uvm_aobj.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 1,553 行 · 第 1/3 页

	if (flags == 0 || (flags & UAO_FLAG_KERNSWAP) != 0) {
		int mflags = (flags & UAO_FLAG_KERNSWAP) != 0 ?
		    M_NOWAIT : M_WAITOK;

		/* allocate hash table or array depending on object size */
		if (UAO_USES_SWHASH(aobj)) {
			aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(aobj),
			    HASH_LIST, M_UVMAOBJ, mflags, &aobj->u_swhashmask);
			if (aobj->u_swhash == NULL)
				panic("uao_create: hashinit swhash failed");
		} else {
			aobj->u_swslots = malloc(pages * sizeof(int),
			    M_UVMAOBJ, mflags);
			if (aobj->u_swslots == NULL)
				panic("uao_create: malloc swslots failed");
			memset(aobj->u_swslots, 0, pages * sizeof(int));
		}

		if (flags) {
			aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */
			return(&aobj->u_obj);
			/* done! */
		}
	}

	/*
 	 * init aobj fields
 	 */
	simple_lock_init(&aobj->u_obj.vmobjlock);
	aobj->u_obj.pgops = &aobj_pager;
	TAILQ_INIT(&aobj->u_obj.memq);
	aobj->u_obj.uo_npages = 0;

	/*
 	 * now that aobj is ready, add it to the global list
 	 */
	simple_lock(&uao_list_lock);
	LIST_INSERT_HEAD(&uao_list, aobj, u_list);
	simple_unlock(&uao_list_lock);

	/*
 	 * done!
 	 */
	return(&aobj->u_obj);
}



/*
 * uao_init: set up aobj pager subsystem
 *
 * => called at boot time from uvm_pager_init()
 */
void
uao_init()
{
	static int uao_initialized;

	if (uao_initialized)
		return;
	uao_initialized = TRUE;

	LIST_INIT(&uao_list);
	simple_lock_init(&uao_list_lock);

	/*
	 * NOTE: Pages fror this pool must not come from a pageable
	 * kernel map!
	 */
	pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt),
	    0, 0, 0, "uaoeltpl", 0, NULL, NULL, M_UVMAOBJ);

	pool_init(&uvm_aobj_pool, sizeof(struct uvm_aobj), 0, 0, 0,
	    "aobjpl", 0,
	    pool_page_alloc_nointr, pool_page_free_nointr, M_UVMAOBJ);
}

/*
 * uao_reference: add a ref to an aobj
 *
 * => aobj must be unlocked
 * => just lock it and call the locked version
 */
void
uao_reference(uobj)
	struct uvm_object *uobj;
{
	simple_lock(&uobj->vmobjlock);
	uao_reference_locked(uobj);
	simple_unlock(&uobj->vmobjlock);
}

/*
 * uao_reference_locked: add a ref to an aobj that is already locked
 *
 * => aobj must be locked
 * this needs to be separate from the normal routine
 * since sometimes we need to add a reference to an aobj when
 * it's already locked.
 */
void
uao_reference_locked(uobj)
	struct uvm_object *uobj;
{
	UVMHIST_FUNC("uao_reference"); UVMHIST_CALLED(maphist);

	/*
 	 * kernel_object already has plenty of references, leave it alone.
 	 */

	if (UVM_OBJ_IS_KERN_OBJECT(uobj))
		return;

	uobj->uo_refs++;		/* bump! */
	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)", 
		    uobj, uobj->uo_refs,0,0);
}


/*
 * uao_detach: drop a reference to an aobj
 *
 * => aobj must be unlocked
 * => just lock it and call the locked version
 */
void
uao_detach(uobj)
	struct uvm_object *uobj;
{
	simple_lock(&uobj->vmobjlock);
	uao_detach_locked(uobj);
}


/*
 * uao_detach_locked: drop a reference to an aobj
 *
 * => aobj must be locked, and is unlocked (or freed) upon return.
 * this needs to be separate from the normal routine
 * since sometimes we need to detach from an aobj when
 * it's already locked.
 */
void
uao_detach_locked(uobj)
	struct uvm_object *uobj;
{
	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
	struct vm_page *pg;
	boolean_t busybody;
	UVMHIST_FUNC("uao_detach"); UVMHIST_CALLED(maphist);

	/*
 	 * detaching from kernel_object is a noop.
 	 */
	if (UVM_OBJ_IS_KERN_OBJECT(uobj)) {
		simple_unlock(&uobj->vmobjlock);
		return;
	}

	UVMHIST_LOG(maphist,"  (uobj=0x%x)  ref=%d", uobj,uobj->uo_refs,0,0);
	uobj->uo_refs--;				/* drop ref! */
	if (uobj->uo_refs) {				/* still more refs? */
		simple_unlock(&uobj->vmobjlock);
		UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0);
		return;
	}

	/*
 	 * remove the aobj from the global list.
 	 */
	simple_lock(&uao_list_lock);
	LIST_REMOVE(aobj, u_list);
	simple_unlock(&uao_list_lock);

	/*
 	 * free all the pages that aren't PG_BUSY,
	 * mark for release any that are.
 	 */
	busybody = FALSE;
	for (pg = TAILQ_FIRST(&uobj->memq);
	     pg != NULL;
	     pg = TAILQ_NEXT(pg, listq)) {
		if (pg->flags & PG_BUSY) {
			pg->flags |= PG_RELEASED;
			busybody = TRUE;
			continue;
		}

		/* zap the mappings, free the swap slot, free the page */
		pmap_page_protect(pg, VM_PROT_NONE);
		uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT);
		uvm_lock_pageq();
		uvm_pagefree(pg);
		uvm_unlock_pageq();
	}

	/*
 	 * if we found any busy pages, we're done for now.
 	 * mark the aobj for death, releasepg will finish up for us.
 	 */
	if (busybody) {
		aobj->u_flags |= UAO_FLAG_KILLME;
		simple_unlock(&aobj->u_obj.vmobjlock);
		return;
	}

	/*
 	 * finally, free the rest.
 	 */
	uao_free(aobj);
}

/*
 * uao_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).
 *	XXXJRT Currently, however, we don't.  In the case of cleaning
 *	XXXJRT a page, we simply just deactivate it.  Should probably
 *	XXXJRT handle this better, in the future (although "flushing"
 *	XXXJRT anonymous memory isn't terribly important).
 * => 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
 *	XXXJRT currently never happens, as we never directly initiate
 *	XXXJRT I/O
 *
 * comment on "cleaning" object and PG_BUSY pages:
 *	this routine is holding the lock on the object.  the only time
 *	that is 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 change 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 is must stary in its object until it is un-busyed.
 *	XXXJRT We never actually do this, as we are "flushing" anonymous
 *	XXXJRT memory, which doesn't have persistent backing store.
 *
 * 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	UAO_HASH_PENALTY 4	/* XXX: a guess */

boolean_t
uao_flush(uobj, start, stop, flags)
	struct uvm_object *uobj;
	voff_t start, stop;
	int flags;
{
	struct uvm_aobj *aobj = (struct uvm_aobj *) uobj;
	struct vm_page *pp, *ppnext;
	boolean_t retval, by_list;
	voff_t curoff;
	UVMHIST_FUNC("uao_flush"); UVMHIST_CALLED(maphist);

	curoff = 0;	/* XXX: shut up gcc */

	retval = TRUE;	/* default to success */

	if (flags & PGO_ALLPAGES) {
		start = 0;
		stop = aobj->u_pages << PAGE_SHIFT;
		by_list = TRUE;		/* always go by the list */
	} else {
		start = trunc_page(start);
		stop = round_page(stop);
		if (stop > (aobj->u_pages << PAGE_SHIFT)) {
			printf("uao_flush: strange, got an out of range "
			    "flush (fixed)\n");
			stop = aobj->u_pages << PAGE_SHIFT;
		}
		by_list = (uobj->uo_npages <=
		    ((stop - start) >> PAGE_SHIFT) * UAO_HASH_PENALTY);
	}

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

	/*
	 * Don't need to do any work here if we're not freeing
	 * or deactivating pages.
	 */
	if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
		UVMHIST_LOG(maphist,
		    "<- done (no work to do)",0,0,0,0);
		return (retval);
	}

	/*
	 * now do it.  note: we must update ppnext in the 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 = uobj->memq.tqh_first;
	} else {
		curoff = start;
		pp = uvm_pagelookup(uobj, curoff);
	}

	ppnext = NULL;	/* XXX: shut up gcc */
	uvm_lock_pageq();	/* page queues locked */

	/* locked: both page queues and uobj */
	for ( ; (by_list && pp != NULL) ||
	    (!by_list && curoff < stop) ; pp = ppnext) {
		if (by_list) {
			ppnext = TAILQ_NEXT(pp, listq);

			/* range check */
			if (pp->offset < start || pp->offset >= stop)
				continue;
		} else {
			curoff += PAGE_SIZE;
			if (curoff < stop)
				ppnext = uvm_pagelookup(uobj, curoff);

			/* null check */
			if (pp == NULL)
				continue;
		}
		
		switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) {
		/*
		 * XXX In these first 3 cases, we always just
		 * XXX deactivate the page.  We may want to
		 * XXX handle the different cases more specifically
		 * XXX in the future.
		 */
		case PGO_CLEANIT|PGO_FREE:
		case PGO_CLEANIT|PGO_DEACTIVATE:
		case PGO_DEACTIVATE:
 deactivate_it:
			/* skip the page if it's loaned or wired */
			if (pp->loan_count != 0 ||
			    pp->wire_count != 0)
				continue;

			/* zap all mappings for the page. */
			pmap_page_protect(pp, VM_PROT_NONE);

			/* ...and deactivate the page. */
			uvm_pagedeactivate(pp);

			continue;

		case PGO_FREE:
			/*
			 * If there are multiple references to
			 * the object, just deactivate the page.
			 */
			if (uobj->uo_refs > 1)
				goto deactivate_it;

			/* XXX skip the page if it's loaned or wired */
			if (pp->loan_count != 0 ||
			    pp->wire_count != 0)
				continue;

			/*
			 * mark the page as released if its busy.
			 */
			if (pp->flags & PG_BUSY) {
				pp->flags |= PG_RELEASED;
				continue;
			}

			/* zap all mappings for the page. */
			pmap_page_protect(pp, VM_PROT_NONE);

			uao_dropswap(uobj, pp->offset >> PAGE_SHIFT);
			uvm_pagefree(pp);

			continue;

		default:
			panic("uao_flush: weird flags");
		}
#ifdef DIAGNOSTIC
		panic("uao_flush: unreachable code");
#endif
	}

	uvm_unlock_pageq();

	UVMHIST_LOG(maphist,
	    "<- done, rv=%d",retval,0,0,0);
	return (retval);
}

/*
 * uao_get: fetch me a page
 *
 * we have three cases:
 * 1: page is resident     -> just return the page.
 * 2: page is zero-fill    -> allocate a new page and zero it.
 * 3: page is swapped out  -> fetch the page from swap.
 *
 * cases 1 and 2 can be handled with PGO_LOCKED, case 3 cannot.
 * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES),
 * then we will need to return VM_PAGER_UNLOCK.
 *
 * => prefer map unlocked (not required)
 * => object must be locked!  we will _unlock_ it before starting any I/O.
 * => flags: PGO_ALLPAGES: get all of the pages
 *           PGO_LOCKED: fault data structures are locked
 * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
 * => NOTE: caller must check for released pages!!
 */
static int
uao_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
	struct uvm_object *uobj;
	voff_t offset;
	struct vm_page **pps;
	int *npagesp;
	int centeridx, advice, flags;
	vm_prot_t access_type;
{
	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
	voff_t current_offset;
	vm_page_t ptmp;
	int lcv, gotpages, maxpages, swslot, rv, pageidx;
	boolean_t done;
	UVMHIST_FUNC("uao_get"); UVMHIST_CALLED(pdhist);

	UVMHIST_LOG(pdhist, "aobj=%p offset=%d, flags=%d",
		    aobj, offset, flags,0);

	/*
 	 * get number of pages
 	 */
	maxpages = *npagesp;

	/*
 	 * step 1: handled the case where fault data structures are locked.
 	 */

	if (flags & PGO_LOCKED) {
		/*
 		 * step 1a: get pages that are already resident.   only do
		 * this if the data structures are locked (i.e. the first
		 * time through).
 		 */

		done = TRUE;	/* be optimistic */
		gotpages = 0;	/* # of pages we got so far */

		for (lcv = 0, current_offset = offset ; lcv < maxpages ;
		    lcv++, current_offset += PAGE_SIZE) {
			/* do we care about this page?  if not, skip it */
			if (pps[lcv] == PGO_DONTCARE)
				continue;

			ptmp = uvm_pagelookup(uobj, current_offset);

			/*
 			 * if page is new, attempt to allocate the page,
			 * zero-fill'd.
 			 */
			if (ptmp == NULL && uao_find_swslot(aobj,
			    current_offset >> PAGE_SHIFT) == 0) {
				ptmp = uvm_pagealloc(uobj, current_offset,
				    NULL, UVM_PGA_ZERO);
				if (ptmp) {
					/* new page */
					ptmp->flags &= ~(PG_BUSY|PG_FAKE);
					ptmp->pqflags |= PQ_AOBJ;
					UVM_PAGE_OWN(ptmp, NULL);
				}
			}

			/*
			 * to be useful must get a non-busy, non-released page
			 */
			if (ptmp == NULL ||
			    (ptmp->flags & (PG_BUSY|PG_RELEASED)) != 0) {
				if (lcv == centeridx ||
				    (flags & PGO_ALLPAGES) != 0)
					/* need to do a wait or I/O! */
					done = FALSE;	
					continue;
			}

			/*
			 * useful page: busy/lock it and plug it in our
			 * result array
			 */
			/* caller must un-busy this page */
			ptmp->flags |= PG_BUSY;	
			UVM_PAGE_OWN(ptmp, "uao_get1");