Whamcloud - gitweb
LU-81 deadlock of changelog adding vs. changelog cancelling
[fs/lustre-release.git] / lustre / mds / mds_log.c
index b582ae1..91ea64c 100644 (file)
@@ -26,7 +26,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  */
 /*
@@ -60,7 +60,7 @@ static int mds_llog_origin_add(struct llog_ctxt *ctxt, struct llog_rec_hdr *rec,
                                struct llog_cookie *logcookies, int numcookies)
 {
         struct obd_device *obd = ctxt->loc_obd;
-        struct obd_device *lov_obd = obd->u.mds.mds_osc_obd;
+        struct obd_device *lov_obd = obd->u.mds.mds_lov_obd;
         struct llog_ctxt *lctxt;
         int rc;
         ENTRY;
@@ -78,7 +78,7 @@ static int mds_llog_origin_connect(struct llog_ctxt *ctxt,
                                    struct obd_uuid *uuid)
 {
         struct obd_device *obd = ctxt->loc_obd;
-        struct obd_device *lov_obd = obd->u.mds.mds_osc_obd;
+        struct obd_device *lov_obd = obd->u.mds.mds_lov_obd;
         struct llog_ctxt *lctxt;
         int rc;
         ENTRY;
@@ -89,11 +89,16 @@ static int mds_llog_origin_connect(struct llog_ctxt *ctxt,
         RETURN(rc);
 }
 
+static struct llog_operations mds_ost_orig_logops = {
+        lop_add:        mds_llog_origin_add,
+        lop_connect:    mds_llog_origin_connect,
+};
+
 static int mds_llog_repl_cancel(struct llog_ctxt *ctxt, struct lov_stripe_md *lsm,
                           int count, struct llog_cookie *cookies, int flags)
 {
         struct obd_device *obd = ctxt->loc_obd;
-        struct obd_device *lov_obd = obd->u.mds.mds_osc_obd;
+        struct obd_device *lov_obd = obd->u.mds.mds_lov_obd;
         struct llog_ctxt *lctxt;
         int rc;
         ENTRY;
@@ -104,36 +109,138 @@ static int mds_llog_repl_cancel(struct llog_ctxt *ctxt, struct lov_stripe_md *ls
         RETURN(rc);
 }
 
-static struct llog_operations mds_ost_orig_logops = {
-        lop_add:        mds_llog_origin_add,
-        lop_connect:    mds_llog_origin_connect,
-};
-
 static struct llog_operations mds_size_repl_logops = {
         lop_cancel:     mds_llog_repl_cancel,
 };
 
+static struct llog_operations changelog_orig_logops;
+
+static int llog_changelog_cancel_cb(struct llog_handle *llh,
+                                    struct llog_rec_hdr *hdr, void *data)
+{
+        struct llog_changelog_rec *rec = (struct llog_changelog_rec *)hdr;
+        struct llog_cookie cookie;
+        long long endrec = *(long long *)data;
+        int rc, err;
+        struct obd_device *obd;
+        void *trans_h;
+        struct inode *inode;
+        ENTRY;
+
+        /* This is always a (sub)log, not the catalog */
+        LASSERT(llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN);
+
+        if (rec->cr.cr_index > endrec)
+                /* records are in order, so we're done */
+                RETURN(LLOG_PROC_BREAK);
+
+        cookie.lgc_lgl = llh->lgh_id;
+        cookie.lgc_index = hdr->lrh_index;
+        obd = llh->lgh_ctxt->loc_exp->exp_obd;
+        inode = llh->lgh_file->f_dentry->d_inode;
+
+        /* XXX This is a workaround for the deadlock of changelog adding vs.
+         * changelog cancelling. Changelog adding always start transaction
+         * before acquiring the catlog lock (lgh_lock), whereas, changelog
+         * cancelling do start transaction after holding catlog lock.
+         *
+         * We start the transaction earlier here to keep the locking ordering:
+         * 'start transaction -> catlog lock'. LU-81. */
+        trans_h = fsfilt_start_log(obd, inode, FSFILT_OP_CANCEL_UNLINK,
+                                   NULL, 1);
+        if (IS_ERR(trans_h)) {
+                CERROR("fsfilt_start_log failed: %ld\n", PTR_ERR(trans_h));
+                RETURN(PTR_ERR(trans_h));
+        }
+
+        /* cancel them one at a time.  I suppose we could store up the cookies
+           and cancel them all at once; probably more efficient, but this is
+           done as a user call, so who cares... */
+        rc = llog_cat_cancel_records(llh->u.phd.phd_cat_handle, 1, &cookie);
+
+        err = fsfilt_commit(obd, inode, trans_h, 0);
+        if (err) {
+                CERROR("fsfilt_commit failed: %d\n", err);
+                rc = (rc >= 0) ? err : rc;
+        }
+
+        RETURN(rc < 0 ? rc : 0);
+}
+
+static int llog_changelog_cancel(struct llog_ctxt *ctxt,
+                                 struct lov_stripe_md *lsm, int count,
+                                 struct llog_cookie *cookies, int flags)
+{
+        struct llog_handle *cathandle = ctxt->loc_handle;
+        int rc;
+        ENTRY;
+
+        /* This should only be called with the catalog handle */
+        LASSERT(cathandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT);
+
+        rc = llog_cat_process(cathandle, llog_changelog_cancel_cb,
+                              (void *)cookies, 0, 0);
+        if (rc >= 0)
+                /* 0 or 1 means we're done */
+                rc = 0;
+        else
+                CERROR("cancel idx %u of catalog "LPX64" rc=%d\n",
+                       cathandle->lgh_last_idx, cathandle->lgh_id.lgl_oid, rc);
+
+        RETURN(rc);
+}
+
+int mds_changelog_llog_init(struct obd_device *obd, struct obd_device *tgt)
+{
+        int rc;
+
+        /* see osc_llog_init */
+        changelog_orig_logops = llog_lvfs_ops;
+        changelog_orig_logops.lop_setup = llog_obd_origin_setup;
+        changelog_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
+        changelog_orig_logops.lop_add = llog_obd_origin_add;
+        changelog_orig_logops.lop_cancel = llog_changelog_cancel;
+
+        rc = llog_setup_named(obd, &obd->obd_olg, LLOG_CHANGELOG_ORIG_CTXT,
+                              tgt, 1, NULL, CHANGELOG_CATALOG,
+                              &changelog_orig_logops);
+        if (rc) {
+                CERROR("changelog llog setup failed %d\n", rc);
+                RETURN(rc);
+        }
+
+        rc = llog_setup_named(obd, &obd->obd_olg, LLOG_CHANGELOG_USER_ORIG_CTXT,
+                              tgt, 1, NULL, CHANGELOG_USERS,
+                              &changelog_orig_logops);
+        if (rc) {
+                CERROR("changelog users llog setup failed %d\n", rc);
+                RETURN(rc);
+        }
+
+        RETURN(rc);
+}
+EXPORT_SYMBOL(mds_changelog_llog_init);
+
 int mds_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
-                  struct obd_device *tgt, int count, struct llog_catid *logid, 
-                  struct obd_uuid *uuid)
+                  struct obd_device *disk_obd, int *index)
 {
-        struct obd_device *lov_obd = obd->u.mds.mds_osc_obd;
+        struct obd_device *lov_obd = obd->u.mds.mds_lov_obd;
         struct llog_ctxt *ctxt;
         int rc;
         ENTRY;
 
         LASSERT(olg == &obd->obd_olg);
-        rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, 0, NULL,
-                        &mds_ost_orig_logops);
+        rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, disk_obd,
+                        0, NULL, &mds_ost_orig_logops);
         if (rc)
                 RETURN(rc);
 
-        rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, 0, NULL,
-                        &mds_size_repl_logops);
+        rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, disk_obd,
+                        0, NULL, &mds_size_repl_logops);
         if (rc)
                 GOTO(err_llog, rc);
 
-        rc = obd_llog_init(lov_obd, &lov_obd->obd_olg, tgt, count, logid, uuid);
+        rc = obd_llog_init(lov_obd, &lov_obd->obd_olg, disk_obd, index);
         if (rc) {
                 CERROR("lov_llog_init err %d\n", rc);
                 GOTO(err_cleanup, rc);
@@ -167,5 +274,79 @@ int mds_llog_finish(struct obd_device *obd, int count)
         if (!rc)
                 rc = rc2;
 
+        ctxt = llog_get_context(obd, LLOG_CHANGELOG_ORIG_CTXT);
+        if (ctxt)
+                rc2 = llog_cleanup(ctxt);
+        if (!rc)
+                rc = rc2;
+
+        ctxt = llog_get_context(obd, LLOG_CHANGELOG_USER_ORIG_CTXT);
+        if (ctxt)
+                rc2 = llog_cleanup(ctxt);
+        if (!rc)
+                rc = rc2;
+
         RETURN(rc);
 }
+
+static int mds_llog_add_unlink(struct obd_device *obd,
+                               struct lov_stripe_md *lsm, obd_count count,
+                               struct llog_cookie *logcookie, int cookies)
+{
+        struct llog_unlink_rec *lur;
+        struct llog_ctxt *ctxt;
+        int rc;
+
+        /* first prepare unlink log record */
+        OBD_ALLOC_PTR(lur);
+        if (!lur)
+                RETURN(rc = -ENOMEM);
+        lur->lur_hdr.lrh_len = lur->lur_tail.lrt_len = sizeof(*lur);
+        lur->lur_hdr.lrh_type = MDS_UNLINK_REC;
+        lur->lur_count = count;
+
+        ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
+        rc = llog_add(ctxt, &lur->lur_hdr, lsm, logcookie, cookies);
+        llog_ctxt_put(ctxt);
+
+        OBD_FREE_PTR(lur);
+        RETURN(rc);
+}
+
+int mds_log_op_unlink(struct obd_device *obd,
+                      struct lov_mds_md *lmm, int lmm_size,
+                      struct llog_cookie *logcookies, int cookies_size)
+{
+        struct mds_obd *mds = &obd->u.mds;
+        struct lov_stripe_md *lsm = NULL;
+        int rc;
+        ENTRY;
+
+        if (IS_ERR(mds->mds_lov_obd))
+                RETURN(PTR_ERR(mds->mds_lov_obd));
+
+        rc = obd_unpackmd(mds->mds_lov_exp, &lsm, lmm, lmm_size);
+        if (rc < 0)
+                RETURN(rc);
+        rc = mds_llog_add_unlink(obd, lsm, 0, logcookies,
+                                 cookies_size / sizeof(struct llog_cookie));
+        obd_free_memmd(mds->mds_lov_exp, &lsm);
+        RETURN(rc);
+}
+EXPORT_SYMBOL(mds_log_op_unlink);
+
+int mds_log_op_orphan(struct obd_device *obd, struct lov_stripe_md *lsm,
+                      obd_count count)
+{
+        struct mds_obd *mds = &obd->u.mds;
+        struct llog_cookie logcookie;
+        int rc;
+        ENTRY;
+
+        if (IS_ERR(mds->mds_lov_obd))
+                RETURN(PTR_ERR(mds->mds_lov_obd));
+
+        rc = mds_llog_add_unlink(obd, lsm, count - 1, &logcookie, 1);
+        RETURN(rc);
+}
+