os_map.c

/*-
 * See the file LICENSE for redistribution information.
 *
 * Copyright (c) 1996-2005
 *      Sleepycat Software.  All rights reserved.
 *
 * $Id: os_map.c,v 12.2 2005/06/16 20:23:29 bostic Exp $
 */

#include "db_config.h"

#include "db_int.h"

static int __os_map
  __P((DB_ENV *, char *, REGINFO *, DB_FH *, size_t, int, int, int, void **));
static int __os_unique_name __P((_TCHAR *, HANDLE, _TCHAR *, size_t));

/*
 * __os_r_sysattach --
 *      Create/join a shared memory region.
 */
int
__os_r_sysattach(dbenv, infop, rp)
        DB_ENV *dbenv;
        REGINFO *infop;
        REGION *rp;
{
        DB_FH *fhp;
        int is_system, ret;

        /*
         * Try to open/create the file.  We DO NOT need to ensure that multiple
         * threads/processes attempting to simultaneously create the region are
         * properly ordered, our caller has already taken care of that.
         */
        if ((ret = __os_open(dbenv, infop->name,
            F_ISSET(infop, REGION_CREATE_OK) ? DB_OSO_CREATE: 0,
            dbenv->db_mode, &fhp)) != 0) {
                __db_err(dbenv, "%s: %s", infop->name, db_strerror(ret));
                return (ret);
        }

        /*
         * On Windows/9X, files that are opened by multiple processes do not
         * share data correctly.  For this reason, the DB_SYSTEM_MEM flag is
         * implied for any application that does not specify the DB_PRIVATE
         * flag.
         */
        is_system = F_ISSET(dbenv, DB_ENV_SYSTEM_MEM) ||
            (!F_ISSET(dbenv, DB_ENV_PRIVATE) && __os_is_winnt() == 0);

        /*
         * Map the file in.  If we're creating an in-system-memory region,
         * specify a segment ID (which is never used again) so that the
         * calling code writes out the REGENV_REF structure to the primary
         * environment file.
         */
        ret = __os_map(dbenv, infop->name, infop, fhp, rp->size,
           1, is_system, 0, &infop->addr);
        if (ret == 0 && is_system == 1)
                rp->segid = 1;

        (void)__os_closehandle(dbenv, fhp);

        return (ret);
}

/*
 * __os_r_sysdetach --
 *      Detach from a shared memory region.
 */
int
__os_r_sysdetach(dbenv, infop, destroy)
        DB_ENV *dbenv;
        REGINFO *infop;
        int destroy;
{
        int ret, t_ret;

        if (infop->wnt_handle != NULL) {
                (void)CloseHandle(infop->wnt_handle);
                infop->wnt_handle = NULL;
        }

        ret = !UnmapViewOfFile(infop->addr) ? __os_get_errno() : 0;
        if (ret != 0)
                __db_err(dbenv, "UnmapViewOfFile: %s", strerror(ret));

        if (!F_ISSET(dbenv, DB_ENV_SYSTEM_MEM) && destroy) {
                if (F_ISSET(dbenv, DB_ENV_OVERWRITE))
                        (void)__db_file_multi_write(dbenv, infop->name);
                if ((t_ret = __os_unlink(dbenv, infop->name)) != 0 && ret == 0)
                        ret = t_ret;
        }

        return (ret);
}

/*
 * __os_mapfile --
 *      Map in a shared memory file.
 */
int
__os_mapfile(dbenv, path, fhp, len, is_rdonly, addr)
        DB_ENV *dbenv;
        char *path;
        DB_FH *fhp;
        int is_rdonly;
        size_t len;
        void **addr;
{
        /* If the user replaced the map call, call through their interface. */
        if (DB_GLOBAL(j_map) != NULL)
                return (DB_GLOBAL(j_map)(path, len, 0, is_rdonly, addr));

        return (__os_map(dbenv, path, NULL, fhp, len, 0, 0, is_rdonly, addr));
}

/*
 * __os_unmapfile --
 *      Unmap the shared memory file.
 */
int
__os_unmapfile(dbenv, addr, len)
        DB_ENV *dbenv;
        void *addr;
        size_t len;
{
        /* If the user replaced the map call, call through their interface. */
        if (DB_GLOBAL(j_unmap) != NULL)
                return (DB_GLOBAL(j_unmap)(addr, len));

        return (!UnmapViewOfFile(addr) ? __os_get_errno() : 0);
}

/*
 * __os_unique_name --
 *      Create a unique identifying name from a pathname (may be absolute or
 *      relative) and/or a file descriptor.
 *
 *      The name returned must be unique (different files map to different
 *      names), and repeatable (same files, map to same names).  It's not
 *      so easy to do by name.  Should handle not only:
 *
 *              foo.bar == ./foo.bar == c:/whatever_path/foo.bar
 *
 *      but also understand that:
 *
 *              foo.bar == Foo.Bar      (FAT file system)
 *              foo.bar != Foo.Bar      (NTFS)
 *
 *      The best solution is to use the file index, found in the file
 *      information structure (similar to UNIX inode #).
 *
 *      When a file is deleted, its file index may be reused,
 *      but if the unique name has not gone from its namespace,
 *      we may get a conflict.  So to ensure some tie in to the
 *      original pathname, we also use the creation time and the
 *      file basename.  This is not a perfect system, but it
 *      should work for all but anamolous test cases.
 *
 */
static int
__os_unique_name(orig_path, hfile, result_path, result_path_len)
        _TCHAR *orig_path, *result_path;
        HANDLE hfile;
        size_t result_path_len;
{
        BY_HANDLE_FILE_INFORMATION fileinfo;
        _TCHAR *basename, *p;

        /*
         * In Windows, pathname components are delimited by '/' or '\', and
         * if neither is present, we need to strip off leading drive letter
         * (e.g. c:foo.txt).
         */
        basename = _tcsrchr(orig_path, '/');
        p = _tcsrchr(orig_path, '\\');
        if (basename == NULL || (p != NULL && p > basename))
                basename = p;
        if (basename == NULL)
                basename = _tcsrchr(orig_path, ':');

        if (basename == NULL)
                basename = orig_path;
        else
                basename++;

        if (!GetFileInformationByHandle(hfile, &fileinfo))
                return (__os_get_errno());

        (void)_sntprintf(result_path, result_path_len,
            _T("__db_shmem.%8.8lx.%8.8lx.%8.8lx.%8.8lx.%8.8lx.%s"),
            fileinfo.dwVolumeSerialNumber,
            fileinfo.nFileIndexHigh,
            fileinfo.nFileIndexLow,
            fileinfo.ftCreationTime.dwHighDateTime,
            fileinfo.ftCreationTime.dwHighDateTime,
            basename);

        return (0);
}

/*
 * __os_map --
 *      The mmap(2) function for Windows.
 */
static int
__os_map(dbenv, path, infop, fhp, len, is_region, is_system, is_rdonly, addr)
        DB_ENV *dbenv;
        REGINFO *infop;
        char *path;
        DB_FH *fhp;
        int is_region, is_system, is_rdonly;
        size_t len;
        void **addr;
{
        HANDLE hMemory;
        int ret, use_pagefile;
        _TCHAR *tpath, shmem_name[MAXPATHLEN];
        void *pMemory;

        ret = 0;
        if (infop != NULL)
                infop->wnt_handle = NULL;

        use_pagefile = is_region && is_system;

        /*
         * If creating a region in system space, get a matching name in the
         * paging file namespace.
         */
        if (use_pagefile) {
                TO_TSTRING(dbenv, path, tpath, ret);
                if (ret != 0)
                        return (ret);
                ret = __os_unique_name(tpath, fhp->handle,
                    shmem_name, sizeof(shmem_name));
                FREE_STRING(dbenv, tpath);
                if (ret != 0)
                        return (ret);
        }

        /*
         * XXX
         * DB: We have not implemented copy-on-write here.
         *
         * If this is an region in system memory, we try to open it using the
         * OpenFileMapping() first, and only call CreateFileMapping() if we're
         * really creating the section.  There are two reasons:
         *
         * 1) We only create the mapping if we have newly created the region.
         *    This avoids a long-running problem caused by Windows reference
         *    counting, where regions that are closed by all processes are
         *    deleted.  It turns out that just checking for a zeroed region
         *    is not good enough. See [#4882] and [#7127] for the details.
         *
         * 2) CreateFileMapping seems to mess up making the commit charge to
         *    the process. It thinks, incorrectly, that when we want to join a
         *    previously existing section, that it should make a commit charge
         *    for the whole section.  In fact, there is no new committed memory
         *    whatever.  The call can fail if there is insufficient memory free
         *    to handle the erroneous commit charge.  So, we find that the
         *    bogus commit is not made if we call OpenFileMapping.
         */
        hMemory = NULL;
        if (use_pagefile) {
                hMemory = OpenFileMapping(
                    is_rdonly ? FILE_MAP_READ : FILE_MAP_ALL_ACCESS,
                    0, shmem_name);

                if (hMemory == NULL && F_ISSET(infop, REGION_CREATE_OK))
                        hMemory = CreateFileMapping((HANDLE)-1, 0,
                            is_rdonly ? PAGE_READONLY : PAGE_READWRITE,
                            0, (DWORD)len, shmem_name);
        } else
                hMemory = CreateFileMapping(fhp->handle, 0,
                    is_rdonly ? PAGE_READONLY : PAGE_READWRITE,
                    0, (DWORD)len, NULL);

        if (hMemory == NULL) {
                ret = __os_get_errno();
                __db_err(dbenv, "OpenFileMapping: %s", strerror(ret));
                return (__db_panic(dbenv, ret));
        }

        pMemory = MapViewOfFile(hMemory,
            (is_rdonly ? FILE_MAP_READ : FILE_MAP_ALL_ACCESS), 0, 0, len);
        if (pMemory == NULL) {
                ret = __os_get_errno();
                __db_err(dbenv, "MapViewOfFile: %s", strerror(ret));
                return (__db_panic(dbenv, ret));
        }

        /*
         * XXX
         * It turns out that the kernel object underlying the named section
         * is reference counted, but that the call to MapViewOfFile() above
         * does NOT increment the reference count! So, if we close the handle
         * here, the kernel deletes the object from the kernel namespace.
         * When a second process comes along to join the region, the kernel
         * happily creates a new object with the same name, but completely
         * different identity. The two processes then have distinct isolated
         * mapped sections, not at all what was wanted. Not closing the handle
         * here fixes this problem.  We carry the handle around in the region
         * structure so we can close it when unmap is called.
         */
        if (use_pagefile && infop != NULL)
                infop->wnt_handle = hMemory;
        else
                CloseHandle(hMemory);

        *addr = pMemory;
        return (ret);
}

---