terom@24: 
terom@24: /*
terom@24:  * A simple PostgreSQL-based filesystem.
terom@24:  */
terom@24: 
terom@26: #include <stdlib.h>
terom@24: #include <string.h>
terom@24: #include <errno.h>
terom@26: #include <assert.h>
terom@24: 
terom@24: #include <event2/event.h>
terom@24: 
terom@24: #include "evfuse.h"
terom@24: #include "evsql.h"
terom@24: #include "dirbuf.h"
terom@24: #include "lib/log.h"
terom@24: #include "lib/signals.h"
terom@24: #include "lib/misc.h"
terom@24: 
terom@24: #define SERROR(val) do { (val); goto error; } while(0)
terom@24: 
terom@24: struct dbfs {
terom@24:     struct event_base *ev_base;
terom@24:     struct signals *signals;
terom@24:     
terom@24:     const char *db_conninfo;
terom@24:     struct evsql *db;
terom@24: 
terom@24:     struct evfuse *ev_fuse;
terom@24: };
terom@24: 
terom@24: #define CONNINFO_DEFAULT "dbname=test"
terom@24: 
terom@24: // XXX: not sure how this should work
terom@24: #define CACHE_TIMEOUT 1.0
terom@24: 
terom@24: mode_t _dbfs_mode (const char *type) {
terom@24:     if (!strcmp(type, "DIR"))
terom@24:         return S_IFDIR;
terom@24: 
terom@24:     if (!strcmp(type, "REG"))
terom@24:         return S_IFREG;
terom@24: 
terom@24:     else {
terom@24:         WARNING("[dbfs] weird mode-type: %s", type);
terom@24:         return 0;
terom@24:     }
terom@24: }
terom@24: 
terom@24: void dbfs_init (void *userdata, struct fuse_conn_info *conn) {
terom@24:     INFO("[dbfs.init] userdata=%p, conn=%p", userdata, conn);
terom@24: 
terom@24: }
terom@24: 
terom@27: void dbfs_destroy (void *arg) {
terom@27:     struct dbfs *ctx = arg;
terom@27:     INFO("[dbfs.destroy %p]", ctx);
terom@24: 
terom@27:     // exit libevent
terom@27:     event_base_loopexit(ctx->ev_base, NULL);
terom@24: }
terom@24: 
terom@25: /*
terom@25:  * Check the result set.
terom@25:  *
terom@25:  * Returns;
terom@27:  *  -1  if the query failed, the columns do not match, or there are too many/few rows (unless rows was zero)
terom@25:  *  0   the results match
terom@25:  *  1   there were no results
terom@25:  */
terom@25: int _dbfs_check_res (const struct evsql_result_info *res, size_t rows, size_t cols) {
terom@25:     int err = 0;
terom@25: 
terom@25:     // check if it failed
terom@25:     if (res->error)
terom@25:         NERROR(evsql_result_error(res));
terom@25:         
terom@25:     // not found?
terom@25:     if (evsql_result_rows(res) == 0)
terom@25:         SERROR(err = 1);
terom@25: 
terom@25:     // duplicate rows?
terom@27:     if (rows && evsql_result_rows(res) != rows)
terom@27:         ERROR("wrong number of rows returned");
terom@25:     
terom@25:     // correct number of columns
terom@27:     if (evsql_result_cols(res) != cols)
terom@25:         ERROR("wrong number of columns: %zu", evsql_result_cols(res));
terom@25: 
terom@25:     // good
terom@25:     return 0;
terom@25: 
terom@25: error:
terom@25:     if (!err)
terom@25:         err = -1;
terom@25: 
terom@25:     return err;
terom@25: }
terom@25: 
terom@25: int _dbfs_stat_info (struct stat *st, const struct evsql_result_info *res, size_t row, size_t col_offset) {
terom@25:     int err = 0;
terom@25:     
terom@25:     uint16_t mode;
terom@25:     uint64_t size, nlink;
terom@25:     const char *type;
terom@25:     
terom@25:     // extract the data
terom@25:     if (0
terom@25:         ||  evsql_result_string(res, row, col_offset + 0, &type,       0 ) // inodes.type
terom@25:         ||  evsql_result_uint16(res, row, col_offset + 1, &mode,       0 ) // inodes.mode
terom@25:         ||  evsql_result_uint64(res, row, col_offset + 2, &size,       0 ) // inodes.size
terom@25:         ||  evsql_result_uint64(res, row, col_offset + 3, &nlink,      0 ) // count(*)
terom@25:     )
terom@25:         EERROR(err = EIO, "invalid db data");
terom@25: 
terom@25:     INFO("\tst_mode=S_IF%s | %ho, st_nlink=%llu, st_size=%llu", type, mode, (long long unsigned int) nlink, (long long unsigned int) size);
terom@25: 
terom@25:     // convert and store
terom@25:     st->st_mode = _dbfs_mode(type) | mode;
terom@25:     st->st_nlink = nlink;
terom@25:     st->st_size = size;
terom@25:     
terom@25:     // good
terom@25:     return 0;
terom@25: 
terom@25: error:
terom@25:     return -1;
terom@25: }
terom@25: 
terom@24: void _dbfs_lookup_result (const struct evsql_result_info *res, void *arg) {
terom@24:     struct fuse_req *req = arg;
terom@24:     struct fuse_entry_param e; ZINIT(e);
terom@24:     int err = 0;
terom@24:     
terom@24:     uint32_t ino;
terom@24:     
terom@25:     // check the results
terom@25:     if ((err = _dbfs_check_res(res, 1, 5)))
terom@25:         SERROR(err = (err ==  1 ? ENOENT : EIO));
terom@24:     
terom@24:     // get the data
terom@24:     if (0
terom@24:         ||  evsql_result_uint32(res, 0, 0, &ino,        0 ) // inodes.ino
terom@24:     )
terom@24:         EERROR(err = EIO, "invalid db data");
terom@25:         
terom@27:     INFO("[dbfs.lookup] -> ino=%u", ino);
terom@24:     
terom@25:     // stat attrs
terom@25:     if (_dbfs_stat_info(&e.attr, res, 0, 1))
terom@25:         goto error;
terom@24: 
terom@25:     // other attrs
terom@24:     e.ino = e.attr.st_ino = ino;
terom@24:     e.attr_timeout = CACHE_TIMEOUT;
terom@24:     e.entry_timeout = CACHE_TIMEOUT;
terom@24:         
terom@24:     // reply
terom@24:     if ((err = fuse_reply_entry(req, &e)))
terom@24:         EERROR(err, "fuse_reply_entry");
terom@24: 
terom@24: error:
terom@24:     if (err && (err = fuse_reply_err(req, err)))
terom@24:         EWARNING(err, "fuse_reply_err");
terom@24: 
terom@24:     // free
terom@24:     evsql_result_free(res);
terom@24: }
terom@24: 
terom@24: void dbfs_lookup (struct fuse_req *req, fuse_ino_t parent, const char *name) {
terom@24:     struct dbfs *ctx = fuse_req_userdata(req);
terom@24:     int err;
terom@24: 
terom@24:     INFO("[dbfs.lookup] parent=%lu name=%s", parent, name);
terom@24:     
terom@24:     // query and params
terom@24:     const char *sql = 
terom@24:         "SELECT"
terom@24:         " inodes.ino, inodes.type, inodes.mode, inodes.size, count(*)"
terom@24:         " FROM file_tree INNER JOIN inodes ON (file_tree.inode = inodes.ino)"
terom@25:         " WHERE file_tree.parent = $1::int4 AND file_tree.name = $2::varchar"
terom@24:         " GROUP BY inodes.ino, inodes.type, inodes.mode, inodes.size";
terom@24:     
terom@24:     static struct evsql_query_params params = EVSQL_PARAMS(EVSQL_FMT_BINARY) {
terom@24:         EVSQL_PARAM ( UINT32 ),
terom@24:         EVSQL_PARAM ( STRING ),
terom@24: 
terom@24:         EVSQL_PARAMS_END
terom@24:     };
terom@24:     
terom@24:     // build params
terom@24:     if (0
terom@24:         ||  evsql_param_uint32(&params, 0, parent)
terom@24:         ||  evsql_param_string(&params, 1, name)
terom@24:     )
terom@24:         EERROR(err = EIO, "evsql_param_*");
terom@24: 
terom@24:     // query
terom@25:     if (evsql_query_params(ctx->db, NULL, sql, &params, _dbfs_lookup_result, req) == NULL)
terom@24:         EERROR(err = EIO, "evsql_query_params");
terom@24: 
terom@24:     // XXX: handle interrupts
terom@24:     
terom@24:     // wait
terom@24:     return;
terom@24: 
terom@24: error:
terom@24:     if ((err = fuse_reply_err(req, err)))
terom@24:         EWARNING(err, "fuse_reply_err");
terom@24: }
terom@24: 
terom@25: void _dbfs_getattr_result (const struct evsql_result_info *res, void *arg) {
terom@25:     struct fuse_req *req = arg;
terom@25:     struct stat st; ZINIT(st);
terom@25:     int err = 0;
terom@25:     
terom@25:     // check the results
terom@25:     if ((err = _dbfs_check_res(res, 1, 4)))
terom@25:         SERROR(err = (err ==  1 ? ENOENT : EIO));
terom@25:         
terom@25:     INFO("[dbfs.getattr %p] -> (stat follows)", req);
terom@25:     
terom@25:     // stat attrs
terom@25:     if (_dbfs_stat_info(&st, res, 0, 0))
terom@25:         goto error;
terom@25: 
terom@25:     // XXX: we don't have the ino
terom@25:     st.st_ino = 0;
terom@25: 
terom@25:     // reply
terom@25:     if ((err = fuse_reply_attr(req, &st, CACHE_TIMEOUT)))
terom@25:         EERROR(err, "fuse_reply_entry");
terom@25: 
terom@25: error:
terom@25:     if (err && (err = fuse_reply_err(req, err)))
terom@25:         EWARNING(err, "fuse_reply_err");
terom@25: 
terom@25:     // free
terom@25:     evsql_result_free(res);
terom@25: }
terom@25: 
terom@25: static void dbfs_getattr (struct fuse_req *req, fuse_ino_t ino, struct fuse_file_info *fi) {
terom@25:     struct dbfs *ctx = fuse_req_userdata(req);
terom@25:     int err;
terom@25:     
terom@25:     (void) fi;
terom@25: 
terom@25:     INFO("[dbfs.getattr %p] ino=%lu", req, ino);
terom@25: 
terom@25:     const char *sql =
terom@25:         "SELECT"
terom@25:         " inodes.type, inodes.mode, inodes.size, count(*)"
terom@25:         " FROM inodes"
terom@27:         " WHERE inodes.ino = $1::int4"
terom@25:         " GROUP BY inodes.type, inodes.mode, inodes.size";
terom@25: 
terom@25:     static struct evsql_query_params params = EVSQL_PARAMS(EVSQL_FMT_BINARY) {
terom@25:         EVSQL_PARAM ( UINT32 ),
terom@25: 
terom@25:         EVSQL_PARAMS_END
terom@25:     };
terom@25: 
terom@25:     // build params
terom@25:     if (0
terom@25:         ||  evsql_param_uint32(&params, 0, ino)
terom@25:     )
terom@25:         SERROR(err = EIO);
terom@25:         
terom@25:     // query
terom@25:     if (evsql_query_params(ctx->db, NULL, sql, &params, _dbfs_getattr_result, req) == NULL)
terom@25:         SERROR(err = EIO);
terom@25: 
terom@25:     // XXX: handle interrupts
terom@25:     
terom@25:     // wait
terom@25:     return;
terom@25: 
terom@25: error:
terom@25:     if ((err = fuse_reply_err(req, err)))
terom@25:         EWARNING(err, "fuse_reply_err");
terom@25: }
terom@25: 
terom@26: struct dbfs_dirop {
terom@27:     struct fuse_file_info fi;
terom@26:     struct fuse_req *req;
terom@26: 
terom@26:     struct evsql_trans *trans;
terom@26:     
terom@27:     // dir/parent dir inodes
terom@27:     uint32_t ino, parent;
terom@27:     
terom@26:     // opendir has returned and releasedir hasn't been called yet
terom@26:     int open;
terom@27: 
terom@27:     // for readdir
terom@27:     struct dirbuf dirbuf;
terom@26: };
terom@26: 
terom@26: /*
terom@26:  * Free the dirop, aborting any in-progress transaction.
terom@26:  *
terom@26:  * req must be NULL.
terom@26:  */
terom@26: static void dbfs_dirop_free (struct dbfs_dirop *dirop) {
terom@27:     assert(dirop);
terom@27:     assert(!dirop->open);
terom@27:     assert(!dirop->req);
terom@26: 
terom@27:     if (dirop->trans) {
terom@27:         WARNING("aborting transaction");
terom@26:         evsql_trans_abort(dirop->trans);
terom@27:     }
terom@27: 
terom@27:     dirbuf_release(&dirop->dirbuf);
terom@26: 
terom@26:     free(dirop);
terom@26: }
terom@26: 
terom@27: static void dbfs_opendir_info_res (const struct evsql_result_info *res, void *arg) {
terom@27:     struct dbfs_dirop *dirop = arg;
terom@27:     struct fuse_req *req = dirop->req; dirop->req = NULL;
terom@27:     int err;
terom@27:     
terom@27:     assert(req != NULL);
terom@27:    
terom@27:     // check the results
terom@27:     if ((err = _dbfs_check_res(res, 1, 2)))
terom@27:         SERROR(err = (err ==  1 ? ENOENT : EIO));
terom@27: 
terom@27:     const char *type;
terom@27: 
terom@27:     // extract the data
terom@27:     if (0
terom@27:         ||  evsql_result_uint32(res, 0, 0, &dirop->parent,  1 ) // file_tree.parent
terom@27:         ||  evsql_result_string(res, 0, 1, &type,           0 ) // inodes.type
terom@27:     )
terom@27:         SERROR(err = EIO);
terom@27: 
terom@27:     // is it a dir?
terom@27:     if (_dbfs_mode(type) != S_IFDIR)
terom@27:         EERROR(err = ENOTDIR, "wrong type: %s", type);
terom@27:     
terom@27:     INFO("[dbfs.opendir %p:%p] -> ino=%lu, parent=%lu, type=%s", dirop, req, (unsigned long int) dirop->ino, (unsigned long int) dirop->parent, type);
terom@27:     
terom@27:     // send the openddir reply
terom@27:     if ((err = fuse_reply_open(req, &dirop->fi)))
terom@27:         EERROR(err, "fuse_reply_open");
terom@27:     
terom@27:     // dirop is now open
terom@27:     dirop->open = 1;
terom@27: 
terom@27:     // ok, wait for the opendir call
terom@27:     return;
terom@27: 
terom@27: error:
terom@27:     if (err) {
terom@27:         // abort the trans
terom@27:         evsql_trans_abort(dirop->trans);
terom@27:         
terom@27:         dirop->trans = NULL;
terom@27: 
terom@27:         if ((err = fuse_reply_err(req, err)))
terom@27:             EWARNING(err, "fuse_reply_err");
terom@27:     }
terom@27:     
terom@27:     // free
terom@27:     evsql_result_free(res);
terom@27: }
terom@27: 
terom@26: /*
terom@26:  * The opendir transaction is ready
terom@26:  */
terom@26: static void dbfs_dirop_ready (struct evsql_trans *trans, void *arg) {
terom@26:     struct dbfs_dirop *dirop = arg;
terom@27:     struct fuse_req *req = dirop->req;
terom@27:     struct dbfs *ctx = fuse_req_userdata(req);
terom@26:     int err;
terom@26: 
terom@27:     assert(req != NULL);
terom@27: 
terom@27:     INFO("[dbfs.opendir %p:%p] -> trans=%p", dirop, req, trans);
terom@26: 
terom@26:     // remember the transaction
terom@26:     dirop->trans = trans;
terom@27:     
terom@27:     // first fetch info about the dir itself
terom@27:     const char *sql =
terom@27:         "SELECT"
terom@27:         " file_tree.parent, inodes.type"
terom@27:         " FROM file_tree LEFT OUTER JOIN inodes ON (file_tree.inode = inodes.ino)"
terom@27:         " WHERE file_tree.inode = $1::int4";
terom@26: 
terom@27:     static struct evsql_query_params params = EVSQL_PARAMS(EVSQL_FMT_BINARY) {
terom@27:         EVSQL_PARAM ( UINT32 ),
terom@26: 
terom@27:         EVSQL_PARAMS_END
terom@27:     };
terom@27: 
terom@27:     // build params
terom@27:     if (0
terom@27:         ||  evsql_param_uint32(&params, 0, dirop->ino)
terom@27:     )
terom@27:         SERROR(err = EIO);
terom@27:         
terom@27:     // query
terom@27:     if (evsql_query_params(ctx->db, dirop->trans, sql, &params, dbfs_opendir_info_res, dirop) == NULL)
terom@27:         SERROR(err = EIO);
terom@27: 
terom@27:     // ok, wait for the info results
terom@26:     return;
terom@26: 
terom@26: error:
terom@27:     // we handle the req
terom@27:     dirop->req = NULL;
terom@27:     
terom@27:     // free the dirop
terom@26:     dbfs_dirop_free(dirop);
terom@27:     
terom@26:     if ((err = fuse_reply_err(req, err)))
terom@26:         EWARNING(err, "fuse_reply_err");
terom@26: }
terom@26: 
terom@26: static void dbfs_dirop_done (struct evsql_trans *trans, void *arg) {
terom@26:     struct dbfs_dirop *dirop = arg;
terom@27:     struct fuse_req *req = dirop->req; dirop->req = NULL;
terom@26:     int err;
terom@26:     
terom@27:     assert(req != NULL);
terom@26: 
terom@27:     INFO("[dbfs.releasedir %p:%p] -> OK", dirop, req);
terom@27: 
terom@27:     // forget trans
terom@27:     dirop->trans = NULL;
terom@27:     
terom@27:     // just reply
terom@27:     if ((err = fuse_reply_err(req, 0)))
terom@27:         EWARNING(err, "fuse_reply_err");
terom@27:     
terom@27:     // we can free dirop
terom@27:     dbfs_dirop_free(dirop);
terom@26: }
terom@26: 
terom@26: static void dbfs_dirop_error (struct evsql_trans *trans, void *arg) {
terom@26:     struct dbfs_dirop *dirop = arg;
terom@26:     int err;
terom@26: 
terom@26:     INFO("[dbfs:dirop %p:%p] evsql transaction error: %s", dirop, dirop->req, evsql_trans_error(trans));
terom@26:     
terom@26:     // deassociate the trans
terom@26:     dirop->trans = NULL;
terom@26:     
terom@26:     // error out and pending req
terom@26:     if (dirop->req) {
terom@26:         if ((err = fuse_reply_err(dirop->req, EIO)))
terom@26:             EWARNING(err, "fuse_erply_err");
terom@26: 
terom@26:         dirop->req = NULL;
terom@26: 
terom@26:         // only free the dirop if it isn't open
terom@26:         if (!dirop->open)
terom@26:             dbfs_dirop_free(dirop);
terom@26:     }
terom@26: }
terom@26: 
terom@26: static void dbfs_opendir (struct fuse_req *req, fuse_ino_t ino, struct fuse_file_info *fi) {
terom@26:     struct dbfs *ctx = fuse_req_userdata(req);
terom@26:     struct dbfs_dirop *dirop = NULL;
terom@26:     int err;
terom@26:     
terom@26:     // allocate it
terom@26:     if ((dirop = calloc(1, sizeof(*dirop))) == NULL && (err = EIO))
terom@26:         ERROR("calloc");
terom@26: 
terom@26:     INFO("[dbfs.opendir %p:%p] ino=%lu, fi=%p", dirop, req, ino, fi);
terom@27:     
terom@26:     // store the dirop
terom@27:     // copy *fi since it's on the stack
terom@27:     dirop->fi = *fi;
terom@27:     dirop->fi.fh = (uint64_t) dirop;
terom@26:     dirop->req = req;
terom@27:     dirop->ino = ino;
terom@26: 
terom@26:     // start a new transaction
terom@27:     if ((dirop->trans = evsql_trans(ctx->db, EVSQL_TRANS_SERIALIZABLE, dbfs_dirop_error, dbfs_dirop_ready, dbfs_dirop_done, dirop)) == NULL)
terom@26:         SERROR(err = EIO);
terom@26:     
terom@26:     // XXX: handle interrupts
terom@26:     
terom@26:     // wait
terom@26:     return;
terom@26: 
terom@26: error:
terom@26:     // we handle the req
terom@26:     dirop->req = NULL;
terom@26: 
terom@26:     dbfs_dirop_free(dirop);
terom@26: 
terom@26:     if ((err = fuse_reply_err(req, err)))
terom@26:         EWARNING(err, "fuse_reply_err");
terom@26: }
terom@26: 
terom@27: static void dbfs_readdir_files_res (const struct evsql_result_info *res, void *arg) {
terom@27:     struct dbfs_dirop *dirop = arg;
terom@27:     struct fuse_req *req = dirop->req; dirop->req = NULL;
terom@27:     int err;
terom@27:     size_t row;
terom@27:     
terom@27:     assert(req != NULL);
terom@27:     
terom@27:     // check the results
terom@27:     if ((err = _dbfs_check_res(res, 0, 4)) < 0)
terom@27:         SERROR(err = EIO);
terom@27:         
terom@27:     INFO("[dbfs.readdir %p:%p] -> files: res_rows=%zu", dirop, req, evsql_result_rows(res));
terom@27:         
terom@27:     // iterate over the rows
terom@27:     for (row = 0; row < evsql_result_rows(res); row++) {
terom@27:         uint32_t off, ino;
terom@27:         const char *name, *type;
terom@27: 
terom@27:         // extract the data
terom@27:         if (0
terom@27:             ||  evsql_result_uint32(res, row, 0, &off,          0 ) // file_tree.offset
terom@27:             ||  evsql_result_string(res, row, 1, &name,         0 ) // file_tree.name
terom@27:             ||  evsql_result_uint32(res, row, 2, &ino,          0 ) // inodes.ino
terom@27:             ||  evsql_result_string(res, row, 3, &type,         0 ) // inodes.type
terom@27:         )
terom@27:             SERROR(err = EIO);
terom@27:         
terom@27:         INFO("\t%zu: off=%lu+2, name=%s, ino=%lu, type=%s", row, (long unsigned int) off, name, (long unsigned int) ino, type);
terom@27: 
terom@27:         // add to the dirbuf
terom@27:         // offsets are just offset + 2
terom@27:         if ((err = dirbuf_add(req, &dirop->dirbuf, off + 2, off + 3, name, ino, _dbfs_mode(type))) < 0 && (err = EIO))
terom@27:             ERROR("failed to add dirent for inode=%lu", (long unsigned int) ino);
terom@27:         
terom@27:         // stop if it's full
terom@27:         if (err > 0)
terom@27:             break;
terom@27:     }
terom@27: 
terom@27:     // send it
terom@27:     if ((err = dirbuf_done(req, &dirop->dirbuf)))
terom@27:         EERROR(err, "failed to send buf");
terom@27:     
terom@27:     // good, fallthrough
terom@27:     err = 0;
terom@27: 
terom@27: error:
terom@27:     if (err) {
terom@27:         // abort the trans
terom@27:         evsql_trans_abort(dirop->trans);
terom@27:         
terom@27:         dirop->trans = NULL;
terom@27: 
terom@27:         // we handle the req
terom@27:         dirop->req = NULL;
terom@27:     
terom@27:         if ((err = fuse_reply_err(req, err)))
terom@27:             EWARNING(err, "fuse_reply_err");
terom@27:     }
terom@27:     
terom@27:     // free
terom@27:     evsql_result_free(res);
terom@27: }
terom@27: 
terom@26: static void dbfs_readdir (struct fuse_req *req, fuse_ino_t ino, size_t size, off_t off, struct fuse_file_info *fi) {
terom@26:     struct dbfs *ctx = fuse_req_userdata(req);
terom@26:     struct dbfs_dirop *dirop = (struct dbfs_dirop *) fi->fh;
terom@26:     int err;
terom@26: 
terom@27:     assert(!dirop->req);
terom@27:     assert(dirop->trans);
terom@27:     assert(dirop->ino == ino);
terom@27:     
terom@26:     INFO("[dbfs.readdir %p:%p] ino=%lu, size=%zu, off=%zu, fi=%p : trans=%p", dirop, req, ino, size, off, fi, dirop->trans);
terom@26: 
terom@26:     // update dirop
terom@26:     dirop->req = req;
terom@27: 
terom@27:     // create the dirbuf
terom@27:     if (dirbuf_init(&dirop->dirbuf, size, off))
terom@27:         SERROR(err = EIO);
terom@27: 
terom@27:     // add . and ..
terom@27:     // we set the next offset to 2, because all dirent offsets will be larger than that
terom@27:     if ((err = (0
terom@27:         ||  dirbuf_add(req, &dirop->dirbuf, 0, 1, ".",   dirop->ino,    S_IFDIR )
terom@27:         ||  dirbuf_add(req, &dirop->dirbuf, 1, 2, "..",  
terom@27:                         dirop->parent ? dirop->parent : dirop->ino,     S_IFDIR )
terom@27:     )) && (err = EIO))
terom@27:         ERROR("failed to add . and .. dirents");
terom@26: 
terom@26:     // select all relevant file entries
terom@26:     const char *sql = 
terom@26:         "SELECT"
terom@27:         " file_tree.\"offset\", file_tree.name, inodes.ino, inodes.type"
terom@26:         " FROM file_tree LEFT OUTER JOIN inodes ON (file_tree.inode = inodes.ino)"
terom@27:         " WHERE file_tree.parent = $1::int4 AND file_tree.\"offset\" >= $2::int4"
terom@26:         " LIMIT $3::int4";
terom@26: 
terom@26:     static struct evsql_query_params params = EVSQL_PARAMS(EVSQL_FMT_BINARY) {
terom@26:         EVSQL_PARAM ( UINT32 ),
terom@26:         EVSQL_PARAM ( UINT32 ),
terom@26:         EVSQL_PARAM ( UINT32 ),
terom@26: 
terom@26:         EVSQL_PARAMS_END
terom@26:     };
terom@26: 
terom@27:     // adjust offset to take . and .. into account
terom@27:     if (off > 2)
terom@27:         off -= 2;
terom@27:     
terom@27:     // build params
terom@27:     if (0
terom@27:         ||  evsql_param_uint32(&params, 0, dirop->ino)
terom@27:         ||  evsql_param_uint32(&params, 1, off)
terom@27:         ||  evsql_param_uint32(&params, 2, dirbuf_estimate(&dirop->dirbuf, 0))
terom@27:     )
terom@27:         SERROR(err = EIO);
terom@27: 
terom@27:     // query
terom@27:     if (evsql_query_params(ctx->db, dirop->trans, sql, &params, dbfs_readdir_files_res, dirop) == NULL)
terom@27:         SERROR(err = EIO);
terom@27: 
terom@27:     // good, wait
terom@27:     return;
terom@27: 
terom@27: error:
terom@27:     // we handle the req
terom@27:     dirop->req = NULL;
terom@27: 
terom@27:     // abort the trans
terom@27:     evsql_trans_abort(dirop->trans); dirop->trans = NULL;
terom@27: 
terom@27:     if ((err = fuse_reply_err(req, err)))
terom@27:         EWARNING(err, "fuse_reply_err");
terom@27: 
terom@26: }
terom@26: 
terom@26: static void dbfs_releasedir (struct fuse_req *req, fuse_ino_t ino, struct fuse_file_info *fi) {
terom@26:     struct dbfs *ctx = fuse_req_userdata(req);
terom@26:     struct dbfs_dirop *dirop = (struct dbfs_dirop *) fi->fh;
terom@26:     int err;
terom@26: 
terom@26:     (void) ctx;
terom@27:     
terom@27:     assert(!dirop->req);
terom@27:     assert(dirop->ino == ino);
terom@26: 
terom@26:     INFO("[dbfs.releasedir %p:%p] ino=%lu, fi=%p : trans=%p", dirop, req, ino, fi, dirop->trans);
terom@26: 
terom@26:     // update dirop. Must keep it open so that dbfs_dirop_error won't free it
terom@27:     // copy *fi since it's on the stack
terom@27:     dirop->fi = *fi;
terom@27:     dirop->fi.fh = (uint64_t) dirop;
terom@26:     dirop->req = req;
terom@27:     
terom@27:     if (dirop->trans) {
terom@27:         // we can commit the transaction, although we didn't make any changes
terom@27:         // if this fails the transaction, then dbfs_dirop_error will take care of sending the error, and dirop->req will be
terom@27:         // NULL
terom@27:         if (evsql_trans_commit(dirop->trans))
terom@27:             SERROR(err = EIO);
terom@26: 
terom@27:     } else {
terom@27:         // trans failed earlier, so have releasedir just succeed
terom@27:         if ((err = fuse_reply_err(req, 0)))
terom@27:             EERROR(err, "fuse_reply_err");
terom@26: 
terom@27:         // req is done
terom@27:         dirop->req = NULL;
terom@27:     }
terom@27: 
terom@27:     // fall-through to cleanup
terom@27:     err = 0;
terom@27: 
terom@27: error:
terom@27:     // the dirop is not open anymore and can be freed once done with
terom@26:     dirop->open = 0;
terom@26: 
terom@27:     // if trans_commit triggered an error but didn't call dbfs_dirop_error, we need to take care of it
terom@27:     if (err && dirop->req) {
terom@27:         int err2;
terom@26: 
terom@26:         // we handle the req
terom@26:         dirop->req = NULL;
terom@26: 
terom@27:         if ((err2 = fuse_reply_err(req, err)))
terom@27:             EWARNING(err2, "fuse_reply_err");
terom@27:     } 
terom@27:     
terom@27:     // same for trans, we need to abort it if trans_commit failed and fs_dirop_error didn't get called
terom@27:     if (err && dirop->trans) {
terom@26:         dbfs_dirop_free(dirop);
terom@27:     
terom@27:     } else
terom@27:       // alternatively, if the trans error'd itself away (now or earlier), we don't need to keep the dirop around
terom@27:       // anymore now that we've checkd its state
terom@27:       if (!dirop->trans) {
terom@27:         dbfs_dirop_free(dirop);
terom@26:     }
terom@26: }
terom@26: 
terom@24: struct fuse_lowlevel_ops dbfs_llops = {
terom@25: 
terom@24:     .init           = dbfs_init,
terom@24:     .destroy        = dbfs_destroy,
terom@24:     
terom@24:     .lookup         = dbfs_lookup,
terom@25: 
terom@25:     .getattr        = dbfs_getattr,
terom@26: 
terom@26:     .opendir        = dbfs_opendir,
terom@27:     .readdir        = dbfs_readdir,
terom@26:     .releasedir     = dbfs_releasedir,
terom@24: };
terom@24: 
terom@24: void dbfs_sql_error (struct evsql *evsql, void *arg) {
terom@24:     struct dbfs *ctx = arg;
terom@24: 
terom@24:     // AAAAAAAAAA.... panic
terom@24:     WARNING("[dbfs] SQL error: BREAKING MAIN LOOP LIEK NAO");
terom@24: 
terom@24:     event_base_loopbreak(ctx->ev_base);
terom@24: }
terom@24: 
terom@24: int main (int argc, char **argv) {
terom@24:     struct fuse_args fuse_args = FUSE_ARGS_INIT(argc, argv);
terom@24:     struct dbfs ctx; ZINIT(ctx);
terom@24:     
terom@24:     // parse args, XXX: fuse_args
terom@24:     ctx.db_conninfo = CONNINFO_DEFAULT;
terom@24:     
terom@24:     // init libevent
terom@24:     if ((ctx.ev_base = event_base_new()) == NULL)
terom@24:         ERROR("event_base_new");
terom@24:     
terom@24:     // setup signals
terom@24:     if ((ctx.signals = signals_default(ctx.ev_base)) == NULL)
terom@24:         ERROR("signals_default");
terom@24: 
terom@24:     // open sql
terom@24:     if ((ctx.db = evsql_new_pq(ctx.ev_base, ctx.db_conninfo, dbfs_sql_error, &ctx)) == NULL)
terom@24:         ERROR("evsql_new_pq");
terom@24: 
terom@24:     // open fuse
terom@24:     if ((ctx.ev_fuse = evfuse_new(ctx.ev_base, &fuse_args, &dbfs_llops, &ctx)) == NULL)
terom@24:         ERROR("evfuse_new");
terom@24: 
terom@24:     // run libevent
terom@24:     INFO("running libevent loop");
terom@24: 
terom@24:     if (event_base_dispatch(ctx.ev_base))
terom@24:         PERROR("event_base_dispatch");
terom@24:     
terom@24:     // clean shutdown
terom@24: 
terom@24: error :
terom@24:     // cleanup
terom@24:     if (ctx.ev_fuse)
terom@27:         evfuse_free(ctx.ev_fuse);
terom@24: 
terom@24:     // XXX: ctx.db
terom@24:     
terom@24:     if (ctx.signals)
terom@24:         signals_free(ctx.signals);
terom@24: 
terom@24:     if (ctx.ev_base)
terom@24:         event_base_free(ctx.ev_base);
terom@24:     
terom@24:     fuse_opt_free_args(&fuse_args);
terom@24: }
terom@24: