void rozofs_ll_lookup_nb(fuse_req_t req, fuse_ino_t parent, const char *name)
{
ientry_t *ie = 0;
ientry_t *nie = 0;
epgw_lookup_arg_t arg;
int ret;
void *buffer_p = NULL;
int trc_idx;
mattr_obj_t mattr_obj;
struct fuse_entry_param fep;
struct stat stbuf;
int allocated = 0;
int len_name;
fuse_ino_t child = 0;
// int local_lookup_success = 0;
uint32_t lookup_flags=0;
int extra_length = 0;
fuse_ino_t ino = 0;
int trace_flag = 0;
/*
** Update the IO statistics
*/
rozofs_thr_cnt_update(rozofs_thr_counter[ROZOFSMOUNT_COUNTER_LOOKUP], 1);
extra_length = rozofs_check_extra_inode_in_lookup((char*)name, &len_name);
if (extra_length !=0)
{
uint8_t *pdata_p;
uint32_t *lookup_flags_p = (uint32_t*)&name[len_name+1];
lookup_flags =*lookup_flags_p;
if (extra_length > 4)
{
pdata_p = (uint8_t*)&name[len_name+1];
pdata_p+=sizeof(uint32_t);
fuse_ino_t *inode_p = (fuse_ino_t*)pdata_p;
child = *inode_p;
ino = child;
}
}
trace_flag = lookup_flags;
if (ino != 0) {trace_flag |=(1<<31);}
trc_idx = rozofs_trc_req_name_flags(srv_rozofs_ll_lookup,parent,(char*)name,(int)(trace_flag));
/*
** allocate a context for saving the fuse parameters
*/
buffer_p = rozofs_fuse_alloc_saved_context();
if (buffer_p == NULL)
{
severe("out of fuse saved context");
errno = ENOMEM;
goto error;
}
SAVE_FUSE_PARAM(buffer_p,req);
SAVE_FUSE_PARAM(buffer_p,parent);
SAVE_FUSE_STRING(buffer_p,name);
SAVE_FUSE_PARAM(buffer_p,trc_idx);
SAVE_FUSE_PARAM(buffer_p,ino);
SAVE_FUSE_PARAM(buffer_p,lookup_flags);
DEBUG("lookup (%lu,%s)\n", (unsigned long int) parent, name);
START_PROFILING_NB(buffer_p,rozofs_ll_lookup);
len_name=strlen(name);
if (len_name > ROZOFS_FILENAME_MAX) {
errno = ENAMETOOLONG;
goto error;
}
/*
** Check the case of NFS that attempt to revalidate an inode that has been flushed from the cache
*/
if ((strcmp(name,"..")== 0) || (strcmp(name,".")==0))
{
fid_t lookup_fid;
rozofs_inode_t *inode_p = (rozofs_inode_t*)lookup_fid;
inode_p->fid[0]=0;
inode_p->s.eid= exportclt.eid;
inode_p->fid[1]=parent;
arg.arg_gw.eid = exportclt.eid;
memcpy(arg.arg_gw.parent,lookup_fid, sizeof (uuid_t));
arg.arg_gw.name = (char*)name;
/*
** Queue the request and attempt to check if there is already the same
** request queued
*/
if (rozofs_lookup_insert_queue(buffer_p,parent,name,req,trc_idx,lookup_flags)== 1)
{
/*
** There is already a pending request, so nothing to send to the export
*/
gprofiler->rozofs_ll_lookup_agg[P_COUNT]++;
rozofs_fuse_release_saved_context(buffer_p);
return;
}
ret = rozofs_expgateway_send_routing_common(arg.arg_gw.eid,lookup_fid,EXPORT_PROGRAM, EXPORT_VERSION,
EP_LOOKUP,(xdrproc_t) xdr_epgw_lookup_arg_t,(void *)&arg,
rozofs_ll_lookup_cbk,buffer_p);
if (ret < 0) goto error;
/*
** no error just waiting for the answer
*/
return;
}
if (!(ie = get_ientry_by_inode(parent))) {
errno = ENOENT;
goto error;
}
/*
** check for direct access
*/
if (strncmp(name,"@rozofs@",8) == 0)
{
ret = rozofs_parse_object_name((char*)(name+8),&mattr_obj);
if (ret == 0)
{
/*
** successful parsing-> attempt to create a fake ientry
*/
//errno = ENOENT;
goto lookup_objectmode;
}
}
/*
** Queue the request and attempt to check if there is already the same
** request queued
*/
if (rozofs_lookup_insert_queue(buffer_p,parent,name,req,trc_idx,lookup_flags)== 1)
{
/*
** There is already a pending request, so nothing to send to the export
*/
gprofiler->rozofs_ll_lookup_agg[P_COUNT]++;
rozofs_fuse_release_saved_context(buffer_p);
return;
}
/*
** check if the lookup is a lookup revalidate. In such a case, the VFS provides*
** the inode. So if the i-node attributes are fine, we do not worry
*/
if ((child != 0) && ((lookup_flags & 0x100) == 0))
{
uint64_t attr_us = rozofs_tmr_get_attr_us(rozofs_is_directory_inode(ino));
nie = get_ientry_by_inode(child);
if (nie != NULL)
{
if (
/* check regular file */
((((nie->timestamp+attr_us) > rozofs_get_ticker_us()) || (rozofs_mode == 1))&&(S_ISREG(nie->attrs.attrs.mode))) ||
/* check directory */
(((nie->pending_getattr_cnt>0)||((nie->timestamp+attr_us) > rozofs_get_ticker_us()))&&(S_ISDIR(nie->attrs.attrs.mode)))
)
{
/*
** check if parent and child are either deleted/deleted or active/active
*/
int trash_state = 0;
if (rozofs_inode_is_del_pending(ie->attrs.attrs.fid)) trash_state = 1;
if (rozofs_inode_is_del_pending(nie->attrs.attrs.fid)) trash_state |= 1<<1;
switch (trash_state)
{
case 2:
if (rozofs_inode_is_trash(nie->attrs.attrs.fid) == 0)
{
errno = ENOENT;
goto error;
}
mattr_to_stat(&nie->attrs, &stbuf,exportclt.bsize);
stbuf.st_ino = child;
/*
** set nlinks to reflect . & ..
*/
stbuf.st_nlink =2;
goto success;
break;
case 0:
case 3:
mattr_to_stat(&nie->attrs, &stbuf,exportclt.bsize);
stbuf.st_ino = child;
goto success;
break;
default:
errno = ENOENT;
goto error;
break;
}
}
}
}
/*
** fill up the structure that will be used for creating the xdr message
*/
arg.arg_gw.eid = exportclt.eid;
memcpy(arg.arg_gw.parent,ie->fid, sizeof (uuid_t));
arg.arg_gw.name = (char*)name;
/*
** now initiates the transaction towards the remote end
*/
/*
** In case the EXPORT LBG is down ans we know this ientry, let's respond to
** the requester with the current available information
*/
#if 1
if ((common_config.client_fast_reconnect) && (child != 0)) {
expgw_tx_routing_ctx_t routing_ctx;
if (expgw_get_export_routing_lbg_info(arg.arg_gw.eid,ie->fid,&routing_ctx) != 0) {
goto error;
}
if (north_lbg_get_state(routing_ctx.lbg_id[0]) != NORTH_LBG_UP) {
if (!(nie = get_ientry_by_inode(child))) {
errno = ENOENT;
goto error;
}
mattr_to_stat(&nie->attrs, &stbuf,exportclt.bsize);
stbuf.st_ino = child;
goto success;
}
}
#endif
#if 1
ret = rozofs_expgateway_send_routing_common(arg.arg_gw.eid,ie->fid,EXPORT_PROGRAM, EXPORT_VERSION,
EP_LOOKUP,(xdrproc_t) xdr_epgw_lookup_arg_t,(void *)&arg,
rozofs_ll_lookup_cbk,buffer_p);
#else
ret = rozofs_export_send_common(&exportclt,EXPORT_PROGRAM, EXPORT_VERSION,
EP_LOOKUP,(xdrproc_t) xdr_epgw_lookup_arg_t,(void *)&arg,
rozofs_ll_lookup_cbk,buffer_p);
#endif
if (ret < 0) {
/*
** In case of fast reconnect mode let's respond with the previously knows
** parameters instead of failing
*/
if (common_config.client_fast_reconnect) {
if (child != 0) {
if (!(nie = get_ientry_by_inode(child))) {
errno = ENOENT;
goto error;
}
mattr_to_stat(&nie->attrs, &stbuf,exportclt.bsize);
stbuf.st_ino = child;
goto success;
}
}
goto error;
}
/*
** no error just waiting for the answer
*/
return;
error:
fuse_reply_err(req, errno);
/*
** remove the context from the lookup queue
*/
if (buffer_p != NULL) ruc_objRemove(buffer_p);
/*
** release the buffer if has been allocated
*/
rozofs_trc_rsp(srv_rozofs_ll_lookup,parent,NULL,1,trc_idx);
out:
STOP_PROFILING_NB(buffer_p,rozofs_ll_lookup);
if (buffer_p != NULL) rozofs_fuse_release_saved_context(buffer_p);
return;
/**
* case of the object mode
*/
lookup_objectmode:
if (!(nie = get_ientry_by_fid(mattr_obj.fid))) {
nie = alloc_ientry(mattr_obj.fid);
allocated=1;
}
/**
* update the timestamp in the ientry context
*/
nie->timestamp = rozofs_get_ticker_us();
if (allocated)
{
/*
** update the attributes in the ientry
*/
memcpy(nie->attrs.attrs.fid, mattr_obj.fid, sizeof(fid_t));
nie->attrs.attrs.cid = mattr_obj.cid;
memcpy(nie->attrs.attrs.sids, mattr_obj.sids, sizeof(sid_t)*ROZOFS_SAFE_MAX);
nie->attrs.attrs.size = mattr_obj.size;
nie->attrs.attrs.nlink = 1;
nie->attrs.attrs.mode = S_IFREG | S_IRWXU | S_IRWXG | S_IRWXO ;
nie->attrs.attrs.uid = 0;
nie->attrs.attrs.gid = 0;
nie->nlookup = 0;
}
mattr_to_stat(&nie->attrs, &stbuf,exportclt.bsize);
stbuf.st_ino = nie->inode;
// info("FDL %d mode %d uid %d gid %d",allocated,nie->attrs.attrs.mode,nie->attrs.attrs.uid,nie->attrs.attrs.gid);
success:
memset(&fep, 0, sizeof (fep));
fep.ino =stbuf.st_ino;
fep.attr_timeout = rozofs_tmr_get_attr(rozofs_is_directory_inode(nie->inode));
fep.entry_timeout = rozofs_tmr_get_entry(rozofs_is_directory_inode(nie->inode));
memcpy(&fep.attr, &stbuf, sizeof (struct stat));
nie->nlookup++;
rozofs_inode_t * finode = (rozofs_inode_t *) nie->attrs.attrs.fid;
fep.generation = finode->fid[0];
rz_fuse_reply_entry(req, &fep);
rozofs_trc_rsp(srv_rozofs_ll_lookup,(nie==NULL)?0:nie->inode,(nie==NULL)?NULL:nie->attrs.attrs.fid,0,trc_idx);
goto out;
}
int export_unlink_multiple(export_t * e, fid_t parent, char *name, fid_t fid,mattr_t * pattrs,lv2_entry_t *plv2) {
int status = -1;
lv2_entry_t *lv2=NULL;
fid_t child_fid;
uint32_t child_type;
uint16_t nlink = 0;
int fdp = -1;
int ret;
rozofs_inode_t *fake_inode_p;
rmfentry_disk_t trash_entry;
int root_dirent_mask = 0;
int unknown_fid = 0;
int deleted_fid_count = 0;
char filename[1024];
char *basename;
int filecount;
char *filename_p;
int k;
/*
** search for rozofs key in the filename
*/
if (strncmp(name,".@rozofs-mf@",12) == 0)
{
ret = rozofs_parse_object_name((char*)(name+12),&basename,&filecount);
if (ret < 0)
{
errno = ENOTSUP;
goto error;
}
}
/*
** load the root_idx bitmap of the old parent
*/
export_dir_load_root_idx_bitmap(e,parent,plv2);
/*
** set global variables associated with the export
*/
fdp = export_open_parent_directory(e,parent);
if (get_mdirentry(plv2->dirent_root_idx_p,fdp, parent, name, child_fid, &child_type,&root_dirent_mask) != 0)
goto out;
if (S_ISDIR(child_type)) {
errno = EISDIR;
goto out;
}
/*
** do a loop on the file count
*/
for (k = 0; k < filecount ; k++)
{
sprintf(filename,"%s.%d",basename,k);
filename_p = filename;
// Delete the mdirentry if exist
ret =del_mdirentry(plv2->dirent_root_idx_p,fdp, parent, name, child_fid, &child_type,root_dirent_mask);
if (ret != 0)
{
if (errno != ENOENT) goto out;
/*
** check the next entry
*/
continue;
}
// Get mattrs of child to delete
if (!(lv2 = export_lookup_fid(e->trk_tb_p,e->lv2_cache, child_fid)))
{
unknown_fid++;
/*
** check the next entry
*/
continue;
}
deleted_fid_count++;
// Compute hash value for this fid
uint32_t hash = 0;
uint8_t *c = 0;
for (c = lv2->attributes.s.attrs.fid; c != lv2->attributes.s.attrs.fid + 16; c++)
hash = *c + (hash << 6) + (hash << 16) - hash;
hash %= RM_MAX_BUCKETS;
/*
** prepare the trash entry
*/
trash_entry.size = lv2->attributes.s.attrs.size;
memcpy(trash_entry.fid, lv2->attributes.s.attrs.fid, sizeof (fid_t));
trash_entry.cid = lv2->attributes.s.attrs.cid;
memcpy(trash_entry.initial_dist_set, lv2->attributes.s.attrs.sids,
sizeof (sid_t) * ROZOFS_SAFE_MAX);
memcpy(trash_entry.current_dist_set, lv2->attributes.s.attrs.sids,
sizeof (sid_t) * ROZOFS_SAFE_MAX);
fake_inode_p = (rozofs_inode_t *)parent;
ret = exp_trash_entry_create(e->trk_tb_p,fake_inode_p->s.usr_id,&trash_entry);
if (ret < 0)
{
/*
** error while inserting entry in trash file
*/
severe("error on trash insertion name %s error %s",name,strerror(errno));
}
/*
** delete the metadata associated with the file
*/
ret = exp_delete_file(e,lv2);
/*
* In case of geo replication, insert a delete request from the 2 sites
*/
if (e->volume->georep)
{
/*
** update the geo replication: set start=end=0 to indicate a deletion
*/
geo_rep_insert_fid(e->geo_replication_tb[0],
lv2->attributes.s.attrs.fid,
0/*start*/,0/*end*/,
e->layout,
lv2->attributes.s.attrs.cid,
lv2->attributes.s.attrs.sids);
/*
** update the geo replication: set start=end=0 to indicate a deletion
*/
geo_rep_insert_fid(e->geo_replication_tb[1],
lv2->attributes.s.attrs.fid,
0/*start*/,0/*end*/,
e->layout,
lv2->attributes.s.attrs.cid,
lv2->attributes.s.attrs.sids);
}
/*
** Preparation of the rmfentry
*/
rmfentry_t *rmfe = xmalloc(sizeof (rmfentry_t));
export_rm_bins_pending_count++;
memcpy(rmfe->fid, trash_entry.fid, sizeof (fid_t));
rmfe->cid = trash_entry.cid;
memcpy(rmfe->initial_dist_set, trash_entry.initial_dist_set,
sizeof (sid_t) * ROZOFS_SAFE_MAX);
memcpy(rmfe->current_dist_set, trash_entry.current_dist_set,
sizeof (sid_t) * ROZOFS_SAFE_MAX);
memcpy(rmfe->trash_inode,trash_entry.trash_inode,sizeof(fid_t));
list_init(&rmfe->list);
/* Acquire lock on bucket trash list
*/
if ((errno = pthread_rwlock_wrlock
(&e->trash_buckets[hash].rm_lock)) != 0) {
severe("pthread_rwlock_wrlock failed: %s", strerror(errno));
// Best effort
}
/*
** Check size of file
*/
if (lv2->attributes.s.attrs.size >= RM_FILE_SIZE_TRESHOLD) {
// Add to front of list
list_push_front(&e->trash_buckets[hash].rmfiles, &rmfe->list);
} else {
// Add to back of list
list_push_back(&e->trash_buckets[hash].rmfiles, &rmfe->list);
}
if ((errno = pthread_rwlock_unlock
(&e->trash_buckets[hash].rm_lock)) != 0) {
severe("pthread_rwlock_unlock failed: %s", strerror(errno));
// Best effort
}
// Update the nb. of blocks
if (export_update_blocks(e,
-(((int64_t) lv2->attributes.s.attrs.size + ROZOFS_BSIZE_BYTES(e->bsize) - 1)
/ ROZOFS_BSIZE_BYTES(e->bsize))) != 0) {
severe("export_update_blocks failed: %s", strerror(errno));
// Best effort
}
// Remove from the cache (will be closed and freed)
lv2_cache_del(e->lv2_cache, child_fid);
}
/*
** al the subfile have been deleted so Update export files
*/
if (export_update_files(e, 0-deleted_fid_count) != 0)
goto out;
// Update parent
plv2->attributes.s.attrs.mtime = plv2->attributes.s.attrs.ctime = time(NULL);
plv2->attributes.s.attrs.children--;
// Write attributes of parents
if (export_lv2_write_attributes(e->trk_tb_p,plv2) != 0)
goto out;
/*
** return the parent attributes
*/
memcpy(pattrs, &plv2->attributes.s.attrs, sizeof (mattr_t));
status = 0;
out:
/*
** check if parent root idx bitmap must be updated
*/
if (plv2 != NULL) export_dir_flush_root_idx_bitmap(e,parent,plv2->dirent_root_idx_p);
if(fdp != -1) close(fdp);
return status;
}
