its LMV */
+static int mdt_restripe_migrate_finish(struct mdt_thread_info *info,
+ struct mdt_object *stripe,
+ struct lmv_mds_md_v1 *lmv)
+{
+ struct mdt_device *mdt = info->mti_mdt;
+ struct lu_buf buf;
+ struct mdt_lock_handle *lh;
+ int rc;
+
+ ENTRY;
+
+ LASSERT(le32_to_cpu(lmv->lmv_magic) == LMV_MAGIC_STRIPE);
+ LASSERT(lmv_is_restriping(lmv));
+
+ lmv->lmv_hash_type &= ~cpu_to_le32(LMV_HASH_FLAG_MIGRATION);
+ buf.lb_buf = lmv;
+ buf.lb_len = sizeof(*lmv);
+
+ lh = &info->mti_lh[MDT_LH_PARENT];
+ mdt_lock_reg_init(lh, LCK_EX);
+ rc = mdt_reint_object_lock(info, stripe, lh, MDS_INODELOCK_XATTR,
+ false);
+ if (!rc)
+ rc = mo_xattr_set(info->mti_env, mdt_object_child(stripe), &buf,
+ XATTR_NAME_LMV, LU_XATTR_REPLACE);
+ mdt_object_unlock(info, stripe, lh, rc);
+ if (rc)
+ CERROR("%s: update "DFID" LMV failed: rc = %d\n",
+ mdt_obd_name(mdt), PFID(mdt_object_fid(stripe)), rc);
+
+ LASSERT(!list_empty(&stripe->mot_restripe_linkage));
+ LASSERT(stripe->mot_restriping);
+
+ spin_lock(&mdt->mdt_lock);
+ stripe->mot_restriping = 0;
+ list_del_init(&stripe->mot_restripe_linkage);
+ spin_unlock(&mdt->mdt_lock);
+
+ mdt_object_put(info->mti_env, stripe);
+
+ RETURN(rc);
+}
+
+static void mdt_restripe_migrate_prep(struct mdt_thread_info *info,
+ const struct lu_fid *fid1,
+ const struct lu_fid *fid2,
+ const struct lu_name *lname,
+ __u16 type,
+ const struct lmv_mds_md_v1 *lmv)
+{
+ struct lu_attr *attr = &info->mti_attr.ma_attr;
+ struct mdt_reint_record *rr = &info->mti_rr;
+ struct md_op_spec *spec = &info->mti_spec;
+ struct lmv_user_md_v1 *lum;
+
+ attr->la_ctime = attr->la_mtime = ktime_get_real_seconds();
+ attr->la_valid = LA_CTIME | LA_MTIME;
+
+ rr->rr_fid1 = fid1;
+ rr->rr_fid2 = fid2;
+ rr->rr_name = *lname;
+
+ lum = &info->mti_mdt->mdt_restriper.mdr_lmv.lmv_user_md;
+ lum->lum_magic = cpu_to_le32(LMV_USER_MAGIC);
+ lum->lum_stripe_offset = cpu_to_le32(LMV_OFFSET_DEFAULT);
+ if (lmv_is_splitting(lmv)) {
+ lum->lum_stripe_count = lmv->lmv_stripe_count;
+ lum->lum_hash_type =
+ lmv->lmv_hash_type & le32_to_cpu(LMV_HASH_TYPE_MASK);
+ } else if (lmv_is_merging(lmv)) {
+ lum->lum_stripe_count = lmv->lmv_merge_offset;
+ lum->lum_hash_type = lmv->lmv_merge_hash;
+ }
+
+ spec->u.sp_ea.eadatalen = sizeof(*lum);
+ spec->u.sp_ea.eadata = lum;
+ spec->sp_cr_flags = MDS_OPEN_HAS_EA;
+ spec->no_create = 0;
+ spec->sp_migrate_close = 0;
+ /* if 'nsonly' is set, don't migrate inode */
+ if (S_ISDIR(type))
+ spec->sp_migrate_nsonly = 1;
+ else
+ spec->sp_migrate_nsonly =
+ info->mti_mdt->mdt_dir_restripe_nsonly;
+}
+
+/* migrate sub-file from @mdr_restripe_offset */
+static int mdt_restripe_migrate(struct mdt_thread_info *info)
+{
+ const struct lu_env *env = info->mti_env;
+ struct mdt_device *mdt = info->mti_mdt;
+ struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
+ struct mdt_object *stripe = NULL;
+ struct mdt_object *master = NULL;
+ struct md_attr *ma = &info->mti_attr;
+ struct lmv_mds_md_v1 *lmv;
+ struct lu_name *lname = &info->mti_name;
+ struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
+ struct lu_fid fid1;
+ struct lu_fid fid2;
+ struct lu_dirpage *dp;
+ struct lu_dirent *ent;
+ const char *name = NULL;
+ int namelen = 0;
+ __u16 type;
+ int idx = 0;
+ int len;
+ int rc;
+
+ ENTRY;
+
+ if (list_empty(&restriper->mdr_migrating))
+ RETURN(0);
+
+ stripe = list_entry(restriper->mdr_migrating.next, typeof(*stripe),
+ mot_restripe_linkage);
+
+ /* get master object FID and stripe name */
+ rc = mdt_attr_get_pfid_name(info, stripe, &fid1, lname);
+ if (rc)
+ GOTO(out, rc);
+
+ snprintf(info->mti_filename, sizeof(info->mti_filename), DFID,
+ PFID(mdt_object_fid(stripe)));
+ len = strlen(info->mti_filename) + 1;
+ if (len >= lname->ln_namelen)
+ GOTO(out, rc = -EBADF);
+
+ while (len < lname->ln_namelen) {
+ if (!isdigit(lname->ln_name[len]))
+ GOTO(out, rc = -EBADF);
+
+ idx = idx * 10 + lname->ln_name[len++] - '0';
+ };
+
+ /* check whether stripe is newly created in split */
+ rc = mdt_stripe_get(info, stripe, ma, XATTR_NAME_LMV);
+ if (rc)
+ GOTO(out, rc);
+
+ if (!(ma->ma_valid & MA_LMV))
+ GOTO(out, rc = -ENODATA);
+
+ lmv = &ma->ma_lmv->lmv_md_v1;
+ if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_STRIPE)
+ GOTO(out, rc = -EBADF);
+
+ if (!lmv_is_restriping(lmv))
+ GOTO(out, rc = -EINVAL);
+
+ if ((lmv_is_splitting(lmv) &&
+ idx >= le32_to_cpu(lmv->lmv_split_offset)) ||
+ (lmv_is_merging(lmv) &&
+ le32_to_cpu(lmv->lmv_hash_type) == LMV_HASH_TYPE_CRUSH &&
+ idx < le32_to_cpu(lmv->lmv_merge_offset))) {
+ /* new stripes doesn't need to migrate sub files in dir
+ * split, neither for target stripes in dir merge if hash type
+ * is CRUSH.
+ */
+ rc = mdt_restripe_migrate_finish(info, stripe, lmv);
+ RETURN(rc);
+ }
+
+ /* get sub file name @mot_restripe_offset.
+ * TODO: read one dirent instead of whole page.
+ */
+ rdpg->rp_hash = stripe->mot_restripe_offset;
+ rdpg->rp_count = PAGE_SIZE;
+ rdpg->rp_npages = 1;
+ rdpg->rp_attrs = LUDA_64BITHASH | LUDA_FID | LUDA_TYPE;
+ rdpg->rp_pages = &restriper->mdr_page;
+ rc = mo_readpage(env, mdt_object_child(stripe), rdpg);
+ if (rc < 0)
+ GOTO(out, rc);
+
+ dp = page_address(restriper->mdr_page);
+ for (ent = lu_dirent_start(dp); ent; ent = lu_dirent_next(ent)) {
+ LASSERT(le64_to_cpu(ent->lde_hash) >= rdpg->rp_hash);
+
+ if (unlikely(!(le32_to_cpu(ent->lde_attrs) & LUDA_TYPE)))
+ GOTO(out, rc = -EINVAL);
+
+ namelen = le16_to_cpu(ent->lde_namelen);
+ if (!namelen)
+ continue;
+
+ if (name_is_dot_or_dotdot(ent->lde_name, namelen))
+ continue;
+
+ name = ent->lde_name;
+ type = lu_dirent_type_get(ent);
+ break;
+ }
+
+ if (!name) {
+ if (le64_to_cpu(dp->ldp_hash_end) == MDS_DIR_END_OFF) {
+ rc = mdt_restripe_migrate_finish(info, stripe, lmv);
+ RETURN(rc);
+ }
+
+ GOTO(out, rc = -EBADF);
+ }
+
+ /* copy name out because it should end with '\0' */
+ memcpy(info->mti_filename, name, namelen);
+ info->mti_filename[namelen] = '\0';
+ lname->ln_name = info->mti_filename;
+ lname->ln_namelen = namelen;
+
+ CDEBUG(D_INFO, "migrate "DFID"/"DNAME" type %ho\n",
+ PFID(&fid1), PNAME(lname), type);
+
+ master = mdt_object_find(env, mdt, &fid1);
+ if (IS_ERR(master))
+ GOTO(out, rc = PTR_ERR(master));
+
+ rc = mdt_fid_alloc(env, mdt, &fid2, master, lname);
+ mdt_object_put(env, master);
+ if (rc < 0)
+ GOTO(out, rc);
+
+ mdt_restripe_migrate_prep(info, &fid1, &fid2, lname, type, lmv);
+
+ rc = mdt_reint_migrate(info, NULL);
+ /* mti_big_buf is allocated in XATTR migration */
+ if (unlikely(info->mti_big_buf.lb_buf))
+ lu_buf_free(&info->mti_big_buf);
+ if (rc == -EALREADY)
+ rc = 0;
+ if (rc)
+ GOTO(out, rc);
+
+ LASSERT(ent);
+ do {
+ ent = lu_dirent_next(ent);
+ if (!ent)
+ break;
+
+ namelen = le16_to_cpu(ent->lde_namelen);
+ } while (namelen == 0); /* Skip dummy record */
+
+ if (ent)
+ stripe->mot_restripe_offset = le64_to_cpu(ent->lde_hash);
+ else
+ stripe->mot_restripe_offset = le64_to_cpu(dp->ldp_hash_end);
+
+ EXIT;
+out:
+ if (rc) {
+ /* -EBUSY: file is opened by others */
+ if (rc != -EBUSY)
+ CERROR("%s: migrate "DFID"/"DNAME" failed: rc = %d\n",
+ mdt_obd_name(mdt), PFID(&fid1), PNAME(lname),
+ rc);
+
+ spin_lock(&mdt->mdt_lock);
+ stripe->mot_restriping = 0;
+ list_del_init(&stripe->mot_restripe_linkage);
+ spin_unlock(&mdt->mdt_lock);
+
+ mdt_object_put(env, stripe);
+ }
+
+ return rc;
+}
+
+static inline bool mdt_restripe_update_pending(struct mdt_thread_info *info)
+{
+ struct mdt_device *mdt = info->mti_mdt;
+
+ if (list_empty(&mdt->mdt_restriper.mdr_updating))
+ return false;
+
+ return mdt->mdt_restriper.mdr_update_time < ktime_get_real_seconds();
+}
+
+static void mdt_restripe_layout_update_prep(struct mdt_thread_info *info,
+ const struct lu_fid *fid,
+ const struct lmv_mds_md_v1 *lmv)
+{
+ struct lu_attr *attr = &info->mti_attr.ma_attr;
+ struct mdt_reint_record *rr = &info->mti_rr;
+ struct lmv_user_md_v1 *lum;
+
+ attr->la_ctime = attr->la_mtime = ktime_get_real_seconds();
+ attr->la_valid = LA_CTIME | LA_MTIME;
+
+ strncpy(info->mti_filename, XATTR_NAME_LMV,
+ sizeof(info->mti_filename));
+
+ lum = &info->mti_mdt->mdt_restriper.mdr_lmv.lmv_user_md;
+ lum->lum_magic = cpu_to_le32(LMV_USER_MAGIC);
+ lum->lum_stripe_offset = cpu_to_le32(LMV_OFFSET_DEFAULT);
+ if (lmv_is_splitting(lmv)) {
+ lum->lum_stripe_count = lmv->lmv_stripe_count;
+ lum->lum_hash_type =
+ lmv->lmv_hash_type & le32_to_cpu(LMV_HASH_TYPE_MASK);
+ } else if (lmv_is_merging(lmv)) {
+ lum->lum_stripe_count = lmv->lmv_merge_offset;
+ lum->lum_hash_type = lmv->lmv_merge_hash;
+ }
+
+ rr->rr_opcode = REINT_SETXATTR;
+ rr->rr_fid1 = fid;
+ rr->rr_name.ln_name = info->mti_filename;
+ rr->rr_name.ln_namelen = strlen(info->mti_filename);
+ rr->rr_eadata = lum;
+ rr->rr_eadatalen = sizeof(*lum);
+}
+
+static int mdt_restripe_layout_update(struct mdt_thread_info *info)
+{
+ const struct lu_env *env = info->mti_env;
+ struct mdt_device *mdt = info->mti_mdt;
+ struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
+ struct md_attr *ma = &info->mti_attr;
+ struct lu_fid *fid = &info->mti_tmp_fid1;
+ struct mdt_object *master;
+ struct mdt_object *stripe;
+ struct lmv_mds_md_v1 *lmv;
+ int i;
+ int rc;
+
+ ENTRY;
+
+ if (list_empty(&restriper->mdr_updating))
+ RETURN(0);
+
+ master = list_entry(restriper->mdr_updating.next, typeof(*master),
+ mot_restripe_linkage);
+
+ rc = mdt_stripe_get(info, master, ma, XATTR_NAME_LMV);
+ if (rc)
+ GOTO(out, rc);
+
+ if (!(ma->ma_valid & MA_LMV))
+ GOTO(out, rc = -ENODATA);
+
+ lmv = &ma->ma_lmv->lmv_md_v1;
+ if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
+ GOTO(out, rc = -EBADF);
+
+ if (!lmv_is_restriping(lmv))
+ GOTO(out, rc = -EINVAL);
+
+ /* use different buffer to store stripe LMV */
+ ma->ma_lmv = &restriper->mdr_lmv;
+ ma->ma_lmv_size = sizeof(restriper->mdr_lmv);
+ for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
+ fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
+ stripe = mdt_object_find(env, mdt, fid);
+ if (IS_ERR(stripe))
+ GOTO(out, rc = PTR_ERR(stripe));
+
+ ma->ma_valid = 0;
+ rc = __mdt_stripe_get(info, stripe, ma, XATTR_NAME_LMV);
+ /* LMV is checked without lock, don't cache it */
+ mo_invalidate(env, mdt_object_child(stripe));
+ mdt_object_put(env, stripe);
+ if (rc)
+ GOTO(out, rc);
+
+ if (!(ma->ma_valid & MA_LMV))
+ GOTO(out, rc = -ENODATA);
+
+ /* check MIGRATION flag cleared on all stripes */
+ if (lmv_is_restriping(&ma->ma_lmv->lmv_md_v1))
+ GOTO(out, rc = -EINPROGRESS);
+ }
+
+ mdt_restripe_layout_update_prep(info, mdt_object_fid(master), lmv);
+
+ rc = mdt_dir_layout_update(info);
+ if (rc) {
+ CERROR("update "DFID" layout failed: rc = %d\n",
+ PFID(mdt_object_fid(master)), rc);
+ GOTO(out, rc);
+ }
+
+out:
+ LASSERT(!list_empty(&master->mot_restripe_linkage));
+ if (rc == -EINPROGRESS) {
+ restriper->mdr_update_time = ktime_get_real_seconds() + 5;
+ } else {
+ spin_lock(&restriper->mdr_lock);
+ master->mot_restriping = 0;
+ list_del_init(&master->mot_restripe_linkage);
+ spin_unlock(&restriper->mdr_lock);
+
+ mdt_object_put(env, master);
+ }
+
+ return rc;
+}
+
+static int mdt_restriper_main(void *arg)
+{
+ struct mdt_thread_info *info = arg;
+ struct mdt_device *mdt = info->mti_mdt;
+ struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
+
+ ENTRY;
+
+ while (({set_current_state(TASK_IDLE);
+ !kthread_should_stop(); })) {
+ if (!list_empty(&restriper->mdr_auto_splitting)) {
+ __set_current_state(TASK_RUNNING);
+ mdt_auto_split(info);
+ cond_resched();
+ } else if (mdt_restripe_update_pending(info)) {
+ __set_current_state(TASK_RUNNING);
+ mdt_restripe_layout_update(info);
+ cond_resched();
+ } else if (!list_empty(&restriper->mdr_migrating)) {
+ __set_current_state(TASK_RUNNING);
+ mdt_restripe_migrate(info);
+ cond_resched();
+ } else {
+ schedule();
+ }
+ }
+ __set_current_state(TASK_RUNNING);
+
+ RETURN(0);
+}
+
+int mdt_restriper_start(struct mdt_device *mdt)
+{
+ struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
+ struct task_struct *task;
+ struct mdt_thread_info *info;
+ struct lu_ucred *uc;
+ int rc;
+
+ ENTRY;
+
+ spin_lock_init(&restriper->mdr_lock);
+ INIT_LIST_HEAD(&restriper->mdr_auto_splitting);
+ INIT_LIST_HEAD(&restriper->mdr_migrating);
+ INIT_LIST_HEAD(&restriper->mdr_updating);
+ restriper->mdr_dir_split_count = DIR_SPLIT_COUNT_DEFAULT;
+ restriper->mdr_dir_split_delta = DIR_SPLIT_DELTA_DEFAULT;
+
+ restriper->mdr_page = alloc_page(GFP_KERNEL);
+ if (!restriper->mdr_page)
+ RETURN(-ENOMEM);
+
+ rc = lu_env_init(&restriper->mdr_env, LCT_MD_THREAD);
+ if (rc)
+ GOTO(out_page, rc);
+
+ rc = lu_context_init(&restriper->mdr_session, LCT_SERVER_SESSION);
+ if (rc)
+ GOTO(out_env, rc);
+
+ lu_context_enter(&restriper->mdr_session);
+ restriper->mdr_env.le_ses = &restriper->mdr_session;
+
+ info = lu_context_key_get(&restriper->mdr_env.le_ctx, &mdt_thread_key);
+ info->mti_env = &restriper->mdr_env;
+ info->mti_mdt = mdt;
+ info->mti_pill = NULL;
+ info->mti_dlm_req = NULL;
+
+ uc = mdt_ucred(info);
+ uc->uc_valid = UCRED_OLD;
+ uc->uc_o_uid = 0;
+ uc->uc_o_gid = 0;
+ uc->uc_o_fsuid = 0;
+ uc->uc_o_fsgid = 0;
+ uc->uc_uid = 0;
+ uc->uc_gid = 0;
+ uc->uc_fsuid = 0;
+ uc->uc_fsgid = 0;
+ uc->uc_suppgids[0] = -1;
+ uc->uc_suppgids[1] = -1;
+ uc->uc_cap = CFS_CAP_FS_MASK;
+ uc->uc_umask = 0644;
+ uc->uc_ginfo = NULL;
+ uc->uc_identity = NULL;
+
+ task = kthread_create(mdt_restriper_main, info, "mdt_restriper_%03d",
+ mdt_seq_site(mdt)->ss_node_id);
+ if (IS_ERR(task)) {
+ rc = PTR_ERR(task);
+ CERROR("%s: Can't start directory restripe thread: rc %d\n",
+ mdt_obd_name(mdt), rc);
+ GOTO(out_ses, rc);
+ }
+ restriper->mdr_task = task;
+ wake_up_process(task);
+
+ RETURN(0);
+
+out_ses:
+ lu_context_exit(restriper->mdr_env.le_ses);
+ lu_context_fini(restriper->mdr_env.le_ses);
+out_env:
+ lu_env_fini(&restriper->mdr_env);
+out_page:
+ __free_page(restriper->mdr_page);
+
+ return rc;
+}
+
+void mdt_restriper_stop(struct mdt_device *mdt)
+{
+ struct mdt_dir_restriper *restriper = &mdt->mdt_restriper;
+ struct lu_env *env = &restriper->mdr_env;
+ struct mdt_object *mo, *next;
+
+ if (!restriper->mdr_task)
+ return;
+
+ kthread_stop(restriper->mdr_task);
+ restriper->mdr_task = NULL;
+
+ list_for_each_entry_safe(mo, next, &restriper->mdr_auto_splitting,
+ mot_restripe_linkage) {
+ list_del_init(&mo->mot_restripe_linkage);
+ mdt_object_put(env, mo);
+ }
+
+ list_for_each_entry_safe(mo, next, &restriper->mdr_migrating,
+ mot_restripe_linkage) {
+ list_del_init(&mo->mot_restripe_linkage);
+ mdt_object_put(env, mo);
+ }
+
+ list_for_each_entry_safe(mo, next, &restriper->mdr_updating,
+ mot_restripe_linkage) {
+ list_del_init(&mo->mot_restripe_linkage);
+ mdt_object_put(env, mo);
+ }
+
+ __free_page(restriper->mdr_page);
+
+ lu_context_exit(env->le_ses);
+ lu_context_fini(env->le_ses);
+ lu_env_fini(env);
+}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff