glass_compact.cc

Go to the documentation of this file.

/* Copyright (C) 2004,2005,2006,2007,2008,2009,2010,2011,2012,2013,2014,2015,2017 Olly Betts
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public License as
 * published by the Free Software Foundation; either version 2 of the
 * License, or (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301
 * USA
 */

#include <config.h>

#include "xapian/compactor.h"
#include "xapian/constants.h"
#include "xapian/error.h"
#include "xapian/types.h"

#include "autoptr.h"
#include <algorithm>
#include <queue>

#include <cerrno>
#include <cstdio>

#include "backends/flint_lock.h"
#include "glass_database.h"
#include "glass_defs.h"
#include "glass_table.h"
#include "glass_cursor.h"
#include "glass_version.h"
#include "filetests.h"
#include "internaltypes.h"
#include "pack.h"
#include "backends/valuestats.h"

#include "../byte_length_strings.h"
#include "../prefix_compressed_strings.h"

using namespace std;

// Put all the helpers in a namespace to avoid symbols colliding with those of
// the same name in other flint-derived backends.
namespace GlassCompact {

static inline bool
is_user_metadata_key(const string & key)
{
    return key.size() > 1 && key[0] == '\0' && key[1] == '\xc0';
}

static inline bool
is_valuestats_key(const string & key)
{
    return key.size() > 1 && key[0] == '\0' && key[1] == '\xd0';
}

static inline bool
is_valuechunk_key(const string & key)
{
    return key.size() > 1 && key[0] == '\0' && key[1] == '\xd8';
}

static inline bool
is_doclenchunk_key(const string & key)
{
    return key.size() > 1 && key[0] == '\0' && key[1] == '\xe0';
}

class PostlistCursor : private GlassCursor {
    Xapian::docid offset;

  public:
    string key, tag;
    Xapian::docid firstdid;
    Xapian::termcount tf, cf;

    PostlistCursor(GlassTable *in, Xapian::docid offset_)
        : GlassCursor(in), offset(offset_), firstdid(0)
    {
        find_entry(string());
        next();
    }

    bool next() {
        if (!GlassCursor::next()) return false;
        // We put all chunks into the non-initial chunk form here, then fix up
        // the first chunk for each term in the merged database as we merge.
        read_tag();
        key = current_key;
        tag = current_tag;
        tf = cf = 0;
        if (is_user_metadata_key(key)) return true;
        if (is_valuestats_key(key)) return true;
        if (is_valuechunk_key(key)) {
            const char * p = key.data();
            const char * end = p + key.length();
            p += 2;
            Xapian::valueno slot;
            if (!unpack_uint(&p, end, &slot))
                throw Xapian::DatabaseCorruptError("bad value key");
            Xapian::docid did;
            if (!unpack_uint_preserving_sort(&p, end, &did))
                throw Xapian::DatabaseCorruptError("bad value key");
            did += offset;

            key.assign("\0\xd8", 2);
            pack_uint(key, slot);
            pack_uint_preserving_sort(key, did);
            return true;
        }

        // Adjust key if this is *NOT* an initial chunk.
        // key is: pack_string_preserving_sort(key, tname)
        // plus optionally: pack_uint_preserving_sort(key, did)
        const char * d = key.data();
        const char * e = d + key.size();
        if (is_doclenchunk_key(key)) {
            d += 2;
        } else {
            string tname;
            if (!unpack_string_preserving_sort(&d, e, tname))
                throw Xapian::DatabaseCorruptError("Bad postlist key");
        }

        if (d == e) {
            // This is an initial chunk for a term, so adjust tag header.
            d = tag.data();
            e = d + tag.size();
            if (!unpack_uint(&d, e, &tf) ||
                !unpack_uint(&d, e, &cf) ||
                !unpack_uint(&d, e, &firstdid)) {
                throw Xapian::DatabaseCorruptError("Bad postlist key");
            }
            ++firstdid;
            tag.erase(0, d - tag.data());
        } else {
            // Not an initial chunk, so adjust key.
            size_t tmp = d - key.data();
            if (!unpack_uint_preserving_sort(&d, e, &firstdid) || d != e)
                throw Xapian::DatabaseCorruptError("Bad postlist key");
            if (is_doclenchunk_key(key)) {
                key.erase(tmp);
            } else {
                key.erase(tmp - 1);
            }
        }
        firstdid += offset;
        return true;
    }
};

class PostlistCursorGt {
  public:
    bool operator()(const PostlistCursor *a, const PostlistCursor *b) const {
        if (a->key > b->key) return true;
        if (a->key != b->key) return false;
        return (a->firstdid > b->firstdid);
    }
};

static string
encode_valuestats(Xapian::doccount freq,
                  const string & lbound, const string & ubound)
{
    string value;
    pack_uint(value, freq);
    pack_string(value, lbound);
    // We don't store or count empty values, so neither of the bounds
    // can be empty.  So we can safely store an empty upper bound when
    // the bounds are equal.
    if (lbound != ubound) value += ubound;
    return value;
}

static void
merge_postlists(Xapian::Compactor * compactor,
                GlassTable * out, vector<Xapian::docid>::const_iterator offset,
                vector<GlassTable*>::const_iterator b,
                vector<GlassTable*>::const_iterator e)
{
    priority_queue<PostlistCursor *, vector<PostlistCursor *>, PostlistCursorGt> pq;
    for ( ; b != e; ++b, ++offset) {
        GlassTable *in = *b;
        if (in->empty()) {
            // Skip empty tables.
            continue;
        }

        pq.push(new PostlistCursor(in, *offset));
    }

    string last_key;
    {
        // Merge user metadata.
        vector<string> tags;
        while (!pq.empty()) {
            PostlistCursor * cur = pq.top();
            const string& key = cur->key;
            if (!is_user_metadata_key(key)) break;

            if (key != last_key) {
                if (!tags.empty()) {
                    if (tags.size() > 1 && compactor) {
                        Assert(!last_key.empty());
                        // FIXME: It would be better to merge all duplicates
                        // for a key in one call, but currently we don't in
                        // multipass mode.
                        const string & resolved_tag =
                            compactor->resolve_duplicate_metadata(last_key,
                                                                  tags.size(),
                                                                  &tags[0]);
                        if (!resolved_tag.empty())
                            out->add(last_key, resolved_tag);
                    } else {
                        Assert(!last_key.empty());
                        out->add(last_key, tags[0]);
                    }
                    tags.resize(0);
                }
                last_key = key;
            }
            tags.push_back(cur->tag);

            pq.pop();
            if (cur->next()) {
                pq.push(cur);
            } else {
                delete cur;
            }
        }
        if (!tags.empty()) {
            if (tags.size() > 1 && compactor) {
                Assert(!last_key.empty());
                const string & resolved_tag =
                    compactor->resolve_duplicate_metadata(last_key,
                                                          tags.size(),
                                                          &tags[0]);
                if (!resolved_tag.empty())
                    out->add(last_key, resolved_tag);
            } else {
                Assert(!last_key.empty());
                out->add(last_key, tags[0]);
            }
        }
    }

    {
        // Merge valuestats.
        Xapian::doccount freq = 0;
        string lbound, ubound;

        while (!pq.empty()) {
            PostlistCursor * cur = pq.top();
            const string& key = cur->key;
            if (!is_valuestats_key(key)) break;
            if (key != last_key) {
                // For the first valuestats key, last_key will be the previous
                // key we wrote, which we don't want to overwrite.  This is the
                // only time that freq will be 0, so check that.
                if (freq) {
                    out->add(last_key, encode_valuestats(freq, lbound, ubound));
                    freq = 0;
                }
                last_key = key;
            }

            const string & tag = cur->tag;

            const char * pos = tag.data();
            const char * end = pos + tag.size();

            Xapian::doccount f;
            string l, u;
            if (!unpack_uint(&pos, end, &f)) {
                if (*pos == 0) throw Xapian::DatabaseCorruptError("Incomplete stats item in value table");
                throw Xapian::RangeError("Frequency statistic in value table is too large");
            }
            if (!unpack_string(&pos, end, l)) {
                if (*pos == 0) throw Xapian::DatabaseCorruptError("Incomplete stats item in value table");
                throw Xapian::RangeError("Lower bound in value table is too large");
            }
            size_t len = end - pos;
            if (len == 0) {
                u = l;
            } else {
                u.assign(pos, len);
            }
            if (freq == 0) {
                freq = f;
                lbound = l;
                ubound = u;
            } else {
                freq += f;
                if (l < lbound) lbound = l;
                if (u > ubound) ubound = u;
            }

            pq.pop();
            if (cur->next()) {
                pq.push(cur);
            } else {
                delete cur;
            }
        }

        if (freq) {
            out->add(last_key, encode_valuestats(freq, lbound, ubound));
        }
    }

    // Merge valuestream chunks.
    while (!pq.empty()) {
        PostlistCursor * cur = pq.top();
        const string & key = cur->key;
        if (!is_valuechunk_key(key)) break;
        Assert(!is_user_metadata_key(key));
        out->add(key, cur->tag);
        pq.pop();
        if (cur->next()) {
            pq.push(cur);
        } else {
            delete cur;
        }
    }

    Xapian::termcount tf = 0, cf = 0; // Initialise to avoid warnings.
    vector<pair<Xapian::docid, string>> tags;
    while (true) {
        PostlistCursor * cur = NULL;
        if (!pq.empty()) {
            cur = pq.top();
            pq.pop();
        }
        Assert(cur == NULL || !is_user_metadata_key(cur->key));
        if (cur == NULL || cur->key != last_key) {
            if (!tags.empty()) {
                string first_tag;
                pack_uint(first_tag, tf);
                pack_uint(first_tag, cf);
                pack_uint(first_tag, tags[0].first - 1);
                string tag = tags[0].second;
                tag[0] = (tags.size() == 1) ? '1' : '0';
                first_tag += tag;
                out->add(last_key, first_tag);

                string term;
                if (!is_doclenchunk_key(last_key)) {
                    const char * p = last_key.data();
                    const char * end = p + last_key.size();
                    if (!unpack_string_preserving_sort(&p, end, term) || p != end)
                        throw Xapian::DatabaseCorruptError("Bad postlist chunk key");
                }

                auto i = tags.begin();
                while (++i != tags.end()) {
                    tag = i->second;
                    tag[0] = (i + 1 == tags.end()) ? '1' : '0';
                    out->add(pack_glass_postlist_key(term, i->first), tag);
                }
            }
            tags.clear();
            if (cur == NULL) break;
            tf = cf = 0;
            last_key = cur->key;
        }
        tf += cur->tf;
        cf += cur->cf;
        tags.push_back(make_pair(cur->firstdid, cur->tag));
        if (cur->next()) {
            pq.push(cur);
        } else {
            delete cur;
        }
    }
}

struct MergeCursor : public GlassCursor {
    explicit MergeCursor(GlassTable *in) : GlassCursor(in) {
        find_entry(string());
        next();
    }
};

struct CursorGt {
    bool operator()(const GlassCursor *a, const GlassCursor *b) const {
        if (b->after_end()) return false;
        if (a->after_end()) return true;
        return (a->current_key > b->current_key);
    }
};

static void
merge_spellings(GlassTable * out,
                vector<GlassTable*>::const_iterator b,
                vector<GlassTable*>::const_iterator e)
{
    priority_queue<MergeCursor *, vector<MergeCursor *>, CursorGt> pq;
    for ( ; b != e; ++b) {
        GlassTable *in = *b;
        if (!in->empty()) {
            pq.push(new MergeCursor(in));
        }
    }

    while (!pq.empty()) {
        MergeCursor * cur = pq.top();
        pq.pop();

        string key = cur->current_key;
        if (pq.empty() || pq.top()->current_key > key) {
            // No need to merge the tags, just copy the (possibly compressed)
            // tag value.
            bool compressed = cur->read_tag(true);
            out->add(key, cur->current_tag, compressed);
            if (cur->next()) {
                pq.push(cur);
            } else {
                delete cur;
            }
            continue;
        }

        // Merge tag values with the same key:
        string tag;
        if (key[0] != 'W') {
            // We just want the union of words, so copy over the first instance
            // and skip any identical ones.
            priority_queue<PrefixCompressedStringItor *,
                           vector<PrefixCompressedStringItor *>,
                           PrefixCompressedStringItorGt> pqtag;
            // Stick all the MergeCursor pointers in a vector because their
            // current_tag members must remain valid while we're merging their
            // tags, but we need to call next() on them all afterwards.
            vector<MergeCursor *> vec;
            vec.reserve(pq.size());

            while (true) {
                cur->read_tag();
                pqtag.push(new PrefixCompressedStringItor(cur->current_tag));
                vec.push_back(cur);
                if (pq.empty() || pq.top()->current_key != key) break;
                cur = pq.top();
                pq.pop();
            }

            PrefixCompressedStringWriter wr(tag);
            string lastword;
            while (!pqtag.empty()) {
                PrefixCompressedStringItor * it = pqtag.top();
                pqtag.pop();
                string word = **it;
                if (word != lastword) {
                    lastword = word;
                    wr.append(lastword);
                }
                ++*it;
                if (!it->at_end()) {
                    pqtag.push(it);
                } else {
                    delete it;
                }
            }

            vector<MergeCursor *>::const_iterator i;
            for (i = vec.begin(); i != vec.end(); ++i) {
                cur = *i;
                if (cur->next()) {
                    pq.push(cur);
                } else {
                    delete cur;
                }
            }
        } else {
            // We want to sum the frequencies from tags for the same key.
            Xapian::termcount tot_freq = 0;
            while (true) {
                cur->read_tag();
                Xapian::termcount freq;
                const char * p = cur->current_tag.data();
                const char * end = p + cur->current_tag.size();
                if (!unpack_uint_last(&p, end, &freq) || freq == 0) {
                    throw Xapian::DatabaseCorruptError("Bad spelling word freq");
                }
                tot_freq += freq;
                if (cur->next()) {
                    pq.push(cur);
                } else {
                    delete cur;
                }
                if (pq.empty() || pq.top()->current_key != key) break;
                cur = pq.top();
                pq.pop();
            }
            tag.resize(0);
            pack_uint_last(tag, tot_freq);
        }
        out->add(key, tag);
    }
}

static void
merge_synonyms(GlassTable * out,
               vector<GlassTable*>::const_iterator b,
               vector<GlassTable*>::const_iterator e)
{
    priority_queue<MergeCursor *, vector<MergeCursor *>, CursorGt> pq;
    for ( ; b != e; ++b) {
        GlassTable *in = *b;
        if (!in->empty()) {
            pq.push(new MergeCursor(in));
        }
    }

    while (!pq.empty()) {
        MergeCursor * cur = pq.top();
        pq.pop();

        string key = cur->current_key;
        if (pq.empty() || pq.top()->current_key > key) {
            // No need to merge the tags, just copy the (possibly compressed)
            // tag value.
            bool compressed = cur->read_tag(true);
            out->add(key, cur->current_tag, compressed);
            if (cur->next()) {
                pq.push(cur);
            } else {
                delete cur;
            }
            continue;
        }

        // Merge tag values with the same key:
        string tag;

        // We just want the union of words, so copy over the first instance
        // and skip any identical ones.
        priority_queue<ByteLengthPrefixedStringItor *,
                       vector<ByteLengthPrefixedStringItor *>,
                       ByteLengthPrefixedStringItorGt> pqtag;
        vector<MergeCursor *> vec;

        while (true) {
            cur->read_tag();
            pqtag.push(new ByteLengthPrefixedStringItor(cur->current_tag));
            vec.push_back(cur);
            if (pq.empty() || pq.top()->current_key != key) break;
            cur = pq.top();
            pq.pop();
        }

        string lastword;
        while (!pqtag.empty()) {
            ByteLengthPrefixedStringItor * it = pqtag.top();
            pqtag.pop();
            if (**it != lastword) {
                lastword = **it;
                tag += uint8_t(lastword.size() ^ MAGIC_XOR_VALUE);
                tag += lastword;
            }
            ++*it;
            if (!it->at_end()) {
                pqtag.push(it);
            } else {
                delete it;
            }
        }

        vector<MergeCursor *>::const_iterator i;
        for (i = vec.begin(); i != vec.end(); ++i) {
            cur = *i;
            if (cur->next()) {
                pq.push(cur);
            } else {
                delete cur;
            }
        }

        out->add(key, tag);
    }
}

static void
multimerge_postlists(Xapian::Compactor * compactor,
                     GlassTable * out, const char * tmpdir,
                     vector<GlassTable *> tmp,
                     vector<Xapian::docid> off)
{
    unsigned int c = 0;
    while (tmp.size() > 3) {
        vector<GlassTable *> tmpout;
        tmpout.reserve(tmp.size() / 2);
        vector<Xapian::docid> newoff;
        newoff.resize(tmp.size() / 2);
        for (unsigned int i = 0, j; i < tmp.size(); i = j) {
            j = i + 2;
            if (j == tmp.size() - 1) ++j;

            string dest = tmpdir;
            char buf[64];
            sprintf(buf, "/tmp%u_%u.", c, i / 2);
            dest += buf;

            GlassTable * tmptab = new GlassTable("postlist", dest, false);

            // Use maximum blocksize for temporary tables.  And don't compress
            // entries in temporary tables, even if the final table would do
            // so.  Any already compressed entries will get copied in
            // compressed form.
            RootInfo root_info;
            root_info.init(65536, 0);
            const int flags = Xapian::DB_DANGEROUS|Xapian::DB_NO_SYNC;
            tmptab->create_and_open(flags, root_info);

            merge_postlists(compactor, tmptab, off.begin() + i,
                            tmp.begin() + i, tmp.begin() + j);
            if (c > 0) {
                for (unsigned int k = i; k < j; ++k) {
                    unlink(tmp[k]->get_path().c_str());
                    delete tmp[k];
                    tmp[k] = NULL;
                }
            }
            tmpout.push_back(tmptab);
            tmptab->flush_db();
            tmptab->commit(1, &root_info);
            AssertRel(root_info.get_blocksize(),==,65536);
        }
        swap(tmp, tmpout);
        swap(off, newoff);
        ++c;
    }
    merge_postlists(compactor, out, off.begin(), tmp.begin(), tmp.end());
    if (c > 0) {
        for (size_t k = 0; k < tmp.size(); ++k) {
            unlink(tmp[k]->get_path().c_str());
            delete tmp[k];
            tmp[k] = NULL;
        }
    }
}

class PositionCursor : private GlassCursor {
    Xapian::docid offset;

  public:
    string key;
    Xapian::docid firstdid;

    PositionCursor(GlassTable *in, Xapian::docid offset_)
        : GlassCursor(in), offset(offset_), firstdid(0) {
        find_entry(string());
        next();
    }

    bool next() {
        if (!GlassCursor::next()) return false;
        read_tag();
        const char * d = current_key.data();
        const char * e = d + current_key.size();
        string term;
        Xapian::docid did;
        if (!unpack_string_preserving_sort(&d, e, term) ||
            !unpack_uint_preserving_sort(&d, e, &did) ||
            d != e) {
            throw Xapian::DatabaseCorruptError("Bad position key");
        }

        key.resize(0);
        pack_string_preserving_sort(key, term);
        pack_uint_preserving_sort(key, did + offset);
        return true;
    }

    const string & get_tag() const {
        return current_tag;
    }
};

class PositionCursorGt {
  public:
    bool operator()(const PositionCursor *a, const PositionCursor *b) const {
        return a->key > b->key;
    }
};

static void
merge_positions(GlassTable *out, const vector<GlassTable*> & inputs,
                const vector<Xapian::docid> & offset)
{
    priority_queue<PositionCursor *, vector<PositionCursor *>, PositionCursorGt> pq;
    for (size_t i = 0; i < inputs.size(); ++i) {
        GlassTable *in = inputs[i];
        if (in->empty()) {
            // Skip empty tables.
            continue;
        }

        pq.push(new PositionCursor(in, offset[i]));
    }

    while (!pq.empty()) {
        PositionCursor * cur = pq.top();
        pq.pop();
        out->add(cur->key, cur->get_tag());
        if (cur->next()) {
            pq.push(cur);
        } else {
            delete cur;
        }
    }
}

static void
merge_docid_keyed(GlassTable *out, const vector<GlassTable*> & inputs,
                  const vector<Xapian::docid> & offset)
{
    for (size_t i = 0; i < inputs.size(); ++i) {
        Xapian::docid off = offset[i];

        GlassTable * in = inputs[i];
        if (in->empty()) continue;

        GlassCursor cur(in);
        cur.find_entry(string());

        string key;
        while (cur.next()) {
            // Adjust the key if this isn't the first database.
            if (off) {
                Xapian::docid did;
                const char * d = cur.current_key.data();
                const char * e = d + cur.current_key.size();
                if (!unpack_uint_preserving_sort(&d, e, &did)) {
                    string msg = "Bad key in ";
                    msg += inputs[i]->get_path();
                    throw Xapian::DatabaseCorruptError(msg);
                }
                did += off;
                key.resize(0);
                pack_uint_preserving_sort(key, did);
                if (d != e) {
                    // Copy over anything extra in the key (e.g. the zero byte
                    // at the end of "used value slots" in the termlist table).
                    key.append(d, e - d);
                }
            } else {
                key = cur.current_key;
            }
            bool compressed = cur.read_tag(true);
            out->add(key, cur.current_tag, compressed);
        }
    }
}

}

using namespace GlassCompact;

void
GlassDatabase::compact(Xapian::Compactor * compactor,
                       const char * destdir,
                       int fd,
                       const vector<Xapian::Database::Internal*> & sources,
                       const vector<Xapian::docid> & offset,
                       size_t block_size,
                       Xapian::Compactor::compaction_level compaction,
                       unsigned flags,
                       Xapian::docid last_docid)
{
    struct table_list {
        // The "base name" of the table.
        char name[9];
        // The type.
        Glass::table_type type;
        // Create tables after position lazily.
        bool lazy;
    };

    static const table_list tables[] = {
        // name         type                    lazy
        { "postlist",   Glass::POSTLIST,        false },
        { "docdata",    Glass::DOCDATA,         true },
        { "termlist",   Glass::TERMLIST,        false },
        { "position",   Glass::POSITION,        true },
        { "spelling",   Glass::SPELLING,        true },
        { "synonym",    Glass::SYNONYM,         true }
    };
    const table_list * tables_end = tables +
        (sizeof(tables) / sizeof(tables[0]));

    const int FLAGS = Xapian::DB_DANGEROUS;

    bool single_file = (flags & Xapian::DBCOMPACT_SINGLE_FILE);
    bool multipass = (flags & Xapian::DBCOMPACT_MULTIPASS);
    if (single_file) {
        // FIXME: Support this combination - we need to put temporary files
        // somewhere.
        multipass = false;
    }

    for (size_t i = 0; i != sources.size(); ++i) {
        auto db = static_cast<const GlassDatabase*>(sources[i]);
        if (db->has_uncommitted_changes()) {
            const char * m =
                "Can't compact from a WritableDatabase with uncommitted "
                "changes - either call commit() first, or create a new "
                "Database object from the filename on disk";
            throw Xapian::InvalidOperationError(m);
        }
    }

    if (block_size < 2048 || block_size > 65536 ||
        (block_size & (block_size - 1)) != 0) {
        block_size = GLASS_DEFAULT_BLOCKSIZE;
    }

    FlintLock lock(destdir ? destdir : "");
    if (!single_file) {
        string explanation;
        FlintLock::reason why = lock.lock(true, false, explanation);
        if (why != FlintLock::SUCCESS) {
            lock.throw_databaselockerror(why, destdir, explanation);
        }
    }

    AutoPtr<GlassVersion> version_file_out;
    if (single_file) {
        if (destdir) {
            fd = open(destdir, O_RDWR|O_CREAT|O_TRUNC|O_BINARY|O_CLOEXEC, 0666);
            if (fd < 0) {
                throw Xapian::DatabaseCreateError("open() failed", errno);
            }
        }
        version_file_out.reset(new GlassVersion(fd));
    } else {
        fd = -1;
        version_file_out.reset(new GlassVersion(destdir));
    }

    version_file_out->create(block_size);
    for (size_t i = 0; i != sources.size(); ++i) {
        GlassDatabase * db = static_cast<GlassDatabase*>(sources[i]);
        version_file_out->merge_stats(db->version_file);
    }

    string fl_serialised;
    if (single_file) {
        GlassFreeList fl;
        fl.set_first_unused_block(1); // FIXME: Assumption?
        fl.pack(fl_serialised);
    }

    vector<GlassTable *> tabs;
    tabs.reserve(tables_end - tables);
    off_t prev_size = block_size;
    for (const table_list * t = tables; t < tables_end; ++t) {
        // The postlist table requires an N-way merge, adjusting the
        // headers of various blocks.  The spelling and synonym tables also
        // need special handling.  The other tables have keys sorted in
        // docid order, so we can merge them by simply copying all the keys
        // from each source table in turn.
        if (compactor)
            compactor->set_status(t->name, string());

        string dest;
        if (!single_file) {
            dest = destdir;
            dest += '/';
            dest += t->name;
            dest += '.';
        }

        bool output_will_exist = !t->lazy;

        // Sometimes stat can fail for benign reasons (e.g. >= 2GB file
        // on certain systems).
        bool bad_stat = false;

        // We can't currently report input sizes if there's a single file DB
        // amongst the inputs.
        bool single_file_in = false;

        off_t in_size = 0;

        vector<GlassTable*> inputs;
        inputs.reserve(sources.size());
        size_t inputs_present = 0;
        for (auto src : sources) {
            GlassDatabase * db = static_cast<GlassDatabase*>(src);
            GlassTable * table;
            switch (t->type) {
                case Glass::POSTLIST:
                    table = &(db->postlist_table);
                    break;
                case Glass::DOCDATA:
                    table = &(db->docdata_table);
                    break;
                case Glass::TERMLIST:
                    table = &(db->termlist_table);
                    break;
                case Glass::POSITION:
                    table = &(db->position_table);
                    break;
                case Glass::SPELLING:
                    table = &(db->spelling_table);
                    break;
                case Glass::SYNONYM:
                    table = &(db->synonym_table);
                    break;
                default:
                    Assert(false);
                    return;
            }

            if (db->single_file()) {
                if (t->lazy && table->empty()) {
                    // Essentially doesn't exist.
                } else {
                    // FIXME: Find actual size somehow?
                    // in_size += table->size() / 1024;
                    single_file_in = true;
                    output_will_exist = true;
                    ++inputs_present;
                }
            } else {
                off_t db_size = file_size(table->get_path());
                if (errno == 0) {
                    in_size += db_size / 1024;
                    output_will_exist = true;
                    ++inputs_present;
                } else if (errno != ENOENT) {
                    // We get ENOENT for an optional table.
                    bad_stat = true;
                    output_will_exist = true;
                    ++inputs_present;
                }
            }
            inputs.push_back(table);
        }

        // If any inputs lack a termlist table, suppress it in the output.
        if (t->type == Glass::TERMLIST && inputs_present != sources.size()) {
            if (inputs_present != 0) {
                if (compactor) {
                    string m = str(inputs_present);
                    m += " of ";
                    m += str(sources.size());
                    m += " inputs present, so suppressing output";
                    compactor->set_status(t->name, m);
                }
                continue;
            }
            output_will_exist = false;
        }

        if (!output_will_exist) {
            if (compactor)
                compactor->set_status(t->name, "doesn't exist");
            continue;
        }

        GlassTable * out;
        if (single_file) {
            out = new GlassTable(t->name, fd, version_file_out->get_offset(),
                                 false, false);
        } else {
            out = new GlassTable(t->name, dest, false, t->lazy);
        }
        tabs.push_back(out);
        RootInfo * root_info = version_file_out->root_to_set(t->type);
        if (single_file) {
            root_info->set_free_list(fl_serialised);
            out->open(FLAGS, version_file_out->get_root(t->type), version_file_out->get_revision());
        } else {
            out->create_and_open(FLAGS, *root_info);
        }

        out->set_full_compaction(compaction != compactor->STANDARD);
        if (compaction == compactor->FULLER) out->set_max_item_size(1);

        switch (t->type) {
            case Glass::POSTLIST: {
                if (multipass && inputs.size() > 3) {
                    multimerge_postlists(compactor, out, destdir,
                                         inputs, offset);
                } else {
                    merge_postlists(compactor, out, offset.begin(),
                                    inputs.begin(), inputs.end());
                }
                break;
            }
            case Glass::SPELLING:
                merge_spellings(out, inputs.begin(), inputs.end());
                break;
            case Glass::SYNONYM:
                merge_synonyms(out, inputs.begin(), inputs.end());
                break;
            case Glass::POSITION:
                merge_positions(out, inputs, offset);
                break;
            default:
                // DocData, Termlist
                merge_docid_keyed(out, inputs, offset);
                break;
        }

        if (out->is_modified()) {
            // Commit as revision 1.
            out->flush_db();
            out->commit(1, root_info);
            out->sync();
        }
        if (single_file) fl_serialised = root_info->get_free_list();

        off_t out_size = 0;
        if (!bad_stat && !single_file_in) {
            off_t db_size;
            if (single_file) {
                db_size = file_size(fd);
            } else {
                db_size = file_size(dest + GLASS_TABLE_EXTENSION);
            }
            if (errno == 0) {
                if (single_file) {
                    off_t old_prev_size = max(prev_size, off_t(block_size));
                    prev_size = db_size;
                    db_size = max(db_size, off_t(block_size));
                    db_size -= old_prev_size;
                }
                out_size = db_size / 1024;
            } else {
                bad_stat = (errno != ENOENT);
            }
        }
        if (bad_stat) {
            if (compactor)
                compactor->set_status(t->name, "Done (couldn't stat all the DB files)");
        } else if (single_file_in) {
            if (compactor)
                compactor->set_status(t->name, "Done (table sizes unknown for single file DB input)");
        } else {
            string status;
            if (out_size == in_size) {
                status = "Size unchanged (";
            } else {
                off_t delta;
                if (out_size < in_size) {
                    delta = in_size - out_size;
                    status = "Reduced by ";
                } else {
                    delta = out_size - in_size;
                    status = "INCREASED by ";
                }
                if (in_size) {
                    status += str(100 * delta / in_size);
                    status += "% ";
                }
                status += str(delta);
                status += "K (";
                status += str(in_size);
                status += "K -> ";
            }
            status += str(out_size);
            status += "K)";
            if (compactor)
                compactor->set_status(t->name, status);
        }
    }

    // If compacting to a single file output and all the tables are empty, pad
    // the output so that it isn't mistaken for a stub database when we try to
    // open it.  For this it needs to be a multiple of 2KB in size.
    if (single_file && prev_size < off_t(block_size)) {
#ifdef HAVE_FTRUNCATE
        if (ftruncate(fd, block_size) < 0) {
            throw Xapian::DatabaseError("Failed to set size of output database", errno);
        }
#else
        const off_t off = block_size - 1;
        if (lseek(fd, off, SEEK_SET) != off || write(fd, "", 1) != 1) {
            throw Xapian::DatabaseError("Failed to set size of output database", errno);
        }
#endif
    }

    if (single_file) {
        if (lseek(fd, version_file_out->get_offset(), SEEK_SET) < 0) {
            throw Xapian::DatabaseError("lseek() failed", errno);
        }
    }
    version_file_out->set_last_docid(last_docid);
    string tmpfile = version_file_out->write(1, FLAGS);
    for (unsigned j = 0; j != tabs.size(); ++j) {
        tabs[j]->sync();
    }
    // Commit with revision 1.
    version_file_out->sync(tmpfile, 1, FLAGS);
    for (unsigned j = 0; j != tabs.size(); ++j) {
        delete tabs[j];
    }

    if (!single_file) lock.release();
}