// virtual file system - transparent access to files in archives; // allows multiple mount points // // Copyright (c) 2004 Jan Wassenberg // // 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. // // Contact info: // Jan.Wassenberg@stud.uni-karlsruhe.de // http://www.stud.uni-karlsruhe.de/~urkt/ #include "precompiled.h" #include "lib.h" #include "res.h" #include "zip.h" #include "file.h" #include "adts.h" #include "hotload.h" // see NO_DIR_WATCH #include #include #include #include #include #include // currently not thread safe. will have to change that if // a prefetch thread is to be used. // not safe to call before main! // we add/cancel directory watches from the VFS mount code for convenience - // it iterates through all subdirectories anyway (*) and provides storage for // a key to identify the watch (obviates separate TDir -> watch mapping). // // define this to strip out that code - removes .watch from struct TDir, // and calls to res_watch_dir / res_cancel_watch. // // *: the add_watch code would need to iterate through subdirs and watch // each one, because the monitor API (e.g. FAM) may only be able to // watch single directories, instead of a whole subdirectory tree. //#define NO_DIR_WATCH // pathnames are case-insensitive. // implementation: // when mounting, we get the exact filenames as reported by the OS; // we allow open requests with mixed case to match those, // but still use the correct case when passing to other libraries // (e.g. the actual open() syscall, called via file_open). // rationale: // necessary, because some exporters output .EXT uppercase extensions // and it's unreasonable to expect that users will always get it right. // rationale for no forcibly-close support: // issue: // we might want to edit files while the game has them open. // usual case: edit file, notify engine that it should be reloaded. // here: need to tell the engine to stop what it's doing and close the file; // only then can the artist write to the file, and trigger a reload. // // work involved: // since closing a file with pending aios results in undefined // behavior on Win32, we would have to keep track of all aios from each file, // and cancel them. we'd also need to notify the higher level resource user // that its read was cancelled, as opposed to failing due to read errors // (which might cause the game to terminate). // // this is just more work than benefit. cases where the game holds on to files // are rare: // - streaming music (artist can use regular commands to stop the current // track, or all music) // - if the engine happens to be reading that file at the moment (expected // to happen only during loading, and these are usually one-shot anway, // i.e. it'll be done soon) // - bug (someone didn't close a file - tough luck, and should be fixed // instead of hacking around it). // - archives (these remain open. allowing reload would mean we'd have to keep // track of all files from an archive, and reload them all. another hassle. // anyway, if files are to be changed in-game, then change the plain-file // version - that's what they're for). ////////////////////////////////////////////////////////////////////////////// // // path // ////////////////////////////////////////////////////////////////////////////// // path types: // p_*: posix (e.g. mount object name or for open()) // v_*: vfs (e.g. mount point) // fn : filename only (e.g. from readdir) // dir_name: directory only, no path (e.g. subdir name) // // all paths must be relative (no leading '/'); components are separated // by '/'; no ':', '\\', "." or ".." allowed; root dir is "". // // grammar: // path ::= dir*file? // dir ::= name/ // file ::= name // name ::= [^/] // if path is invalid (see source for criteria), print a diagnostic message // (indicating line number of the call that failed) and // return a negative error code. used by CHECK_PATH. static int path_validate(const uint line, const char* const path) { size_t path_len = 0; // counted as we go; checked against max. const char* msg = 0; // error occurred <==> != 0 int err = -1; // what we pass to caller int c = 0, last_c; // used for ./ detection // disallow "/", because it would create a second 'root' (with name = ""). // root dir is "". if(path[0] == '/') { msg = "starts with '/'"; goto fail; } // scan each char in path string; count length. for(;;) { last_c = c; c = path[path_len++]; // whole path is too long if(path_len >= VFS_MAX_PATH) { msg = "path too long"; goto fail; } // disallow: // - ".." (prevent going above the VFS root dir) // - "./" (security hole when mounting and not supported on Windows). // allow "/.", because CVS backup files include it. if(last_c == '.' && (c == '.' || c == '/')) { msg = "contains '..' or './'"; goto fail; } // disallow OS-specific dir separators if(c == '\\' || c == ':') { msg = "contains OS-specific dir separator (e.g. '\\', ':')"; goto fail; } // end of string, all is well. if(c == '\0') goto ok; } // failed somewhere - err is the error code, // or -1 if not set specifically above. fail: debug_out("path_validate at line %d failed: %s (error code %d)\n", line, msg, err); debug_warn("path_validate failed"); return err; ok: return 0; } #define CHECK_PATH(path) CHECK_ERR(path_validate(__LINE__, path)) // combine and into one path, and write to . // if necessary, a directory separator is added between the paths. // each may be empty, filenames, or full paths. // total path length (including '\0') must not exceed VFS_MAX_PATH. static int path_append(char* dst, const char* path1, const char* path2) { const size_t len1 = strlen(path1); const size_t len2 = strlen(path2); size_t total_len = len1 + len2 + 1; // includes '\0' // check if we need to add '/' between path1 and path2 bool need_separator = false; const bool first_no_slash = (len1 == 0) || (path1[len1-1] != '/'); // note: the second can't start with '/' (not allowed by path_validate) if(first_no_slash) { total_len++; // for '/' need_separator = true; } if(total_len+1 > VFS_MAX_PATH) return ERR_PATH_LENGTH; strcpy(dst, path1); dst += len1; if(need_separator) *dst++ = '/'; strcpy(dst, path2); return 0; } // strip from the start of , prepend , // and write to . // used when converting VFS <--> real paths. static int path_replace(char* dst, const char* src, const char* remove, const char* replace) { // remove doesn't match start of const size_t remove_len = strlen(remove); if(strncmp(src, remove, remove_len) != 0) return -1; // get rid of trailing / in src (must not be included in remove) const char* start = src+remove_len; if(*start == '/' || *start == DIR_SEP) start++; // prepend replace. CHECK_ERR(path_append(dst, replace, start)); return 0; } /////////////////////////////////////////////////////////////////////////////// // // file system tree storing information about each file // /////////////////////////////////////////////////////////////////////////////// // TLoc = location of a file in the tree. // TFile = all information about a file stored in the tree. // TDir = container holding TFile-s representing a dir. in the tree. // // the VFS stores the location (archive or directory) of each file; // this allows multiple search paths without having to check each one // when opening a file (slow). // // one TLoc is allocated for each archive or directory mounted. // therefore, files only /point/ to a (possibly shared) TLoc. // if a file's location changes (e.g. after mounting a higher-priority // directory), the VFS entry will point to the new TLoc; the priority // of both locations is unchanged. // // allocate via mnt_create, passing the location. do not free! // we keep track of all Locs allocated; they are freed at exit, // and by mnt_free_all (useful when rebuilding the VFS). // this is much easier and safer than walking the VFS tree and // freeing every location we find. // case-independent string for comparing file/directory names. // modified from GotW #29. struct ci_char_traits : public std::char_traits // inherit all the other functions that we don't need to override. { static bool eq(char c1, char c2) { return tolower(c1) == tolower(c2); } static bool ne(char c1, char c2) { return tolower(c1) != tolower(c2); } static bool lt( char c1, char c2 ) { return tolower(c1) < tolower(c2); } static int compare(const char* s1, const char* s2, size_t n) { return stricmp(s1, s2); } static const char* find(const char* s, int n, char a) { while(n-- > 0 && tolower(*s) != tolower(a)) ++s; return s; } }; typedef std::basic_string ci_string; // location of a file: either archive or a real directory. // not many instances => don't worry about struct size / alignment. struct TLoc { Handle archive; // not freed in dtor, so that users don't have to avoid // TLoc temporary objects (that would free the archive) const std::string v_mount_point; const std::string p_real_path; uint pri; TLoc(Handle _archive, const char* _v_mount_point, const char* _p_real_path, uint _pri) : v_mount_point(_v_mount_point), p_real_path(_p_real_path) { archive = _archive; pri = _pri; } }; // container must not invalidate iterators after insertion! // (we keep and pass around pointers to Mount.archive_locs elements) // see below. typedef std::list TLocs; typedef TLocs::iterator TLocIt; // rationale for separate file / subdir containers: // problems: // - duplicates find/add routines // - makes ordered output of all dirents difficult; // irrelevant, since dirs and files are enumerated separately. // advantages: // - simplifies code: no need to always check if a node is a dir; // no messing around with union TNode // - negligibly reduces memory use (sizeof TDir > sizeof TFile) // // add_file aborts if a subdir of the same name already exists, // and vice versa. struct TFile { // required: const TLoc* loc; // allocated and owned by caller (mount code) // set by ctor if struct stat is available: time_t mtime; off_t size; // these can also be extracted from TLoc, // but better cache coherency when accessing them here. u16 pri; u32 in_archive : 1; // c_str() pointer to the TFiles container's key for this node. // set by TDir::find_file; used by callers needing the exact case, // e.g. for case-sensitive syscalls; see rationale above. const char* exact_fn; TFile(const TLoc* _loc = 0, const struct stat* s = 0) { loc = _loc; if(s) mtime = s->st_mtime, size = s->st_size; else mtime = 0, size = 0; pri = _loc->pri; in_archive = _loc->archive > 0; exact_fn = 0; // safety } }; typedef std::pair TFilePair; typedef std:: map TFiles; typedef TFiles::iterator TFileIt; struct TDir; typedef std::pair TDirPair; typedef std:: map TDirs; typedef TDirs::iterator TDirIt; struct TDir { #ifndef NO_DIR_WATCH intptr_t watch; #endif TDirs subdirs; TFiles files; // if exactly one real directory is mounted into this virtual dir, // this points to its location. used to add files to VFS when writing. // // the TLoc is actually in the mount info and is invalid when // that's unmounted, but the VFS would then be rebuilt anyway. // // = 0 if no real dir mounted here; = -1 if more than one. const TLoc* loc; // c_str() pointer to the TDirs container's key for this node. // set by TDir::find_dir; used by callers needing the exact case, // e.g. for case-sensitive syscalls; see rationale above. const char* exact_dir_name; TDir() { #ifndef NO_DIR_WATCH watch = -1; #endif loc = 0; exact_dir_name = 0; // safety } int add_file(const char* fn, const TLoc* loc, const struct stat* s); TFile* find_file(const char* fn); int add_subdir(const char* d_name); TDir* find_subdir(const char* d_name); void clearR(); void displayR(int indent_level = 0); int TDir::addR(const char* p_path, const TLoc* dir_loc, TLocs* archive_locs = 0); }; int TDir::add_subdir(const char* const dir_name) { if(find_file(dir_name)) { debug_warn("TDir::add_subdir: file of same name already exists"); return -1; } TDirPair pair = std::make_pair(dir_name, TDir()); std::pair ret = subdirs.insert(pair); return ret.second? 0 : 1; } TDir* TDir::find_subdir(const char* dir_name) { TDirIt it = subdirs.find(dir_name); if(it == subdirs.end()) return 0; TDir* dir= &it->second; dir->exact_dir_name = it->first.c_str(); // see decl return dir; } int TDir::add_file(const char* fn, const TLoc* loc, const struct stat* s) { if(find_subdir(fn)) { debug_warn("TDir::add_file: subdir of same name already exists"); return -1; } TFilePair pair = std::make_pair(fn, TFile(loc, s)); std::pair ret = files.insert(pair); return ret.second? 0 : 1; } TFile* TDir::find_file(const char* fn) { TFileIt it = files.find(fn); if(it == files.end()) return 0; TFile* file = &it->second; file->exact_fn = it->first.c_str(); // see decl return file; } // empty this directory and all subdirectories; used when rebuilding VFS. void TDir::clearR() { // recurse for all subdirs // (preorder traversal - need to do this before clearing the list) TDirIt it; for(it = subdirs.begin(); it != subdirs.end(); ++it) { TDir& subdir = it->second; subdir.clearR(); } // wipe out this directory subdirs.clear(); files.clear(); #ifndef NO_DIR_WATCH res_cancel_watch(watch); watch = 0; #endif } void TDir::displayR(int indent_level) { const char indent[] = " "; // list all files in this dir for(TFileIt file_it = files.begin(); file_it != files.end(); ++file_it) { const char* name = file_it->first.c_str(); TFile* file = &file_it->second; char is_archive = file->in_archive? 'A' : 'L'; char* timestamp = ctime(&file->mtime); timestamp[24] = '\0'; // remove '\n' const off_t size = file->size; for(int i = 0; i < indent_level; i++) printf(indent); char fmt[25]; int chars = 80 - indent_level*(sizeof(indent)-1); sprintf(fmt, "%%-%d.%ds (%%c; %%6d; %%s)\n", chars, chars); // build format string: tell it how long the filename may be, // so that it takes up all space before file info column. printf(fmt, name, is_archive, size, timestamp); } // recurse over all subdirs for(TDirIt dir_it = subdirs.begin(); dir_it != subdirs.end(); ++dir_it) { TDir* const subdir = &dir_it->second; const char* subdir_name = dir_it->first.c_str(); // write subdir's name // note: do it now, instead of in recursive call so that: // - we don't have to pass dir_name parameter; // - the VFS root node isn't displayed. for(int i = 0; i < indent_level; i++) printf(indent); printf("[%s/]\n", subdir_name); subdir->displayR(indent_level+1); } } /////////////////////////////////////////////////////////////////////////////// static TDir vfs_root; static inline void tree_clear() { vfs_root.clearR(); } // write a representation of the VFS tree to stdout. static inline void tree_display() { vfs_root.displayR(); } enum TreeLookupFlags { LF_CREATE_MISSING = 1 }; // starting at VFS root, traverse and pass back information // for its last directory component. // // if & LF_CREATE_MISSING, all missing subdirectory components are // added to the VFS. // if != 0, it receives a copy of with // the exact case of each component as returned by the OS // (useful for calling external case-sensitive code). // // can be to a file or dir (in which case it must end in '/', // to make sure the last component is treated as a directory). // // return 0 on success, or a negative error code // (in which case output params are undefined). static int tree_lookup_dir(const char* path, TDir** pdir, uint flags = 0, char* exact_path = 0) { CHECK_PATH(path); assert(!(flags & ~LF_CREATE_MISSING)); // no undefined bits set // can't check if path ends in '/' here - we're called via tree_lookup. const bool create_missing = !!(flags & LF_CREATE_MISSING); // copy into (writeable) buffer so we can 'tokenize' path components // by replacing '/' with '\0'. length check done by CHECK_PATH. char v_path[VFS_MAX_PATH]; strcpy(v_path, path); char* cur_component = v_path; TDir* cur_dir = &vfs_root; // successively navigate to the next subdirectory in . for(;;) { // "extract" cur_component string (0-terminate by replacing '/') char* slash = (char*)strchr(cur_component, '/'); // .. it's a filename (since it doesn't end in '/') => done. if(!slash) break; *slash = '\0'; // create subdir (no-op if it already exists) if(create_missing) CHECK_ERR(cur_dir->add_subdir(cur_component)); cur_dir = cur_dir->find_subdir(cur_component); if(!cur_dir) return ERR_PATH_NOT_FOUND; if(exact_path) exact_path += sprintf(exact_path, "%s/", cur_dir->exact_dir_name); // no length check needed: length is the same as path cur_component = slash+1; } // success. *pdir = cur_dir; return 0; } // pass back file information for (relative to VFS root). // // if & LF_CREATE_MISSING, the file is added to VFS unless // a higher-priority file of the same name already exists // (used by VFile_reload when opening for writing). // if != 0, it receives a copy of with // the exact case of each component as returned by the OS // (useful for calling external case-sensitive code). // // return 0 on success, or a negative error code // (in which case output params are undefined). static int tree_lookup(const char* path, TFile** pfile, uint flags = 0, char* exact_path = 0) { // path and flags checked by tree_lookup_dir TDir* dir; CHECK_ERR(tree_lookup_dir(path, &dir, flags, exact_path)); const bool create_missing = !!(flags & LF_CREATE_MISSING); // strip away path to file const char* fn = path; const char* slash = strrchr(path, '/'); if(slash) fn = slash+1; // empty filename ("" or "dir/" etc.) if(fn[0] == '\0') return -1; if(create_missing) { // not exactly one real dir mounted here => // we don't have a location where to add the file. if(!dir->loc || dir->loc == (TLoc*)-1) return -1; CHECK_ERR(dir->add_file(fn, dir->loc, (struct stat*)0)); // we don't know stat (e.g. size, mtime) yet - // it's updated when file is closed. } TFile* file = dir->find_file(fn); // the file (still) doesn't exist if(!file) return ERR_FILE_NOT_FOUND; if(exact_path) strcat(exact_path, file->exact_fn); *pfile = file; return 0; } ////////////////////////////////////////////////////////////////////////////// // passed to add_dirent_cb struct FileCBParams { // tree dir into which the dirent is to be added TDir* const dir; // .. specifying this location const TLoc* const cur_loc; // if the dirent is an archive, its TLoc is added here. TLocs* const archive_locs; FileCBParams(TDir* _dir, const TLoc* _cur_loc, TLocs* _archive_locs = 0) : dir(_dir), cur_loc(_cur_loc), archive_locs(_archive_locs) {} // no copy ctor, since members are const private: FileCBParams& operator=(const FileCBParams&); }; // we're adding , but (same name) already exists. // decide which is more important; return 1 iff is to be replaced. // // note: if "priority" is the same, replace! // this makes sure mods/patches etc. actually replace files. // // called by add_dirent_cb. static int should_replace(TFile* old_file, TFile* new_file) { // older is higher priority - keep. if(old_file->pri > new_file->pri) return 0; // assume they're the same if size and last-modified time match. bool is_same = (old_file->size == new_file->size); is_same &= fabs(difftime(old_file->mtime, new_file->mtime)) <= 2.0; // (FAT timestamp has 2 second resolution) // strategy: // s = "is same", oa = "old is in archive", na = "new is in archive" // s oa na result // ---------------- // 0 0 0 replace // 0 0 0 replace // 0 1 0 replace // 0 1 1 replace // 1 0 0 replace // 1 0 1 replace // 1 1 0 keep // 1 1 1 replace // // in english: always replace, unless both are the same, // old is archived (fast), and new is loose (slow). if(is_same && old_file->in_archive && !new_file->in_archive) return 0; return 1; } // either called by: // - add_dirent_cb's zip_enum for each file in archive, or // - TDir::addR's file_enum for each entry in a real directory // when mounting. // // if called for a real directory, it is added to VFS. // else if called for a loose file that is a valid archive (*), // it is mounted (all of its files are added) // else the file is added to VFS. // // * we only perform this check in the directory being mounted, // i.e. passed in by tree_add_dir. to determine if a file is an archive, // we have to open it and read the header, which is slow. // can't just check extension, because it might not be .zip (e.g. Quake3 .pk3). static int add_dirent_cb(const char* const n_name, const struct stat* s, const uintptr_t user) { const FileCBParams* const params = (FileCBParams*)user; TDir* cur_dir = params->dir; const TLoc* cur_loc = params->cur_loc; TLocs* archive_locs = params->archive_locs; // = 0 <==> this is the directory being added by tree_add_dir // loose (i.e. not in archive) directory: add it. // note: when called by zip_enum, we only get files in the archive. if(S_ISDIR(s->st_mode)) return cur_dir->add_subdir(n_name); const char* fn = n_name; // we get full path for files in archive; will strip below. // loose file if(cur_loc->archive <= 0) { // at first level of nesting (i.e. dir being mounted): test if file // is an archive; if so, mount it (we only do this at the first // nesting level to avoid testing every single file - slow) if(archive_locs) { char path[PATH_MAX]; path_append(path, cur_loc->p_real_path.c_str(), n_name); // don't bother checking extension - archives won't // necessarily be called .zip (e.g. Quake III .pk3). Handle archive = zip_archive_open(path); if(archive > 0) { archive_locs->push_back(TLoc(archive, "", "", cur_loc->pri)); FileCBParams params(cur_dir, &archive_locs->back()); zip_enum(archive, add_dirent_cb, (uintptr_t)¶ms); return 0; // don't add as a file } } } // archived file else { // look up directory in archive from full pathname. CHECK_ERR(tree_lookup_dir(n_name, &cur_dir, LF_CREATE_MISSING)); // we have to create them if missing, since we can't rely on the // archiver placing directories before subdirs or files that // reference them (WinZip doesn't). // get filename only (no path) for TDir::add_file. const char* slash = strrchr(n_name, '/'); if(slash) fn = slash+1; } int ret = cur_dir->add_file(fn, cur_loc, s); // wasn't in dir yet - it's added and we're done. if(ret == 0) return 0; // replace old file with newer one? TFile new_file(cur_loc, s); TFile* old_file = cur_dir->find_file(fn); if(should_replace(old_file, &new_file) == 1) *old_file = new_file; return 0; } int TDir::addR(const char* p_path, const TLoc* dir_loc, TLocs* archive_locs) { // more than one real dir mounted into VFS dir // (=> can't create files for writing here) if(loc) loc = (TLoc*)-1; else loc = dir_loc; #ifndef NO_DIR_WATCH res_watch_dir(p_path, &watch); #endif // add files and subdirs to vdir const FileCBParams params(this, dir_loc, archive_locs); file_enum(p_path, add_dirent_cb, (uintptr_t)¶ms); // recurse over all subdirs for(TDirIt it = subdirs.begin(); it != subdirs.end(); ++it) { TDir* subdir = &it->second; const char* d_subdir_name = (it->first).c_str(); // don't clutter the tree with versioning system dirs. // only applicable for normal dirs; the archive builder // takes care of removing these there. if(!strcmp(d_subdir_name, "CVS") || !strcmp(d_subdir_name, ".svn")) continue; char p_subdir_path[PATH_MAX]; CHECK_ERR(path_append(p_subdir_path, p_path, d_subdir_name)); subdir->addR(p_subdir_path, dir_loc, 0); } return 0; } static int tree_add_dir(TDir* dir, const char* p_path, const TLoc* dir_loc, TLocs* archive_locs) { return dir->addR(p_path, dir_loc, archive_locs); } /////////////////////////////////////////////////////////////////////////////// // // mount directories into the VFS // /////////////////////////////////////////////////////////////////////////////// struct Mount { // note: we basically duplicate the mount information in dir_loc. // it's no big deal - there won't be many mountings. // // reason is, we need this info in Mount when remounting, // but also in TLoc when getting real file path. // accessing everything via TDir's TLoc is ugly. // mounting into this VFS directory; // must end in '/' (unless if root dir, i.e. "") const std::string v_mount_point; // real directory being mounted const std::string p_real_path; uint pri; // storage for all Locs ensuing from this mounting. // it's safe to store pointers to them: the Mount and Locs containers // are std::lists, and all pointers are reset after unmounting something. TLoc dir_loc; // referenced by TDir::dir_loc (used when creating files for writing) TLocs archive_locs; // contains one TLoc for every archive in this directory that // was mounted - in alphabetical order! // // multiple archives per dir support is required for patches. Mount(const char* _v_mount_point, const char* _p_real_path, uint _pri) : v_mount_point(_v_mount_point), p_real_path(_p_real_path), dir_loc(0, _v_mount_point, _p_real_path, _pri), archive_locs() { pri = _pri; } }; typedef std::list Mounts; typedef Mounts::iterator MountIt; static Mounts mounts; // actually mount the specified entry. split out of vfs_mount, // because when invalidating (reloading) the VFS, we need to // be able to mount without changing the mount list. static int remount(Mount& m) { const char* v_mount_point = m.v_mount_point.c_str(); const char* p_real_path = m.p_real_path.c_str(); const uint pri = m.pri; TLoc* dir_loc = &m.dir_loc; TLocs& archive_locs = m.archive_locs; // callers have a tendency to forget required trailing '/'; // complain if it's not there, unless path = "" (root dir). #ifndef NDEBUG size_t len = strlen(v_mount_point); if(len && v_mount_point[len-1] != '/') debug_warn("remount: path doesn't end in '/'"); #endif TDir* dir; CHECK_ERR(tree_lookup_dir(v_mount_point, &dir, LF_CREATE_MISSING)); // add all loose files and subdirectories (recursive). // also mounts all archives in p_real_path and adds to archive_locs. return tree_add_dir(dir, p_real_path, dir_loc, &archive_locs); } // don't do this in dtor, to allow use of temporary Mount objects. static int unmount(Mount& m) { for(TLocIt it = m.archive_locs.begin(); it != m.archive_locs.end(); ++it) zip_archive_close(it->archive); m.archive_locs.clear(); return 0; } // trivial, but used by vfs_shutdown and vfs_rebuild static inline void unmount_all(void) { std::for_each(mounts.begin(), mounts.end(), unmount); } static inline void remount_all() { std::for_each(mounts.begin(), mounts.end(), remount); } // is s2 a subpath of s1, or vice versa? // used by vfs_mount. static bool is_subpath(const char* s1, const char* s2) { // make sure s1 is the shorter string if(strlen(s1) > strlen(s2)) std::swap(s1, s2); int c1 = 0, last_c1, c2; for(;;) { last_c1 = c1; c1 = *s1++, c2 = *s2++; // end of s1 reached: if(c1 == '\0') { // s1 matched s2 up until: if(c2 == '\0' || // its end (i.e. they're equal length) c2 == '/' || // start of next component last_c1 == '/') // ", but both have a trailing slash // => is subpath return true; } // mismatch => is not subpath if(c1 != c2) return false; } } // mount into the VFS at , // which is created if it does not yet exist. // files in that directory override the previous VFS contents if // (ority) is not lower. // all archives in are also mounted, in alphabetical order. // // p_real_path = "." or "./" isn't allowed - see implementation for rationale. int vfs_mount(const char* v_mount_point, const char* p_real_path, const uint pri) { // make sure it's not already mounted, i.e. in mounts. // also prevents mounting a parent directory of a previously mounted // directory, or vice versa. example: mount $install/data and then // $install/data/mods/official - mods/official would also be accessible // from the first mount point - bad. // no matter if it's an archive - still shouldn't be a "subpath". for(MountIt it = mounts.begin(); it != mounts.end(); ++it) if(is_subpath(p_real_path, it->p_real_path.c_str())) { debug_warn("vfs_mount: already mounted"); return -1; } // disallow "." because "./" isn't supported on Windows. // it would also create a loophole for the parent dir check above. // "./" and "/." are caught by CHECK_PATH. if(!strcmp(p_real_path, ".")) { debug_warn("vfs_mount: mounting . not allowed"); return -1; } mounts.push_back(Mount(v_mount_point, p_real_path, pri)); // actually mount the entry Mount& m = mounts.back(); return remount(m); } // rebuild the VFS, i.e. re-mount everything. open files are not affected. // necessary after loose files or directories change, so that the VFS // "notices" the changes and updates file locations. res calls this after // dir_watch reports changes; can also be called from the console after a // rebuild command. there is no provision for updating single VFS dirs - // it's not worth the trouble. int vfs_rebuild() { tree_clear(); unmount_all(); remount_all(); return 0; } // unmount a previously mounted item, and rebuild the VFS afterwards. int vfs_unmount(const char* p_real_path) { for(MountIt it = mounts.begin(); it != mounts.end(); ++it) // found the corresponding entry if(it->p_real_path == p_real_path) { Mount& m = *it; unmount(m); mounts.erase(it); return vfs_rebuild(); } return ERR_PATH_NOT_FOUND; } // if or its ancestors are mounted, // return a VFS path that accesses it. // used when receiving paths from external code. int vfs_make_vfs_path(const char* const path, char* const vfs_path) { for(MountIt it = mounts.begin(); it != mounts.end(); ++it) { const char* remove = it->p_real_path.c_str(); const char* replace = it->v_mount_point.c_str(); if(path_replace(vfs_path, path, remove, replace) == 0) return 0; } return -1; } // given and the file's location, // return the actual filename. // used by vfs_realpath and VFile_reopen. static int make_file_path(char* path, const char* vfs_path, const TLoc* loc) { assert(loc->archive == 0); const char* remove = loc->v_mount_point.c_str(); const char* replace = loc->p_real_path.c_str(); return path_replace(path, vfs_path, remove, replace); } /////////////////////////////////////////////////////////////////////////////// // // directory // /////////////////////////////////////////////////////////////////////////////// struct VDir { // we need to cache the complete contents of the directory: // if we reference the real directory and it changes, // the c_str pointers may become invalid, and some files // may be returned out of order / not at all. // we copy the directory's subdirectory and file containers. // // note: allocate dynamically to make sure they fit into control block. TDirs* subdirs; TDirIt subdir_it; TFiles* files; TFileIt file_it; }; H_TYPE_DEFINE(VDir); static void VDir_init(VDir* vd, va_list args) { UNUSED(vd); UNUSED(args); } static void VDir_dtor(VDir* vd) { delete vd->subdirs; delete vd->files; } static int VDir_reload(VDir* vd, const char* path, Handle) { if(vd->subdirs) { debug_warn("VDir_reload called when already loaded - why?"); return 0; } // add required trailing slash if not already present, // to make caller's life easier. char path_slash[PATH_MAX]; CHECK_ERR(path_append(path_slash, path, "")); TDir* dir; CHECK_ERR(tree_lookup_dir(path_slash, &dir)); // rationale for copying the dir's data: see VDir definition // note: bad_alloc exception handled by h_alloc. vd->subdirs = new TDirs(dir->subdirs); vd->subdir_it = vd->subdirs->begin(); vd->files = new TFiles(dir->files); vd->file_it = vd->files->begin(); return 0; } // open a directory for reading its entries via vfs_next_dirent. // need not end in '/'; we add it if not present. // directory contents are cached here; subsequent changes to the dir // are not returned by this handle. rationale: see VDir definition. Handle vfs_open_dir(const char* const v_dir) { return h_alloc(H_VDir, v_dir, RES_NO_CACHE); // must not cache, since the position in file array // is advanced => not copy-equivalent. } // close the handle to a directory. // all vfsDirEnt.name strings are now invalid. int vfs_close_dir(Handle& hd) { return h_free(hd, H_VDir); } // retrieve the next dir entry (in alphabetical order) matching . // return 0 on success, ERR_DIR_END if no matching entry was found, // or a negative error code on failure. // filter values: // - 0: any file; // - "/": any subdirectory // - anything else: pattern for name (may include '?' and '*' wildcards) // // rationale: the filename is currently stored internally as // std::string (=> less manual memory allocation). we don't want to // return a reference, because that would break C compatibility. // we're trying to avoid fixed-size buffers, so that is out as well. // finally, allocating a copy is not so good because it has to be // freed by the user (won't happen). returning a volatile pointer // to the string itself via c_str is the only remaining option. int vfs_next_dirent(const Handle hd, vfsDirEnt* ent, const char* const filter) { H_DEREF(hd, VDir, vd); // caller wants a subdirectory; return the next one. if(filter && filter[0] == '/' && filter[1] == '\0') { if(vd->subdir_it == vd->subdirs->end()) return ERR_DIR_END; ent->name = vd->subdir_it->first.c_str(); ++vd->subdir_it; return 0; } // caller wants a file; loop until one matches or end of list. for(;;) { if(vd->file_it == vd->files->end()) return ERR_DIR_END; ent->name = vd->file_it->first.c_str(); ++vd->file_it; if(!filter || match_wildcard(ent->name, filter)) return 0; } } // return actual path to the specified file: // "/fn" or "/fn". int vfs_realpath(const char* v_path, char* realpath) { TFile* file; char v_exact_path[VFS_MAX_PATH]; CHECK_ERR(tree_lookup(v_path, &file, 0, v_exact_path)); if(file->in_archive) { const char* archive_fn = h_filename(file->loc->archive); if(!archive_fn) return -1; CHECK_ERR(path_append(realpath, archive_fn, v_exact_path)); } // file is in normal directory else CHECK_ERR(make_file_path(realpath, v_exact_path, file->loc)); return 0; } // does the specified file exist? return false on error. // useful because a "file not found" warning is not raised, unlike vfs_stat. bool vfs_exists(const char* v_fn) { TFile* file; return (tree_lookup(v_fn, &file) == 0); } // get file status (mode, size, mtime). output param is zeroed on error. int vfs_stat(const char* v_path, struct stat* s) { TFile* file; CHECK_ERR(tree_lookup(v_path, &file)); // all stat members currently supported are stored in TFile, // so we can return that without having to call file_stat(). s->st_mode = S_IFREG; s->st_size = file->size; s->st_mtime = file->mtime; return 0; } /////////////////////////////////////////////////////////////////////////////// // // file // /////////////////////////////////////////////////////////////////////////////// // // logging // static int file_listing_enabled; // tristate; -1 = already shut down static FILE* file_list; static void file_listing_shutdown() { if(file_listing_enabled == 1) { fclose(file_list); file_listing_enabled = -1; } } static void file_listing_add(const char* v_fn) { // we've already shut down - complain. if(file_listing_enabled == -1) { debug_warn("file_listing_add: called after file_listing_shutdown atexit"); return; } // listing disabled. if(file_listing_enabled == 0) return; if(!file_list) { file_list = fopen("../logs/filelist.txt", "w"); if(!file_list) return; } fputs(v_fn, file_list); fputc('\n', file_list); } void vfs_enable_file_listing(bool want_enabled) { // already shut down - don't allow enabling if(file_listing_enabled == -1 && want_enabled) { debug_warn("vfs_enable_file_listing: enabling after shutdown"); return; } file_listing_enabled = (int)want_enabled; } /////////////////////////////////////////////////////////////////////////////// enum { // internal file state flags // make sure these don't conflict with vfs.h flags VF_OPEN = 0x100, VF_ZIP = 0x200 }; struct VFile { // cached contents of file from vfs_load // (can't just use pointer - may be freed behind our back) Handle hm; off_t ofs; union { File f; ZFile zf; }; // be aware when adding fields that this struct is quite large, // and may require increasing the control block size limit. // (especially in PARANOIA builds, which add a member!) }; H_TYPE_DEFINE(VFile); // with #define PARANOIA, File and ZFile get an additional member, // and VFile was exceeding HDATA_USER_SIZE. flags and size (required // in File as well as VFile) are now moved into the union. // use the functions below to insulate against change a bit. static off_t& vf_size(VFile* vf) { assert(offsetof(struct File, size) == offsetof(struct ZFile, ucsize)); return vf->f.size; } static uint& vf_flags(VFile* vf) { assert(offsetof(struct File, flags) == offsetof(struct ZFile, flags)); return vf->f.flags; } static void VFile_init(VFile* vf, va_list args) { uint flags = va_arg(args, int); vf_flags(vf) = flags; } static void VFile_dtor(VFile* vf) { uint& flags = vf_flags(vf); if(flags & VF_OPEN) { if(flags & VF_ZIP) zip_close(&vf->zf); else file_close(&vf->f); flags &= ~(VF_OPEN); } mem_free_h(vf->hm); } static int VFile_reload(VFile* vf, const char* v_path, Handle) { uint& flags = vf_flags(vf); // note: no matter if flags are assigned and the function later // fails: the Handle will be closed anyway. // we're done if file is already open. need to check this because // reload order (e.g. if resource opens a file) is unspecified. if(flags & VF_OPEN) return 0; file_listing_add(v_path); TFile* file; char v_exact_path[VFS_MAX_PATH]; uint lf = (flags & FILE_WRITE)? LF_CREATE_MISSING : 0; int err = tree_lookup(v_path, &file, lf, v_exact_path); if(err < 0) { // don't CHECK_ERR - this happens often and the dialog is annoying debug_out("tree_lookup failed for %s\n", v_path); return err; } if(file->in_archive) { if(flags & FILE_WRITE) { debug_warn("requesting write access to file in archive"); return -1; } CHECK_ERR(zip_open(file->loc->archive, v_exact_path, &vf->zf)); flags |= VF_ZIP; } // normal file else { char p_path[PATH_MAX]; CHECK_ERR(make_file_path(p_path, v_exact_path, file->loc)); CHECK_ERR(file_open(p_path, flags, &vf->f)); } // success flags |= VF_OPEN; return 0; } // return the size of an already opened file, or a negative error code. ssize_t vfs_size(Handle hf) { H_DEREF(hf, VFile, vf); return vf_size(vf); } // open the file for synchronous or asynchronous IO. write access is // requested via FILE_WRITE flag, and is not possible for files in archives. Handle vfs_open(const char* v_fn, uint file_flags /* = 0 */) { // keeping files open doesn't make sense in most cases (because the // file is used to load resources, which are cached at a higher level). uint res_flags = RES_NO_CACHE; if(file_flags & FILE_CACHE) res_flags = 0; Handle h = h_alloc(H_VFile, v_fn, res_flags, file_flags); // pass file flags to init #ifdef PARANOIA debug_out("vfs_open fn=%s %llx\n", v_fn, h); #endif return h; } // close the handle to a file. int vfs_close(Handle& h) { #ifdef PARANOIA debug_out("vfs_close %llx\n", h); #endif return h_free(h, H_VFile); } // transfer the next bytes to/from the given file. // (read or write access was chosen at file-open time). // // if non-NULL, is called for each block transferred, passing . // it returns how much data was actually transferred, or a negative error // code (in which case we abort the transfer and return that value). // the callback mechanism is useful for user progress notification or // processing data while waiting for the next I/O to complete // (quasi-parallel, without the complexity of threads). // // p (value-return) indicates the buffer mode: // - *p == 0: read into buffer we allocate; set *p. // caller should mem_free it when no longer needed. // - *p != 0: read into or write into the buffer *p. // - p == 0: only read into temp buffers. useful if the callback // is responsible for processing/copying the transferred blocks. // since only temp buffers can be added to the cache, // this is the preferred read method. // // return number of bytes transferred (see above), or a negative error code. ssize_t vfs_io(const Handle hf, const size_t size, void** p, FileIOCB cb, uintptr_t ctx) { #ifdef PARANOIA debug_out("vfs_io size=%d\n", size); #endif H_DEREF(hf, VFile, vf); off_t ofs = vf->ofs; vf->ofs += (off_t)size; void* buf = 0; // assume temp buffer (p == 0) if(p) // user-specified buffer if(*p) buf = *p; // we allocate else { buf = mem_alloc(round_up(size, 4096), 4096); if(!buf) return ERR_NO_MEM; *p = buf; } // (vfs_open makes sure it's not opened for writing if zip) if(vf_flags(vf) & VF_ZIP) return zip_read(&vf->zf, ofs, size, buf, cb, ctx); // normal file: // let file_io alloc the buffer if the caller didn't (i.e. p = 0), // because it knows about alignment / padding requirements return file_io(&vf->f, ofs, size, buf, cb, ctx); } #include "timer.h" static double dt; static double totaldata; void dump() { debug_out("TOTAL TIME IN VFS_IO: %f\nthroughput: %f MB/s\n\n", dt, totaldata/dt/1e6); } static ssize_t vfs_timed_io(const Handle hf, const size_t size, void** p, FileIOCB cb = 0, uintptr_t ctx = 0) { ONCE(atexit(dump)); double t1=get_time(); totaldata += size; ssize_t nread = vfs_io(hf, size, p, cb, ctx); double t2=get_time(); if(t2-t1 < 1.0) dt += t2-t1; return nread; } // load the entire file into memory; return a handle to the memory // and the buffer address/size. output parameters are zeroed on failure. // in addition to the regular file cache, the entire buffer is kept in memory // if flags & FILE_CACHE. // // note: we need the Handle return value for Tex.hm - the data pointer // must be protected against being accidentally free-d in that case. Handle vfs_load(const char* const v_fn, void*& p, size_t& size, uint flags /* default 0 */) { #ifdef PARANOIA debug_out("vfs_load v_fn=%s\n", v_fn); #endif p = 0; size = 0; // zeroed in case vfs_open or H_DEREF fails Handle hf = vfs_open(v_fn, flags); H_DEREF(hf, VFile, vf); Handle hm = 0; // return value - handle to memory or error code size = vf_size(vf); // already read into mem - return existing mem handle // TODO: what if mapped? if(vf->hm > 0) { p = mem_get_ptr(vf->hm, &size); if(p) { assert(vf_size(vf) == (off_t)size && "vfs_load: mismatch between File and Mem size"); hm = vf->hm; goto ret; } else debug_warn("vfs_load: invalid MEM attached to vfile (0 pointer)"); // happens if someone frees the pointer. not an error! } // allocate memory. does expose implementation details of File // (padding), but it greatly simplifies returning the Handle // (if we allow File to alloc, have to make sure the Handle references // the actual data address, not that of the padding). { const size_t BLOCK_SIZE = 64*KB; p = mem_alloc(size, BLOCK_SIZE, 0, &hm); if(!p) goto ret; } { ssize_t nread = vfs_timed_io(hf, size, &p); // failed if(nread < 0) { mem_free_h(hm); hm = nread; // error code } else { if(flags & FILE_CACHE) vf->hm = hm; } } ret: vfs_close(hf); // if FILE_CACHE, it's kept open // if we fail, make sure these are set to 0 // (they may have been assigned values above) if(hm <= 0) p = 0, size = 0; if (hm == 0) debug_out("hm == 0!!\n"); return hm; } // caveat: pads file to next max(4kb, sector_size) boundary // (due to limitation of Win32 FILE_FLAG_NO_BUFFERING I/O). // if that's a problem, specify FILE_NO_AIO when opening. int vfs_store(const char* const v_fn, void* p, const size_t size, uint flags /* default 0 */) { Handle hf = vfs_open(v_fn, flags|FILE_WRITE); if(hf <= 0) return (int)hf; // error code H_DEREF(hf, VFile, vf); const int ret = vfs_io(hf, size, &p); vfs_close(hf); return ret; } /////////////////////////////////////////////////////////////////////////////// // // asynchronous I/O // /////////////////////////////////////////////////////////////////////////////// struct IO { union { FileIO fio; ZipIO zio; }; bool is_zip; // necessary if we have separate File and Zip IO structures // default is false, since control block is 0-initialized }; H_TYPE_DEFINE(IO); // don't support forcibly closing files => don't need to keep track of // all IOs pending for each file. too much work, little benefit. static void IO_init(IO*, va_list) { } static void IO_dtor(IO* io) { if(io->is_zip) zip_discard_io(&io->zio); else file_discard_io(&io->fio); } // we don't support transparent read resume after file invalidation. // if the file has changed, we'd risk returning inconsistent data. // doesn't look possible without controlling the AIO implementation: // when we cancel, we can't prevent the app from calling // aio_result, which would terminate the read. static int IO_reload(IO* io, const char*, Handle) { return 0; } // begin transferring bytes, starting at . get result // with vfs_wait_io; when no longer needed, free via vfs_discard_io. Handle vfs_start_io(Handle hf, size_t size, void* buf) { Handle hio = h_alloc(H_IO, 0); H_DEREF(hio, IO, io); H_DEREF(hf, VFile, vf); off_t ofs = vf->ofs; vf->ofs += (off_t)size; if(vf_flags(vf) & VF_ZIP) { io->is_zip = true; return zip_start_io(&vf->zf, ofs, size, buf, &io->zio); } return file_start_io(&vf->f, ofs, size, buf, &io->fio); } // indicates if the IO referenced by has completed. // return value: 0 if pending, 1 if complete, < 0 on error. int vfs_io_complete(Handle hio) { H_DEREF(hio, IO, io); if(io->is_zip) return zip_io_complete(&io->zio); else return file_io_complete(&io->fio); } // wait until the transfer completes, and return its buffer. // output parameters are zeroed on error. int vfs_wait_io(Handle hio, void*& p, size_t& size) { H_DEREF(hio, IO, io); if(io->is_zip) return zip_wait_io(&io->zio, p, size); else return file_wait_io(&io->fio, p, size); } // finished with transfer - free its buffer (returned by vfs_wait_io) int vfs_discard_io(Handle& hio) { return h_free(hio, H_IO); } /////////////////////////////////////////////////////////////////////////////// // // memory mapping // /////////////////////////////////////////////////////////////////////////////// // map the entire (uncompressed!) file into memory. if currently // already mapped, return the previous mapping (reference-counted). // output parameters are zeroed on failure. // // the mapping will be removed (if still open) when its file is closed. // however, map/unmap calls should still be paired so that the mapping // may be removed when no longer needed. int vfs_map(const Handle hf, const uint flags, void*& p, size_t& size) { UNUSED(flags); p = 0; size = 0; // need to zero these here in case H_DEREF fails H_DEREF(hf, VFile, vf); if(vf_flags(vf) & VF_ZIP) return zip_map(&vf->zf, p, size); else return file_map(&vf->f, p, size); } // decrement the reference count for the mapping belonging to file . // fail if there are no references; remove the mapping if the count reaches 0. // // the mapping will be removed (if still open) when its file is closed. // however, map/unmap calls should still be paired so that the mapping // may be removed when no longer needed. int vfs_unmap(const Handle hf) { H_DEREF(hf, VFile, vf); if(vf_flags(vf) & VF_ZIP) return zip_unmap(&vf->zf); else return file_unmap(&vf->f); } // write a representation of the VFS tree to stdout. inline void vfs_display() { tree_display(); } void vfs_shutdown() { file_listing_shutdown(); tree_clear(); unmount_all(); }