* Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2012, 2014 Intel Corporation.
+ * Copyright (c) 2012, 2015, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#define DEBUG_SUBSYSTEM S_LOG
+#include <dt_object.h>
+#include <llog_swab.h>
+#include <lustre_fid.h>
#include <obd.h>
#include <obd_class.h>
-#include <lustre_fid.h>
-#include <dt_object.h>
#include "llog_internal.h"
#include "local_storage.h"
}
/**
+ * Implementation of the llog_operations::lop_exist
+ *
+ * This function checks that llog exists on storage.
+ *
+ * \param[in] handle llog handle of the current llog
+ *
+ * \retval true if llog object exists and is not just destroyed
+ * \retval false if llog doesn't exist or just destroyed
+ */
+static int llog_osd_exist(struct llog_handle *handle)
+{
+ LASSERT(handle->lgh_obj);
+ return dt_object_exists(handle->lgh_obj) &&
+ !lu_object_is_dying(handle->lgh_obj->do_lu.lo_header);
+}
+
+static void *rec_tail(struct llog_rec_hdr *rec)
+{
+ return (void *)((char *)rec + rec->lrh_len -
+ sizeof(struct llog_rec_tail));
+}
+
+/**
* Write a padding record to the llog
*
* This function writes a padding record to the end of llog. That may
handle->lgh_hdr->llh_flags |= (flags & LLOG_F_EXT_MASK);
handle->lgh_last_idx = LLOG_HDR_TAIL(handle->lgh_hdr)->lrt_index;
+ handle->lgh_write_offset = lgi->lgi_attr.la_size;
RETURN(0);
}
struct dt_object *o;
__u32 chunk_size;
size_t left;
-
+ __u32 orig_last_idx;
+ __u64 orig_write_offset;
ENTRY;
- LASSERT(env);
llh = loghandle->lgh_hdr;
- LASSERT(llh);
o = loghandle->lgh_obj;
- LASSERT(o);
- LASSERT(th);
chunk_size = llh->llh_hdr.lrh_len;
CDEBUG(D_OTHER, "new record %x to "DFID"\n",
rec->lrh_type, PFID(lu_object_fid(&o->do_lu)));
+ if (!llog_osd_exist(loghandle))
+ RETURN(-ENOENT);
+
/* record length should not bigger than */
if (reclen > loghandle->lgh_hdr->llh_hdr.lrh_len)
RETURN(-E2BIG);
+ /* sanity check for fixed-records llog */
+ if (idx != LLOG_HEADER_IDX && (llh->llh_flags & LLOG_F_IS_FIXSIZE)) {
+ LASSERT(llh->llh_size != 0);
+ LASSERT(llh->llh_size == reclen);
+ }
+
rc = dt_attr_get(env, o, &lgi->lgi_attr);
if (rc)
RETURN(rc);
POSTID(&loghandle->lgh_id.lgl_oi), idx,
rec->lrh_len, (long long)lgi->lgi_off);
} else if (llh->llh_flags & LLOG_F_IS_FIXSIZE) {
- if (llh->llh_size == 0 ||
- llh->llh_size != rec->lrh_len) {
- CERROR("%s: wrong record size, llh_size is %u"
- " but record size is %u\n",
- o->do_lu.lo_dev->ld_obd->obd_name,
- llh->llh_size, rec->lrh_len);
- RETURN(-EINVAL);
- }
- lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen;
+ lgi->lgi_off = llh->llh_hdr.lrh_len +
+ (idx - 1) * reclen;
} else {
/* This can be result of lgh_cur_idx is not set during
* llog processing or llh_size is not set to proper
* process them page-at-a-time if needed. If it will cross a chunk
* boundary, write in a fake (but referenced) entry to pad the chunk.
*/
+
+
+ /* simulate ENOSPC when new plain llog is being added to the
+ * catalog */
+ if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LLOG_CREATE_FAILED2) &&
+ llh->llh_flags & LLOG_F_IS_CAT)
+ RETURN(-ENOSPC);
+
LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
+ orig_last_idx = loghandle->lgh_last_idx;
+ orig_write_offset = loghandle->lgh_write_offset;
lgi->lgi_off = lgi->lgi_attr.la_size;
left = chunk_size - (lgi->lgi_off & (chunk_size - 1));
/* NOTE: padding is a record, but no bit is set */
rc = llog_osd_pad(env, o, &lgi->lgi_off, left, index, th);
if (rc)
RETURN(rc);
+
+ if (dt_object_remote(o))
+ loghandle->lgh_write_offset = lgi->lgi_off;
+
loghandle->lgh_last_idx++; /* for pad rec */
}
- /* if it's the last idx in log file, then return -ENOSPC */
- if (loghandle->lgh_last_idx >= LLOG_HDR_BITMAP_SIZE(llh) - 1)
- RETURN(-ENOSPC);
+ /* if it's the last idx in log file, then return -ENOSPC
+ * or wrap around if a catalog */
+ if (llog_is_full(loghandle) ||
+ unlikely(llh->llh_flags & LLOG_F_IS_CAT &&
+ OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS) &&
+ loghandle->lgh_last_idx >= cfs_fail_val)) {
+ if (llh->llh_flags & LLOG_F_IS_CAT)
+ loghandle->lgh_last_idx = 0;
+ else
+ RETURN(-ENOSPC);
+ }
/* increment the last_idx along with llh_tail index, they should
* be equal for a llog lifetime */
}
llh->llh_count++;
- if (llh->llh_flags & LLOG_F_IS_FIXSIZE) {
- LASSERT(llh->llh_size == reclen);
- } else {
+ if (!(llh->llh_flags & LLOG_F_IS_FIXSIZE)) {
/* Update the minimum size of the llog record */
if (llh->llh_size == 0)
llh->llh_size = reclen;
if (rc)
GOTO(out, rc);
- rc = dt_attr_get(env, o, &lgi->lgi_attr);
- if (rc)
- GOTO(out, rc);
+ /* computed index can be used to determine offset for fixed-size
+ * records. This also allows to handle Catalog wrap around case */
+ if (llh->llh_flags & LLOG_F_IS_FIXSIZE) {
+ lgi->lgi_off = llh->llh_hdr.lrh_len + (index - 1) * reclen;
+ } else if (dt_object_remote(o)) {
+ lgi->lgi_off = max_t(__u64, loghandle->lgh_write_offset,
+ lgi->lgi_off);
+ } else {
+ rc = dt_attr_get(env, o, &lgi->lgi_attr);
+ if (rc)
+ GOTO(out, rc);
+
+ LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
+ lgi->lgi_off = max_t(__u64, lgi->lgi_attr.la_size,
+ lgi->lgi_off);
+ }
- LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
- lgi->lgi_off = max_t(__u64, lgi->lgi_attr.la_size, lgi->lgi_off);
lgi->lgi_buf.lb_len = reclen;
lgi->lgi_buf.lb_buf = rec;
rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
if (rc < 0)
GOTO(out, rc);
- CDEBUG(D_OTHER, "added record "DOSTID": idx: %u, %u off"LPU64"\n",
- POSTID(&loghandle->lgh_id.lgl_oi), index, rec->lrh_len,
+ if (dt_object_remote(o))
+ loghandle->lgh_write_offset = lgi->lgi_off;
+
+ CDEBUG(D_HA, "added record "DFID": idx: %u, %u off%llu\n",
+ PFID(lu_object_fid(&o->do_lu)), index, rec->lrh_len,
lgi->lgi_off);
if (reccookie != NULL) {
reccookie->lgc_lgl = loghandle->lgh_id;
mutex_unlock(&loghandle->lgh_hdr_mutex);
/* restore llog last_idx */
- loghandle->lgh_last_idx--;
+ if (dt_object_remote(o)) {
+ loghandle->lgh_last_idx = orig_last_idx;
+ loghandle->lgh_write_offset = orig_write_offset;
+ } else if (--loghandle->lgh_last_idx == 0 &&
+ (llh->llh_flags & LLOG_F_IS_CAT) && llh->llh_cat_idx != 0) {
+ /* catalog had just wrap-around case */
+ loghandle->lgh_last_idx = LLOG_HDR_BITMAP_SIZE(llh) - 1;
+ }
+
LLOG_HDR_TAIL(llh)->lrt_index = loghandle->lgh_last_idx;
RETURN(rc);
if (len == 0 || len & (chunk_size - 1))
RETURN(-EINVAL);
- CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off "LPU64")\n",
- next_idx, *cur_idx, *cur_offset);
-
LASSERT(loghandle);
LASSERT(loghandle->lgh_ctxt);
if (rc)
GOTO(out, rc);
+ CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off"
+ "%llu), size %llu\n", next_idx, *cur_idx,
+ *cur_offset, lgi->lgi_attr.la_size);
+
while (*cur_offset < lgi->lgi_attr.la_size) {
struct llog_rec_hdr *rec, *last_rec;
struct llog_rec_tail *tail;
goto retry;
CERROR("%s: can't read llog block from log "DFID
- " offset "LPU64": rc = %d\n",
+ " offset %llu: rc = %d\n",
o->do_lu.lo_dev->ld_obd->obd_name,
PFID(lu_object_fid(&o->do_lu)), *cur_offset,
rc);
goto retry;
CERROR("%s: invalid llog block at log id "DOSTID"/%u "
- "offset "LPU64"\n",
+ "offset %llu\n",
o->do_lu.lo_dev->ld_obd->obd_name,
POSTID(&loghandle->lgh_id.lgl_oi),
loghandle->lgh_id.lgl_ogen, *cur_offset);
if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
lustre_swab_llog_rec(last_rec);
- LASSERT(last_rec->lrh_index == tail->lrt_index);
+
+ if (last_rec->lrh_index != tail->lrt_index) {
+ CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
+ "offset %llu last_rec idx %u tail idx %u\n",
+ o->do_lu.lo_dev->ld_obd->obd_name,
+ POSTID(&loghandle->lgh_id.lgl_oi),
+ loghandle->lgh_id.lgl_ogen, *cur_offset,
+ last_rec->lrh_index, tail->lrt_index);
+ GOTO(out, rc = -EINVAL);
+ }
*cur_idx = tail->lrt_index;
/* this shouldn't happen */
if (tail->lrt_index == 0) {
CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
- "offset "LPU64"\n",
+ "offset %llu bytes %d\n",
o->do_lu.lo_dev->ld_obd->obd_name,
POSTID(&loghandle->lgh_id.lgl_oi),
- loghandle->lgh_id.lgl_ogen, *cur_offset);
+ loghandle->lgh_id.lgl_ogen, *cur_offset, rc);
GOTO(out, rc = -EINVAL);
}
if (tail->lrt_index < next_idx) {
rc = dt_read(env, o, &lgi->lgi_buf, &cur_offset);
if (rc < 0) {
CERROR("%s: can't read llog block from log "DFID
- " offset "LPU64": rc = %d\n",
+ " offset %llu: rc = %d\n",
o->do_lu.lo_dev->ld_obd->obd_name,
PFID(lu_object_fid(&o->do_lu)), cur_offset, rc);
GOTO(out, rc);
if (rc < sizeof(*tail)) {
CERROR("%s: invalid llog block at log id "DOSTID"/%u "
- "offset "LPU64"\n",
+ "offset %llu\n",
o->do_lu.lo_dev->ld_obd->obd_name,
POSTID(&loghandle->lgh_id.lgl_oi),
loghandle->lgh_id.lgl_ogen, cur_offset);
/* this shouldn't happen */
if (tail->lrt_index == 0) {
CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
- "offset "LPU64"\n",
+ "offset %llu\n",
o->do_lu.lo_dev->ld_obd->obd_name,
POSTID(&loghandle->lgh_id.lgl_oi),
loghandle->lgh_id.lgl_ogen, cur_offset);
struct ls_device *ls;
struct local_oid_storage *los = NULL;
int rc = 0;
+ bool new_id = false;
ENTRY;
/* generate fid for new llog */
rc = local_object_fid_generate(env, los,
&lgi->lgi_fid);
+ new_id = true;
}
if (rc < 0)
GOTO(out, rc);
} else {
LASSERTF(open_param & LLOG_OPEN_NEW, "%#x\n", open_param);
/* generate fid for new llog */
+generate:
rc = local_object_fid_generate(env, los, &lgi->lgi_fid);
if (rc < 0)
GOTO(out, rc);
+ new_id = true;
}
o = ls_locate(env, ls, &lgi->lgi_fid, NULL);
if (IS_ERR(o))
GOTO(out_name, rc = PTR_ERR(o));
+ if (dt_object_exists(o) && new_id) {
+ /* llog exists with just generated ID, e.g. some old llog file
+ * still is in use or is orphan, drop a warn and skip it. */
+ CDEBUG(D_INFO, "%s: llog exists with the same FID: "DFID
+ ", skipping\n",
+ o->do_lu.lo_dev->ld_obd->obd_name,
+ PFID(lu_object_fid(&o->do_lu)));
+ lu_object_put(env, &o->do_lu);
+ /* just skip this llog ID, we shouldn't delete it because we
+ * don't know exactly what is its purpose and state. */
+ goto generate;
+ }
+
after_open:
/* No new llog is expected but doesn't exist */
if (open_param != LLOG_OPEN_NEW && !dt_object_exists(o))
}
/**
- * Implementation of the llog_operations::lop_exist
- *
- * This function checks that llog exists on storage.
- *
- * \param[in] handle llog handle of the current llog
- *
- * \retval true if llog object exists and is not just destroyed
- * \retval false if llog doesn't exist or just destroyed
- */
-static int llog_osd_exist(struct llog_handle *handle)
-{
- LASSERT(handle->lgh_obj);
- return (dt_object_exists(handle->lgh_obj) &&
- !lu_object_is_dying(handle->lgh_obj->do_lu.lo_header));
-}
-
-/**
* Get dir for regular fid log object
*
* Get directory for regular fid log object, and these regular fid log