env_region.c

这是linux下运行的mysql软件包,可用于linux 下安装 php + mysql + apach 的网络配置

/*-
 * See the file LICENSE for redistribution information.
 *
 * Copyright (c) 1996-2002
 *	Sleepycat Software.  All rights reserved.
 */

#include "db_config.h"

#ifndef lint
static const char revid[] = "$Id: env_region.c,v 11.64 2002/07/17 15:09:19 bostic Exp $";
#endif /* not lint */

#ifndef NO_SYSTEM_INCLUDES
#include <sys/types.h>

#include <ctype.h>
#include <string.h>
#include <unistd.h>
#endif

#include "db_int.h"
#include "dbinc/db_shash.h"
#include "dbinc/lock.h"
#include "dbinc/log.h"
#include "dbinc/mp.h"
#include "dbinc/rep.h"
#include "dbinc/txn.h"

static int  __db_des_destroy __P((DB_ENV *, REGION *));
static int  __db_des_get __P((DB_ENV *, REGINFO *, REGINFO *, REGION **));
static int  __db_e_remfile __P((DB_ENV *));
static int  __db_faultmem __P((DB_ENV *, void *, size_t, int));
static void __db_region_destroy __P((DB_ENV *, REGINFO *));

/*
 * __db_e_attach
 *	Join/create the environment
 *
 * PUBLIC: int __db_e_attach __P((DB_ENV *, u_int32_t *));
 */
int
__db_e_attach(dbenv, init_flagsp)
	DB_ENV *dbenv;
	u_int32_t *init_flagsp;
{
	REGENV *renv;
	REGENV_REF ref;
	REGINFO *infop;
	REGION *rp, tregion;
	size_t size;
	size_t nrw;
	u_int32_t mbytes, bytes;
	int retry_cnt, ret, segid;
	char buf[sizeof(DB_REGION_FMT) + 20];

#if !defined(HAVE_MUTEX_THREADS)
	/*
	 * !!!
	 * If we don't have spinlocks, we need a file descriptor for fcntl(2)
	 * locking.  We use the file handle from the REGENV file for this
	 * purpose.
	 *
	 * Since we may be using shared memory regions, e.g., shmget(2), and
	 * not a mapped-in regular file, the backing file may be only a few
	 * bytes in length.  So, this depends on the ability to call fcntl to
	 * lock file offsets much larger than the actual physical file.  I
	 * think that's safe -- besides, very few systems actually need this
	 * kind of support, SunOS is the only one still in wide use of which
	 * I'm aware.
	 *
	 * The error case is if an application lacks spinlocks and wants to be
	 * threaded.  That doesn't work because fcntl may lock the underlying
	 * process, including all its threads.
	 */
	if (F_ISSET(dbenv, DB_ENV_THREAD)) {
		__db_err(dbenv,
"architecture lacks fast mutexes: applications cannot be threaded");
		return (EINVAL);
	}
#endif

	/* Initialization */
	retry_cnt = 0;

	/* Repeated initialization. */
loop:	renv = NULL;

	/* Set up the DB_ENV's REG_INFO structure. */
	if ((ret = __os_calloc(dbenv, 1, sizeof(REGINFO), &infop)) != 0)
		return (ret);
	infop->type = REGION_TYPE_ENV;
	infop->id = REGION_ID_ENV;
	infop->mode = dbenv->db_mode;
	infop->flags = REGION_JOIN_OK;
	if (F_ISSET(dbenv, DB_ENV_CREATE))
		F_SET(infop, REGION_CREATE_OK);

	/*
	 * We have to single-thread the creation of the REGENV region.  Once
	 * it exists, we can do locking using locks in the region, but until
	 * then we have to be the only player in the game.
	 *
	 * If this is a private environment, we are only called once and there
	 * are no possible race conditions.
	 *
	 * If this is a public environment, we use the filesystem to ensure
	 * the creation of the environment file is single-threaded.
	 */
	if (F_ISSET(dbenv, DB_ENV_PRIVATE)) {
		if ((ret = __os_strdup(dbenv,
		    "process-private", &infop->name)) != 0)
			goto err;
		goto creation;
	}

	/* Build the region name. */
	(void)snprintf(buf, sizeof(buf), "%s", DB_REGION_ENV);
	if ((ret = __db_appname(dbenv,
	    DB_APP_NONE, buf, 0, NULL, &infop->name)) != 0)
		goto err;

	/*
	 * Try to create the file, if we have the authority.  We have to ensure
	 * that multiple threads/processes attempting to simultaneously create
	 * the file are properly ordered.  Open using the O_CREAT and O_EXCL
	 * flags so that multiple attempts to create the region will return
	 * failure in all but one.  POSIX 1003.1 requires that EEXIST be the
	 * errno return value -- I sure hope they're right.
	 */
	if (F_ISSET(dbenv, DB_ENV_CREATE)) {
		if ((ret = __os_open(dbenv, infop->name,
		    DB_OSO_CREATE | DB_OSO_DIRECT | DB_OSO_EXCL | DB_OSO_REGION,
		    dbenv->db_mode, dbenv->lockfhp)) == 0)
			goto creation;
		if (ret != EEXIST) {
			__db_err(dbenv,
			    "%s: %s", infop->name, db_strerror(ret));
			goto err;
		}
	}

	/*
	 * If we couldn't create the file, try and open it.  (If that fails,
	 * we're done.)
	 */
	if ((ret = __os_open(dbenv, infop->name, DB_OSO_REGION | DB_OSO_DIRECT,
	    dbenv->db_mode, dbenv->lockfhp)) != 0)
		goto err;

	/*
	 * !!!
	 * The region may be in system memory not backed by the filesystem
	 * (more specifically, not backed by this file), and we're joining
	 * it.  In that case, the process that created it will have written
	 * out a REGENV_REF structure as its only contents.  We read that
	 * structure before we do anything further, e.g., we can't just map
	 * that file in and then figure out what's going on.
	 *
	 * All of this noise is because some systems don't have a coherent VM
	 * and buffer cache, and what's worse, when you mix operations on the
	 * VM and buffer cache, half the time you hang the system.
	 *
	 * If the file is the size of an REGENV_REF structure, then we know
	 * the real region is in some other memory.  (The only way you get a
	 * file that size is to deliberately write it, as it's smaller than
	 * any possible disk sector created by writing a file or mapping the
	 * file into memory.)  In which case, retrieve the structure from the
	 * file and use it to acquire the referenced memory.
	 *
	 * If the structure is larger than a REGENV_REF structure, then this
	 * file is backing the shared memory region, and we just map it into
	 * memory.
	 *
	 * And yes, this makes me want to take somebody and kill them.  (I
	 * digress -- but you have no freakin' idea.  This is unbelievably
	 * stupid and gross, and I've probably spent six months of my life,
	 * now, trying to make different versions of it work.)
	 */
	if ((ret = __os_ioinfo(dbenv, infop->name,
	    dbenv->lockfhp, &mbytes, &bytes, NULL)) != 0) {
		__db_err(dbenv, "%s: %s", infop->name, db_strerror(ret));
		goto err;
	}

	/*
	 * !!!
	 * A size_t is OK -- regions get mapped into memory, and so can't
	 * be larger than a size_t.
	 */
	size = mbytes * MEGABYTE + bytes;

	/*
	 * If the size is less than the size of a REGENV_REF structure, the
	 * region (or, possibly, the REGENV_REF structure) has not yet been
	 * completely written.  Wait awhile and try again.
	 *
	 * Otherwise, if the size is the size of a REGENV_REF structure,
	 * read it into memory and use it as a reference to the real region.
	 */
	if (size <= sizeof(ref)) {
		if (size != sizeof(ref))
			goto retry;

		if ((ret = __os_read(dbenv, dbenv->lockfhp, &ref,
		    sizeof(ref), &nrw)) != 0 || nrw < (size_t)sizeof(ref)) {
			if (ret == 0)
				ret = EIO;
			__db_err(dbenv,
		    "%s: unable to read system-memory information from: %s",
			    infop->name, db_strerror(ret));
			goto err;
		}
		size = ref.size;
		segid = ref.segid;

		F_SET(dbenv, DB_ENV_SYSTEM_MEM);
	} else if (F_ISSET(dbenv, DB_ENV_SYSTEM_MEM)) {
		ret = EINVAL;
		__db_err(dbenv,
		    "%s: existing environment not created in system memory: %s",
		    infop->name, db_strerror(ret));
		goto err;
	} else
		segid = INVALID_REGION_SEGID;

	/*
	 * If not doing thread locking, we need to save the file handle for
	 * fcntl(2) locking.  Otherwise, discard the handle, we no longer
	 * need it, and the less contact between the buffer cache and the VM,
	 * the better.
	 */
#ifdef HAVE_MUTEX_THREADS
	 __os_closehandle(dbenv, dbenv->lockfhp);
#endif

	/* Call the region join routine to acquire the region. */
	memset(&tregion, 0, sizeof(tregion));
	tregion.size = (roff_t)size;
	tregion.segid = segid;
	if ((ret = __os_r_attach(dbenv, infop, &tregion)) != 0)
		goto err;

	/*
	 * The environment's REGENV structure has to live at offset 0 instead
	 * of the usual shalloc information.  Set the primary reference and
	 * correct the "addr" value to reference the shalloc region.  Note,
	 * this means that all of our offsets (R_ADDR/R_OFFSET) get shifted
	 * as well, but that should be fine.
	 */
	infop->primary = R_ADDR(infop, 0);
	infop->addr = (u_int8_t *)infop->addr + sizeof(REGENV);

	/*
	 * Check if the environment has had a catastrophic failure.
	 *
	 * Check the magic number to ensure the region is initialized.  If the
	 * magic number isn't set, the lock may not have been initialized, and
	 * an attempt to use it could lead to random behavior.
	 *
	 * The panic and magic values aren't protected by any lock, so we never
	 * use them in any check that's more complex than set/not-set.
	 *
	 * !!!
	 * I'd rather play permissions games using the underlying file, but I
	 * can't because Windows/NT filesystems won't open files mode 0.
	 */
	renv = infop->primary;
	if (renv->envpanic && !F_ISSET(dbenv, DB_ENV_NOPANIC)) {
		ret = __db_panic_msg(dbenv);
		goto err;
	}
	if (renv->magic != DB_REGION_MAGIC)
		goto retry;

	/* Make sure the region matches our build. */
	if (renv->majver != DB_VERSION_MAJOR ||
	    renv->minver != DB_VERSION_MINOR ||
	    renv->patch != DB_VERSION_PATCH) {
		__db_err(dbenv,
	"Program version %d.%d.%d doesn't match environment version %d.%d.%d",
		    DB_VERSION_MAJOR, DB_VERSION_MINOR, DB_VERSION_PATCH,
		    renv->majver, renv->minver, renv->patch);
#ifndef DIAGNOSTIC
		ret = EINVAL;
		goto err;
#endif
	}

	/* Lock the environment. */
	MUTEX_LOCK(dbenv, &renv->mutex);

	/*
	 * Finally!  We own the environment now.  Repeat the panic check, it's
	 * possible that it was set while we waited for the lock.
	 */
	if (renv->envpanic && !F_ISSET(dbenv, DB_ENV_NOPANIC)) {
		ret = __db_panic_msg(dbenv);
		goto err_unlock;
	}

	/*
	 * Get a reference to the underlying REGION information for this
	 * environment.
	 */
	if ((ret = __db_des_get(dbenv, infop, infop, &rp)) != 0 || rp == NULL) {
		MUTEX_UNLOCK(dbenv, &renv->mutex);
		goto find_err;
	}
	infop->rp = rp;

	/*
	 * There's still a possibility for inconsistent data.  When we acquired
	 * the size of the region and attached to it, it might have still been
	 * growing as part of its creation.  We can detect this by checking the
	 * size we originally found against the region's current size.  (The
	 * region's current size has to be final, the creator finished growing
	 * it before releasing the environment for us to lock.)
	 */
	if (rp->size != size) {
err_unlock:	MUTEX_UNLOCK(dbenv, &renv->mutex);
		goto retry;
	}

	/* Increment the reference count. */
	++renv->refcnt;

	/*
	 * If our caller wants them, return the flags this environment was
	 * initialized with.
	 */
	if (init_flagsp != NULL)
		*init_flagsp = renv->init_flags;

	/* Discard our lock. */
	MUTEX_UNLOCK(dbenv, &renv->mutex);

	/*
	 * Fault the pages into memory.  Note, do this AFTER releasing the
	 * lock, because we're only reading the pages, not writing them.
	 */
	(void)__db_faultmem(dbenv, infop->primary, rp->size, 0);

	/* Everything looks good, we're done. */
	dbenv->reginfo = infop;
	return (0);

creation:
	/* Create the environment region. */
	F_SET(infop, REGION_CREATE);

	/*
	 * Allocate room for 50 REGION structures plus overhead (we're going
	 * to use this space for last-ditch allocation requests), although we
	 * should never need anything close to that.
	 *
	 * Encryption passwds are stored in the env region.  Add that in too.
	 */
	memset(&tregion, 0, sizeof(tregion));
	tregion.size = (roff_t)(50 * sizeof(REGION) +
	    dbenv->passwd_len + 2048);
	tregion.segid = INVALID_REGION_SEGID;
	if ((ret = __os_r_attach(dbenv, infop, &tregion)) != 0)
		goto err;

	/*
	 * Fault the pages into memory.  Note, do this BEFORE we initialize
	 * anything, because we're writing the pages, not just reading them.
	 */
	(void)__db_faultmem(dbenv, infop->addr, tregion.size, 1);

	/*
	 * The first object in the region is the REGENV structure.  This is
	 * different from the other regions, and, from everything else in
	 * this region, where all objects are allocated from the pool, i.e.,
	 * there aren't any fixed locations.  The remaining space is made
	 * available for later allocation.
	 *
	 * The allocation space must be size_t aligned, because that's what
	 * the initialization routine is going to store there.  To make sure
	 * that happens, the REGENV structure was padded with a final size_t.
	 * No other region needs to worry about it because all of them treat
	 * the entire region as allocation space.
	 *
	 * Set the primary reference and correct the "addr" value to reference
	 * the shalloc region.  Note, this requires that we "uncorrect" it at
	 * region detach, and that all of our offsets (R_ADDR/R_OFFSET) will be
	 * shifted as well, but that should be fine.
	 */
	infop->primary = R_ADDR(infop, 0);
	infop->addr = (u_int8_t *)infop->addr + sizeof(REGENV);
	__db_shalloc_init(infop->addr, tregion.size - sizeof(REGENV));

	/*
	 * Initialize the rest of the REGENV structure, except for the magic
	 * number which validates the file/environment.
	 */
	renv = infop->primary;
	renv->envpanic = 0;
	db_version(&renv->majver, &renv->minver, &renv->patch);
	SH_LIST_INIT(&renv->regionq);
	renv->refcnt = 1;
	renv->cipher_off = INVALID_ROFF;
	renv->rep_off = INVALID_ROFF;

	/*
	 * Initialize init_flags to store the flags that any other environment
	 * handle that uses DB_JOINENV to join this environment will need.
	 */
	renv->init_flags = (init_flagsp == NULL) ? 0 : *init_flagsp;

	/*
	 * Lock the environment.
	 *
	 * Check the lock call return.  This is the first lock we initialize
	 * and acquire, and we have to know if it fails.  (It CAN fail, e.g.,
	 * SunOS, when using fcntl(2) for locking and using an in-memory
	 * filesystem as the database home.  But you knew that, I'm sure -- it
	 * probably wasn't even worth mentioning.)