os_map.c revision 1
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * See the file LICENSE for redistribution information.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Copyright (c) 1996, 1997, 1998
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Sleepycat Software. All rights reserved.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chinstatic const char sccsid[] = "@(#)os_map.c 10.24 (Sleepycat) 10/12/98";
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin#endif /* not lint */
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chinstatic int __os_map __P((char *, int, size_t, int, int, int, void **));
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * __db_mapanon_ok --
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Return if this OS can support anonymous memory regions.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * PUBLIC: int __db_mapanon_ok __P((int));
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * If we don't have spinlocks, we have to have a file descriptor
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * for fcntl(2) locking, which implies using mmap(2) to map in a
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * regular file. Theoretically, we could probably find ways to
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * get a file descriptor to lock other types of shared regions,
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * but I don't see any reason to do so.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * If need_names is set, the application wants to share anonymous
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * memory among multiple processes, so we have to have a way to
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * name it. This requires shmget(2), on UNIX systems.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin#endif /* HAVE_MMAP */
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin#endif /* HAVE_SPINLOCKS */
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * __db_mapinit --
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Return if shared regions need to be initialized.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * PUBLIC: int __db_mapinit __P((void));
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Historically, some systems required that all of the bytes of the
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * region be written before it could be mmapped and accessed randomly.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * We have the option of setting REGION_INIT_NEEDED at configuration
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * time if we're running on one of those systems.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin return (1);
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin return (0);
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * __db_mapregion --
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Attach to a shared memory region.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * PUBLIC: int __db_mapregion __P((char *, REGINFO *));
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin /* If the user replaces the map call, call through their interface. */
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin 1, F_ISSET(infop, REGION_ANONYMOUS), 0, &infop->addr));
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * If we're creating anonymous regions:
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * If it's private, we use mmap(2). The problem with using
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * shmget(2) is that we may be creating a region of which the
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * application isn't aware, and if the application crashes
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * we'll have no way to remove the system resources for the
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * If it's not private, we use the shmget(2) interface if it's
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * available, because it allows us to name anonymous memory.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * If shmget(2) isn't available, use the mmap(2) calls.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * In the case of anonymous memory, using mmap(2) means the
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * memory isn't named and only the single process and its
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * threads can access the region.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * If we're trying to join an unnamed anonymous region, fail --
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * that's not possible.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin "cannot join region in unnamed anonymous memory");
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * If we're creating normal regions, we use the mmap(2)
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * interface if it's available because it's POSIX 1003.1
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * standard and we trust it more than we do shmget(2).
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin /* Mmap(2) regions that aren't anonymous can grow. */
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * __db_unmapregion --
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Detach from the shared memory region.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * PUBLIC: int __db_unmapregion __P((REGINFO *));
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * __db_unlinkregion --
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Remove the shared memory region.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * PUBLIC: int __db_unlinkregion __P((char *, REGINFO *));
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin ret = shmctl(infop->segid, IPC_RMID, NULL) ? errno : 0;
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * __db_mapfile --
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Map in a shared memory file.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * PUBLIC: int __db_mapfile __P((char *, int, size_t, int, void **));
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin return (__db_jump.j_map(path, fd, len, 0, 0, is_rdonly, addr));
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin return (__os_map(path, fd, len, 0, 0, is_rdonly, addr));
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * __db_unmapfile --
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Unmap the shared memory file.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * PUBLIC: int __db_unmapfile __P((void *, size_t));
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * __os_map --
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Call the mmap(2) function.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin__os_map(path, fd, len, is_region, is_anonymous, is_rdonly, addr)
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * If it's read-only, it's private, and if it's not, it's shared.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Don't bother with an additional parameter.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * BSD derived systems use MAP_ANON; Digital Unix and HP/UX
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * use MAP_ANONYMOUS.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Historically, MAP_FILE was required for mapping regular
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * files, even though it was the default. Some systems have
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * it, some don't, some that have it set it to 0.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * I know of no systems that implement the flag to tell the system
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * that the region contains semaphores, but it's not an unreasonable
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * thing to do, and has been part of the design since forever. I
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * don't think anyone will object, but don't set it for read-only
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * files, it doesn't make sense.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Work around a bug in the VMS V7.1 mmap() implementation. To map a file
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * into memory on VMS it needs to be opened in a certain way, originally.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * To get the file opened in that certain way, the VMS mmap() closes the
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * file and re-opens it. When it does this, it doesn't flush any caches
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * out to disk before closing. The problem this causes us is that when the
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * memory cache doesn't get written out, the file isn't big enough to match
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * the memory chunk and the mmap() call fails. This call to fsync() fixes
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * the problem. DEC thinks this isn't a bug because of language in XPG5
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * discussing user responsibility for on-disk and in-memory synchronization.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin /* MAP_FAILED was not defined in early mmap implementations. */
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin mmap(NULL, len, prot, flags, fd, (off_t)0)) == (void *)MAP_FAILED)
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin return (0);
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * __os_shmget --
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * Call the shmget(2) family of functions.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin (infop->segid = shmget(0, infop->size, IPC_PRIVATE | 0600)) == -1)
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin if ((infop->addr = shmat(infop->segid, NULL, 0)) == (void *)-1) {
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * If we're trying to join the region and failing, assume
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin * that there was a reboot and the region no longer exists.
da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968chin return (0);