* GPL HEADER END
*/
/*
- * Copyright (c) 2013, 2014, Intel Corporation.
+ * Copyright (c) 2013, 2015, 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"
struct osp_object *obj,
void *data, int index, int rc)
{
- if (rc != 0) {
+ if (rc != 0 && rc != -EEXIST) {
obj->opo_obj.do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
obj->opo_non_exist = 1;
}
+
+ /* Invalid the opo cache for the object after the object
+ * is being created, so attr_get will try to get attr
+ * from the remote object. XXX this can be improved when
+ * we have object lock/cache invalidate mechanism in OSP
+ * layer */
+ if (obj->opo_ooa != NULL) {
+ spin_lock(&obj->opo_lock);
+ obj->opo_ooa->ooa_attr.la_valid = 0;
+ spin_unlock(&obj->opo_lock);
+ }
+
return 0;
}
/**
* Implementation of dt_object_operations::do_declare_create
*
- * Create the dt_update_request to track the update for this OSP
+ * Create the osp_update_request to track the update for this OSP
* in the transaction.
*
* \param[in] env execution environment
struct dt_object_format *dof,
struct thandle *th)
{
+ struct osp_object *obj = dt2osp_obj(dt);
+ int rc;
+
+ if (obj->opo_ooa == NULL) {
+ rc = osp_oac_init(obj);
+ if (rc != 0)
+ return rc;
+ }
+
return osp_trans_update_request_create(th);
}
return ptr;
}
-int osp_extend_update_buffer(const struct lu_env *env,
- struct update_buffer *ubuf)
-{
- struct object_update_request *ureq;
- size_t new_size = ubuf->ub_req_size + OUT_UPDATE_BUFFER_SIZE_ADD;
-
- /* enlarge object update request size */
- if (new_size > OUT_UPDATE_BUFFER_SIZE_MAX)
- return -E2BIG;
-
- OBD_ALLOC_LARGE(ureq, new_size);
- if (ureq == NULL)
- return -ENOMEM;
-
- memcpy(ureq, ubuf->ub_req, ubuf->ub_req_size);
-
- OBD_FREE_LARGE(ubuf->ub_req, ubuf->ub_req_size);
-
- ubuf->ub_req = ureq;
- ubuf->ub_req_size = new_size;
-
- return 0;
-}
-
/**
* Implementation of dt_object_operations::do_create
*
struct lu_attr *attr, struct dt_allocation_hint *hint,
struct dt_object_format *dof, struct thandle *th)
{
- struct dt_update_request *update;
+ struct osp_update_request *update;
+ struct osp_object *obj = dt2osp_obj(dt);
int rc;
- update = thandle_to_dt_update_request(th);
+ update = thandle_to_osp_update_request(th);
LASSERT(update != NULL);
+ LASSERT(attr->la_valid & LA_TYPE);
rc = osp_update_rpc_pack(env, create, update, OUT_CREATE,
lu_object_fid(&dt->do_lu), attr, hint, dof);
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(obj->opo_ooa != NULL);
+ obj->opo_ooa->ooa_attr = *attr;
out:
return rc;
}
/**
* Implementation of dt_object_operations::do_declare_ref_del
*
- * Create the dt_update_request to track the update for this OSP
+ * Create the osp_update_request to track the update for this OSP
* in the transaction.
*
* \param[in] env execution environment
static int osp_md_ref_del(const struct lu_env *env, struct dt_object *dt,
struct thandle *th)
{
- struct dt_update_request *update;
+ 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 = osp_update_rpc_pack(env, ref_del, update, OUT_REF_DEL,
/**
* Implementation of dt_object_operations::do_declare_ref_del
*
- * Create the dt_update_request to track the update for this OSP
+ * Create the osp_update_request to track the update for this OSP
* in the transaction.
*
* \param[in] env execution environment
static int osp_md_ref_add(const struct lu_env *env, struct dt_object *dt,
struct thandle *th)
{
- struct dt_update_request *update;
+ 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 = osp_update_rpc_pack(env, ref_add, update, OUT_REF_ADD,
/**
* Implementation of dt_object_operations::do_declare_attr_get
*
- * Create the dt_update_request to track the update for this OSP
+ * Create the osp_update_request to track the update for this OSP
* in the transaction.
*
* \param[in] env execution environment
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;
+ 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 = osp_update_rpc_pack(env, attr_set, update, OUT_ATTR_SET,
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));
if (req != NULL)
ptlrpc_req_finished(req);
- dt_update_request_destroy(update);
+ osp_update_request_destroy(update);
return rc;
}
/**
* Implementation of dt_index_operations::dio_declare_insert
*
- * Create the dt_update_request to track the update for this OSP
+ * Create the osp_update_request to track the update for this OSP
* in the transaction.
*
* \param[in] env execution environment
struct thandle *th,
int ignore_quota)
{
- struct osp_thandle *oth = thandle_to_osp_thandle(th);
- struct dt_update_request *update = oth->ot_dur;
+ struct osp_update_request *update;
int rc;
+ update = thandle_to_osp_update_request(th);
+ LASSERT(update != NULL);
+
rc = osp_update_rpc_pack(env, index_insert, update, OUT_INDEX_INSERT,
lu_object_fid(&dt->do_lu), rec, key);
return rc;
/**
* Implementation of dt_index_operations::dio_declare_delete
*
- * Create the dt_update_request to track the update for this OSP
+ * Create the osp_update_request to track the update for this OSP
* in the transaction.
*
* \param[in] env execution environment
const struct dt_key *key,
struct thandle *th)
{
- struct dt_update_request *update;
+ 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 = osp_update_rpc_pack(env, index_delete, update, OUT_INDEX_DELETE,
* \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.
*/
* 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 lu_device *top_device;
struct ptlrpc_request *req;
int rc = 0;
- __u64 flags = 0;
- ldlm_mode_t mode;
+ __u64 flags = LDLM_FL_NO_LRU;
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_nonblock)
+ flags |= LDLM_FL_BLOCK_NOWAIT;
+ 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));
+ /* During recovery, it needs to let OSP send enqueue
+ * without checking recoverying status, in case the
+ * other target is being recovered at the same time,
+ * and if we wait here for the import to be recovered,
+ * it might cause deadlock */
+ top_device = dt_dev->dd_lu_dev.ld_site->ls_top_dev;
+ if (top_device->ld_obd->obd_recovering)
+ req->rq_allow_replay = 1;
+
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);
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;
}
/**
- * Interpreter call for object destroy
- *
- * Object destroy interpreter, which will be called after destroying
- * the remote object to set flags and status.
- *
- * \param[in] env execution environment
- * \param[in] reply update reply
- * \param[in] req ptlrpc update request for destroying object
- * \param[in] obj object to be destroyed
- * \param[in] data data used in this function.
- * \param[in] index index(position) of destroy update in the whole
- * updates
- * \param[in] rc update result on the remote MDT.
- *
- * \retval only return 0 for now
- */
-static int osp_md_object_destroy_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)
-{
- /* not needed in cache any more */
- set_bit(LU_OBJECT_HEARD_BANSHEE,
- &obj->opo_obj.do_lu.lo_header->loh_flags);
- return 0;
-}
-
-/**
* Implement OSP layer dt_object_operations::do_destroy() interface.
*
* Pack the destroy update into the RPC buffer, which will be sent
{
struct osp_object *o = dt2osp_obj(dt);
struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
- struct dt_update_request *update;
+ struct osp_update_request *update;
int rc = 0;
ENTRY;
o->opo_non_exist = 1;
LASSERT(osp->opd_connect_mdt);
- update = thandle_to_dt_update_request(th);
+ update = thandle_to_osp_update_request(th);
LASSERT(update != NULL);
rc = osp_update_rpc_pack(env, object_destroy, update, OUT_DESTROY,
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,
- osp_md_object_destroy_interpreter);
+ NULL);
+
RETURN(rc);
}
/**
* Implementation of dt_body_operations::dbo_declare_write
*
- * Create the dt_update_request to track the update for this OSP
+ * Create the osp_update_request to track the update for this OSP
* in the transaction.
*
* \param[in] env execution environment
const struct lu_buf *buf,
loff_t pos, struct thandle *th)
{
- return osp_trans_update_request_create(th);
+ struct osp_device *osp = dt2osp_dev(th->th_dev);
+ int rc;
+
+ rc = osp_trans_update_request_create(th);
+ if (rc != 0)
+ return rc;
+
+ if (osp->opd_update == NULL)
+ return 0;
+
+ if (dt2osp_obj(dt)->opo_stale)
+ return -ESTALE;
+
+ return 0;
}
/**
const struct lu_buf *buf, loff_t *pos,
struct thandle *th, int ignore_quota)
{
- struct dt_update_request *update;
+ 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_dt_update_request(th);
+ update = thandle_to_osp_update_request(th);
LASSERT(update != NULL);
+ CDEBUG(D_INFO, "write "DFID" offset = "LPU64" length = %zu\n",
+ PFID(lu_object_fid(&dt->do_lu)), *pos, buf->lb_len);
+
rc = osp_update_rpc_pack(env, write, update, OUT_WRITE,
lu_object_fid(&dt->do_lu), buf, *pos);
if (rc < 0)
- return rc;
+ 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_ooa != NULL &&
+ obj->opo_ooa->ooa_attr.la_valid & LA_SIZE &&
+ obj->opo_ooa->ooa_attr.la_size < *pos)
+ obj->opo_ooa->ooa_attr.la_size = *pos;
+
+ RETURN(buf->lb_len);
+}
+
+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;
+ __u32 left_size;
+ int nbufs;
+ int i;
+ 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, read, update, OUT_READ,
+ 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);
+
+ nbufs = (rbuf->lb_len + OUT_BULK_BUFFER_SIZE - 1) /
+ OUT_BULK_BUFFER_SIZE;
+ /* allocate bulk descriptor */
+ desc = ptlrpc_prep_bulk_imp(req, nbufs, 1,
+ PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KVEC,
+ MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
+ if (desc == NULL)
+ GOTO(out, rc = -ENOMEM);
+
+ /* split the buffer into small chunk size */
+ left_size = rbuf->lb_len;
+ for (i = 0; i < nbufs; i++) {
+ int read_size;
+
+ read_size = left_size > OUT_BULK_BUFFER_SIZE ?
+ OUT_BULK_BUFFER_SIZE : left_size;
+ desc->bd_frag_ops->add_iov_frag(desc, ptr, read_size);
+
+ ptr += read_size;
+ }
+
+ /* This will only be called with read-only update, and these updates
+ * might be used to retrieve update log during recovery process, so
+ * it will be allowed to send during recovery process */
+ req->rq_allow_replay = 1;
+ 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(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,
};