Whamcloud - gitweb
LU-7340 mdd: changelogs garbage collection
[fs/lustre-release.git] / lustre / obdclass / llog.c
index 0fdd8fe..d3b1e04 100644 (file)
  *
  * You should have received a copy of the GNU General Public License
  * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
  *
  * GPL HEADER END
  */
@@ -27,7 +23,7 @@
  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2012, 2015, Intel Corporation.
+ * Copyright (c) 2012, 2016, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -47,6 +43,7 @@
 
 #define DEBUG_SUBSYSTEM S_LOG
 
+#include <linux/pid_namespace.h>
 #include <linux/kthread.h>
 #include <llog_swab.h>
 #include <lustre_log.h>
@@ -207,8 +204,8 @@ int llog_cancel_rec(const struct lu_env *env, struct llog_handle *loghandle,
 
        ENTRY;
 
-       CDEBUG(D_RPCTRACE, "Canceling %d in log "DOSTID"\n", index,
-              POSTID(&loghandle->lgh_id.lgl_oi));
+       CDEBUG(D_RPCTRACE, "Canceling %d in log "DFID"\n", index,
+              PFID(&loghandle->lgh_id.lgl_oi.oi_fid));
 
        if (index == 0) {
                CERROR("Can't cancel index 0 which is header\n");
@@ -229,8 +226,11 @@ int llog_cancel_rec(const struct lu_env *env, struct llog_handle *loghandle,
        if (rc < 0)
                GOTO(out_trans, rc);
 
-       if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY))
+       if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY)) {
                rc = llog_declare_destroy(env, loghandle, th);
+               if (rc < 0)
+                       GOTO(out_trans, rc);
+       }
 
        th->th_wait_submit = 1;
        rc = dt_trans_start_local(env, dt, th);
@@ -262,18 +262,18 @@ int llog_cancel_rec(const struct lu_env *env, struct llog_handle *loghandle,
             (loghandle->u.phd.phd_cat_handle != NULL &&
              loghandle->u.phd.phd_cat_handle->u.chd.chd_current_log !=
                loghandle))) {
+               /* never try to destroy it again */
+               llh->llh_flags &= ~LLOG_F_ZAP_WHEN_EMPTY;
                rc = llog_trans_destroy(env, loghandle, th);
                if (rc < 0) {
                        /* Sigh, can not destroy the final plain llog, but
                         * the bitmap has been clearly, so the record can not
                         * be accessed anymore, let's return 0 for now, and
                         * the orphan will be handled by LFSCK. */
-                       CERROR("%s: can't destroy empty llog #"DOSTID
-                              "#%08x: rc = %d\n",
+                       CERROR("%s: can't destroy empty llog "DFID": rc = %d\n",
                               loghandle->lgh_ctxt->loc_obd->obd_name,
-                              POSTID(&loghandle->lgh_id.lgl_oi),
-                              loghandle->lgh_id.lgl_ogen, rc);
-                       GOTO(out_unlock, rc);
+                              PFID(&loghandle->lgh_id.lgl_oi.oi_fid), rc);
+                       GOTO(out_unlock, rc = 0);
                }
                rc = LLOG_DEL_PLAIN;
        }
@@ -314,14 +314,13 @@ int llog_read_header(const struct lu_env *env, struct llog_handle *handle,
 
                /* lrh_len should be initialized in llog_init_handle */
                handle->lgh_last_idx = 0; /* header is record with index 0 */
-               handle->lgh_write_offset = 0;
                llh->llh_count = 1;         /* for the header record */
                llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
                LASSERT(handle->lgh_ctxt->loc_chunk_size >=
                                                LLOG_MIN_CHUNK_SIZE);
                llh->llh_hdr.lrh_len = handle->lgh_ctxt->loc_chunk_size;
                llh->llh_hdr.lrh_index = 0;
-               llh->llh_timestamp = cfs_time_current_sec();
+               llh->llh_timestamp = ktime_get_real_seconds();
                if (uuid)
                        memcpy(&llh->llh_tgtuuid, uuid,
                               sizeof(llh->llh_tgtuuid));
@@ -426,6 +425,7 @@ static int llog_process_thread(void *arg)
        int                              rc = 0, index = 1, last_index;
        int                              saved_index = 0;
        int                              last_called_index = 0;
+       bool                             repeated = false;
 
        ENTRY;
 
@@ -453,9 +453,10 @@ static int llog_process_thread(void *arg)
 
        while (rc == 0) {
                struct llog_rec_hdr *rec;
-               off_t chunk_offset;
+               off_t chunk_offset = 0;
                unsigned int buf_offset = 0;
                bool partial_chunk;
+               int     lh_last_idx;
 
                /* skip records not set in bitmap */
                while (index <= last_index &&
@@ -472,8 +473,19 @@ static int llog_process_thread(void *arg)
 repeat:
                /* get the buf with our target record; avoid old garbage */
                memset(buf, 0, chunk_size);
+               /* the record index for outdated chunk data */
+               lh_last_idx = loghandle->lgh_last_idx + 1;
                rc = llog_next_block(lpi->lpi_env, loghandle, &saved_index,
                                     index, &cur_offset, buf, chunk_size);
+               if (repeated && rc)
+                       CDEBUG(D_OTHER, "cur_offset %llu, chunk_offset %llu,"
+                              " buf_offset %u, rc = %d\n", cur_offset,
+                              (__u64)chunk_offset, buf_offset, rc);
+               /* we`ve tried to reread the chunk, but there is no
+                * new records */
+               if (rc == -EIO && repeated && (chunk_offset + buf_offset) ==
+                   cur_offset)
+                       GOTO(out, rc = 0);
                if (rc != 0)
                        GOTO(out, rc);
 
@@ -482,7 +494,7 @@ repeat:
                 * The absolute offset of the current chunk is calculated
                 * from cur_offset value and stored in chunk_offset variable.
                 */
-               if (cur_offset % chunk_size != 0) {
+               if ((cur_offset & (chunk_size - 1)) != 0) {
                        partial_chunk = true;
                        chunk_offset = cur_offset & ~(chunk_size - 1);
                } else {
@@ -506,14 +518,38 @@ repeat:
                        CDEBUG(D_OTHER, "after swabbing, type=%#x idx=%d\n",
                               rec->lrh_type, rec->lrh_index);
 
+                       /* the bitmap could be changed during processing
+                        * records from the chunk. For wrapped catalog
+                        * it means we can read deleted record and try to
+                        * process it. Check this case and reread the chunk. */
+
                        /* for partial chunk the end of it is zeroed, check
                         * for index 0 to distinguish it. */
-                       if (partial_chunk && rec->lrh_index == 0) {
+                       if ((partial_chunk && rec->lrh_index == 0) ||
+                            (index == lh_last_idx &&
+                             lh_last_idx != (loghandle->lgh_last_idx + 1))) {
                                /* concurrent llog_add() might add new records
                                 * while llog_processing, check this is not
                                 * the case and re-read the current chunk
                                 * otherwise. */
-                               if (index > loghandle->lgh_last_idx)
+                               int records;
+                               /* lgh_last_idx could be less then index
+                                * for catalog, if catalog is wrapped */
+                               if ((index > loghandle->lgh_last_idx &&
+                                   !(loghandle->lgh_hdr->llh_flags &
+                                     LLOG_F_IS_CAT)) || repeated ||
+                                   (loghandle->lgh_obj != NULL &&
+                                    dt_object_remote(loghandle->lgh_obj)))
+                                       GOTO(out, rc = 0);
+                               /* <2 records means no more records
+                                * if the last record we processed was
+                                * the final one, then the underlying
+                                * object might have been destroyed yet.
+                                * we better don't access that.. */
+                               mutex_lock(&loghandle->lgh_hdr_mutex);
+                               records = loghandle->lgh_hdr->llh_count;
+                               mutex_unlock(&loghandle->lgh_hdr_mutex);
+                               if (records <= 1)
                                        GOTO(out, rc = 0);
                                CDEBUG(D_OTHER, "Re-read last llog buffer for "
                                       "new records, index %u, last %u\n",
@@ -521,13 +557,20 @@ repeat:
                                /* save offset inside buffer for the re-read */
                                buf_offset = (char *)rec - (char *)buf;
                                cur_offset = chunk_offset;
+                               repeated = true;
                                goto repeat;
                        }
 
+                       repeated = false;
+
                        if (rec->lrh_len == 0 || rec->lrh_len > chunk_size) {
-                               CWARN("invalid length %d in llog record for "
-                                     "index %d/%d\n", rec->lrh_len,
-                                     rec->lrh_index, index);
+                               CWARN("%s: invalid length %d in llog "DFID
+                                     "record for index %d/%d\n",
+                                      loghandle->lgh_ctxt->loc_obd->obd_name,
+                                      rec->lrh_len,
+                                      PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
+                                      rec->lrh_index, index);
+
                                GOTO(out, rc = -EINVAL);
                        }
 
@@ -538,9 +581,10 @@ repeat:
                        }
 
                        if (rec->lrh_index != index) {
-                               CERROR("%s: Invalid record: index %u but "
-                                      "expected %u\n",
+                               CERROR("%s: "DFID" Invalid record: index %u"
+                                      " but expected %u\n",
                                       loghandle->lgh_ctxt->loc_obd->obd_name,
+                                      PFID(&loghandle->lgh_id.lgl_oi.oi_fid),
                                       rec->lrh_index, index);
                                GOTO(out, rc = -ERANGE);
                        }
@@ -568,6 +612,12 @@ repeat:
                                }
                                if (rc)
                                        GOTO(out, rc);
+                               /* some stupid callbacks directly cancel records
+                                * and delete llog. Check it and stop
+                                * processing. */
+                               if (loghandle->lgh_hdr == NULL ||
+                                   loghandle->lgh_hdr->llh_count == 1)
+                                       GOTO(out, rc = 0);
                        }
                        /* exit if the last index is reached */
                        if (index >= last_index)
@@ -595,7 +645,14 @@ out:
                         * llog file, probably I/O error or the log got
                         * corrupted to be able to finally release the log we
                         * discard any remaining bits in the header */
-                       CERROR("Local llog found corrupted\n");
+                       CERROR("%s: Local llog found corrupted #"DOSTID":%x"
+                              " %s index %d count %d\n",
+                              loghandle->lgh_ctxt->loc_obd->obd_name,
+                              POSTID(&loghandle->lgh_id.lgl_oi),
+                              loghandle->lgh_id.lgl_ogen,
+                              ((llh->llh_flags & LLOG_F_IS_CAT) ? "catalog" :
+                               "plain"), index, llh->llh_count);
+
                        while (index <= last_index) {
                                if (ext2_test_bit(index,
                                                  LLOG_HDR_BITMAP(llh)) != 0)
@@ -608,8 +665,8 @@ out:
        }
 
        OBD_FREE_LARGE(buf, chunk_size);
-        lpi->lpi_rc = rc;
-        return 0;
+       lpi->lpi_rc = rc;
+       return 0;
 }
 
 static int llog_process_thread_daemonize(void *arg)
@@ -617,6 +674,22 @@ static int llog_process_thread_daemonize(void *arg)
        struct llog_process_info        *lpi = arg;
        struct lu_env                    env;
        int                              rc;
+       struct nsproxy                  *new_ns, *curr_ns = current->nsproxy;
+
+       task_lock(lpi->lpi_reftask);
+       new_ns = lpi->lpi_reftask->nsproxy;
+       if (curr_ns != new_ns) {
+               get_nsproxy(new_ns);
+
+               current->nsproxy = new_ns;
+               /* XXX: we should call put_nsproxy() instead of
+                * atomic_dec(&ns->count) directly. But put_nsproxy() cannot be
+                * used outside of the kernel itself, because it calls
+                * free_nsproxy() which is not exported by the kernel
+                * (defined in kernel/nsproxy.c) */
+               atomic_dec(&curr_ns->count);
+       }
+       task_unlock(lpi->lpi_reftask);
 
        unshare_fs_struct();
 
@@ -643,15 +716,15 @@ int llog_process_or_fork(const struct lu_env *env,
 
         ENTRY;
 
-        OBD_ALLOC_PTR(lpi);
-        if (lpi == NULL) {
-                CERROR("cannot alloc pointer\n");
-                RETURN(-ENOMEM);
-        }
-        lpi->lpi_loghandle = loghandle;
-        lpi->lpi_cb        = cb;
-        lpi->lpi_cbdata    = data;
-        lpi->lpi_catdata   = catdata;
+       OBD_ALLOC_PTR(lpi);
+       if (lpi == NULL) {
+               CERROR("cannot alloc pointer\n");
+               RETURN(-ENOMEM);
+       }
+       lpi->lpi_loghandle = loghandle;
+       lpi->lpi_cb        = cb;
+       lpi->lpi_cbdata    = data;
+       lpi->lpi_catdata   = catdata;
 
        if (fork) {
                struct task_struct *task;
@@ -660,6 +733,10 @@ int llog_process_or_fork(const struct lu_env *env,
                 * init the new one in llog_process_thread_daemonize. */
                lpi->lpi_env = NULL;
                init_completion(&lpi->lpi_completion);
+               /* take reference to current, so that
+                * llog_process_thread_daemonize() can use it to switch to
+                * namespace associated with current  */
+               lpi->lpi_reftask = current;
                task = kthread_run(llog_process_thread_daemonize, lpi,
                                   "llog_process_thread");
                if (IS_ERR(task)) {
@@ -892,11 +969,29 @@ int llog_write_rec(const struct lu_env *env, struct llog_handle *handle,
 
        ENTRY;
 
+       /* API sanity checks */
+       if (handle == NULL) {
+               CERROR("loghandle is missed\n");
+               RETURN(-EPROTO);
+       } else if (handle->lgh_obj == NULL) {
+               CERROR("loghandle %p with NULL object\n",
+                       handle);
+               RETURN(-EPROTO);
+       } else if (th == NULL) {
+               CERROR("%s: missed transaction handle\n",
+                       handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name);
+               RETURN(-EPROTO);
+       } else if (handle->lgh_hdr == NULL) {
+               CERROR("%s: loghandle %p with no header\n",
+                       handle->lgh_obj->do_lu.lo_dev->ld_obd->obd_name,
+                       handle);
+               RETURN(-EPROTO);
+       }
+
        rc = llog_handle2ops(handle, &lop);
        if (rc)
                RETURN(rc);
 
-       LASSERT(lop);
        if (lop->lop_write_rec == NULL)
                RETURN(-EOPNOTSUPP);
 
@@ -1244,3 +1339,22 @@ out_close:
        RETURN(rc);
 }
 EXPORT_SYMBOL(llog_backup);
+
+/* Get size of llog */
+__u64 llog_size(const struct lu_env *env, struct llog_handle *llh)
+{
+       int rc;
+       struct lu_attr la;
+
+       rc = llh->lgh_obj->do_ops->do_attr_get(env, llh->lgh_obj, &la);
+       if (rc) {
+               CERROR("%s: attr_get failed for "DFID": rc = %d\n",
+                      llh->lgh_ctxt->loc_obd->obd_name,
+                      PFID(&llh->lgh_id.lgl_oi.oi_fid), rc);
+               return 0;
+       }
+
+       return la.la_size;
+}
+EXPORT_SYMBOL(llog_size);
+