* GPL HEADER END
*/
/*
- * Copyright (c) 2013, 2014, Intel Corporation.
+ * Copyright (c) 2013, 2017, Intel Corporation.
*/
/*
* lustre/osp/osp_md_object.c
#define DEBUG_SUBSYSTEM S_MDS
+#include <llog_swab.h>
#include <lustre_log.h>
#include "osp_internal.h"
-static const char dot[] = ".";
-static const char dotdot[] = "..";
+#define OUT_UPDATE_BUFFER_SIZE_ADD 4096
+#define OUT_UPDATE_BUFFER_SIZE_MAX (256 * 4096) /* 1M update size now */
/**
- * Add OUT_CREATE sub-request into the OUT RPC.
+ * Interpreter call for object creation
*
- * Note: if the object has already been created, we must add object
- * destroy sub-request ahead of the create, so it will destroy then
- * re-create the object.
+ * Object creation interpreter, which will be called after creating
+ * the remote object to set flags and status.
*
* \param[in] env execution environment
- * \param[in] dt object to be created
- * \param[in] attr attribute of the created object
- * \param[in] hint creation hint
- * \param[in] dof creation format information
- * \param[in] th the transaction handle
+ * \param[in] reply update reply
+ * \param[in] req ptlrpc update request for creating object
+ * \param[in] obj object to be created
+ * \param[in] data data used in this function.
+ * \param[in] index index(position) of create update in the whole
+ * updates
+ * \param[in] rc update result on the remote MDT.
*
* \retval only return 0 for now
*/
-static int __osp_md_declare_object_create(const struct lu_env *env,
- struct dt_object *dt,
- struct lu_attr *attr,
- struct dt_allocation_hint *hint,
- struct dt_object_format *dof,
- struct thandle *th)
+static int osp_create_interpreter(const struct lu_env *env,
+ struct object_update_reply *reply,
+ struct ptlrpc_request *req,
+ struct osp_object *obj,
+ void *data, int index, int rc)
{
- struct dt_update_request *update;
- int rc;
-
- update = thandle_to_dt_update_request(th);
- LASSERT(update != NULL);
-
- if (lu_object_exists(&dt->do_lu)) {
- /* If the object already exists, we needs to destroy
- * this orphan object first.
- *
- * The scenario might happen in this case
- *
- * 1. client send remote create to MDT0.
- * 2. MDT0 send create update to MDT1.
- * 3. MDT1 finished create synchronously.
- * 4. MDT0 failed and reboot.
- * 5. client resend remote create to MDT0.
- * 6. MDT0 tries to resend create update to MDT1,
- * but find the object already exists
- */
- CDEBUG(D_HA, "%s: object "DFID" exists, destroy this orphan\n",
- dt->do_lu.lo_dev->ld_obd->obd_name,
- PFID(lu_object_fid(&dt->do_lu)));
-
- rc = out_ref_del_pack(env, &update->dur_buf,
- lu_object_fid(&dt->do_lu),
- update->dur_batchid);
- if (rc != 0)
- GOTO(out, rc);
-
- if (S_ISDIR(lu_object_attr(&dt->do_lu))) {
- /* decrease for ".." */
- rc = out_ref_del_pack(env, &update->dur_buf,
- lu_object_fid(&dt->do_lu),
- update->dur_batchid);
- if (rc != 0)
- GOTO(out, rc);
- }
-
- rc = out_object_destroy_pack(env, &update->dur_buf,
- lu_object_fid(&dt->do_lu),
- update->dur_batchid);
- if (rc != 0)
- GOTO(out, rc);
-
- dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
- /* Increase batchid to add this orphan object deletion
- * to separate transaction */
- update_inc_batchid(update);
+ if (rc != 0 && rc != -EEXIST) {
+ obj->opo_obj.do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
+ obj->opo_non_exist = 1;
}
- rc = out_create_pack(env, &update->dur_buf,
- lu_object_fid(&dt->do_lu), attr, hint, dof,
- update->dur_batchid);
- if (rc != 0)
- GOTO(out, rc);
-out:
- if (rc)
- CERROR("%s: Insert update error: rc = %d\n",
- dt->do_lu.lo_dev->ld_obd->obd_name, rc);
+ /*
+ * invalidate opo cache for the object after the object is created, so
+ * attr_get will try to get attr from remote object.
+ */
+ osp_obj_invalidate_cache(obj);
- return rc;
+ return 0;
}
/**
* Implementation of dt_object_operations::do_declare_create
*
- * For non-remote transaction, it will add an OUT_CREATE sub-request
- * into the OUT RPC that will be flushed when the transaction start.
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
*
* \param[in] env execution environment
* \param[in] dt remote object to be created
* \param[in] dof creation format information
* \param[in] th the transaction handle
*
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
+ * \retval 0 if preparation succeeds.
+ * \retval negative errno if preparation fails.
*/
-int osp_md_declare_object_create(const struct lu_env *env,
- struct dt_object *dt,
- struct lu_attr *attr,
- struct dt_allocation_hint *hint,
- struct dt_object_format *dof,
- struct thandle *th)
+int osp_md_declare_create(const struct lu_env *env, struct dt_object *dt,
+ struct lu_attr *attr, struct dt_allocation_hint *hint,
+ struct dt_object_format *dof, struct thandle *th)
{
- int rc = 0;
+ return osp_trans_update_request_create(th);
+}
- if (!is_only_remote_trans(th)) {
- rc = __osp_md_declare_object_create(env, dt, attr, hint,
- dof, th);
+struct object_update *
+update_buffer_get_update(struct object_update_request *request,
+ unsigned int index)
+{
+ void *ptr;
+ int i;
- CDEBUG(D_INFO, "declare create md_object "DFID": rc = %d\n",
- PFID(&dt->do_lu.lo_header->loh_fid), rc);
- }
+ if (index > request->ourq_count)
+ return NULL;
- return rc;
+ ptr = &request->ourq_updates[0];
+ for (i = 0; i < index; i++)
+ ptr += object_update_size(ptr);
+
+ return ptr;
}
/**
* Implementation of dt_object_operations::do_create
*
- * For remote transaction, it will add an OUT_CREATE sub-request into
- * the OUT RPC that will be flushed when the transaction stop.
- *
- * It sets necessary flags for created object. In DNE phase I,
- * remote updates are actually executed during transaction start,
- * i.e. the object has already been created when calling this method.
+ * It adds an OUT_CREATE sub-request into the OUT RPC that will be flushed
+ * when the transaction stop, and sets necessary flags for created object.
*
* \param[in] env execution environment
* \param[in] dt object to be created
* \param[in] dof creation format information
* \param[in] th the transaction handle
*
- * \retval only return 0 for now
+ * \retval 0 if packing creation succeeds.
+ * \retval negative errno if packing creation fails.
*/
-int osp_md_object_create(const struct lu_env *env, struct dt_object *dt,
- struct lu_attr *attr, struct dt_allocation_hint *hint,
- struct dt_object_format *dof, struct thandle *th)
+int osp_md_create(const struct lu_env *env, struct dt_object *dt,
+ struct lu_attr *attr, struct dt_allocation_hint *hint,
+ struct dt_object_format *dof, struct thandle *th)
{
- int rc = 0;
-
- if (is_only_remote_trans(th)) {
- rc = __osp_md_declare_object_create(env, dt, attr, hint,
- dof, th);
+ struct osp_update_request *update;
+ struct osp_object *obj = dt2osp_obj(dt);
+ int rc;
- CDEBUG(D_INFO, "create md_object "DFID": rc = %d\n",
- PFID(&dt->do_lu.lo_header->loh_fid), rc);
- }
+ update = thandle_to_osp_update_request(th);
+ LASSERT(update != NULL);
- if (rc == 0) {
- dt->do_lu.lo_header->loh_attr |= LOHA_EXISTS |
- (attr->la_mode & S_IFMT);
- dt2osp_obj(dt)->opo_non_exist = 0;
- }
+ LASSERT(attr->la_valid & LA_TYPE);
+ rc = OSP_UPDATE_RPC_PACK(env, out_create_pack, update,
+ lu_object_fid(&dt->do_lu), attr, hint, dof);
+ if (rc != 0)
+ GOTO(out, rc);
- return rc;
-}
+ rc = osp_insert_update_callback(env, update, dt2osp_obj(dt), NULL,
+ osp_create_interpreter);
-/**
- * Add OUT_REF_DEL sub-request into the OUT RPC.
- *
- * \param[in] env execution environment
- * \param[in] dt object to decrease the reference count.
- * \param[in] th the transaction handle of refcount decrease.
- *
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
- */
-static int __osp_md_ref_del(const struct lu_env *env, struct dt_object *dt,
- struct thandle *th)
-{
- struct dt_update_request *update;
- int rc;
+ if (rc < 0)
+ GOTO(out, rc);
- update = thandle_to_dt_update_request(th);
- LASSERT(update != NULL);
+ dt->do_lu.lo_header->loh_attr |= LOHA_EXISTS | (attr->la_mode & S_IFMT);
+ dt2osp_obj(dt)->opo_non_exist = 0;
+ obj->opo_stale = 0;
- rc = out_ref_del_pack(env, &update->dur_buf,
- lu_object_fid(&dt->do_lu),
- update->dur_batchid);
+ obj->opo_attr = *attr;
+out:
return rc;
}
/**
* Implementation of dt_object_operations::do_declare_ref_del
*
- * For non-remote transaction, it will add an OUT_REF_DEL sub-request
- * into the OUT RPC that will be flushed when the transaction start.
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
*
* \param[in] env execution environment
* \param[in] dt object to decrease the reference count.
* \param[in] th the transaction handle of refcount decrease.
*
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
+ * \retval 0 if preparation succeeds.
+ * \retval negative errno if preparation fails.
*/
static int osp_md_declare_ref_del(const struct lu_env *env,
struct dt_object *dt, struct thandle *th)
{
- int rc = 0;
-
- if (!is_only_remote_trans(th)) {
- rc = __osp_md_ref_del(env, dt, th);
-
- CDEBUG(D_INFO, "declare ref del "DFID": rc = %d\n",
- PFID(&dt->do_lu.lo_header->loh_fid), rc);
- }
-
- return rc;
+ return osp_trans_update_request_create(th);
}
/**
* Implementation of dt_object_operations::do_ref_del
*
- * For remote transaction, it will add an OUT_REF_DEL sub-request into
- * the OUT RPC that will be flushed when the transaction stop.
+ * Add an OUT_REF_DEL sub-request into the OUT RPC that will be
+ * flushed when the transaction stop.
*
* \param[in] env execution environment
* \param[in] dt object to decrease the reference count
* \param[in] th the transaction handle
*
- * \retval only return 0 for now
+ * \retval 0 if packing ref_del succeeds.
+ * \retval negative errno if packing fails.
*/
static int osp_md_ref_del(const struct lu_env *env, struct dt_object *dt,
struct thandle *th)
{
- int rc = 0;
+ struct osp_update_request *update;
+ int rc;
- if (is_only_remote_trans(th)) {
- rc = __osp_md_ref_del(env, dt, th);
-
- CDEBUG(D_INFO, "ref del "DFID": rc = %d\n",
- PFID(&dt->do_lu.lo_header->loh_fid), rc);
- }
-
- return rc;
-}
-
-/**
- * Add OUT_REF_ADD sub-request into the OUT RPC.
- *
- * \param[in] env execution environment
- * \param[in] dt object on which to increase the reference count.
- * \param[in] th the transaction handle.
- *
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
- */
-static int __osp_md_ref_add(const struct lu_env *env, struct dt_object *dt,
- struct thandle *th)
-{
- struct dt_update_request *update;
- int rc;
-
- update = thandle_to_dt_update_request(th);
+ update = thandle_to_osp_update_request(th);
LASSERT(update != NULL);
- rc = out_ref_add_pack(env, &update->dur_buf,
- lu_object_fid(&dt->do_lu),
- update->dur_batchid);
-
+ rc = OSP_UPDATE_RPC_PACK(env, out_ref_del_pack, update,
+ lu_object_fid(&dt->do_lu));
return rc;
}
/**
* Implementation of dt_object_operations::do_declare_ref_del
*
- * For non-remote transaction, it will add an OUT_REF_ADD sub-request
- * into the OUT RPC that will be flushed when the transaction start.
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
*
* \param[in] env execution environment
* \param[in] dt object on which to increase the reference count.
* \param[in] th the transaction handle.
*
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
+ * \retval 0 if preparation succeeds.
+ * \retval negative errno if preparation fails.
*/
static int osp_md_declare_ref_add(const struct lu_env *env,
struct dt_object *dt, struct thandle *th)
{
- int rc = 0;
-
- if (!is_only_remote_trans(th)) {
- rc = __osp_md_ref_add(env, dt, th);
-
- CDEBUG(D_INFO, "declare ref add "DFID": rc = %d\n",
- PFID(&dt->do_lu.lo_header->loh_fid), rc);
- }
-
- return rc;
+ return osp_trans_update_request_create(th);
}
/**
* Implementation of dt_object_operations::do_ref_add
*
- * For remote transaction, it will add an OUT_REF_ADD sub-request into
- * the OUT RPC that will be flushed when the transaction stop.
+ * Add an OUT_REF_ADD sub-request into the OUT RPC that will be flushed
+ * when the transaction stop.
*
* \param[in] env execution environment
* \param[in] dt object on which to increase the reference count
* \param[in] th the transaction handle
*
- * \retval only return 0 for now
+ * \retval 0 if packing ref_add succeeds.
+ * \retval negative errno if packing fails.
*/
static int osp_md_ref_add(const struct lu_env *env, struct dt_object *dt,
struct thandle *th)
{
- int rc = 0;
+ struct osp_update_request *update;
+ int rc;
- if (is_only_remote_trans(th)) {
- rc = __osp_md_ref_add(env, dt, th);
-
- CDEBUG(D_INFO, "ref add "DFID": rc = %d\n",
- PFID(&dt->do_lu.lo_header->loh_fid), rc);
- }
+ update = thandle_to_osp_update_request(th);
+ LASSERT(update != NULL);
+ rc = OSP_UPDATE_RPC_PACK(env, out_ref_add_pack, update,
+ lu_object_fid(&dt->do_lu));
return rc;
}
}
/**
- * Add OUT_ATTR_SET sub-request into the OUT RPC.
- *
- * \param[in] env execution environment
- * \param[in] dt object on which to set attributes
- * \param[in] attr attributes to be set
- * \param[in] th the transaction handle
- *
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
- */
-int __osp_md_attr_set(const struct lu_env *env, struct dt_object *dt,
- const struct lu_attr *attr, struct thandle *th)
-{
- struct dt_update_request *update;
- int rc;
-
- update = thandle_to_dt_update_request(th);
- LASSERT(update != NULL);
-
- rc = out_attr_set_pack(env, &update->dur_buf,
- lu_object_fid(&dt->do_lu), attr,
- update->dur_batchid);
-
- return rc;
-}
-
-/**
* Implementation of dt_object_operations::do_declare_attr_get
*
- * Declare setting attributes to the specified remote object.
- *
- * If the transaction is a non-remote transaction, then add the OUT_ATTR_SET
- * sub-request into the OUT RPC that will be flushed when the transaction start.
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
*
* \param[in] env execution environment
* \param[in] dt object on which to set attributes
* \param[in] attr attributes to be set
* \param[in] th the transaction handle
*
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
+ * \retval 0 if preparation succeeds.
+ * \retval negative errno if preparation fails.
*/
int osp_md_declare_attr_set(const struct lu_env *env, struct dt_object *dt,
const struct lu_attr *attr, struct thandle *th)
{
- int rc = 0;
-
- if (!is_only_remote_trans(th)) {
- rc = __osp_md_attr_set(env, dt, attr, th);
-
- CDEBUG(D_INFO, "declare attr set md_object "DFID": rc = %d\n",
- PFID(&dt->do_lu.lo_header->loh_fid), rc);
- }
-
- return rc;
+ return osp_trans_update_request_create(th);
}
/**
*
* Set attributes to the specified remote object.
*
- * If the transaction is a remote transaction, then add the OUT_ATTR_SET
- * sub-request into the OUT RPC that will be flushed when the transaction stop.
+ * Add the OUT_ATTR_SET sub-request into the OUT RPC that will be flushed
+ * when the transaction stop.
*
* \param[in] env execution environment
* \param[in] dt object to set attributes
* \param[in] attr attributes to be set
* \param[in] th the transaction handle
*
- * \retval only return 0 for now
+ * \retval 0 if packing attr_set succeeds.
+ * \retval negative errno if packing fails.
*/
int osp_md_attr_set(const struct lu_env *env, struct dt_object *dt,
const struct lu_attr *attr, struct thandle *th)
{
- int rc = 0;
-
- if (is_only_remote_trans(th)) {
- rc = __osp_md_attr_set(env, dt, attr, th);
+ struct osp_update_request *update;
+ int rc;
- CDEBUG(D_INFO, "attr set md_object "DFID": rc = %d\n",
- PFID(&dt->do_lu.lo_header->loh_fid), rc);
- }
+ update = thandle_to_osp_update_request(th);
+ LASSERT(update != NULL);
+ rc = OSP_UPDATE_RPC_PACK(env, out_attr_set_pack, update,
+ lu_object_fid(&dt->do_lu), attr);
return rc;
}
/**
* Implementation of dt_object_operations::do_read_lock
*
- * osp_md_object_{read,write}_lock() will only lock the remote object in the
+ * osp_md_{read,write}_lock() will only lock the remote object in the
* local cache, which uses the semaphore (opo_sem) inside the osp_object to
* lock the object. Note: it will not lock the object in the whole cluster,
* which relies on the LDLM lock.
*
* \param[in] env execution environment
* \param[in] dt object to be locked
- * \param[in] role lock role from MDD layer, see mdd_object_role().
+ * \param[in] role lock role from MDD layer, see dt_object_role().
*/
-static void osp_md_object_read_lock(const struct lu_env *env,
- struct dt_object *dt, unsigned role)
+static void osp_md_read_lock(const struct lu_env *env, struct dt_object *dt,
+ unsigned role)
{
struct osp_object *obj = dt2osp_obj(dt);
*
* \param[in] env execution environment
* \param[in] dt object to be locked
- * \param[in] role lock role from MDD layer, see mdd_object_role().
+ * \param[in] role lock role from MDD layer, see dt_object_role().
*/
-static void osp_md_object_write_lock(const struct lu_env *env,
- struct dt_object *dt, unsigned role)
+static void osp_md_write_lock(const struct lu_env *env, struct dt_object *dt,
+ unsigned role)
{
struct osp_object *obj = dt2osp_obj(dt);
* \param[in] env execution environment
* \param[in] dt object to be unlocked
*/
-static void osp_md_object_read_unlock(const struct lu_env *env,
- struct dt_object *dt)
+static void osp_md_read_unlock(const struct lu_env *env, struct dt_object *dt)
{
struct osp_object *obj = dt2osp_obj(dt);
* \param[in] env execution environment
* \param[in] dt object to be unlocked
*/
-static void osp_md_object_write_unlock(const struct lu_env *env,
- struct dt_object *dt)
+static void osp_md_write_unlock(const struct lu_env *env, struct dt_object *dt)
{
struct osp_object *obj = dt2osp_obj(dt);
* \param[in] env execution environment
* \param[in] dt object to be tested
*/
-static int osp_md_object_write_locked(const struct lu_env *env,
- struct dt_object *dt)
+static int osp_md_write_locked(const struct lu_env *env, struct dt_object *dt)
{
struct osp_object *obj = dt2osp_obj(dt);
struct lu_buf *lbuf = &osp_env_info(env)->osi_lb2;
struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
struct dt_device *dt_dev = &osp->opd_dt_dev;
- struct dt_update_request *update;
+ struct osp_update_request *update;
struct object_update_reply *reply;
struct ptlrpc_request *req = NULL;
struct lu_fid *fid;
* just create an update buffer, instead of attaching the
* update_remote list of the thandle.
*/
- update = dt_update_request_create(dt_dev);
+ update = osp_update_request_create(dt_dev);
if (IS_ERR(update))
RETURN(PTR_ERR(update));
- rc = out_index_lookup_pack(env, &update->dur_buf,
- lu_object_fid(&dt->do_lu), rec, key);
+ rc = OSP_UPDATE_RPC_PACK(env, out_index_lookup_pack, update,
+ lu_object_fid(&dt->do_lu), rec, key);
if (rc != 0) {
CERROR("%s: Insert update error: rc = %d\n",
dt_dev->dd_lu_dev.ld_obd->obd_name, rc);
GOTO(out, rc);
}
- rc = osp_remote_sync(env, osp, update, &req, false);
+ rc = osp_remote_sync(env, osp, update, &req);
if (rc < 0)
GOTO(out, rc);
if (req != NULL)
ptlrpc_req_finished(req);
- dt_update_request_destroy(update);
-
- return rc;
-}
-
-/**
- * Add OUT_INDEX_INSERT sub-request into the OUT RPC.
- *
- * \param[in] env execution environment
- * \param[in] dt object for which to insert index
- * \param[in] rec record of the index which will be inserted
- * \param[in] key key of the index which will be inserted
- * \param[in] th the transaction handle
- *
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
- */
-static int __osp_md_index_insert(const struct lu_env *env,
- struct dt_object *dt,
- const struct dt_rec *rec,
- const struct dt_key *key,
- struct thandle *th)
-{
- struct osp_thandle *oth = thandle_to_osp_thandle(th);
- struct dt_update_request *update = oth->ot_dur;
- int rc;
-
-
- rc = out_index_insert_pack(env, &update->dur_buf,
- lu_object_fid(&dt->do_lu), rec, key,
- update->dur_batchid);
- if (rc != 0)
- return rc;
-
- /* Before async update is allowed, if it will insert remote
- * name entry, it should make sure the local object is created,
- * i.e. the remote update RPC should be sent after local
- * update(create object) */
- oth->ot_send_updates_after_local_trans = true;
+ osp_update_request_destroy(env, update);
return rc;
}
/**
* Implementation of dt_index_operations::dio_declare_insert
*
- * For non-remote transaction, it will add an OUT_INDEX_INSERT sub-request
- * into the OUT RPC that will be flushed when the transaction start.
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
*
* \param[in] env execution environment
* \param[in] dt object for which to insert index
* \param[in] key key of the index which will be inserted
* \param[in] th the transaction handle
*
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
+ * \retval 0 if preparation succeeds.
+ * \retval negative errno if preparation fails.
*/
static int osp_md_declare_index_insert(const struct lu_env *env,
struct dt_object *dt,
const struct dt_key *key,
struct thandle *th)
{
- int rc = 0;
-
- if (!is_only_remote_trans(th)) {
- rc = __osp_md_index_insert(env, dt, rec, key, th);
-
- CDEBUG(D_INFO, "declare index insert "DFID" key %s, rec "DFID
- ": rc = %d\n", PFID(&dt->do_lu.lo_header->loh_fid),
- (char *)key,
- PFID(((struct dt_insert_rec *)rec)->rec_fid), rc);
- }
-
- return rc;
+ return osp_trans_update_request_create(th);
}
/**
* Implementation of dt_index_operations::dio_insert
*
- * For remote transaction, it will add an OUT_INDEX_INSERT sub-request
- * into the OUT RPC that will be flushed when the transaction stop.
+ * Add an OUT_INDEX_INSERT sub-request into the OUT RPC that will
+ * be flushed when the transaction stop.
*
* \param[in] env execution environment
* \param[in] dt object for which to insert index
* \param[in] rec record of the index to be inserted
* \param[in] key key of the index to be inserted
* \param[in] th the transaction handle
- * \param[in] ignore_quota quota enforcement for insert
*
- * \retval only return 0 for now
+ * \retval 0 if packing index insert succeeds.
+ * \retval negative errno if packing fails.
*/
-static int osp_md_index_insert(const struct lu_env *env,
- struct dt_object *dt,
+static int osp_md_index_insert(const struct lu_env *env, struct dt_object *dt,
const struct dt_rec *rec,
- const struct dt_key *key,
- struct thandle *th,
- int ignore_quota)
-{
- int rc = 0;
-
- if (is_only_remote_trans(th)) {
- rc = __osp_md_index_insert(env, dt, rec, key, th);
-
- CDEBUG(D_INFO, "index insert "DFID" key %s, rec "DFID
- ": rc = %d\n", PFID(&dt->do_lu.lo_header->loh_fid),
- (char *)key,
- PFID(((struct dt_insert_rec *)rec)->rec_fid), rc);
- }
-
- return rc;
-}
-
-/**
- * Add OUT_INDEX_DELETE sub-request into the OUT RPC.
- *
- * \param[in] env execution environment
- * \param[in] dt object for which to delete index
- * \param[in] key key of the index
- * \param[in] th the transaction handle
- *
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
- */
-static int __osp_md_index_delete(const struct lu_env *env,
- struct dt_object *dt,
- const struct dt_key *key,
- struct thandle *th)
+ const struct dt_key *key, struct thandle *th)
{
- struct dt_update_request *update;
- int rc;
+ struct osp_update_request *update;
+ int rc;
- update = thandle_to_dt_update_request(th);
+ update = thandle_to_osp_update_request(th);
LASSERT(update != NULL);
- rc = out_index_delete_pack(env, &update->dur_buf,
- lu_object_fid(&dt->do_lu), key,
- update->dur_batchid);
+ rc = OSP_UPDATE_RPC_PACK(env, out_index_insert_pack, update,
+ lu_object_fid(&dt->do_lu), rec, key);
return rc;
}
/**
* Implementation of dt_index_operations::dio_declare_delete
*
- * For non-remote transaction, it will add an OUT_INDEX_DELETE sub-request
- * into the OUT RPC that will be flushed when the transaction start.
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
*
* \param[in] env execution environment
* \param[in] dt object for which to delete index
* \param[in] key key of the index
* \param[in] th the transaction handle
*
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
+ * \retval 0 if preparation succeeds.
+ * \retval negative errno if preparation fails.
*/
static int osp_md_declare_index_delete(const struct lu_env *env,
struct dt_object *dt,
const struct dt_key *key,
struct thandle *th)
{
- int rc = 0;
-
- if (!is_only_remote_trans(th)) {
- rc = __osp_md_index_delete(env, dt, key, th);
-
- CDEBUG(D_INFO, "declare index delete "DFID" %s: rc = %d\n",
- PFID(&dt->do_lu.lo_header->loh_fid), (char *)key, rc);
- }
-
- return rc;
+ return osp_trans_update_request_create(th);
}
/**
* Implementation of dt_index_operations::dio_delete
*
- * For remote transaction, it will add an OUT_INDEX_DELETE sub-request
- * into the OUT RPC that will be flushed when the transaction stop.
+ * Add an OUT_INDEX_DELETE sub-request into the OUT RPC that will
+ * be flushed when the transaction stop.
*
* \param[in] env execution environment
* \param[in] dt object for which to delete index
* \param[in] key key of the index which will be deleted
* \param[in] th the transaction handle
*
- * \retval only return 0 for now
+ * \retval 0 if packing index delete succeeds.
+ * \retval negative errno if packing fails.
*/
static int osp_md_index_delete(const struct lu_env *env,
struct dt_object *dt,
const struct dt_key *key,
struct thandle *th)
{
- int rc = 0;
+ struct osp_update_request *update;
+ int rc;
- if (is_only_remote_trans(th)) {
- rc = __osp_md_index_delete(env, dt, key, th);
+ update = thandle_to_osp_update_request(th);
+ LASSERT(update != NULL);
- CDEBUG(D_INFO, "index delete "DFID" %s: rc = %d\n",
- PFID(&dt->do_lu.lo_header->loh_fid), (char *)key, rc);
- }
+ rc = OSP_UPDATE_RPC_PACK(env, out_index_delete_pack, update,
+ lu_object_fid(&dt->do_lu), key);
return rc;
}
* \param[in] env execution environment
* \param[in] di iterator of this iteration
*
- * \retval 0 if the pointer is advanced successfuly.
+ * \retval 0 if the pointer is advanced successfully.
* \retval 1 if it reaches to the end of the index object.
* \retval negative errno if the pointer cannot be advanced.
*/
};
/**
+ * Implement OSP layer dt_object_operations::do_xattr_list() interface.
+ *
+ * List extended attribute from the specified MDT/OST object, result is not
+ * cached because this is called by directory migration only.
+ *
+ * \param[in] env pointer to the thread context
+ * \param[in] dt pointer to the OSP layer dt_object
+ * \param[out] buf pointer to the lu_buf to hold the extended attribute
+ *
+ * \retval positive bytes used/required in the buffer
+ * \retval negative error number on failure
+ */
+static int osp_md_xattr_list(const struct lu_env *env, struct dt_object *dt,
+ const struct lu_buf *buf)
+{
+ struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
+ struct osp_object *obj = dt2osp_obj(dt);
+ struct dt_device *dev = &osp->opd_dt_dev;
+ struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2;
+ struct osp_update_request *update = NULL;
+ struct ptlrpc_request *req = NULL;
+ struct object_update_reply *reply;
+ const char *dname = dt->do_lu.lo_dev->ld_obd->obd_name;
+ int rc = 0;
+
+ ENTRY;
+
+ LASSERT(buf);
+
+ if (unlikely(obj->opo_non_exist))
+ RETURN(-ENOENT);
+
+ update = osp_update_request_create(dev);
+ if (IS_ERR(update))
+ RETURN(PTR_ERR(update));
+
+ rc = OSP_UPDATE_RPC_PACK(env, out_xattr_list_pack, update,
+ lu_object_fid(&dt->do_lu), buf->lb_len);
+ if (rc) {
+ CERROR("%s: Insert update error "DFID": rc = %d\n",
+ dname, PFID(lu_object_fid(&dt->do_lu)), rc);
+ GOTO(out, rc);
+ }
+
+ rc = osp_remote_sync(env, osp, update, &req);
+ if (rc < 0) {
+ if (rc == -ENOENT) {
+ dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
+ obj->opo_non_exist = 1;
+ }
+ GOTO(out, rc);
+ }
+
+ reply = req_capsule_server_sized_get(&req->rq_pill,
+ &RMF_OUT_UPDATE_REPLY,
+ OUT_UPDATE_REPLY_SIZE);
+ if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
+ DEBUG_REQ(D_ERROR, req,
+ "%s: Wrong version %x expected %x "DFID": rc = %d",
+ dname, reply->ourp_magic, UPDATE_REPLY_MAGIC,
+ PFID(lu_object_fid(&dt->do_lu)), -EPROTO);
+
+ GOTO(out, rc = -EPROTO);
+ }
+
+ rc = object_update_result_data_get(reply, rbuf, 0);
+ if (rc < 0)
+ GOTO(out, rc);
+
+ if (!buf->lb_buf)
+ GOTO(out, rc);
+
+ if (unlikely(buf->lb_len < rbuf->lb_len))
+ GOTO(out, rc = -ERANGE);
+
+ memcpy(buf->lb_buf, rbuf->lb_buf, rbuf->lb_len);
+ EXIT;
+
+out:
+ if (req)
+ ptlrpc_req_finished(req);
+
+ if (update && !IS_ERR(update))
+ osp_update_request_destroy(env, update);
+
+ return rc;
+}
+
+/**
* Implementation of dt_object_operations::do_index_try
*
* Try to initialize the index API pointer for the given object. This
* Implementation of dt_object_operations::do_object_lock
*
* Enqueue a lock (by ldlm_cli_enqueue()) of remote object on the remote MDT,
- * which will lock the object in the global namespace.
+ * which will lock the object in the global namespace. And because the
+ * cross-MDT locks are relatively rare compared with normal local MDT operation,
+ * let's release it right away, instead of putting it into the LRU list.
*
* \param[in] env execution environment
* \param[in] dt object to be locked
struct dt_object *dt,
struct lustre_handle *lh,
struct ldlm_enqueue_info *einfo,
- ldlm_policy_data_t *policy)
+ union ldlm_policy_data *policy)
{
struct ldlm_res_id *res_id;
- struct dt_device *dt_dev = lu2dt_dev(dt->do_lu.lo_dev);
- struct osp_device *osp = dt2osp_dev(dt_dev);
+ struct osp_device *osp = dt2osp_dev(lu2dt_dev(dt->do_lu.lo_dev));
struct ptlrpc_request *req;
int rc = 0;
- __u64 flags = 0;
- ldlm_mode_t mode;
+ __u64 flags = LDLM_FL_NO_LRU;
+ ENTRY;
res_id = einfo->ei_res_id;
LASSERT(res_id != NULL);
- mode = ldlm_lock_match(osp->opd_obd->obd_namespace,
- LDLM_FL_BLOCK_GRANTED, res_id,
- einfo->ei_type, policy,
- einfo->ei_mode, lh, 0);
- if (mode > 0)
- return ELDLM_OK;
+ if (einfo->ei_mode & (LCK_EX | LCK_PW))
+ flags |= LDLM_FL_COS_INCOMPAT;
req = ldlm_enqueue_pack(osp->opd_exp, 0);
if (IS_ERR(req))
RETURN(PTR_ERR(req));
+ osp_set_req_replay(osp, req);
rc = ldlm_cli_enqueue(osp->opd_exp, &req, einfo, res_id,
- (const ldlm_policy_data_t *)policy,
+ (const union ldlm_policy_data *)policy,
&flags, NULL, 0, LVB_T_NONE, lh, 0);
ptlrpc_req_finished(req);
- return rc == ELDLM_OK ? 0 : -EIO;
+ RETURN(rc == ELDLM_OK ? 0 : -EIO);
}
/**
static int osp_md_object_unlock(const struct lu_env *env,
struct dt_object *dt,
struct ldlm_enqueue_info *einfo,
- ldlm_policy_data_t *policy)
+ union ldlm_policy_data *policy)
{
struct lustre_handle *lockh = einfo->ei_cbdata;
return 0;
}
+/**
+ * Implement OSP layer dt_object_operations::do_declare_destroy() interface.
+ *
+ * Create the dt_update_request to track the update for this OSP
+ * in the transaction.
+ *
+ * \param[in] env pointer to the thread context
+ * \param[in] dt pointer to the OSP layer dt_object to be destroyed
+ * \param[in] th pointer to the transaction handler
+ *
+ * \retval 0 for success
+ * \retval negative error number on failure
+ */
+int osp_md_declare_destroy(const struct lu_env *env, struct dt_object *dt,
+ struct thandle *th)
+{
+ return osp_trans_update_request_create(th);
+}
+
+/**
+ * Implement OSP layer dt_object_operations::do_destroy() interface.
+ *
+ * Pack the destroy update into the RPC buffer, which will be sent
+ * to the remote MDT during transaction stop.
+ *
+ * It also marks the object as non-cached.
+ *
+ * \param[in] env pointer to the thread context
+ * \param[in] dt pointer to the OSP layer dt_object to be destroyed
+ * \param[in] th pointer to the transaction handler
+ *
+ * \retval 0 for success
+ * \retval negative error number on failure
+ */
+int osp_md_destroy(const struct lu_env *env, struct dt_object *dt,
+ struct thandle *th)
+{
+ struct osp_object *o = dt2osp_obj(dt);
+ struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
+ struct osp_update_request *update;
+ int rc = 0;
+
+ ENTRY;
+ o->opo_non_exist = 1;
+
+ LASSERT(osp->opd_connect_mdt);
+ update = thandle_to_osp_update_request(th);
+ LASSERT(update != NULL);
+
+ rc = OSP_UPDATE_RPC_PACK(env, out_destroy_pack, update,
+ lu_object_fid(&dt->do_lu));
+ if (rc != 0)
+ RETURN(rc);
+
+ set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
+ rc = osp_insert_update_callback(env, update, dt2osp_obj(dt), NULL,
+ NULL);
+
+ RETURN(rc);
+}
+
struct dt_object_operations osp_md_obj_ops = {
- .do_read_lock = osp_md_object_read_lock,
- .do_write_lock = osp_md_object_write_lock,
- .do_read_unlock = osp_md_object_read_unlock,
- .do_write_unlock = osp_md_object_write_unlock,
- .do_write_locked = osp_md_object_write_locked,
- .do_declare_create = osp_md_declare_object_create,
- .do_create = osp_md_object_create,
+ .do_read_lock = osp_md_read_lock,
+ .do_write_lock = osp_md_write_lock,
+ .do_read_unlock = osp_md_read_unlock,
+ .do_write_unlock = osp_md_write_unlock,
+ .do_write_locked = osp_md_write_locked,
+ .do_declare_create = osp_md_declare_create,
+ .do_create = osp_md_create,
.do_declare_ref_add = osp_md_declare_ref_add,
.do_ref_add = osp_md_ref_add,
.do_declare_ref_del = osp_md_declare_ref_del,
.do_ref_del = osp_md_ref_del,
- .do_declare_destroy = osp_declare_object_destroy,
- .do_destroy = osp_object_destroy,
+ .do_declare_destroy = osp_md_declare_destroy,
+ .do_destroy = osp_md_destroy,
.do_ah_init = osp_md_ah_init,
.do_attr_get = osp_attr_get,
.do_declare_attr_set = osp_md_declare_attr_set,
.do_attr_set = osp_md_attr_set,
.do_xattr_get = osp_xattr_get,
+ .do_xattr_list = osp_md_xattr_list,
.do_declare_xattr_set = osp_declare_xattr_set,
.do_xattr_set = osp_xattr_set,
.do_declare_xattr_del = osp_declare_xattr_del,
.do_index_try = osp_md_index_try,
.do_object_lock = osp_md_object_lock,
.do_object_unlock = osp_md_object_unlock,
+ .do_invalidate = osp_invalidate,
};
/**
* Implementation of dt_body_operations::dbo_declare_write
*
- * Declare an object write. In DNE phase I, it will pack the write
- * object update into the RPC.
- *
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
+ *
* \param[in] env execution environment
* \param[in] dt object to be written
* \param[in] buf buffer to write which includes an embedded size field
* \param[in] pos offet in the object to start writing at
* \param[in] th transaction handle
*
- * \retval 0 if the insertion succeeds.
- * \retval negative errno if the insertion fails.
+ * \retval 0 if preparation succeeds.
+ * \retval negative errno if preparation fails.
*/
static ssize_t osp_md_declare_write(const struct lu_env *env,
struct dt_object *dt,
const struct lu_buf *buf,
loff_t pos, struct thandle *th)
{
- struct dt_update_request *update;
- ssize_t rc;
+ struct osp_device *osp = dt2osp_dev(th->th_dev);
+ int rc;
- update = thandle_to_dt_update_request(th);
- LASSERT(update != NULL);
+ rc = osp_trans_update_request_create(th);
+ if (rc != 0)
+ return rc;
- rc = out_write_pack(env, &update->dur_buf, lu_object_fid(&dt->do_lu),
- buf, pos, update->dur_batchid);
+ if (osp->opd_update == NULL)
+ return 0;
- return rc;
+ if (dt2osp_obj(dt)->opo_stale)
+ return -ESTALE;
+ return 0;
}
/**
* Implementation of dt_body_operations::dbo_write
*
- * Return the buffer size. In DNE phase I, remote updates
- * are actually executed during transaction start, the buffer has
- * already been written when this method is being called.
+ * Pack the write object update into the RPC buffer, which will be sent
+ * to the remote MDT during transaction stop.
*
* \param[in] env execution environment
* \param[in] dt object to be written
* \param[in] buf buffer to write which includes an embedded size field
* \param[in] pos offet in the object to start writing at
* \param[in] th transaction handle
- * \param[in] ignore_quota quota enforcement for this write
*
- * \retval the buffer size in bytes.
+ * \retval the buffer size in bytes if packing succeeds.
+ * \retval negative errno if packing fails.
*/
static ssize_t osp_md_write(const struct lu_env *env, struct dt_object *dt,
const struct lu_buf *buf, loff_t *pos,
- struct thandle *handle, int ignore_quota)
+ struct thandle *th)
{
+ struct osp_object *obj = dt2osp_obj(dt);
+ struct osp_update_request *update;
+ struct osp_thandle *oth = thandle_to_osp_thandle(th);
+ ssize_t rc;
+ ENTRY;
+
+ update = thandle_to_osp_update_request(th);
+ LASSERT(update != NULL);
+
+ CDEBUG(D_INFO, "write "DFID" offset = %llu length = %zu\n",
+ PFID(lu_object_fid(&dt->do_lu)), *pos, buf->lb_len);
+
+ rc = OSP_UPDATE_RPC_PACK(env, out_write_pack, update,
+ lu_object_fid(&dt->do_lu), buf, *pos);
+ if (rc < 0)
+ RETURN(rc);
+
+ rc = osp_check_and_set_rpc_version(oth, obj);
+ if (rc < 0)
+ RETURN(rc);
+
+ /* XXX: how about the write error happened later? */
*pos += buf->lb_len;
- return buf->lb_len;
+
+ if (obj->opo_attr.la_valid & LA_SIZE && obj->opo_attr.la_size < *pos)
+ obj->opo_attr.la_size = *pos;
+
+ spin_lock(&obj->opo_lock);
+ if (list_empty(&obj->opo_invalidate_cb_list)) {
+ lu_object_get(&obj->opo_obj.do_lu);
+
+ list_add_tail(&obj->opo_invalidate_cb_list,
+ &update->our_invalidate_cb_list);
+ }
+ spin_unlock(&obj->opo_lock);
+
+ RETURN(buf->lb_len);
+}
+
+static inline void orr_le_to_cpu(struct out_read_reply *orr_dst,
+ const struct out_read_reply *orr_src)
+{
+ orr_dst->orr_size = le32_to_cpu(orr_src->orr_size);
+ orr_dst->orr_padding = le32_to_cpu(orr_src->orr_padding);
+ orr_dst->orr_offset = le64_to_cpu(orr_dst->orr_offset);
+}
+
+
+
+static ssize_t osp_md_read(const struct lu_env *env, struct dt_object *dt,
+ struct lu_buf *rbuf, loff_t *pos)
+{
+ struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
+ struct dt_device *dt_dev = &osp->opd_dt_dev;
+ struct lu_buf *lbuf = &osp_env_info(env)->osi_lb2;
+ char *ptr = rbuf->lb_buf;
+ struct osp_update_request *update;
+ struct ptlrpc_request *req = NULL;
+ struct out_read_reply *orr;
+ struct ptlrpc_bulk_desc *desc;
+ struct object_update_reply *reply;
+ int pages;
+ int rc;
+ ENTRY;
+
+ /* Because it needs send the update buffer right away,
+ * just create an update buffer, instead of attaching the
+ * update_remote list of the thandle. */
+ update = osp_update_request_create(dt_dev);
+ if (IS_ERR(update))
+ RETURN(PTR_ERR(update));
+
+ rc = OSP_UPDATE_RPC_PACK(env, out_read_pack, update,
+ lu_object_fid(&dt->do_lu),
+ rbuf->lb_len, *pos);
+ if (rc != 0) {
+ CERROR("%s: cannot insert update: rc = %d\n",
+ dt_dev->dd_lu_dev.ld_obd->obd_name, rc);
+ GOTO(out_update, rc);
+ }
+
+ CDEBUG(D_INFO, "%s "DFID" read offset %llu size %zu\n",
+ dt_dev->dd_lu_dev.ld_obd->obd_name,
+ PFID(lu_object_fid(&dt->do_lu)), *pos, rbuf->lb_len);
+ rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import, update,
+ &req);
+ if (rc != 0)
+ GOTO(out_update, rc);
+
+ /* First *and* last might be partial pages, hence +1 */
+ pages = DIV_ROUND_UP(rbuf->lb_len, PAGE_SIZE) + 1;
+
+ /* allocate bulk descriptor */
+ desc = ptlrpc_prep_bulk_imp(req, pages, 1,
+ PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
+ MDS_BULK_PORTAL,
+ &ptlrpc_bulk_kiov_nopin_ops);
+ if (desc == NULL)
+ GOTO(out, rc = -ENOMEM);
+
+ desc->bd_frag_ops->add_iov_frag(desc, ptr, rbuf->lb_len);
+
+ osp_set_req_replay(osp, req);
+ req->rq_bulk_read = 1;
+ /* send request to master and wait for RPC to complete */
+ rc = ptlrpc_queue_wait(req);
+ if (rc != 0)
+ GOTO(out, rc);
+
+ rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
+ req->rq_bulk->bd_nob_transferred);
+ if (rc < 0)
+ GOTO(out, rc);
+
+ reply = req_capsule_server_sized_get(&req->rq_pill,
+ &RMF_OUT_UPDATE_REPLY,
+ OUT_UPDATE_REPLY_SIZE);
+
+ if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
+ CERROR("%s: invalid update reply magic %x expected %x:"
+ " rc = %d\n", dt_dev->dd_lu_dev.ld_obd->obd_name,
+ reply->ourp_magic, UPDATE_REPLY_MAGIC, -EPROTO);
+ GOTO(out, rc = -EPROTO);
+ }
+
+ rc = object_update_result_data_get(reply, lbuf, 0);
+ if (rc < 0)
+ GOTO(out, rc);
+
+ if (lbuf->lb_len < sizeof(*orr))
+ GOTO(out, rc = -EPROTO);
+
+ orr = lbuf->lb_buf;
+ orr_le_to_cpu(orr, orr);
+ rc = orr->orr_size;
+ *pos = orr->orr_offset;
+out:
+ ptlrpc_req_finished(req);
+
+out_update:
+ osp_update_request_destroy(env, update);
+
+ RETURN(rc);
}
/* These body operation will be used to write symlinks during migration etc */
struct dt_body_operations osp_md_body_ops = {
.dbo_declare_write = osp_md_declare_write,
.dbo_write = osp_md_write,
+ .dbo_read = osp_md_read,
};
git://git.whamcloud.com / fs / lustre-release.git / blobdiff