[paraslash.git] / osl.c

/*
 * Copyright (C) 2007 Andre Noll <maan@systemlinux.org>
 *
 * Licensed under the GPL v2. For licencing details see COPYING.
 */

/** \file osl.c Object storage layer functions. */
#include "para.h"
#include "error.h"
#include "list.h"
#include "osl_core.h"
#include <dirent.h> /* readdir() */
#include <assert.h>

//#define FMT_OFF_T "%li"


/**
 * A wrapper for lseek(2).
 *
 * \param fd The filedescriptor whose offset is to be to repositioned.
 * \param offset A value-result parameter.
 * \param whence Usual repositioning directive.
 *
 * Reposition the offset of the file descriptor \a fd to the argument \a offset
 * according to the directive \a whence. Upon successful return, \a offset
 * contains the resulting offset location as measured in bytes from the
 * beginning of the file.
 *
 * \return Positive on success. Otherwise, the function returns \p -E_LSEEK.
 *
 * \sa lseek(2).
 */
int para_lseek(int fd, off_t *offset, int whence)
{
        *offset = lseek(fd, *offset, whence);
        int ret = -E_LSEEK;
        if (*offset == -1)
                return ret;
        return 1;
}

/**
 * Waraper for the write system call.
 *
 * \param fd The file descriptor to write to.
 * \param buf The buffer to write.
 * \param size The length of \a buf in bytes.
 *
 * This function writes out the given bufffer and retries if an interrupt
 * occured during the write.
 *
 * \return On success, the number of bytes written is returned, otherwise, the
 * function returns \p -E_WRITE.
 *
 * \sa write(2).
 */
ssize_t para_write(int fd, const void *buf, size_t size)
{
        ssize_t ret;

        for (;;) {
                ret = write(fd, buf, size);
                if ((ret < 0) && (errno == EAGAIN || errno == EINTR))
                        continue;
                return ret >= 0? ret : -E_WRITE;
        }
}

/**
 * Write the whole buffer to a file descriptor.
 *
 * \param fd The file descriptor to write to.
 * \param buf The buffer to write.
 * \param size The length of \a buf in bytes.
 *
 * This function writes the given buffer and continues on short writes and
 * when interrupted by a signal.
 *
 * \return Positive on success, negative on errors. Possible errors: any
 * errors returned by para_write().
 *
 * \sa para_write().
 */
ssize_t para_write_all(int fd, const void *buf, size_t size)
{
//      PARA_DEBUG_LOG("writing %zu bytes\n", size);
        const char *b = buf;
        while (size) {
                ssize_t ret = para_write(fd, b, size);
//              PARA_DEBUG_LOG("ret: %zd\n", ret);
                if (ret < 0)
                        return ret;
                b += ret;
                size -= ret;
        }
        return 1;
}
/**
 * Wrapper for the open(2) system call.
 *
 * \param path The filename.
 * \param flags The usual open(2) flags.
 * \param mode Specifies the permissions to use.
 *
 * The mode parameter must be specified when O_CREAT is in the flags, and is ignored
 * otherwise.
 *
 * \return Positive on success, negative on errors. Possible errors: \p
 * E_EXIST, \p E_ISDIR, \p E_NOENT, \p E_OSL_PERM.
 *
 * \sa open(2).
 */
int para_open(const char *path, int flags, mode_t mode)
{
        PARA_DEBUG_LOG("opening %s\n", path);
        int ret = open(path, flags, mode);

        if (ret >= 0)
                return ret;
        switch (errno) {
        case EEXIST:
                ret = -E_EXIST;
                break;
        case EISDIR:
                ret = -E_ISDIR;
                break;
        case ENOENT:
                ret = -E_NOENT;
                break;
        case EPERM:
                ret = -E_OSL_PERM;
                break;
        };
        PARA_ERROR_LOG("failed to open %s: %s\n", path, strerror(errno));
        return ret;
}

/**
 * Open a file, write the given buffer and close the file.
 *
 * \param filename Full path to the file to open.
 * \param buf The buffer to write to the file.
 * \param size The size of \a buf.
 *
 * \return Positive on success, negative on errors. Possible errors include:
 * any errors from para_open() or para_write().
 *
 * \sa para_open(), para_write().
 */
int para_write_file(const char *filename, const void *buf, size_t size)
{
        int ret, fd;

        ret = para_open(filename, O_WRONLY | O_CREAT | O_EXCL, 0644);
        if (ret < 0)
                return ret;
        fd = ret;
        ret = para_write_all(fd, buf, size);
        if (ret < 0)
                goto out;
        ret = 1;
out:
        close(fd);
        return ret;
}

static int append_file(const char *filename, char *header, size_t header_size,
        char *data, size_t data_size, uint32_t *new_pos)
{
        int ret, fd;

//      PARA_DEBUG_LOG("appending %zu  + %zu bytes\n", header_size, data_size);
        ret = para_open(filename, O_WRONLY | O_CREAT | O_APPEND, 0644);
        if (ret < 0)
                return ret;
        fd = ret;
        if (header && header_size) {
                ret = para_write_all(fd, header, header_size);
                if (ret < 0)
                        goto out;
        }
        ret = para_write_all(fd, data, data_size);
        if (ret < 0)
                goto out;
        if (new_pos) {
                off_t offset = 0;
                ret = para_lseek(fd, &offset, SEEK_END);
                if (ret < 0)
                        goto out;
//              PARA_DEBUG_LOG("new file size: " FMT_OFF_T "\n", offset);
                *new_pos = offset;
        }
        ret = 1;
out:
        close(fd);
        return ret;
}

/**
 * Map a file into memory.
 *
 * \param path Name of the regular file to map.
 * \param open_mode Either \p O_RDONLY or \p O_RDWR.
 * \param obj On success, the mapping is returned here.
 *
 * \return Positive on success, negative on errors. Possible errors include: \p
 * E_FSTAT, any errors returned by para_open(), \p E_EMPTY, \p E_MMAP.
 *
 * \sa para_open(), mmap(2).
 */
int mmap_full_file(const char *path, int open_mode, struct osl_object *obj)
{
        int fd, ret, mmap_prot, mmap_flags;
        struct stat file_status;

        if (open_mode == O_RDONLY) {
                mmap_prot = PROT_READ;
                mmap_flags = MAP_PRIVATE;
        } else {
                mmap_prot = PROT_READ | PROT_WRITE;
                mmap_flags = MAP_SHARED;
        }
        ret = para_open(path, open_mode, 0);
        if (ret < 0)
                return ret;
        fd = ret;
        ret = -E_FSTAT;
        if (fstat(fd, &file_status) < 0)
                goto out;
        obj->size = file_status.st_size;
        ret = -E_EMPTY;
        PARA_DEBUG_LOG("%s: size %zu\n", path, obj->size);
        if (!obj->size)
                goto out;
        obj->data = mmap(NULL, obj->size, mmap_prot, mmap_flags, fd, 0);
        if (obj->data == MAP_FAILED) {
                obj->data = NULL;
                ret = -E_MMAP;
                goto out;
        }
        ret = 1;
out:
        close(fd);
        return ret;
}

/**
 * Traverse the given directory recursively.
 *
 * \param dirname The directory to traverse.
 * \param func The function to call for each entry.
 * \param private_data Pointer to an arbitrary data structure.
 *
 * For each regular file  in \a dirname, the supplied function \a func is
 * called.  The full path of the regular file and the \a private_data pointer
 * are passed to \a func.
 *
 * \return On success, 1 is returned. Otherwise, this function returns a
 * negative value which indicates the kind of the error.
 */
int for_each_file_in_dir(const char *dirname,
                int (*func)(const char *, const void *), const void *private_data)
{
        DIR *dir = NULL;
        struct dirent *entry;
        /*
         * Opening the current directory (".") and calling fchdir() to return
         * is usually faster and more reliable than saving cwd in some buffer
         * and calling chdir() afterwards (see man 3 getcwd).
         */
        int cwd_fd = open(".", O_RDONLY);
        struct stat s;
        int ret = -1;

//      PARA_DEBUG_LOG("dirname: %s\n", dirname);
        if (cwd_fd < 0)
                return -E_OSL_GETCWD;
        ret = -E_OSL_CHDIR;
        if (chdir(dirname) < 0)
                goto out;
        ret = -E_OSL_OPENDIR;
        dir = opendir(".");
        if (!dir)
                goto out;
        /* scan cwd recursively */
        while ((entry = readdir(dir))) {
                mode_t m;
                char *tmp;

                if (!strcmp(entry->d_name, "."))
                        continue;
                if (!strcmp(entry->d_name, ".."))
                        continue;
                ret = -E_OSL_LSTAT;
                if (lstat(entry->d_name, &s) == -1)
                        continue;
                m = s.st_mode;
                if (!S_ISREG(m) && !S_ISDIR(m))
                        continue;
                tmp = make_message("%s/%s", dirname, entry->d_name);
                if (!S_ISDIR(m)) {
                        ret = func(tmp, private_data);
                        free(tmp);
                        if (ret < 0)
                                goto out;
                        continue;
                }
                /* directory */
                ret = for_each_file_in_dir(tmp, func, private_data);
                free(tmp);
                if (ret < 0)
                        goto out;
        }
        ret = 1;
out:
        if (dir)
                closedir(dir);
        if (fchdir(cwd_fd) < 0 && ret >= 0)
                ret = -E_OSL_CHDIR;
        close(cwd_fd);
        return ret;
}

int para_mkdir(const char *path, mode_t mode)
{
        if (!mkdir(path, mode))
                return 1;
        if (errno == EEXIST)
                return -E_EXIST;
        if (errno == ENOSPC)
                return -E_NOSPC;
        if (errno == ENOTDIR)
                return -E_NOTDIR;
        if (errno == EPERM)
                return E_OSL_PERM;
        return -E_MKDIR;
}

static int verify_basename(const char *name)
{
        if (!name)
                return -E_BAD_BASENAME;
        if (!*name)
                return -E_BAD_BASENAME;
        if (strchr(name, '/'))
                return -E_BAD_BASENAME;
        if (!strcmp(name, ".."))
                return -E_BAD_BASENAME;
        if (!strcmp(name, "."))
                return -E_BAD_BASENAME;
        return 1;
}

/**
 * Compare two osl objects pointing to unsigned integers of 32 bit size.
 *
 * \param obj1 Pointer to the first integer.
 * \param obj2 Pointer to the second integer.
 *
 * \return The values required for an osl compare function.
 *
 * \sa osl_compare_func, osl_hash_compare().
 */
int uint32_compare(const struct osl_object *obj1, const struct osl_object *obj2)
{
        uint32_t d1 = read_u32((const char *)obj1->data);
        uint32_t d2 = read_u32((const char *)obj2->data);

        if (d1 < d2)
                return 1;
        if (d1 > d2)
                return -1;
        return 0;
}

/**
 * Compare two osl objects pointing to hash values.
 *
 * \param obj1 Pointer to the first hash object.
 * \param obj2 Pointer to the second hash object.
 *
 * \return The values required for an osl compare function.
 *
 * \sa osl_compare_func, uint32_compare().
 */
int osl_hash_compare(const struct osl_object *obj1, const struct osl_object *obj2)
{
        return hash_compare((HASH_TYPE *)obj1->data, (HASH_TYPE *)obj2->data);
}

static char *disk_storage_dirname(const struct osl_table *t, unsigned col_num,
                const char *ds_name)
{
        char *dirname, *column_name = column_filename(t, col_num);

        if (!(t->desc->flags & OSL_LARGE_TABLE))
                return column_name;
        dirname = make_message("%s/%.2s", column_name, ds_name);
        free(column_name);
        return dirname;
}

static char *disk_storage_name_of_object(const struct osl_table *t,
        const struct osl_object *obj)
{
        HASH_TYPE hash[HASH_SIZE];
        hash_object(obj, hash);
        return disk_storage_name_of_hash(t, hash);
}

static int disk_storage_name_of_row(const struct osl_table *t,
                const struct osl_row *row, char **name)
{
        struct osl_object obj;
        int ret = osl_get_object(t, row, t->disk_storage_name_column, &obj);

        if (ret < 0)
                return ret;
        *name = disk_storage_name_of_object(t, &obj);
        return 1;
}

static void column_name_hash(const char *col_name, HASH_TYPE *hash)
{
        return hash_function(col_name, strlen(col_name), hash);
}

static int init_column_descriptions(struct osl_table *t)
{
        int i, j, ret;
        const struct osl_column_description *cd;

        ret = -E_BAD_TABLE_DESC;
        ret = verify_basename(t->desc->name);
        if (ret < 0)
                goto err;
        ret = -E_BAD_DB_DIR;
        if (!t->desc->dir)
                goto err;
        /* the size of the index header without column descriptions */
        t->index_header_size = IDX_COLUMN_DESCRIPTIONS;
        FOR_EACH_COLUMN(i, t->desc, cd) {
                struct osl_column *col = t->columns + i;
                if (cd->storage_flags & OSL_RBTREE) {
                        if (!cd->compare_function)
                                return -E_NO_COMPARE_FUNC;
                }
                if (cd->storage_type == OSL_NO_STORAGE)
                        continue;
                ret = -E_NO_COLUMN_NAME;
                if (!cd->name || !cd->name[0])
                        goto err;
                ret = verify_basename(cd->name);
                if (ret < 0)
                        goto err;
                t->index_header_size += index_column_description_size(cd->name);
                column_name_hash(cd->name, col->name_hash);
                ret = -E_DUPLICATE_COL_NAME;
                for (j = i + 1; j < t->desc->num_columns; j++) {
                        const char *name2 = get_column_description(t->desc,
                                j)->name;
                        if (cd->name && name2 && !strcmp(cd->name, name2))
                                goto err;
                }
        }
        return 1;
err:
        return ret;
}

/**
 * Initialize a struct table from given table description.
 *
 * \param desc The description of the osl table.
 * \param table_ptr Result is returned here.
 *
 * This function performs several sanity checks on \p desc and returns if any
 * of these tests fail. On success, a struct \p osl_table is allocated and
 * initialized with data derived from \p desc.
 *
 * \return Positive on success, negative on errors. Possible errors include: \p
 * E_BAD_TABLE_DESC, \p E_NO_COLUMN_DESC, \p E_NO_COLUMNS, \p
 * E_BAD_STORAGE_TYPE, \p E_BAD_STORAGE_FLAGS, \p E_BAD_STORAGE_SIZE, \p
 * E_NO_UNIQUE_RBTREE_COLUMN, \p E_NO_RBTREE_COL.
 *
 * \sa struct osl_table.
 */
int init_table_structure(const struct osl_table_description *desc,
                struct osl_table **table_ptr)
{
        const struct osl_column_description *cd;
        struct osl_table *t = para_calloc(sizeof(*t));
        int i, ret = -E_BAD_TABLE_DESC, have_disk_storage_name_column = 0;

        if (!desc)
                goto err;
        PARA_DEBUG_LOG("creating table structure for '%s' from table "
                "description\n", desc->name);
        ret = -E_NO_COLUMN_DESC;
        if (!desc->column_descriptions)
                goto err;
        ret = -E_NO_COLUMNS;
        if (!desc->num_columns)
                goto err;
        t->columns = para_calloc(desc->num_columns * sizeof(struct osl_column));
        t->desc = desc;
        FOR_EACH_COLUMN(i, t->desc, cd) {
                enum osl_storage_type st = cd->storage_type;
                enum osl_storage_flags sf = cd->storage_flags;
                struct osl_column *col = &t->columns[i];

                ret = -E_BAD_STORAGE_TYPE;
                if (st != OSL_MAPPED_STORAGE && st != OSL_DISK_STORAGE
                                && st != OSL_NO_STORAGE)
                        goto err;
                ret = -E_BAD_STORAGE_FLAGS;
                if (st == OSL_DISK_STORAGE && sf & OSL_RBTREE)
                        goto err;
                ret = -E_BAD_STORAGE_SIZE;
                if (sf & OSL_FIXED_SIZE && !cd->data_size)
                        goto err;
                switch (st) {
                case OSL_DISK_STORAGE:
                        t->num_disk_storage_columns++;
                        break;
                case OSL_MAPPED_STORAGE:
                        t->num_mapped_columns++;
                        col->index_offset = t->index_entry_size;
                        t->index_entry_size += 8;
                        break;
                case OSL_NO_STORAGE:
                        col->volatile_num = t->num_volatile_columns;
                        t->num_volatile_columns++;
                        break;
                }
                if (sf & OSL_RBTREE) {
                        col->rbtree_num = t->num_rbtrees;
                        t->num_rbtrees++;
                        if ((sf & OSL_UNIQUE) && (st == OSL_MAPPED_STORAGE)) {
                                if (!have_disk_storage_name_column)
                                        t->disk_storage_name_column = i;
                                have_disk_storage_name_column = 1;
                        }
                }
        }
        ret = -E_NO_UNIQUE_RBTREE_COLUMN;
        if (t->num_disk_storage_columns && !have_disk_storage_name_column)
                goto err;
        ret = -E_NO_RBTREE_COL;
        if (!t->num_rbtrees)
                goto err;
        /* success */
        PARA_DEBUG_LOG("OK. Index entry size: %u\n", t->index_entry_size);
        ret = init_column_descriptions(t);
        if (ret < 0)
                goto err;
        *table_ptr = t;
        return 1;
err:
        free(t->columns);
        free(t);
        return ret;
}

/**
 * Read the table description from index header.
 *
 * \param map The memory mapping of the index file.
 * \param desc The values found in the index header are returned here.
 *
 * Read the index header, check for the paraslash magic string and the table version number.
 * Read all information stored in the index header into \a desc.
 *
 * \return Positive on success, negative on errors.
 *
 * \sa struct osl_table_description, osl_create_table.
 */
int read_table_desc(struct osl_object *map, struct osl_table_description *desc)
{
        char *buf = map->data;
        uint8_t version;
        uint16_t header_size;
        int ret, i;
        unsigned offset;
        struct osl_column_description *cd;

        if (map->size < MIN_INDEX_HEADER_SIZE(1))
                return -E_SHORT_TABLE;
        if (strncmp(buf + IDX_PARA_MAGIC, PARA_MAGIC, strlen(PARA_MAGIC)))
                return -E_NO_MAGIC;
        version = read_u8(buf + IDX_VERSION);
        if (version < MIN_TABLE_VERSION || version > MAX_TABLE_VERSION)
                return -E_VERSION_MISMATCH;
        desc->num_columns = read_u8(buf + IDX_TABLE_FLAGS);
        desc->flags = read_u8(buf + IDX_TABLE_FLAGS);
        desc->num_columns = read_u16(buf + IDX_NUM_COLUMNS);
        PARA_DEBUG_LOG("%u columns\n", desc->num_columns);
        if (!desc->num_columns)
                return -E_NO_COLUMNS;
        header_size = read_u16(buf + IDX_HEADER_SIZE);
        if (map->size < header_size)
                return -E_BAD_SIZE;
        desc->column_descriptions = para_calloc(desc->num_columns
                * sizeof(struct osl_column_description));
        offset = IDX_COLUMN_DESCRIPTIONS;
        FOR_EACH_COLUMN(i, desc, cd) {
                char *null_byte;

                ret = -E_SHORT_TABLE;
                if (map->size < offset + MIN_IDX_COLUMN_DESCRIPTION_SIZE) {
                        PARA_ERROR_LOG("map size = %zu < %u = offset + min desc size\n",
                                map->size, offset + MIN_IDX_COLUMN_DESCRIPTION_SIZE);
                        goto err;
                }
                cd->storage_type = read_u16(buf + offset + IDX_CD_STORAGE_TYPE);
                cd->storage_flags = read_u16(buf + offset +
                        IDX_CD_STORAGE_FLAGS);
                cd->data_size = read_u32(buf + offset + IDX_CD_DATA_SIZE);
                null_byte = memchr(buf + offset + IDX_CD_NAME, '\0',
                        map->size - offset - IDX_CD_NAME);
                ret = -E_INDEX_CORRUPTION;
                if (!null_byte)
                        goto err;
                cd->name = para_strdup(buf + offset + IDX_CD_NAME);
                offset += index_column_description_size(cd->name);
        }
        if (offset != header_size) {
                ret = -E_INDEX_CORRUPTION;
                PARA_ERROR_LOG("real header size = %u != %u = stored header size\n",
                        offset, header_size);
                goto err;
        }
        return 1;
err:
        FOR_EACH_COLUMN(i, desc, cd)
                free(cd->name);
        return ret;
}

/*
 * check whether the table description given by \p t->desc matches the on-disk
 * table structure stored in the index of \a t.
 */
static int compare_table_descriptions(struct osl_table *t)
{
        int i, ret;
        struct osl_table_description desc;
        const struct osl_column_description *cd1, *cd2;

        /* read the on-disk structure into desc */
        ret = read_table_desc(&t->index_map, &desc);
        if (ret < 0)
                return ret;
        ret = -E_BAD_TABLE_FLAGS;
        if (desc.flags != t->desc->flags)
                goto out;
        ret = E_BAD_COLUMN_NUM;
        if (desc.num_columns != t->desc->num_columns)
                goto out;
        FOR_EACH_COLUMN(i, t->desc, cd1) {
                cd2 = get_column_description(&desc, i);
                ret = -E_BAD_STORAGE_TYPE;
                if (cd1->storage_type != cd2->storage_type)
                        goto out;
                ret = -E_BAD_STORAGE_FLAGS;
                if (cd1->storage_flags != cd2->storage_flags) {
                        PARA_ERROR_LOG("sf1 = %u != %u = sf2\n",
                                cd1->storage_flags, cd2->storage_flags);
                        goto out;
                }
                ret = -E_BAD_DATA_SIZE;
                if (cd1->storage_flags & OSL_FIXED_SIZE)
                        if (cd1->data_size != cd2->data_size)
                                goto out;
                ret = -E_BAD_COLUMN_NAME;
                if (strcmp(cd1->name, cd2->name))
                        goto out;
        }
        PARA_DEBUG_LOG("table description of '%s' matches on-disk data, good\n",
                t->desc->name);
        ret = 1;
out:
        FOR_EACH_COLUMN(i, &desc, cd1)
                free(cd1->name);
        free(desc.column_descriptions);
        return ret;
}

static int create_table_index(struct osl_table *t)
{
        char *buf, *filename;
        int i, ret;
        size_t size = t->index_header_size;
        const struct osl_column_description *cd;
        unsigned offset;

        PARA_INFO_LOG("creating %zu byte index for table %s\n", size,
                t->desc->name);
        buf = para_calloc(size);
        sprintf(buf + IDX_PARA_MAGIC, "%s", PARA_MAGIC);
        write_u8(buf + IDX_TABLE_FLAGS, t->desc->flags);
        write_u8(buf + IDX_DIRTY_FLAG, 0);
        write_u8(buf + IDX_VERSION, CURRENT_TABLE_VERSION);
        write_u16(buf + IDX_NUM_COLUMNS, t->desc->num_columns);
        write_u16(buf + IDX_HEADER_SIZE, t->index_header_size);
        offset = IDX_COLUMN_DESCRIPTIONS;
        FOR_EACH_COLUMN(i, t->desc, cd) {
                write_u16(buf + offset + IDX_CD_STORAGE_TYPE,
                        cd->storage_type);
                write_u16(buf + offset + IDX_CD_STORAGE_FLAGS,
                        cd->storage_flags);
                if (cd->storage_flags & OSL_FIXED_SIZE)
                        write_u32(buf + offset + IDX_CD_DATA_SIZE,
                                cd->data_size);
                strcpy(buf + offset + IDX_CD_NAME, cd->name);
                offset += index_column_description_size(cd->name);
        }
        assert(offset = size);
        filename = index_filename(t->desc);
        ret = para_write_file(filename, buf, size);
        free(buf);
        free(filename);
        return ret;
}

/**
 * Create a new osl table.
 *
 * \param desc Pointer to the table description.
 *
 * \return Positive on success, negative on errors. Possible errors include: \p
 * E_BAD_TABLE_DESC, \p E_BAD_DB_DIR, \p E_BAD_BASENAME, \p E_NO_COMPARE_FUNC, \p
 * E_NO_COLUMN_NAME, \p E_DUPLICATE_COL_NAME, \p E_MKDIR, any errors returned
 * by para_open().
 */
int osl_create_table(const struct osl_table_description *desc)
{
        const struct osl_column_description *cd;
        char *table_dir = NULL, *filename;
        struct osl_table *t;
        int i, ret = init_table_structure(desc, &t);

        if (ret < 0)
                return ret;
        PARA_INFO_LOG("creating %s\n", desc->name);
        FOR_EACH_COLUMN(i, t->desc, cd) {
                if (cd->storage_type == OSL_NO_STORAGE)
                        continue;
                if (!table_dir) {
                        ret = para_mkdir(desc->dir, 0777);
                        if (ret < 0 && ret != -E_EXIST)
                                goto out;
                        table_dir = make_message("%s/%s", desc->dir,
                                desc->name);
                        ret = para_mkdir(table_dir, 0777);
                        if (ret < 0)
                                goto out;
                }
                filename = column_filename(t, i);
                PARA_INFO_LOG("filename: %s\n", filename);
                if (cd->storage_type == OSL_MAPPED_STORAGE) {
                        ret = para_open(filename, O_RDWR | O_CREAT | O_EXCL,
                                0644);
                        free(filename);
                        if (ret < 0)
                                goto out;
                        close(ret);
                        continue;
                }
                /* DISK STORAGE */
                ret = para_mkdir(filename, 0777);
                free(filename);
                if (ret < 0)
                        goto out;
        }
        if (t->num_mapped_columns) {
                ret = create_table_index(t);
                if (ret < 0)
                        goto out;
        }
        ret = 1;
out:
        free(table_dir);
        free(t->columns);
        free(t);
        return ret;
}

static int table_is_dirty(struct osl_table *t)
{
        char *buf = (char *)t->index_map.data + IDX_DIRTY_FLAG;
        uint8_t dirty = read_u8(buf) & 0x1;
        return !!dirty;
}

static void mark_table_dirty(struct osl_table *t)
{
        char *buf = (char *)t->index_map.data + IDX_DIRTY_FLAG;
        write_u8(buf, read_u8(buf) | 1);
}

static void mark_table_clean(struct osl_table *t)
{
        char *buf = (char *)t->index_map.data + IDX_DIRTY_FLAG;
        write_u8(buf, read_u8(buf) & 0xfe);
}

static void unmap_column(struct osl_table *t, unsigned col_num)
{
        struct osl_object map = t->columns[col_num].data_map;
        int ret;
        if (!map.data)
                return;
        ret = para_munmap(map.data, map.size);
        assert(ret > 0);
        map.data = NULL;
}

/**
 * Unmap all mapped files of an osl table.
 *
 * \param t Pointer to a mapped table.
 * \param flags Options for unmapping.
 *
 * \return Positive on success, negative on errors. Possible errors include:
 * E_NOT_MAPPED, E_MUNMAP.
 *
 * \sa map_table(), enum osl_close_flags, para_munmap().
 */
int unmap_table(struct osl_table *t, enum osl_close_flags flags)
{
        unsigned i;
        const struct osl_column_description *cd;
        int ret;

        if (!t->num_mapped_columns) /* can this ever happen? */
                return 1;
        PARA_DEBUG_LOG("unmapping table '%s'\n", t->desc->name);
        if (!t->index_map.data)
                return -E_NOT_MAPPED;
        if (flags & OSL_MARK_CLEAN)
                mark_table_clean(t);
        ret = para_munmap(t->index_map.data, t->index_map.size);
        if (ret < 0)
                return ret;
        t->index_map.data = NULL;
        if (!t->num_rows)
                return 1;
        FOR_EACH_MAPPED_COLUMN(i, t, cd)
                unmap_column(t, i);
        return 1;
}

static int map_column(struct osl_table *t, unsigned col_num)
{
        struct stat statbuf;
        char *filename = column_filename(t, col_num);
        int ret = -E_STAT;
        if (stat(filename, &statbuf) < 0) {
                free(filename);
                return ret;
        }
        if (!(S_IFREG & statbuf.st_mode)) {
                free(filename);
                return ret;
        }
        ret = mmap_full_file(filename, O_RDWR,
                &t->columns[col_num].data_map);
        free(filename);
        return ret;
}

/**
 * Map the index file and all columns of type \p OSL_MAPPED_STORAGE into memory.
 *
 * \param t Pointer to an initialized table structure.
 * \param flags Mapping options.
 *
 * \return Negative return value on errors; on success the number of rows
 * (including invalid rows) is returned.
 *
 * \sa unmap_table(), enum map_table_flags, osl_open_table(), mmap(2).
 */
int map_table(struct osl_table *t, enum map_table_flags flags)
{
        char *filename;
        const struct osl_column_description *cd;
        int i = 0, ret, num_rows = 0;

        if (!t->num_mapped_columns)
                return 0;
        if (t->index_map.data)
                return -E_ALREADY_MAPPED;
        filename = index_filename(t->desc);
        PARA_DEBUG_LOG("mapping table '%s' (index: %s)\n", t->desc->name, filename);
        ret = mmap_full_file(filename, flags & MAP_TBL_FL_MAP_RDONLY?
                O_RDONLY : O_RDWR, &t->index_map);
        free(filename);
        if (ret < 0)
                return ret;
        if (flags & MAP_TBL_FL_VERIFY_INDEX) {
                ret = compare_table_descriptions(t);
                if (ret < 0)
                        goto err;
        }
        ret = -E_BUSY;
        if (!(flags & MAP_TBL_FL_IGNORE_DIRTY)) {
                if (table_is_dirty(t)) {
                        PARA_ERROR_LOG("%s is dirty\n", t->desc->name);
                        goto err;
                }
        }
        mark_table_dirty(t);
        num_rows = table_num_rows(t);
        if (!num_rows)
                return num_rows;
        /* map data files */
        FOR_EACH_MAPPED_COLUMN(i, t, cd) {
                ret = map_column(t, i);
                if (ret < 0)
                        goto err;
        }
        return num_rows;
err:    /* unmap what is already mapped */
        for (i--; i >= 0; i--) {
                struct osl_object map = t->columns[i].data_map;
                para_munmap(map.data, map.size);
                map.data = NULL;
        }
        para_munmap(t->index_map.data, t->index_map.size);
        t->index_map.data = NULL;
        return ret;
}

/**
 * Retrieve a mapped object by row and column number.
 *
 * \param t Pointer to an open osl table.
 * \param col_num Number of the mapped column containing the object to retrieve.
 * \param row_num Number of the row containing the object to retrieve.
 * \param obj The result is returned here.
 *
 * It is considered an error if \a col_num does not refer to a column
 * of storage type \p OSL_MAPPED_STORAGE.
 *
 * \return Positive on success, negative on errors. Possible errors include:
 * \p E_BAD_ID, \p E_INVALID_OBJECT.
 *
 * \sa osl_storage_type.
 */
int get_mapped_object(const struct osl_table *t, unsigned col_num,
        uint32_t row_num, struct osl_object *obj)
{
        struct osl_column *col = &t->columns[col_num];
        uint32_t offset;
        char *header;
        char *index_entry;
        int ret;

        if (t->num_rows <= row_num)
                return -E_BAD_ID;
        ret = get_index_entry(t, row_num, col_num, &index_entry);
        if (ret < 0)
                return ret;
        offset = read_u32(index_entry);
        obj->size = read_u32(index_entry + 4) - 1;
        header = col->data_map.data + offset;
        obj->data = header + 1;
        if (read_u8(header) == 0xff) {
                PARA_ERROR_LOG("col %u, size %zu, offset %u\n", col_num,
                        obj->size, offset);
                return -E_INVALID_OBJECT;
        }
        return 1;
}

static int search_rbtree(const struct osl_object *obj,
                const struct osl_table *t, unsigned col_num,
                struct rb_node **result, struct rb_node ***rb_link)
{
        struct osl_column *col = &t->columns[col_num];
        struct rb_node **new = &col->rbtree.rb_node, *parent = NULL;
        const struct osl_column_description *cd =
                get_column_description(t->desc, col_num);
        enum osl_storage_type st = cd->storage_type;
        while (*new) {
                struct osl_row *this_row = get_row_pointer(*new,
                        col->rbtree_num);
                int ret;
                struct osl_object this_obj;
                parent = *new;
                if (st == OSL_MAPPED_STORAGE) {
                        ret = get_mapped_object(t, col_num, this_row->id,
                                &this_obj);
                        if (ret < 0)
                                return ret;
                } else
                        this_obj = this_row->volatile_objects[col->volatile_num];
                ret = cd->compare_function(obj, &this_obj);
                if (!ret) {
                        if (result)
                                *result = get_rb_node_pointer(this_row,
                                        col->rbtree_num);
                        return 1;
                }
                if (ret < 0)
                        new = &((*new)->rb_left);
                else
                        new = &((*new)->rb_right);
        }
        if (result)
                *result = parent;
        if (rb_link)
                *rb_link = new;
        return -E_RB_KEY_NOT_FOUND;
}

static int insert_rbtree(struct osl_table *t, unsigned col_num,
        const struct osl_row *row, const struct osl_object *obj)
{
        struct rb_node *parent, **rb_link;
        unsigned rbtree_num;
        struct rb_node *n;
        int ret = search_rbtree(obj, t, col_num, &parent, &rb_link);

        if (ret > 0)
                return -E_RB_KEY_EXISTS;
        rbtree_num = t->columns[col_num].rbtree_num;
        n = get_rb_node_pointer(row, rbtree_num);
        rb_link_node(n, parent, rb_link);
        rb_insert_color(n, &t->columns[col_num].rbtree);
        return 1;
}

static void remove_rb_node(struct osl_table *t, unsigned col_num,
                const struct osl_row *row)
{
        struct osl_column *col = &t->columns[col_num];
        const struct osl_column_description *cd =
                get_column_description(t->desc, col_num);
        enum osl_storage_flags sf = cd->storage_flags;
        struct rb_node *victim, *splice_out_node, *tmp;
        if (!(sf & OSL_RBTREE))
                return;
        /*
         * Which node is removed/spliced out actually depends on how many
         * children the victim node has: If it has no children, it gets
         * deleted. If it has one child, it gets spliced out. If it has two
         * children, its successor (which has at most a right child) gets
         * spliced out.
         */
        victim = get_rb_node_pointer(row, col->rbtree_num);
        if (victim->rb_left && victim->rb_right)
                splice_out_node = rb_next(victim);
        else
                splice_out_node = victim;
        /* Go up to the root and decrement the size of each node in the path. */
        for (tmp = splice_out_node; tmp; tmp = rb_parent(tmp))
                tmp->size--;
        rb_erase(victim, &col->rbtree);
}

static int add_row_to_rbtrees(struct osl_table *t, uint32_t id,
                struct osl_object *volatile_objs, struct osl_row **row_ptr)
{
        unsigned i;
        int ret;
        struct osl_row *row = allocate_row(t->num_rbtrees);
        const struct osl_column_description *cd;

        row->id = id;
        row->volatile_objects = volatile_objs;
        FOR_EACH_RBTREE_COLUMN(i, t, cd) {
                if (cd->storage_type == OSL_MAPPED_STORAGE) {
                        struct osl_object obj;
                        ret = get_mapped_object(t, i, id, &obj);
                        if (ret < 0)
                                goto err;
                        ret = insert_rbtree(t, i, row, &obj);
                } else { /* volatile */
                        const struct osl_object *obj
                                = volatile_objs + t->columns[i].volatile_num;
                        ret = insert_rbtree(t, i, row, obj);
                }
                if (ret < 0)
                        goto err;
        }
        if (row_ptr)
                *row_ptr = row;
        return 1;
err: /* rollback changes, i.e. remove added entries from rbtrees */
        while (i)
                remove_rb_node(t, i--, row);
        free(row);
        return ret;
}

static void free_volatile_objects(const struct osl_table *t,
                enum osl_close_flags flags)
{
        int i, j;
        struct rb_node *n;
        struct osl_column *rb_col;
        const struct osl_column_description *cd;

        if (!t->num_volatile_columns)
                return;
        /* find the first rbtree column (any will do) */
        FOR_EACH_RBTREE_COLUMN(i, t, cd)
                break;
        rb_col = t->columns + i;
        /* walk that rbtree and free all volatile objects */
        for (n = rb_first(&rb_col->rbtree); n; n = rb_next(n)) {
                struct osl_row *r = get_row_pointer(n, rb_col->rbtree_num);
                if (flags & OSL_FREE_VOLATILE)
                        for (j = 0; j < t->num_volatile_columns; j++)
                                free(r->volatile_objects[j].data);
                free(r->volatile_objects);
        }
}

/**
 * Erase all rbtree nodes and free resources.
 *
 * \param t Pointer to an open osl table.
 *
 * This function is called by osl_close_table().
 */
void clear_rbtrees(struct osl_table *t)
{
        const struct osl_column_description *cd;
        unsigned i, rbtrees_cleared = 0;

        FOR_EACH_RBTREE_COLUMN(i, t, cd) {
                struct osl_column *col = &t->columns[i];
                struct rb_node *n;
                rbtrees_cleared++;
                for (n = rb_first(&col->rbtree); n;) {
                        struct osl_row *r;
                        rb_erase(n, &col->rbtree);
                        if (rbtrees_cleared == t->num_rbtrees) {
                                r = get_row_pointer(n, col->rbtree_num);
                                n = rb_next(n);
                                free(r);
                        } else
                                n = rb_next(n);
                }
        }

}

/**
 * Close an osl table.
 *
 * \param t Pointer to the table to be closed.
 * \param flags Options for what should be cleaned up.
 *
 * If osl_open_table() succeeds, the resulting table pointer must later be
 * passed to this function in order to flush all changes to the filesystem and
 * to free the resources that were allocated by osl_open_table().
 *
 * \return Positive on success, negative on errors. Possible errors: \p E_BAD_TABLE,
 * errors returned by unmap_table().
 *
 * \sa osl_open_table(), unmap_table().
 */
int osl_close_table(struct osl_table *t, enum osl_close_flags flags)
{
        int ret;

        if (!t)
                return -E_BAD_TABLE;
        free_volatile_objects(t, flags);
        clear_rbtrees(t);
        ret = unmap_table(t, flags);
        if (ret < 0)
                PARA_ERROR_LOG("unmap_table failed: %d\n", ret);
        free(t->columns);
        free(t);
        return ret;
}

/**
 * Find out whether the given row number corresponds to an invalid row.
 *
 * \param t Pointer to the osl table.
 * \param row_num The number of the row in question.
 *
 * By definition, a row is considered invalid if all its index entries
 * are invalid.
 *
 * \return Positive if \a row_num corresponds to an invalid row,
 * zero if it corresponds to a valid row, negative on errors.
 */
int row_is_invalid(struct osl_table *t, uint32_t row_num)
{
        char *index_entry;
        int i, ret = get_index_entry_start(t, row_num, &index_entry);

        if (ret < 0)
                return ret;
        for (i = 0; i < t->index_entry_size; i++) {
                if ((unsigned char)index_entry[i] != 0xff)
                        return 0;
        }
        PARA_INFO_LOG("row %d is invalid\n", row_num);
        return 1;
}

/**
 * Invalidate a row of an osl table.
 *
 * \param t Pointer to an open osl table.
 * \param row_num Number of the row to mark as invalid.
 *
 * This function marks each mapped object in the index entry of \a row as
 * invalid.
 *
 * \return Positive on success, negative on errors.
 */
int mark_row_invalid(struct osl_table *t, uint32_t row_num)
{
        char *index_entry;
        int i, ret = get_index_entry_start(t, row_num, &index_entry);

        PARA_INFO_LOG("marking row %d as invalid\n", row_num);
        if (ret < 0)
                return ret;
        for (i = 0; i < t->index_entry_size; i++)
                index_entry[i] = 0xff;
        return 1;
}

/**
 * Initialize all rbtrees and compute number of invalid rows.
 *
 * \param t The table containing the rbtrees to be initialized.
 *
 * \return Positive on success, negative on errors.
 */
int init_rbtrees(struct osl_table *t)
{
        int i, ret;
        const struct osl_column_description *cd;

        /* create rbtrees */
        FOR_EACH_RBTREE_COLUMN(i, t, cd)
                t->columns[i].rbtree = RB_ROOT;
        /* add valid rows to rbtrees */
        t->num_invalid_rows = 0;
        for (i = 0; i < t->num_rows; i++) {
                ret = row_is_invalid(t, i);
                if (ret < 0)
                        return ret;
                if (ret) {
                        t->num_invalid_rows++;
                        continue;
                }
                ret = add_row_to_rbtrees(t, i, NULL, NULL);
                if (ret < 0)
                        return ret;
        }
        return 1;
}

/**
 * Open an osl table.
 *
 * Each osl table must be opened before its data can be accessed.
 *
 * \param table_desc Describes the table to be opened.
 * \param result Contains a pointer to the open table on success.
 *
 * The table description given by \a desc should coincide with the
 * description used at creation time.
 *
 * \return Positive on success, negative on errors. Possible errors include:
 * errors returned by init_table_structure(), \p E_NOENT, \p E_STAT, \p \p
 * E_NOTDIR, \p E_BAD_TABLE_DESC, \p E_BAD_DB_DIR, \p E_NO_COMPARE_FUNC, \p
 * E_NO_COLUMN_NAME, errors returned by init_rbtrees().
 */
int osl_open_table(const struct osl_table_description *table_desc,
                struct osl_table **result)
{
        int i, ret;
        struct osl_table *t;
        const struct osl_column_description *cd;

        PARA_INFO_LOG("opening table %s\n", table_desc->name);
        ret = init_table_structure(table_desc, &t);
        if (ret < 0)
                return ret;
        FOR_EACH_DISK_STORAGE_COLUMN(i, t, cd) {
                /* check if directory exists */
                char *dirname = column_filename(t, i);
                struct stat statbuf;
                ret = stat(dirname, &statbuf);
                free(dirname);
                if (ret < 0) {
                        if (errno == ENOENT)
                                ret = -E_NOENT;
                        else
                                ret = -E_STAT;
                        goto err;
                }
                ret = -E_NOTDIR;
                if (!S_ISDIR(statbuf.st_mode))
                        goto err;
        }
        ret = map_table(t, MAP_TBL_FL_VERIFY_INDEX);
        if (ret < 0)
                goto err;
        t->num_rows = ret;
        PARA_DEBUG_LOG("num rows: %d\n", t->num_rows);
        ret = init_rbtrees(t);
        if (ret < 0) {
                osl_close_table(t, OSL_MARK_CLEAN); /* ignore further errors */
                return ret;
        }
        *result = t;
        return 1;
err:
        free(t->columns);
        free(t);
        return ret;
}

static int create_disk_storage_object_dir(const struct osl_table *t,
                unsigned col_num, const char *ds_name)
{
        char *dirname;
        int ret;

        if (!(t->desc->flags & OSL_LARGE_TABLE))
                return 1;
        dirname = disk_storage_dirname(t, col_num, ds_name);
        ret = para_mkdir(dirname, 0777);
        free(dirname);
        if (ret < 0 && ret != -E_EXIST)
                return ret;
        return 1;
}

static int write_disk_storage_file(const struct osl_table *t, unsigned col_num,
        const struct osl_object *obj, const char *ds_name)
{
        int ret;
        char *filename;

        ret = create_disk_storage_object_dir(t, col_num, ds_name);
        if (ret < 0)
                return ret;
        filename = disk_storage_path(t, col_num, ds_name);
        ret = para_write_file(filename, obj->data, obj->size);
        free(filename);
        return ret;
}

static int append_map_file(const struct osl_table *t, unsigned col_num,
        const struct osl_object *obj, uint32_t *new_size)
{
        char *filename = column_filename(t, col_num);
        int ret;
        char header = 0; /* zero means valid object */

//      PARA_DEBUG_LOG("appending %zu + 1 byte\n", obj->size);
        ret = append_file(filename, &header, 1, obj->data, obj->size,
                new_size);
        free(filename);
        return ret;
}

static int append_index_entry(const struct osl_table *t, char *new_index_entry)
{
        char *filename;
        int ret;

        if (!t->num_mapped_columns)
                return 1;
        filename = index_filename(t->desc);
//      PARA_DEBUG_LOG("appending %u bytes\n", t->index_entry_size);
        ret = append_file(filename, NULL, 0, new_index_entry,
                t->index_entry_size, NULL);
        free(filename);
        return ret;
}

/**
 * A wrapper for truncate(2)
 *
 * \param path Name of the regular file to truncate
 * \param size Number of bytes to \b shave \b off
 *
 * Truncate the regular file named by \a path by \a size bytes.
 *
 * \return Positive on success, negative on errors. Possible errors include: \p
 * E_STAT, \p E_BAD_SIZE, \p E_TRUNC.
 *
 * \sa truncate(2)
 */
int para_truncate(const char *path, off_t size)
{
        int ret;
        struct stat statbuf;

        ret = -E_STAT;
        if (stat(path, &statbuf) < 0)
                goto out;
        ret = -E_BAD_SIZE;
        if (statbuf.st_size < size)
                goto out;
        ret = -E_TRUNC;
        if (truncate(path, statbuf.st_size - size) < 0)
                goto out;
        ret = 1;
out:
        return ret;
}

static int truncate_mapped_file(const struct osl_table *t, unsigned col_num,
                off_t size)
{
        char *filename = column_filename(t, col_num);
        int ret = para_truncate(filename, size);
        free(filename);
        return ret;
}

static int delete_disk_storage_file(const struct osl_table *t, unsigned col_num,
                const char *ds_name)
{
        char *dirname, *filename = disk_storage_path(t, col_num, ds_name);
        int ret = unlink(filename);

        PARA_INFO_LOG("deleted %s\n", filename);
        free(filename);
        if (ret < 0) {
                if (errno == ENOENT)
                        return -E_NOENT;
                return -E_UNLINK;
        }
        if (!(t->desc->flags & OSL_LARGE_TABLE))
                return 1;
        dirname = disk_storage_dirname(t, col_num, ds_name);
        rmdir(dirname);
        free(dirname);
        return 1;
}

/**
 * Add a new row to an osl table and retrieve this row.
 *
 * \param t Pointer to an open osl table.
 * \param objects Array of objects to be added.
 * \param row Result pointer.
 *
 * The \a objects parameter must point to an array containing one object per
 * column.  The order of the objects in the array is given by the table
 * description of \a table. Several sanity checks are performed during object
 * insertion and the function returns without modifying the table if any of
 * these tests fail.  In fact, it is atomic in the sense that it either
 * succeeds or leaves the table unchanged (i.e. either all or none of the
 * objects are added to the table).
 *
 * It is considered an error if an object is added to a column with associated
 * rbtree if this object is equal to an object already contained in that column
 * (i.e. the compare function for the column's rbtree returns zero).
 *
 * Possible errors include: \p E_RB_KEY_EXISTS, \p E_BAD_DATA_SIZE.
 *
 * \return Positive on success, negative on errors.
 *
 * \sa struct osl_table_description, osl_compare_func, osl_add_row().
 */
int osl_add_and_get_row(struct osl_table *t, struct osl_object *objects,
                struct osl_row **row)
{
        int i, ret;
        char *ds_name = NULL;
        struct rb_node **rb_parents = NULL, ***rb_links = NULL;
        char *new_index_entry = NULL;
        struct osl_object *volatile_objs = NULL;
        const struct osl_column_description *cd;

        if (!t)
                return -E_BAD_TABLE;
        rb_parents = para_malloc(t->num_rbtrees * sizeof(struct rn_node*));
        rb_links = para_malloc(t->num_rbtrees * sizeof(struct rn_node**));
        if (t->num_mapped_columns)
                new_index_entry = para_malloc(t->index_entry_size);
        /* pass 1: sanity checks */
//      PARA_DEBUG_LOG("sanity tests: %p:%p\n", objects[0].data,
//              objects[1].data);
        FOR_EACH_COLUMN(i, t->desc, cd) {
                enum osl_storage_type st = cd->storage_type;
                enum osl_storage_flags sf = cd->storage_flags;

//              ret = -E_NULL_OBJECT;
//              if (!objects[i])
//                      goto out;
                if (st == OSL_DISK_STORAGE)
                        continue;
                if (sf & OSL_RBTREE) {
                        unsigned rbtree_num = t->columns[i].rbtree_num;
                        ret = -E_RB_KEY_EXISTS;
//                      PARA_DEBUG_LOG("checking whether %p exists\n",
//                              objects[i].data);
                        if (search_rbtree(objects + i, t, i,
                                        &rb_parents[rbtree_num],
                                        &rb_links[rbtree_num]) > 0)
                                goto out;
                }
                if (sf & OSL_FIXED_SIZE) {
//                      PARA_DEBUG_LOG("fixed size. need: %zu, have: %d\n",
//                              objects[i].size, cd->data_size);
                        ret = -E_BAD_DATA_SIZE;
                        if (objects[i].size != cd->data_size)
                                goto out;
                }
        }
        if (t->num_disk_storage_columns)
                ds_name = disk_storage_name_of_object(t,
                        &objects[t->disk_storage_name_column]);
        ret = unmap_table(t, OSL_MARK_CLEAN);
        if (ret < 0)
                goto out;
//      PARA_DEBUG_LOG("sanity tests passed%s\n", "");
        /* pass 2: create data files, append map data */
        FOR_EACH_COLUMN(i, t->desc, cd) {
                enum osl_storage_type st = cd->storage_type;
                if (st == OSL_NO_STORAGE)
                        continue;
                if (st == OSL_MAPPED_STORAGE) {
                        uint32_t new_size;
                        struct osl_column *col = &t->columns[i];
//                      PARA_DEBUG_LOG("appending object of size %zu\n",
//                              objects[i].size);
                        ret = append_map_file(t, i, objects + i, &new_size);
                        if (ret < 0)
                                goto rollback;
                        update_index_entry(new_index_entry, col, new_size,
                                objects[i].size);
                        continue;
                }
                /* DISK_STORAGE */
                ret = write_disk_storage_file(t, i, objects + i, ds_name);
                if (ret < 0)
                        goto rollback;
        }
        ret = append_index_entry(t, new_index_entry);
        if (ret < 0)
                goto rollback;
        ret = map_table(t, MAP_TBL_FL_VERIFY_INDEX);
        if (ret < 0) { /* truncate index and rollback changes */
                char *filename = index_filename(t->desc);
                para_truncate(filename, t->index_entry_size);
                free(filename);
                goto rollback;
        }
        /* pass 3: add entry to rbtrees */
        if (t->num_volatile_columns) {
                volatile_objs = para_calloc(t->num_volatile_columns
                        * sizeof(struct osl_object));
                FOR_EACH_VOLATILE_COLUMN(i, t, cd)
                        volatile_objs[t->columns[i].volatile_num] = objects[i];
        }
        t->num_rows++;
//      PARA_DEBUG_LOG("adding new entry as row #%d\n", t->num_rows - 1);
        ret = add_row_to_rbtrees(t, t->num_rows - 1, volatile_objs, row);
        if (ret < 0)
                goto out;
//      PARA_DEBUG_LOG("added new entry as row #%d\n", t->num_rows - 1);
        ret = 1;
        goto out;
rollback: /* rollback all changes made, ignore further errors */
        for (i--; i >= 0; i--) {
                cd = get_column_description(t->desc, i);
                enum osl_storage_type st = cd->storage_type;
                if (st == OSL_NO_STORAGE)
                        continue;

                if (st == OSL_MAPPED_STORAGE)
                        truncate_mapped_file(t, i, objects[i].size);
                else /* disk storage */
                        delete_disk_storage_file(t, i, ds_name);
        }
        /* ignore error and return previous error value */
        map_table(t, MAP_TBL_FL_VERIFY_INDEX);
out:
        free(new_index_entry);
        free(ds_name);
        free(rb_parents);
        free(rb_links);
        return ret;
}

/**
 * Add a new row to an osl table.
 *
 * \param t Same meaning as osl_add_and_get_row().
 * \param objects Same meaning as osl_add_and_get_row().
 *
 * \return The return value of the underlying call to osl_add_and_get_row().
 *
 * This is equivalent to osl_add_and_get_row(t, objects, NULL).
 */
int osl_add_row(struct osl_table *t, struct osl_object *objects)
{
        return osl_add_and_get_row(t, objects, NULL);
}

/**
 * Retrieve an object identified by row and column
 *
 * \param t Pointer to an open osl table.
 * \param r Pointer to the row.
 * \param col_num The column number.
 * \param object The result pointer.
 *
 * The column determined by \a col_num must be of type \p OSL_MAPPED_STORAGE
 * or \p OSL_NO_STORAGE, i.e. no disk storage objects may be retrieved by this
 * function.
 *
 * \return Positive if object was found, negative on errors. Possible errors
 * include: \p E_BAD_TABLE, \p E_BAD_STORAGE_TYPE.
 *
 * \sa osl_storage_type, osl_open_disk_object().
 */
int osl_get_object(const struct osl_table *t, const struct osl_row *r,
        unsigned col_num, struct osl_object *object)
{
        const struct osl_column_description *cd;

        if (!t)
                return -E_BAD_TABLE;
        cd = get_column_description(t->desc, col_num);
        /* col must not be disk storage */
        if (cd->storage_type == OSL_DISK_STORAGE)
                return -E_BAD_STORAGE_TYPE;
        if (cd->storage_type == OSL_MAPPED_STORAGE)
                return get_mapped_object(t, col_num, r->id, object);
        /* volatile */
        *object = r->volatile_objects[t->columns[col_num].volatile_num];
        return 1;
}

static int mark_mapped_object_invalid(const struct osl_table *t, uint32_t id,
                unsigned col_num)
{
        struct osl_object obj;
        char *p;
        int ret = get_mapped_object(t, col_num, id, &obj);

        if (ret < 0)
                return ret;
        p = obj.data;
        p--;
        *p = 0xff;
        return 1;
}

/**
 * Delete a row from an osl table.
 *
 * \param t Pointer to an open osl table.
 * \param row Pointer to the row to delete.
 *
 * This removes all disk storage objects, removes all rbtree nodes,  and frees
 * all volatile objects belonging to the given row. For mapped columns, the
 * data is merely marked invalid and may be pruned from time to time by
 * para_fsck.
 *
 * \return Positive on success, negative on errors. Possible errors include:
 * \p E_BAD_TABLE, errors returned by osl_get_object().
 */
int osl_del_row(struct osl_table *t, struct osl_row *row)
{
        struct osl_row *r = row;
        int i, ret;
        const struct osl_column_description *cd;

        if (!t)
                return -E_BAD_TABLE;
        PARA_INFO_LOG("deleting row %p\n", row);

        if (t->num_disk_storage_columns) {
                char *ds_name;
                ret = disk_storage_name_of_row(t, r, &ds_name);
                if (ret < 0)
                        goto out;
                FOR_EACH_DISK_STORAGE_COLUMN(i, t, cd)
                        delete_disk_storage_file(t, i, ds_name);
                free(ds_name);
        }
        FOR_EACH_COLUMN(i, t->desc, cd) {
                struct osl_column *col = t->columns + i;
                enum osl_storage_type st = cd->storage_type;
                remove_rb_node(t, i, r);
                if (st == OSL_MAPPED_STORAGE) {
                        mark_mapped_object_invalid(t, r->id, i);
                        continue;
                }
                if (st == OSL_NO_STORAGE)
                        free(r->volatile_objects[col->volatile_num].data);
        }
        if (t->num_mapped_columns) {
                ret = mark_row_invalid(t, r->id);
                if (ret < 0)
                        goto out;
                t->num_invalid_rows++;
        } else
                t->num_rows--;
        ret = 1;
out:
        free(r->volatile_objects);
        free(r);
        return ret;
}

/* test if column has an rbtree */
static int check_rbtree_col(const struct osl_table *t, unsigned col_num,
                struct osl_column **col)
{
        if (!t)
                return -E_BAD_TABLE;
        if (!(get_column_description(t->desc, col_num)->storage_flags & OSL_RBTREE))
                return -E_BAD_STORAGE_FLAGS;
        *col = t->columns + col_num;
        return 1;
}

/**
 * Get the row that contains the given object.
 *
 * \param t Pointer to an open osl table.
 * \param col_num The number of the column to be searched.
 * \param obj The object to be looked up.
 * \param result Points to the row containing \a obj.
 *
 * Lookup \a obj in \a t and return the row containing \a obj. The column
 * specified by \a col_num must have an associated rbtree.
 *
 * \return Positive on success, negative on errors. If an error occured, \a
 * result is set to \p NULL. Possible errors include: \p E_BAD_TABLE, \p
 * E_BAD_STORAGE_FLAGS, errors returned by get_mapped_object(), \p
 * E_RB_KEY_NOT_FOUND.
 *
 * \sa osl_storage_flags
 */
int osl_get_row(const struct osl_table *t, unsigned col_num,
                const struct osl_object *obj, struct osl_row **result)
{
        int ret;
        struct rb_node *node;
        struct osl_row *row;
        struct osl_column *col;

        *result = NULL;
        ret = check_rbtree_col(t, col_num, &col);
        if (ret < 0)
                return ret;
        ret = search_rbtree(obj, t, col_num, &node, NULL);
        if (ret < 0)
                return ret;
        row = get_row_pointer(node, t->columns[col_num].rbtree_num);
        *result = row;
        return 1;
}

static int rbtree_loop(struct osl_column *col,  void *private_data,
                osl_rbtree_loop_func *func)
{
        struct rb_node *n;

        for (n = rb_first(&col->rbtree); n; n = rb_next(n)) {
                struct osl_row *r = get_row_pointer(n, col->rbtree_num);
                int ret = func(r, private_data);
                if (ret < 0)
                        return ret;
        }
        return 1;
}

static int rbtree_loop_reverse(struct osl_column *col,  void *private_data,
                osl_rbtree_loop_func *func)
{
        struct rb_node *n;

        for (n = rb_last(&col->rbtree); n; n = rb_prev(n)) {
                struct osl_row *r = get_row_pointer(n, col->rbtree_num);
                int ret = func(r, private_data);
                if (ret < 0)
                        return ret;
        }
        return 1;
}

/**
 * Loop over all nodes in an rbtree.
 *
 * \param t Pointer to an open osl table.
 * \param col_num The column to use for iterating over the elements.
 * \param private_data Pointer that gets passed to \a func.
 * \param func The function to be called for each node in the rbtree.
 *
 * This function does an in-order walk of the rbtree associated with \a
 * col_num. It is an error if the \p OSL_RBTREE flag is not set for this
 * column. For each node in the rbtree, the given function \a func is called
 * with two \p void* pointers as arguments: The first argument points to the
 * row that contains the object corresponding to the rbtree node currently
 * traversed, and the \a private_data pointer is passed to \a func as the
 * second argument. The loop terminates either if \a func returns a negative
 * value, or if all nodes of the tree have been visited.
 *
 *
 * \return Positive on success, negative on errors. If the termination of the
 * loop was caused by \a func returning a negative value, this value is
 * returned.
 *
 * \sa osl_storage_flags, osl_rbtree_loop_reverse(), osl_compare_func.
 */
int osl_rbtree_loop(const struct osl_table *t, unsigned col_num,
        void *private_data, osl_rbtree_loop_func *func)
{
        struct osl_column *col;

        int ret = check_rbtree_col(t, col_num, &col);
        if (ret < 0)
                return ret;
        return rbtree_loop(col, private_data, func);
}

/**
 * Loop over all nodes in an rbtree in reverse order.
 *
 * \param t Identical meaning as in \p osl_rbtree_loop().
 * \param col_num Identical meaning as in \p osl_rbtree_loop().
 * \param private_data Identical meaning as in \p osl_rbtree_loop().
 * \param func Identical meaning as in \p osl_rbtree_loop().
 *
 * This function is identical to \p osl_rbtree_loop(), the only difference
 * is that the tree is walked in reverse order.
 *
 * \return The same return value as \p osl_rbtree_loop().
 *
 * \sa osl_rbtree_loop().
 */
int osl_rbtree_loop_reverse(const struct osl_table *t, unsigned col_num,
        void *private_data, osl_rbtree_loop_func *func)
{
        struct osl_column *col;

        int ret = check_rbtree_col(t, col_num, &col);
        if (ret < 0)
                return ret;
        return rbtree_loop_reverse(col, private_data, func);
}

/* TODO: Rollback changes on errors */
static int rename_disk_storage_objects(struct osl_table *t,
                struct osl_object *old_obj, struct osl_object *new_obj)
{
        int i, ret;
        const struct osl_column_description *cd;
        char *old_ds_name, *new_ds_name;

        if (!t->num_disk_storage_columns)
                return 1; /* nothing to do */
        if (old_obj->size == new_obj->size && !memcmp(new_obj->data,
                        old_obj->data, new_obj->size))
                return 1; /* object did not change */
        old_ds_name = disk_storage_name_of_object(t, old_obj);
        new_ds_name = disk_storage_name_of_object(t, new_obj);
        FOR_EACH_DISK_STORAGE_COLUMN(i, t, cd) {
                char *old_filename, *new_filename;
                ret = create_disk_storage_object_dir(t, i, new_ds_name);
                if (ret < 0)
                        goto out;
                old_filename = disk_storage_path(t, i, old_ds_name);
                new_filename = disk_storage_path(t, i, new_ds_name);
                ret = para_rename(old_filename, new_filename);
                free(old_filename);
                free(new_filename);
                if (ret < 0)
                        goto out;
        }
        ret = 1;
out:
        free(old_ds_name);
        free(new_ds_name);
        return ret;

}

/**
 * Change an object in an osl table.
 *
 * \param t Pointer to an open osl table.
 * \param r Pointer to the row containing the object to be updated.
 * \param col_num Number of the column containing the object to be updated.
 * \param obj Pointer to the replacement object.
 *
 * This function  gets rid of all references to the old object. This includes
 * removal of the rbtree node in case there is an rbtree associated with \a
 * col_num. It then inserts \a obj into the table and the rbtree if neccessary.
 *
 * If the \p OSL_RBTREE flag is set for \a col_num, you \b MUST call this
 * function in order to change the contents of an object, even for volatile or
 * mapped columns of constant size (which may be updated directly if \p
 * OSL_RBTREE is not set).  Otherwise the rbtree might become corrupted.
 *
 * \return Positive on success, negative on errors. Possible errors include: \p
 * E_BAD_TABLE, \p E_RB_KEY_EXISTS, \p E_BAD_SIZE, \p E_NOENT, \p E_UNLINK,
 * errors returned by para_write_file(), \p E_MKDIR.
 */
int osl_update_object(struct osl_table *t, const struct osl_row *r,
                unsigned col_num, struct osl_object *obj)
{
        struct osl_column *col;
        const struct osl_column_description *cd;
        int ret;

        if (!t)
                return -E_BAD_TABLE;
        col = &t->columns[col_num];
        cd = get_column_description(t->desc, col_num);
        if (cd->storage_flags & OSL_RBTREE) {
                if (search_rbtree(obj, t, col_num, NULL, NULL) > 0)
                        return -E_RB_KEY_EXISTS;
        }
        if (cd->storage_flags & OSL_FIXED_SIZE) {
                if (obj->size != cd->data_size)
                        return -E_BAD_SIZE;
        }
        remove_rb_node(t, col_num, r);
        if (cd->storage_type == OSL_NO_STORAGE) { /* TODO: If fixed size, reuse object? */
                free(r->volatile_objects[col->volatile_num].data);
                r->volatile_objects[col->volatile_num] = *obj;
        } else if (cd->storage_type == OSL_DISK_STORAGE) {
                char *ds_name;
                ret = disk_storage_name_of_row(t, r, &ds_name);
                if (ret < 0)
                        return ret;
                ret = delete_disk_storage_file(t, col_num, ds_name);
                if (ret < 0 && ret != -E_NOENT) {
                        free(ds_name);
                        return ret;
                }
                ret = write_disk_storage_file(t, col_num, obj, ds_name);
                free(ds_name);
                if (ret < 0)
                        return ret;
        } else { /* mapped storage */
                struct osl_object old_obj;
                ret = get_mapped_object(t, col_num, r->id, &old_obj);
                if (ret < 0)
                        return ret;
                /*
                 * If the updated column is the disk storage name column, the
                 * disk storage name changes, so we have to rename all disk
                 * storage objects accordingly.
                 */
                if (col_num == t->disk_storage_name_column) {
                        ret = rename_disk_storage_objects(t, &old_obj, obj);
                        if (ret < 0)
                                return ret;
                }
                if (cd->storage_flags & OSL_FIXED_SIZE)
                        memcpy(old_obj.data, obj->data, cd->data_size);
                else { /* TODO: if the size doesn't change, use old space */
                        uint32_t new_data_map_size;
                        char *index_entry;
                        ret = get_index_entry_start(t, r->id, &index_entry);
                        if (ret < 0)
                                return ret;
                        ret = mark_mapped_object_invalid(t, r->id, col_num);
                        if (ret < 0)
                                return ret;
                        unmap_column(t, col_num);
                        ret = append_map_file(t, col_num, obj,
                                &new_data_map_size);
                        if (ret < 0)
                                return ret;
                        ret = map_column(t, col_num);
                        if (ret < 0)
                                return ret;
                        update_index_entry(index_entry, col, new_data_map_size,
                                obj->size);
                }
        }
        if (cd->storage_flags & OSL_RBTREE) {
                ret = insert_rbtree(t, col_num, r, obj);
                if (ret < 0)
                        return ret;
        }
        return 1;
}

/**
 * Retrieve an object of type \p OSL_DISK_STORAGE by row and column.
 *
 * \param t Pointer to an open osl table.
 * \param r Pointer to the row containing the object.
 * \param col_num The column number.
 * \param obj Points to the result upon successful return.
 *
 * For columns of type \p OSL_DISK_STORAGE, this function must be used to
 * retrieve one of its containing objects. Afterwards, osl_close_disk_object()
 * must be called in order to deallocate the resources.
 *
 * \return Positive on success, negative on errors. Possible errors include:
 * \p E_BAD_TABLE, \p E_BAD_STORAGE_TYPE, errors returned by osl_get_object().
 *
 * \sa osl_get_object(), osl_storage_type, osl_close_disk_object().
 */
int osl_open_disk_object(const struct osl_table *t, const struct osl_row *r,
                unsigned col_num, struct osl_object *obj)
{
        const struct osl_column_description *cd;
        char *ds_name, *filename;
        int ret;

        if (!t)
                return -E_BAD_TABLE;
        cd = get_column_description(t->desc, col_num);
        if (cd->storage_type != OSL_DISK_STORAGE)
                return -E_BAD_STORAGE_TYPE;

        ret = disk_storage_name_of_row(t, r, &ds_name);
        if (ret < 0)
                return ret;
        filename = disk_storage_path(t, col_num, ds_name);
        free(ds_name);
        PARA_DEBUG_LOG("filename: %s\n", filename);
        ret = mmap_full_file(filename, O_RDONLY, obj);
        free(filename);
        return ret;
}

/**
 * Free resources that were allocated during osl_open_disk_object().
 *
 * \param obj Pointer to the object previously returned by open_disk_object().
 *
 * \return The return value of the underlying call to para_munmap().
 *
 * \sa para_munmap().
 */
int osl_close_disk_object(struct osl_object *obj)
{
        return para_munmap(obj->data, obj->size);
}

/**
 * Get the number of rows of the given table.
 *
 * \param t Pointer to an open osl table.
 * \param num_rows Result is returned here.
 *
 * The number of rows returned via \a num_rows excluding any invalid rows.
 *
 * \return Positive on success, \p -E_BAD_TABLE if \a t is \p NULL.
 */
int osl_get_num_rows(const struct osl_table *t, unsigned *num_rows)
{
        if (!t)
                return -E_BAD_TABLE;
        assert(t->num_rows >= t->num_invalid_rows);
        *num_rows = t->num_rows - t->num_invalid_rows;
        return 1;
}

/**
 * Get the rank of a row.
 *
 * \param t An open osl table.
 * \param r The row to get the rank of.
 * \param col_num The number of an rbtree column.
 * \param rank Result pointer.
 *
 * The rank is, by definition, the position of the row in the linear order
 * determined by an inorder tree walk of the rbtree associated with column
 * number \a col_num of \a table.
 *
 * \return Positive on success, negative on errors.
 *
 * \sa osl_get_nth_row().
 */
int osl_get_rank(const struct osl_table *t, struct osl_row *r,
                unsigned col_num, unsigned *rank)
{
        struct osl_object obj;
        struct osl_column *col;
        struct rb_node *node;
        int ret = check_rbtree_col(t, col_num, &col);

        if (ret < 0)
                return ret;
        ret = osl_get_object(t, r, col_num, &obj);
        if (ret < 0)
                return ret;
        ret = search_rbtree(&obj, t, col_num, &node, NULL);
        if (ret < 0)
                return ret;
        ret = rb_rank(node, rank);
        if (ret < 0)
                return -E_BAD_ROW;
        return 1;
}

/**
 * Get the row with n-th greatest value.
 *
 * \param t Pointer to an open osl table.
 * \param col_num The column number.
 * \param n The rank of the desired row.
 * \param result Row is returned here.
 *
 * Retrieve the n-th order statistic with respect to the compare function
 * of the rbtree column \a col_num. In other words, get that row with
 * \a n th greatest value in column \a col_num. It's an error if
 * \a col_num is not a rbtree column, or if \a n is larger than the
 * number of rows in the table.
 *
 * \return Positive on success, negative on errors. Possible errors:
 * \p E_BAD_TABLE, \p E_BAD_STORAGE_FLAGS, \p E_RB_KEY_NOT_FOUND.
 *
 * \sa osl_storage_flags, osl_compare_func, osl_get_row(),
 * osl_rbtree_last_row(), osl_rbtree_first_row(), osl_get_rank().
 */
int osl_get_nth_row(const struct osl_table *t, unsigned col_num,
                unsigned n, struct osl_row **result)
{
        struct osl_column *col;
        struct rb_node *node;
        int ret = check_rbtree_col(t, col_num, &col);

        if (ret < 0)
                return ret;
        node = rb_nth(col->rbtree.rb_node, n);
        if (!node)
                return -E_RB_KEY_NOT_FOUND;
        *result = get_row_pointer(node, col->rbtree_num);
        return 1;
}

/**
 * Get the row corresponding to the smallest rbtree node of a column.
 *
 * \param t An open rbtree table.
 * \param col_num The number of the rbtree column.
 * \param result A pointer to the first row is returned here.
 *
 * The rbtree node of the smallest object (with respect to the corresponding
 * compare function) is selected and the row containing this object is
 * returned. It is an error if \a col_num refers to a column without an
 * associated rbtree.
 *
 * \return Positive on success, negative on errors.
 *
 * \sa osl_get_nth_row(), osl_rbtree_last_row().
 */
int osl_rbtree_first_row(const struct osl_table *t, unsigned col_num,
                struct osl_row **result)
{
        return osl_get_nth_row(t, col_num, 1, result);
}

/**
 * Get the row corresponding to the greatest rbtree node of a column.
 *
 * \param t The same meaning as in \p osl_rbtree_first_row().
 * \param col_num The same meaning as in \p osl_rbtree_first_row().
 * \param result The same meaning as in \p osl_rbtree_first_row().
 *
 * This function works just like osl_rbtree_first_row(), the only difference
 * is that the row containing the greatest rather than the smallest object is
 * returned.
 *
 * \return Positive on success, negative on errors.
 *
 * \sa osl_get_nth_row(), osl_rbtree_first_row().
 */
int osl_rbtree_last_row(const struct osl_table *t, unsigned col_num,
                struct osl_row **result)
{
        unsigned num_rows;
        int ret = osl_get_num_rows(t, &num_rows);

        if (ret < 0)
                return ret;
        return osl_get_nth_row(t, col_num, num_rows, result);
}