uvm_vnode.c

基于组件方式开发操作系统的OSKIT源代码

		} else {

			/*
			 * freeing: nuke all mappings so we can sync
			 * PG_CLEAN bit with no race
			 */
			if ((pp->flags & PG_CLEAN) != 0 && 
			    (flags & PGO_FREE) != 0 &&
			    (pp->pqflags & PQ_ACTIVE) != 0)
				pmap_page_protect(pp, VM_PROT_NONE);
			if ((pp->flags & PG_CLEAN) != 0 &&
			    pmap_is_modified(pp))
				pp->flags &= ~(PG_CLEAN);
			pp->flags |= PG_CLEANCHK;
			needs_clean = ((pp->flags & PG_CLEAN) == 0);
		}

		/*
		 * if we don't need a clean... load ppnext and dispose of pp
		 */
		if (!needs_clean) {
			if (by_list)
				ppnext = TAILQ_NEXT(pp, listq);
			else {
				if (curoff < stop)
					ppnext = uvm_pagelookup(uobj, curoff);
			}

			if (flags & PGO_DEACTIVATE) {
				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
				    pp->wire_count == 0) {
					pmap_page_protect(pp, VM_PROT_NONE);
					uvm_pagedeactivate(pp);
				}

			} else if (flags & PGO_FREE) {
				if (pp->flags & PG_BUSY) {
					pp->flags |= PG_RELEASED;
				} else {
					pmap_page_protect(pp, VM_PROT_NONE);
					uvm_pagefree(pp);
				}
			}
			/* ppnext is valid so we can continue... */
			continue;
		}

		/*
		 * pp points to a page in the locked object that we are
		 * working on.  if it is !PG_CLEAN,!PG_BUSY and we asked
		 * for cleaning (PGO_CLEANIT).  we clean it now.
		 *
		 * let uvm_pager_put attempted a clustered page out.
		 * note: locked: uobj and page queues.
		 */

		wasclean = FALSE;
		pp->flags |= PG_BUSY;	/* we 'own' page now */
		UVM_PAGE_OWN(pp, "uvn_flush");
		pmap_page_protect(pp, VM_PROT_READ);
		pp_version = pp->version;
ReTry:
		ppsp = pps;
		npages = sizeof(pps) / sizeof(struct vm_page *);

		/* locked: page queues, uobj */
		result = uvm_pager_put(uobj, pp, &ppsp, &npages, 
				       flags | PGO_DOACTCLUST, start, stop);
		/* unlocked: page queues, uobj */

		/*
		 * at this point nothing is locked.   if we did an async I/O
		 * it is remotely possible for the async i/o to complete and 
		 * the page "pp" be freed or what not before we get a chance 
		 * to relock the object.   in order to detect this, we have
		 * saved the version number of the page in "pp_version".
		 */

		/* relock! */
		simple_lock(&uobj->vmobjlock);
		uvm_lock_pageq();

		/*
		 * VM_PAGER_AGAIN: given the structure of this pager, this 
		 * can only happen when  we are doing async I/O and can't
		 * map the pages into kernel memory (pager_map) due to lack
		 * of vm space.   if this happens we drop back to sync I/O.
		 */

		if (result == VM_PAGER_AGAIN) {

			/*
			 * it is unlikely, but page could have been released
			 * while we had the object lock dropped.   we ignore
			 * this now and retry the I/O.  we will detect and
			 * handle the released page after the syncio I/O
			 * completes.
			 */
#ifdef DIAGNOSTIC
			if (flags & PGO_SYNCIO)
	panic("uvn_flush: PGO_SYNCIO return 'try again' error (impossible)");
#endif
			flags |= PGO_SYNCIO;
			goto ReTry;
		}

		/*
		 * the cleaning operation is now done.   finish up.  note that
		 * on error (!OK, !PEND) uvm_pager_put drops the cluster for us.
		 * if success (OK, PEND) then uvm_pager_put returns the cluster
		 * to us in ppsp/npages.
		 */

		/*
		 * for pending async i/o if we are not deactivating/freeing
		 * we can move on to the next page.
		 */

		if (result == VM_PAGER_PEND &&
		    (flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {

			/*
			 * no per-page ops: refresh ppnext and continue
			 */
			if (by_list) {
				if (pp->version == pp_version)
					ppnext = TAILQ_NEXT(pp, listq);
				else
					ppnext = TAILQ_FIRST(&uobj->memq);
			} else {
				if (curoff < stop)
					ppnext = uvm_pagelookup(uobj, curoff);
			}
			continue;
		}

		/*
		 * need to look at each page of the I/O operation.  we defer 
		 * processing "pp" until the last trip through this "for" loop 
		 * so that we can load "ppnext" for the main loop after we
		 * play with the cluster pages [thus the "npages + 1" in the 
		 * loop below].
		 */

		for (lcv = 0 ; lcv < npages + 1 ; lcv++) {

			/*
			 * handle ppnext for outside loop, and saving pp
			 * until the end.
			 */
			if (lcv < npages) {
				if (ppsp[lcv] == pp)
					continue; /* skip pp until the end */
				ptmp = ppsp[lcv];
			} else {
				ptmp = pp;

				/* set up next page for outer loop */
				if (by_list) {
					if (pp->version == pp_version)
						ppnext = TAILQ_NEXT(pp, listq);
					else
						ppnext = TAILQ_FIRST(
						    &uobj->memq);
				} else {
					if (curoff < stop)
						ppnext = uvm_pagelookup(uobj,
						    curoff);
				}
			}

			/*
			 * verify the page wasn't moved while obj was
			 * unlocked
			 */
			if (result == VM_PAGER_PEND && ptmp->uobject != uobj)
				continue;

			/*
			 * unbusy the page if I/O is done.   note that for
			 * pending I/O it is possible that the I/O op
			 * finished before we relocked the object (in
			 * which case the page is no longer busy).
			 */

			if (result != VM_PAGER_PEND) {
				if (ptmp->flags & PG_WANTED) {
					/* still holding object lock */
					wakeup(ptmp);
				}
				ptmp->flags &= ~(PG_WANTED|PG_BUSY);
				UVM_PAGE_OWN(ptmp, NULL);
				if (ptmp->flags & PG_RELEASED) {
					uvm_unlock_pageq();
					if (!uvn_releasepg(ptmp, NULL)) {
						UVMHIST_LOG(maphist,
							    "released %p",
							    ptmp, 0,0,0);
						return (TRUE);
					}
					uvm_lock_pageq();
					continue;
				} else {
					if ((flags & PGO_WEAK) == 0 &&
					    !(result == VM_PAGER_ERROR &&
					      curproc == uvm.pagedaemon_proc)) {
						ptmp->flags |=
							(PG_CLEAN|PG_CLEANCHK);
						if ((flags & PGO_FREE) == 0) {
							pmap_clear_modify(ptmp);
						}
					}
				}
			}
	  
			/*
			 * dispose of page
			 */

			if (flags & PGO_DEACTIVATE) {
				if ((pp->pqflags & PQ_INACTIVE) == 0 &&
				    pp->wire_count == 0) {
					pmap_page_protect(ptmp, VM_PROT_NONE);
					uvm_pagedeactivate(ptmp);
				}
			} else if (flags & PGO_FREE) {
				if (result == VM_PAGER_PEND) {
					if ((ptmp->flags & PG_BUSY) != 0)
						/* signal for i/o done */
						ptmp->flags |= PG_RELEASED;
				} else {
					if (result != VM_PAGER_OK) {
						printf("uvn_flush: obj=%p, "
						   "offset=0x%llx.  error %d\n",
						    pp->uobject,
						    (long long)pp->offset,
						    result);
						printf("uvn_flush: WARNING: "
						    "changes to page may be "
						    "lost!\n");
						retval = FALSE;
					}
					pmap_page_protect(ptmp, VM_PROT_NONE);
					uvm_pagefree(ptmp);
				}
			}
		}		/* end of "lcv" for loop */
	}		/* end of "pp" for loop */

	uvm_unlock_pageq();
	if ((flags & PGO_CLEANIT) && all && wasclean &&
	    LIST_FIRST(&vp->v_dirtyblkhd) == NULL &&
	    (vp->v_flag & VONWORKLST)) {
		vp->v_flag &= ~VONWORKLST;
		LIST_REMOVE(vp, v_synclist);
	}
	if (need_iosync) {
		UVMHIST_LOG(maphist,"  <<DOING IOSYNC>>",0,0,0,0);

		/*
		 * XXX this doesn't use the new two-flag scheme,
		 * but to use that, all i/o initiators will have to change.
		 */

		s = splbio();
		while (vp->v_numoutput != 0) {
			UVMHIST_LOG(ubchist, "waiting for vp %p num %d",
				    vp, vp->v_numoutput,0,0);

			vp->v_flag |= VBWAIT;
			UVM_UNLOCK_AND_WAIT(&vp->v_numoutput,
					    &uvn->u_obj.vmobjlock, 
					    FALSE, "uvn_flush",0);
			simple_lock(&uvn->u_obj.vmobjlock);
		}
		splx(s);
	}

	/* return, with object locked! */
	UVMHIST_LOG(maphist,"<- done (retval=0x%x)",retval,0,0,0);
	return(retval);
}

/*
 * uvn_cluster
 *
 * we are about to do I/O in an object at offset.   this function is called
 * to establish a range of offsets around "offset" in which we can cluster
 * I/O.
 *
 * - currently doesn't matter if obj locked or not.
 */

static void
uvn_cluster(uobj, offset, loffset, hoffset)
	struct uvm_object *uobj;
	voff_t offset;
	voff_t *loffset, *hoffset; /* OUT */
{
	struct uvm_vnode *uvn = (struct uvm_vnode *)uobj;

	*loffset = offset;
	*hoffset = min(offset + MAXBSIZE, round_page(uvn->u_size));
}

/*
 * uvn_put: flush page data to backing store.
 *
 * => object must be locked!   we will _unlock_ it before starting I/O.
 * => flags: PGO_SYNCIO -- use sync. I/O
 * => note: caller must set PG_CLEAN and pmap_clear_modify (if needed)
 */

static int
uvn_put(uobj, pps, npages, flags)
	struct uvm_object *uobj;
	struct vm_page **pps;
	int npages, flags;
{
	struct vnode *vp = (struct vnode *)uobj;
	int error;

	error = VOP_PUTPAGES(vp, pps, npages, flags, NULL);
	return uvm_errno2vmerror(error);
}


/*
 * uvn_get: get pages (synchronously) from backing store
 *
 * => 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
uvn_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags)
	struct uvm_object *uobj;
	voff_t offset;
	struct vm_page **pps;		/* IN/OUT */
	int *npagesp;			/* IN (OUT if PGO_LOCKED) */
	int centeridx;
	vm_prot_t access_type;
	int advice, flags;
{
	struct vnode *vp = (struct vnode *)uobj;
	int error;
	UVMHIST_FUNC("uvn_get"); UVMHIST_CALLED(ubchist);

	UVMHIST_LOG(ubchist, "vp %p off 0x%x", vp, (int)offset, 0,0);
	error = VOP_GETPAGES(vp, offset, pps, npagesp, centeridx,
			     access_type, advice, flags);
	return uvm_errno2vmerror(error);
}
#endif /*OSKIT*/

/*
 * uvn_findpages:
 * return the page for the uobj and offset requested, allocating if needed.
 * => uobj must be locked.
 * => returned page will be BUSY.
 */

void
uvn_findpages(uobj, offset, npagesp, pps, flags)
	struct uvm_object *uobj;
	voff_t offset;
	int *npagesp;
	struct vm_page **pps;
	int flags;
{
	int i, rv, npages;

	rv = 0;
	npages = *npagesp;
	for (i = 0; i < npages; i++, offset += PAGE_SIZE) {
		rv += uvn_findpage(uobj, offset, &pps[i], flags);
	}
	*npagesp = rv;
}

static int
uvn_findpage(uobj, offset, pgp, flags)
	struct uvm_object *uobj;
	voff_t offset;
	struct vm_page **pgp;
	int flags;
{
	struct vm_page *pg;
	UVMHIST_FUNC("uvn_findpage"); UVMHIST_CALLED(ubchist);
	UVMHIST_LOG(ubchist, "vp %p off 0x%lx", uobj, offset,0,0);

	if (*pgp != NULL) {
		UVMHIST_LOG(ubchist, "dontcare", 0,0,0,0);
		return 0;
	}
	for (;;) {
		/* look for an existing page */
		pg = uvm_pagelookup(uobj, offset);

		/* nope?   allocate one now */
		if (pg == NULL) {
			if (flags & UFP_NOALLOC) {
				UVMHIST_LOG(ubchist, "noalloc", 0,0,0,0);
				return 0;
			}
#ifndef OSKIT			/* XXX??? */
			if (uvmexp.vnodepages > 
			    (uvmexp.active + uvmexp.inactive + uvmexp.wired +
			     uvmexp.free) * 7 / 8) {
				pg = NULL;
			} else 
#endif
			{
				pg = uvm_pagealloc(uobj, offset, NULL, 0);
			}
			if (pg == NULL) {
				if (flags & UFP_NOWAIT) {
					UVMHIST_LOG(ubchist, "nowait",0,0,0,0);
					return 0;
				}
				simple_unlock(&uobj->vmobjlock);
				uvm_wait("uvn_fp1");
				simple_lock(&uobj->vmobjlock);
				continue;
			}
			uvmexp.vnodepages++;
			UVMHIST_LOG(ubchist, "alloced",0,0,0,0);
			break;
		} else if (flags & UFP_NOCACHE) {
			UVMHIST_LOG(ubchist, "nocache",0,0,0,0);
			return 0;
		}

		/* page is there, see if we need to wait on it */
		if ((pg->flags & (PG_BUSY|PG_RELEASED)) != 0) {
			if (flags & UFP_NOWAIT) {
				UVMHIST_LOG(ubchist, "nowait",0,0,0,0);
				return 0;
			}
			pg->flags |= PG_WANTED;
			UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0,
					    "uvn_fp2", 0);
			simple_lock(&uobj->vmobjlock);
			continue;
		}
			
		/* skip PG_RDONLY pages if requested */
		if ((flags & UFP_NORDONLY) && (pg->flags & PG_RDONLY)) {
			UVMHIST_LOG(ubchist, "nordonly",0,0,0,0);
			return 0;
		}

		/* mark the page BUSY and we're done. */
		pg->flags |= PG_BUSY;
		UVM_PAGE_OWN(pg, "uvn_findpage");
		UVMHIST_LOG(ubchist, "found",0,0,0,0);
		break;
	}
	*pgp = pg;
	return 1;
}

#ifndef OSKIT
/*
 * uvm_vnp_setsize: grow or shrink a vnode uvn
 *
 * grow   => just update size value
 * shrink => toss un-needed pages
 *
 * => we assume that the caller has a reference of some sort to the 
 *	vnode in question so that it will not be yanked out from under
 *	us.
 *
 * called from:
 *  => truncate fns (ext2fs_truncate, ffs_truncate, detrunc[msdos])
 *  => "write" fns (ext2fs_write, WRITE [ufs/ufs], msdosfs_write, nfs_write)
 *  => ffs_balloc [XXX: why? doesn't WRITE handle?]
 *  => NFS: nfs_loadattrcache, nfs_getattrcache, nfs_setattr
 *  => union fs: union_newsize
 */

void
uvm_vnp_setsize(vp, newsize)
	struct vnode *vp;
	voff_t newsize;
{
	struct uvm_vnode *uvn = &vp->v_uvm;
	UVMHIST_FUNC("uvm_vnp_setsize"); UVMHIST_CALLED(ubchist);

	simple_lock(&uvn->u_obj.vmobjlock);

	UVMHIST_LOG(ubchist, "old 0x%x new 0x%x", uvn->u_size, newsize, 0,0);

	/*
	 * now check if the size has changed: if we shrink we had better
	 * toss some pages...
	 */

	if (uvn->u_size > newsize && uvn->u_size != VSIZENOTSET) {
		(void) uvn_flush(&uvn->u_obj, newsize, uvn->u_size, PGO_FREE);
	}
	uvn->u_size = newsize;
	simple_unlock(&uvn->u_obj.vmobjlock);
}
#endif /*OSKIT*/

/*
 * uvm_vnp_zerorange:  set a range of bytes in a file to zero.
 */

#ifndef OSKIT
void
uvm_vnp_zerorange(vp, off, len)
	struct vnode *vp;
	off_t off;
	size_t len;
{
        void *win;

        /*
         * XXXUBC invent kzero() and use it
         */

        while (len) {
                vsize_t bytelen = len;

                win = ubc_alloc(&vp->v_uvm.u_obj, off, &bytelen, UBC_WRITE);
                memset(win, 0, bytelen);
                ubc_release(win, 0);

                off += bytelen;
                len -= bytelen;
        }
}
#endif /*OSKIT*/