uvm_vnode.c

基于组件方式开发操作系统的OSKIT源代码

/*	$NetBSD: uvm_vnode.c,v 1.39 2000/12/06 03:37:30 chs Exp $	*/

/*
 * Copyright (c) 1997 Charles D. Cranor and Washington University.
 * Copyright (c) 1991, 1993
 *      The Regents of the University of California.  
 * Copyright (c) 1990 University of Utah.
 *
 * All rights reserved.
 *
 * This code is derived from software contributed to Berkeley by
 * the Systems Programming Group of the University of Utah Computer
 * Science Department.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 * 3. All advertising materials mentioning features or use of this software
 *    must display the following acknowledgement:
 *      This product includes software developed by Charles D. Cranor,
 *	Washington University, the University of California, Berkeley and 
 *	its contributors.
 * 4. Neither the name of the University nor the names of its contributors
 *    may be used to endorse or promote products derived from this software
 *    without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 *
 *      @(#)vnode_pager.c       8.8 (Berkeley) 2/13/94
 * from: Id: uvm_vnode.c,v 1.1.2.26 1998/02/02 20:38:07 chuck Exp
 */

#include "fs_nfs.h"
#include "opt_uvmhist.h"
#include "opt_ddb.h"

/*
 * uvm_vnode.c: the vnode pager.
 */

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/proc.h>
#include <sys/malloc.h>
#include <sys/vnode.h>
#include <sys/disklabel.h>
#include <sys/ioctl.h>
#include <sys/fcntl.h>
#include <sys/conf.h>
#include <sys/pool.h>
#include <sys/mount.h>

#include <miscfs/specfs/specdev.h>

#include <uvm/uvm.h>
#include <uvm/uvm_vnode.h>

#ifndef OSKIT
/*
 * functions
 */

static void		uvn_cluster __P((struct uvm_object *, voff_t, voff_t *,
					 voff_t *));
static void		uvn_detach __P((struct uvm_object *));
#endif
static int		uvn_findpage __P((struct uvm_object *, voff_t,
					  struct vm_page **, int));
#ifndef OSKIT
static boolean_t	uvn_flush __P((struct uvm_object *, voff_t, voff_t,
				       int));
static int		uvn_get __P((struct uvm_object *, voff_t, vm_page_t *,
				     int *, int, vm_prot_t, int, int));
static int		uvn_put __P((struct uvm_object *, vm_page_t *, int,
				     boolean_t));
static void		uvn_reference __P((struct uvm_object *));
static boolean_t	uvn_releasepg __P((struct vm_page *,
					   struct vm_page **));

/*
 * master pager structure
 */

struct uvm_pagerops uvm_vnodeops = {
	NULL,
	uvn_reference,
	uvn_detach,
	NULL,
	uvn_flush,
	uvn_get,
	uvn_put,
	uvn_cluster,
	uvm_mk_pcluster,
	uvn_releasepg,
};
#endif /*OSKIT*/

#ifndef OSKIT
/*
 * the ops!
 */

/*
 * uvn_attach
 *
 * attach a vnode structure to a VM object.  if the vnode is already
 * attached, then just bump the reference count by one and return the
 * VM object.   if not already attached, attach and return the new VM obj.
 * the "accessprot" tells the max access the attaching thread wants to
 * our pages.
 *
 * => caller must _not_ already be holding the lock on the uvm_object.
 * => in fact, nothing should be locked so that we can sleep here.
 * => note that uvm_object is first thing in vnode structure, so their
 *    pointers are equiv.
 */

struct uvm_object *
uvn_attach(arg, accessprot)
	void *arg;
	vm_prot_t accessprot;
{
	struct vnode *vp = arg;
	struct uvm_vnode *uvn = &vp->v_uvm;
	struct vattr vattr;
	int result;
	struct partinfo pi;
	voff_t used_vnode_size;
	UVMHIST_FUNC("uvn_attach"); UVMHIST_CALLED(maphist);

	UVMHIST_LOG(maphist, "(vn=0x%x)", arg,0,0,0);
	used_vnode_size = (voff_t)0;

	/*
	 * first get a lock on the uvn.
	 */
	simple_lock(&uvn->u_obj.vmobjlock);
	while (uvn->u_flags & VXLOCK) {
		uvn->u_flags |= VXWANT;
		UVMHIST_LOG(maphist, "  SLEEPING on blocked vn",0,0,0,0);
		UVM_UNLOCK_AND_WAIT(uvn, &uvn->u_obj.vmobjlock, FALSE,
		    "uvn_attach", 0);
		simple_lock(&uvn->u_obj.vmobjlock);
		UVMHIST_LOG(maphist,"  WOKE UP",0,0,0,0);
	}

	/*
	 * if we're mapping a BLK device, make sure it is a disk.
	 */
	if (vp->v_type == VBLK && bdevsw[major(vp->v_rdev)].d_type != D_DISK) {
		simple_unlock(&uvn->u_obj.vmobjlock);
		UVMHIST_LOG(maphist,"<- done (VBLK not D_DISK!)", 0,0,0,0);
		return(NULL);
	}

#ifdef DIAGNOSTIC
	if (vp->v_type != VREG) {
		panic("uvn_attach: vp %p not VREG", vp);
	}
#endif

	/*
	 * set up our idea of the size
	 * if this hasn't been done already.
	 */
	if (uvn->u_size == VSIZENOTSET) {

	uvn->u_flags |= VXLOCK;
	simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock in case we sleep */
		/* XXX: curproc? */
	if (vp->v_type == VBLK) {
		/*
		 * We could implement this as a specfs getattr call, but:
		 *
		 *	(1) VOP_GETATTR() would get the file system
		 *	    vnode operation, not the specfs operation.
		 *
		 *	(2) All we want is the size, anyhow.
		 */
		result = (*bdevsw[major(vp->v_rdev)].d_ioctl)(vp->v_rdev,
		    DIOCGPART, (caddr_t)&pi, FREAD, curproc);
		if (result == 0) {
			/* XXX should remember blocksize */
			used_vnode_size = (voff_t)pi.disklab->d_secsize *
			    (voff_t)pi.part->p_size;
		}
	} else {
		result = VOP_GETATTR(vp, &vattr, curproc->p_ucred, curproc);
		if (result == 0)
			used_vnode_size = vattr.va_size;
	}

	/* relock object */
	simple_lock(&uvn->u_obj.vmobjlock);

	if (uvn->u_flags & VXWANT)
		wakeup(uvn);
	uvn->u_flags &= ~(VXLOCK|VXWANT);

	if (result != 0) {
		simple_unlock(&uvn->u_obj.vmobjlock); /* drop lock */
		UVMHIST_LOG(maphist,"<- done (VOP_GETATTR FAILED!)", 0,0,0,0);
		return(NULL);
	}
	uvn->u_size = used_vnode_size;

	}

	/* unlock and return */
	simple_unlock(&uvn->u_obj.vmobjlock);
	UVMHIST_LOG(maphist,"<- done, refcnt=%d", uvn->u_obj.uo_refs,
	    0, 0, 0);
	return (&uvn->u_obj);
}


/*
 * uvn_reference
 *
 * duplicate a reference to a VM object.  Note that the reference
 * count must already be at least one (the passed in reference) so 
 * there is no chance of the uvn being killed or locked out here.
 *
 * => caller must call with object unlocked.  
 * => caller must be using the same accessprot as was used at attach time
 */


static void
uvn_reference(uobj)
	struct uvm_object *uobj;
{
	VREF((struct vnode *)uobj);
}

/*
 * uvn_detach
 *
 * remove a reference to a VM object.
 *
 * => caller must call with object unlocked and map locked.
 * => this starts the detach process, but doesn't have to finish it
 *    (async i/o could still be pending).
 */
static void
uvn_detach(uobj)
	struct uvm_object *uobj;
{
	vrele((struct vnode *)uobj);
}
#endif /*OSKIT*/

/*
 * uvn_releasepg: handled a released page in a uvn
 *
 * => "pg" is a PG_BUSY [caller owns it], PG_RELEASED page that we need
 *	to dispose of.
 * => caller must handled PG_WANTED case
 * => called with page's object locked, pageq's unlocked
 * => returns TRUE if page's object is still alive, FALSE if we
 *	killed the page's object.    if we return TRUE, then we
 *	return with the object locked.
 * => if (nextpgp != NULL) => we return the next page on the queue, and return
 *				with the page queues locked [for pagedaemon]
 * => if (nextpgp == NULL) => we return with page queues unlocked [normal case]
 * => we kill the uvn if it is not referenced and we are suppose to
 *	kill it ("relkill").
 */

boolean_t
uvn_releasepg(pg, nextpgp)
	struct vm_page *pg;
	struct vm_page **nextpgp;	/* OUT */
{
	KASSERT(pg->flags & PG_RELEASED);
	
	/*
	 * dispose of the page [caller handles PG_WANTED]
	 */
	pmap_page_protect(pg, VM_PROT_NONE);
	uvm_lock_pageq();
	if (nextpgp)
		*nextpgp = TAILQ_NEXT(pg, pageq);
	uvm_pagefree(pg);
	if (!nextpgp)
		uvm_unlock_pageq();

	return (TRUE);
}

#ifndef OSKIT
/*
 * NOTE: currently we have to use VOP_READ/VOP_WRITE because they go
 * through the buffer cache and allow I/O in any size.  These VOPs use
 * synchronous i/o.  [vs. VOP_STRATEGY which can be async, but doesn't
 * go through the buffer cache or allow I/O sizes larger than a
 * block].  we will eventually want to change this.
 *
 * issues to consider:
 *   uvm provides the uvm_aiodesc structure for async i/o management.
 * there are two tailq's in the uvm. structure... one for pending async
 * i/o and one for "done" async i/o.   to do an async i/o one puts
 * an aiodesc on the "pending" list (protected by splbio()), starts the
 * i/o and returns VM_PAGER_PEND.    when the i/o is done, we expect
 * some sort of "i/o done" function to be called (at splbio(), interrupt
 * time).   this function should remove the aiodesc from the pending list
 * and place it on the "done" list and wakeup the daemon.   the daemon
 * will run at normal spl() and will remove all items from the "done"
 * list and call the "aiodone" hook for each done request (see uvm_pager.c).
 * [in the old vm code, this was done by calling the "put" routine with
 * null arguments which made the code harder to read and understand because
 * you had one function ("put") doing two things.]  
 *
 * so the current pager needs: 
 *   int uvn_aiodone(struct uvm_aiodesc *)
 *
 * => return KERN_SUCCESS (aio finished, free it).  otherwise requeue for
 *	later collection.
 * => called with pageq's locked by the daemon.
 *
 * general outline:
 * - "try" to lock object.   if fail, just return (will try again later)
 * - drop "u_nio" (this req is done!)
 * - if (object->iosync && u_naio == 0) { wakeup &uvn->u_naio }
 * - get "page" structures (atop?).
 * - handle "wanted" pages
 * - handle "released" pages [using pgo_releasepg]
 *   >>> pgo_releasepg may kill the object
 * dont forget to look at "object" wanted flag in all cases.
 */


/*
 * uvn_flush: flush pages out of a uvm object.
 *
 * => object should be locked by caller.   we may _unlock_ the object
 *	if (and only if) we need to clean a page (PGO_CLEANIT).
 *	we return with the object locked.
 * => if PGO_CLEANIT is set, we may block (due to I/O).   thus, a caller
 *	might want to unlock higher level resources (e.g. vm_map)
 *	before calling flush.
 * => if PGO_CLEANIT is not set, then we will neither unlock the object
 *	or block.
 * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
 *	for flushing.
 * => NOTE: we rely on the fact that the object's memq is a TAILQ and
 *	that new pages are inserted on the tail end of the list.   thus,
 *	we can make a complete pass through the object in one go by starting
 *	at the head and working towards the tail (new pages are put in
 *	front of us).
 * => NOTE: we are allowed to lock the page queues, so the caller
 *	must not be holding the lock on them [e.g. pagedaemon had
 *	better not call us with the queues locked]
 * => we return TRUE unless we encountered some sort of I/O error
 *
 * comment on "cleaning" object and PG_BUSY pages:
 *	this routine is holding the lock on the object.   the only time
 *	that it can run into a PG_BUSY page that it does not own is if
 *	some other process has started I/O on the page (e.g. either
 *	a pagein, or a pageout).    if the PG_BUSY page is being paged
 *	in, then it can not be dirty (!PG_CLEAN) because no one has
 *	had a chance to modify it yet.    if the PG_BUSY page is being
 *	paged out then it means that someone else has already started
 *	cleaning the page for us (how nice!).    in this case, if we 
 *	have syncio specified, then after we make our pass through the
 *	object we need to wait for the other PG_BUSY pages to clear 
 *	off (i.e. we need to do an iosync).   also note that once a
 *	page is PG_BUSY it must stay in its object until it is un-busyed.
 *
 * note on page traversal:
 *	we can traverse the pages in an object either by going down the
 *	linked list in "uobj->memq", or we can go over the address range
 *	by page doing hash table lookups for each address.    depending
 *	on how many pages are in the object it may be cheaper to do one 
 *	or the other.   we set "by_list" to true if we are using memq.
 *	if the cost of a hash lookup was equal to the cost of the list
 *	traversal we could compare the number of pages in the start->stop
 *	range to the total number of pages in the object.   however, it
 *	seems that a hash table lookup is more expensive than the linked
 *	list traversal, so we multiply the number of pages in the 
 *	start->stop range by a penalty which we define below.
 */

#define UVN_HASH_PENALTY 4	/* XXX: a guess */

static boolean_t
uvn_flush(uobj, start, stop, flags)
	struct uvm_object *uobj;
	voff_t start, stop;
	int flags;
{
	struct uvm_vnode *uvn = (struct uvm_vnode *)uobj;
	struct vnode *vp = (struct vnode *)uobj;
	struct vm_page *pp, *ppnext, *ptmp;
	struct vm_page *pps[256], **ppsp;
	int s;
	int npages, result, lcv;
	boolean_t retval, need_iosync, by_list, needs_clean, all, wasclean;
	voff_t curoff;
	u_short pp_version;
	UVMHIST_FUNC("uvn_flush"); UVMHIST_CALLED(maphist);
	UVMHIST_LOG(maphist, "uobj %p start 0x%x stop 0x%x flags 0x%x",
		    uobj, start, stop, flags);
	KASSERT(flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE));

#ifdef DEBUG
	if (uvn->u_size == VSIZENOTSET) {
		printf("uvn_flush: size not set vp %p\n", uvn);
		vprint("uvn_flush VSIZENOTSET", vp);
		flags |= PGO_ALLPAGES;
	}
#endif

	/*
	 * get init vals and determine how we are going to traverse object
	 */

	curoff = 0;
	need_iosync = FALSE;
	retval = TRUE;
	wasclean = TRUE;
	if (flags & PGO_ALLPAGES) {
		all = TRUE;
		by_list = TRUE;
	} else {
		start = trunc_page(start);
		stop = round_page(stop);
#ifdef DEBUG
		if (stop > round_page(uvn->u_size)) {
			printf("uvn_flush: oor vp %p start 0x%x stop 0x%x "
			       "size 0x%x\n", uvn, (int)start, (int)stop,
			       (int)round_page(uvn->u_size));
		}
#endif
		all = FALSE;
		by_list = (uobj->uo_npages <= 
		    ((stop - start) >> PAGE_SHIFT) * UVN_HASH_PENALTY);
	}

	UVMHIST_LOG(maphist,
	    " flush start=0x%x, stop=0x%x, by_list=%d, flags=0x%x",
	    start, stop, by_list, flags);

	/*
	 * PG_CLEANCHK: this bit is used by the pgo_mk_pcluster function as
	 * a _hint_ as to how up to date the PG_CLEAN bit is.   if the hint
	 * is wrong it will only prevent us from clustering... it won't break
	 * anything.   we clear all PG_CLEANCHK bits here, and pgo_mk_pcluster
	 * will set them as it syncs PG_CLEAN.   This is only an issue if we
	 * are looking at non-inactive pages (because inactive page's PG_CLEAN
	 * bit is always up to date since there are no mappings).
	 * [borrowed PG_CLEANCHK idea from FreeBSD VM]
	 */

	if ((flags & PGO_CLEANIT) != 0 &&
	    uobj->pgops->pgo_mk_pcluster != NULL) {
		if (by_list) {
			TAILQ_FOREACH(pp, &uobj->memq, listq) {
				if (!all &&
				    (pp->offset < start || pp->offset >= stop))
					continue;
				pp->flags &= ~PG_CLEANCHK;
			}

		} else {   /* by hash */
			for (curoff = start ; curoff < stop;
			    curoff += PAGE_SIZE) {
				pp = uvm_pagelookup(uobj, curoff);
				if (pp)
					pp->flags &= ~PG_CLEANCHK;
			}
		}
	}

	/*
	 * now do it.   note: we must update ppnext in body of loop or we
	 * will get stuck.  we need to use ppnext because we may free "pp"
	 * before doing the next loop.
	 */

	if (by_list) {
		pp = TAILQ_FIRST(&uobj->memq);
	} else {
		curoff = start;
		pp = uvm_pagelookup(uobj, curoff);
	}

	ppnext = NULL;
	ppsp = NULL;
	uvm_lock_pageq();

	/* locked: both page queues and uobj */
	for ( ; (by_list && pp != NULL) || 
		      (!by_list && curoff < stop) ; pp = ppnext) {
		if (by_list) {
			if (!all &&
			    (pp->offset < start || pp->offset >= stop)) {
				ppnext = TAILQ_NEXT(pp, listq);
				continue;
			}
		} else {
			curoff += PAGE_SIZE;
			if (pp == NULL) {
				if (curoff < stop)
					ppnext = uvm_pagelookup(uobj, curoff);
				continue;
			}
		}

		/*
		 * handle case where we do not need to clean page (either
		 * because we are not clean or because page is not dirty or
		 * is busy):
		 * 
		 * NOTE: we are allowed to deactivate a non-wired active
		 * PG_BUSY page, but once a PG_BUSY page is on the inactive
		 * queue it must stay put until it is !PG_BUSY (so as not to
		 * confuse pagedaemon).
		 */

		if ((flags & PGO_CLEANIT) == 0 || (pp->flags & PG_BUSY) != 0) {
			needs_clean = FALSE;
			if ((flags & (PGO_CLEANIT|PGO_SYNCIO)) ==
			             (PGO_CLEANIT|PGO_SYNCIO))
				need_iosync = TRUE;