uvm_pager.c

来自「基于组件方式开发操作系统的OSKIT源代码」· C语言 代码 · 共 915 行 · 第 1/2 页

/*	$NetBSD: uvm_pager.c,v 1.38 2000/12/09 23:26:27 chs Exp $	*/

/*
 *
 * Copyright (c) 1997 Charles D. Cranor and Washington University.
 * All rights reserved.
 *
 * 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 and
 *      Washington University.
 * 4. The name of the author may not be used to endorse or promote products
 *    derived from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``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 AUTHOR 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.
 *
 * from: Id: uvm_pager.c,v 1.1.2.23 1998/02/02 20:38:06 chuck Exp
 */

#include "opt_uvmhist.h"

/*
 * uvm_pager.c: generic functions used to assist the pagers.
 */

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/proc.h>
#include <sys/malloc.h>
#include <sys/pool.h>
#include <sys/vnode.h>

#define UVM_PAGER
#include <uvm/uvm.h>

struct pool *uvm_aiobuf_pool;

/*
 * list of uvm pagers in the system
 */

extern struct uvm_pagerops uvm_deviceops;
extern struct uvm_pagerops uvm_vnodeops;
extern struct uvm_pagerops ubc_pager;

struct uvm_pagerops *uvmpagerops[] = {
	&aobj_pager,
#ifndef OSKIT
	&uvm_deviceops,
#endif
	&uvm_vnodeops,
#ifndef OSKIT
	&ubc_pager,
#endif
};

/*
 * the pager map: provides KVA for I/O
 */

vm_map_t pager_map;		/* XXX */
simple_lock_data_t pager_map_wanted_lock;
boolean_t pager_map_wanted;	/* locked by pager map */
static vaddr_t emergva;
static boolean_t emerginuse;

/*
 * uvm_pager_init: init pagers (at boot time)
 */

void
uvm_pager_init()
{
	int lcv;

	/*
	 * init pager map
	 */

	pager_map = uvm_km_suballoc(kernel_map, &uvm.pager_sva, &uvm.pager_eva,
	 			    PAGER_MAP_SIZE, 0, FALSE, NULL);
	simple_lock_init(&pager_map_wanted_lock);
	pager_map_wanted = FALSE;
	emergva = uvm_km_valloc(kernel_map, MAXBSIZE);
	emerginuse = FALSE;

	/*
	 * init ASYNC I/O queue
	 */
	
	TAILQ_INIT(&uvm.aio_done);

	/*
	 * call pager init functions
	 */
	for (lcv = 0 ; lcv < sizeof(uvmpagerops)/sizeof(struct uvm_pagerops *);
	    lcv++) {
		if (uvmpagerops[lcv]->pgo_init)
			uvmpagerops[lcv]->pgo_init();
	}
}

/*
 * uvm_pagermapin: map pages into KVA (pager_map) for I/O that needs mappings
 *
 * we basically just map in a blank map entry to reserve the space in the
 * map and then use pmap_enter() to put the mappings in by hand.
 */

vaddr_t
uvm_pagermapin(pps, npages, flags)
	struct vm_page **pps;
	int npages;
	int flags;
{
	vsize_t size;
	vaddr_t kva;
	vaddr_t cva;
	struct vm_page *pp;
	vm_prot_t prot;
	UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist);

	UVMHIST_LOG(maphist,"(pps=0x%x, npages=%d)", pps, npages,0,0);

	/*
	 * compute protection.  outgoing I/O only needs read
	 * access to the page, whereas incoming needs read/write.
	 */

	prot = VM_PROT_READ;
	if (flags & UVMPAGER_MAPIN_READ)
		prot |= VM_PROT_WRITE;

ReStart:
	size = npages << PAGE_SHIFT;
	kva = 0;			/* let system choose VA */

	if (uvm_map(pager_map, &kva, size, NULL, 
	      UVM_UNKNOWN_OFFSET, 0, UVM_FLAG_NOMERGE) != KERN_SUCCESS) {
		if (curproc == uvm.pagedaemon_proc) {
			simple_lock(&pager_map_wanted_lock);
			if (emerginuse) {
				UVM_UNLOCK_AND_WAIT(&emergva,
				    &pager_map_wanted_lock, FALSE,
				    "emergva", 0);
				goto ReStart;
			}
			emerginuse = TRUE;
			simple_unlock(&pager_map_wanted_lock);
			kva = emergva;
			KASSERT(npages <= MAXBSIZE >> PAGE_SHIFT);
			goto enter;
		}
		if ((flags & UVMPAGER_MAPIN_WAITOK) == 0) {
			UVMHIST_LOG(maphist,"<- NOWAIT failed", 0,0,0,0);
			return(0);
		}
		simple_lock(&pager_map_wanted_lock);
		pager_map_wanted = TRUE; 
		UVMHIST_LOG(maphist, "  SLEEPING on pager_map",0,0,0,0);
		UVM_UNLOCK_AND_WAIT(pager_map, &pager_map_wanted_lock, FALSE, 
		    "pager_map", 0);
		goto ReStart;
	}

enter:
	/* got it */
	for (cva = kva ; size != 0 ; size -= PAGE_SIZE, cva += PAGE_SIZE) {
		pp = *pps++;
		KASSERT(pp->flags & PG_BUSY);
		pmap_enter(vm_map_pmap(pager_map), cva, VM_PAGE_TO_PHYS(pp),
		    prot, PMAP_WIRED | ((pp->flags & PG_FAKE) ? prot :
					VM_PROT_READ));
	}

	UVMHIST_LOG(maphist, "<- done (KVA=0x%x)", kva,0,0,0);
	return(kva);
}

/*
 * uvm_pagermapout: remove pager_map mapping
 *
 * we remove our mappings by hand and then remove the mapping (waking
 * up anyone wanting space).
 */

void
uvm_pagermapout(kva, npages)
	vaddr_t kva;
	int npages;
{
	vsize_t size = npages << PAGE_SHIFT;
	vm_map_entry_t entries;
	UVMHIST_FUNC("uvm_pagermapout"); UVMHIST_CALLED(maphist);

	UVMHIST_LOG(maphist, " (kva=0x%x, npages=%d)", kva, npages,0,0);

	/*
	 * duplicate uvm_unmap, but add in pager_map_wanted handling.
	 */

	if (kva == emergva) {
		simple_lock(&pager_map_wanted_lock);
		emerginuse = FALSE;
		wakeup(&emergva);
		simple_unlock(&pager_map_wanted_lock);
		entries = NULL;
		goto remove;
	}

	vm_map_lock(pager_map);
	(void) uvm_unmap_remove(pager_map, kva, kva + size, &entries);
	simple_lock(&pager_map_wanted_lock);
	if (pager_map_wanted) {
		pager_map_wanted = FALSE;
		wakeup(pager_map);
	}
	simple_unlock(&pager_map_wanted_lock);
	vm_map_unlock(pager_map);
remove:
	pmap_remove(pmap_kernel(), kva, kva + (npages << PAGE_SHIFT));
	if (entries)
		uvm_unmap_detach(entries, 0);

	UVMHIST_LOG(maphist,"<- done",0,0,0,0);
}

/*
 * uvm_mk_pcluster
 *
 * generic "make 'pager put' cluster" function.  a pager can either
 * [1] set pgo_mk_pcluster to NULL (never cluster), [2] set it to this
 * generic function, or [3] set it to a pager specific function.
 *
 * => caller must lock object _and_ pagequeues (since we need to look
 *    at active vs. inactive bits, etc.)
 * => caller must make center page busy and write-protect it
 * => we mark all cluster pages busy for the caller
 * => the caller must unbusy all pages (and check wanted/released
 *    status if it drops the object lock)
 * => flags:
 *      PGO_ALLPAGES:  all pages in object are valid targets
 *      !PGO_ALLPAGES: use "lo" and "hi" to limit range of cluster
 *      PGO_DOACTCLUST: include active pages in cluster.
 *        NOTE: the caller should clear PG_CLEANCHK bits if PGO_DOACTCLUST.
 *              PG_CLEANCHK is only a hint, but clearing will help reduce
 *		the number of calls we make to the pmap layer.
 */

struct vm_page **
uvm_mk_pcluster(uobj, pps, npages, center, flags, mlo, mhi)
	struct uvm_object *uobj;	/* IN */
	struct vm_page **pps, *center;  /* IN/OUT, IN */
	int *npages, flags;		/* IN/OUT, IN */
	voff_t mlo, mhi;		/* IN (if !PGO_ALLPAGES) */
{
	struct vm_page **ppsp, *pclust;
	voff_t lo, hi, curoff;
	int center_idx, forward, incr;
	UVMHIST_FUNC("uvm_mk_pcluster"); UVMHIST_CALLED(maphist);

	/* 
	 * center page should already be busy and write protected.  XXX:
	 * suppose page is wired?  if we lock, then a process could
	 * fault/block on it.  if we don't lock, a process could write the
	 * pages in the middle of an I/O.  (consider an msync()).  let's
	 * lock it for now (better to delay than corrupt data?).
	 */

	/*
	 * get cluster boundaries, check sanity, and apply our limits as well.
	 */

	uobj->pgops->pgo_cluster(uobj, center->offset, &lo, &hi);
	if ((flags & PGO_ALLPAGES) == 0) {
		if (lo < mlo)
			lo = mlo;
		if (hi > mhi)
			hi = mhi;
	}
	if ((hi - lo) >> PAGE_SHIFT > *npages) { /* pps too small, bail out! */
#ifdef DIAGNOSTIC
		printf("uvm_mk_pcluster uobj %p npages %d lo 0x%llx hi 0x%llx "
		       "flags 0x%x\n", uobj, *npages, (long long)lo,
		       (long long)hi, flags);
#endif
		pps[0] = center;
		*npages = 1;
		return(pps);
	}

	/*
	 * now determine the center and attempt to cluster around the
	 * edges
	 */

	center_idx = (center->offset - lo) >> PAGE_SHIFT;
	pps[center_idx] = center;	/* plug in the center page */
	ppsp = &pps[center_idx];
	*npages = 1;

	/*
	 * attempt to cluster around the left [backward], and then 
	 * the right side [forward].    
	 *
	 * note that for inactive pages (pages that have been deactivated)
	 * there are no valid mappings and PG_CLEAN should be up to date.
	 * [i.e. there is no need to query the pmap with pmap_is_modified
	 * since there are no mappings].
	 */

	for (forward  = 0 ; forward <= 1 ; forward++) {
		incr = forward ? PAGE_SIZE : -PAGE_SIZE;
		curoff = center->offset + incr;
		for ( ;(forward == 0 && curoff >= lo) ||
		       (forward && curoff < hi);
		      curoff += incr) {

			pclust = uvm_pagelookup(uobj, curoff); /* lookup page */
			if (pclust == NULL) {
				break;			/* no page */
			}
			/* handle active pages */
			/* NOTE: inactive pages don't have pmap mappings */
			if ((pclust->pqflags & PQ_INACTIVE) == 0) {
				if ((flags & PGO_DOACTCLUST) == 0) {
					/* dont want mapped pages at all */
					break;
				}

				/* make sure "clean" bit is sync'd */
				if ((pclust->flags & PG_CLEANCHK) == 0) {
					if ((pclust->flags & (PG_CLEAN|PG_BUSY))
					   == PG_CLEAN &&
					   pmap_is_modified(pclust))
						pclust->flags &= ~PG_CLEAN;

					/* now checked */
					pclust->flags |= PG_CLEANCHK;
				}
			}

			/* is page available for cleaning and does it need it */
			if ((pclust->flags & (PG_CLEAN|PG_BUSY)) != 0) {
				break;	/* page is already clean or is busy */
			}

			/* yes!   enroll the page in our array */
			pclust->flags |= PG_BUSY;		/* busy! */
			UVM_PAGE_OWN(pclust, "uvm_mk_pcluster");

			/* XXX: protect wired page?   see above comment. */
			pmap_page_protect(pclust, VM_PROT_READ);
			if (!forward) {
				ppsp--;			/* back up one page */
				*ppsp = pclust;
			} else {
				/* move forward one page */
				ppsp[*npages] = pclust;
			}
			(*npages)++;
		}
	}
	
	/*
	 * done!  return the cluster array to the caller!!!
	 */

	UVMHIST_LOG(maphist, "<- done",0,0,0,0);
	return(ppsp);
}

/*
 * uvm_pager_put: high level pageout routine
 *
 * we want to pageout page "pg" to backing store, clustering if
 * possible.
 *
 * => page queues must be locked by caller
 * => if page is not swap-backed, then "uobj" points to the object
 *	backing it.   this object should be locked by the caller.
 * => if page is swap-backed, then "uobj" should be NULL.
 * => "pg" should be PG_BUSY (by caller), and !PG_CLEAN
 *    for swap-backed memory, "pg" can be NULL if there is no page
 *    of interest [sometimes the case for the pagedaemon]
 * => "ppsp_ptr" should point to an array of npages vm_page pointers
 *	for possible cluster building
 * => flags (first two for non-swap-backed pages)
 *	PGO_ALLPAGES: all pages in uobj are valid targets
 *	PGO_DOACTCLUST: include "PQ_ACTIVE" pages as valid targets
 *	PGO_SYNCIO: do SYNC I/O (no async)
 *	PGO_PDFREECLUST: pagedaemon: drop cluster on successful I/O
 * => start/stop: if (uobj && !PGO_ALLPAGES) limit targets to this range
 *		  if (!uobj) start is the (daddr_t) of the starting swapblk
 * => return state:
 *	1. we return the VM_PAGER status code of the pageout
 *	2. we return with the page queues unlocked
 *	3. if (uobj != NULL) [!swap_backed] we return with
 *		uobj locked _only_ if PGO_PDFREECLUST is set 
 *		AND result != VM_PAGER_PEND.   in all other cases
 *		we return with uobj unlocked.   [this is a hack
 *		that allows the pagedaemon to save one lock/unlock
 *		pair in the !swap_backed case since we have to
 *		lock the uobj to drop the cluster anyway]
 *	4. on errors we always drop the cluster.   thus, if we return
 *		!PEND, !OK, then the caller only has to worry about
 *		un-busying the main page (not the cluster pages).
 *	5. on success, if !PGO_PDFREECLUST, we return the cluster
 *		with all pages busy (caller must un-busy and check
 *		wanted/released flags).
 */

int
uvm_pager_put(uobj, pg, ppsp_ptr, npages, flags, start, stop)
	struct uvm_object *uobj;	/* IN */
	struct vm_page *pg, ***ppsp_ptr;/* IN, IN/OUT */
	int *npages;			/* IN/OUT */
	int flags;			/* IN */
	voff_t start, stop;		/* IN, IN */
{
	int result;
	daddr_t swblk;
	struct vm_page **ppsp = *ppsp_ptr;
	UVMHIST_FUNC("uvm_pager_put"); UVMHIST_CALLED(ubchist);

	/*
	 * note that uobj is null  if we are doing a swap-backed pageout.
	 * note that uobj is !null if we are doing normal object pageout.
	 * note that the page queues must be locked to cluster.
	 */

	if (uobj) {	/* if !swap-backed */

		/*
		 * attempt to build a cluster for pageout using its
		 * make-put-cluster function (if it has one).
		 */

		if (uobj->pgops->pgo_mk_pcluster) {
			ppsp = uobj->pgops->pgo_mk_pcluster(uobj, ppsp,
			    npages, pg, flags, start, stop);