Index: head/sys/fs/nfsclient/nfs_clbio.c =================================================================== --- head/sys/fs/nfsclient/nfs_clbio.c (revision 222585) +++ head/sys/fs/nfsclient/nfs_clbio.c (revision 222586) @@ -1,1823 +1,1816 @@ /*- * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * Rick Macklem at The University of Guelph. * * 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. * 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. * * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include extern int newnfs_directio_allow_mmap; extern struct nfsstats newnfsstats; extern struct mtx ncl_iod_mutex; extern int ncl_numasync; extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON]; extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON]; extern int newnfs_directio_enable; int ncl_pbuf_freecnt = -1; /* start out unlimited */ static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td); static int nfs_directio_write(struct vnode *vp, struct uio *uiop, struct ucred *cred, int ioflag); /* * Vnode op for VM getpages. */ int ncl_getpages(struct vop_getpages_args *ap) { int i, error, nextoff, size, toff, count, npages; struct uio uio; struct iovec iov; vm_offset_t kva; struct buf *bp; struct vnode *vp; struct thread *td; struct ucred *cred; struct nfsmount *nmp; vm_object_t object; vm_page_t *pages; struct nfsnode *np; vp = ap->a_vp; np = VTONFS(vp); td = curthread; /* XXX */ cred = curthread->td_ucred; /* XXX */ nmp = VFSTONFS(vp->v_mount); pages = ap->a_m; count = ap->a_count; if ((object = vp->v_object) == NULL) { ncl_printf("nfs_getpages: called with non-merged cache vnode??\n"); return (VM_PAGER_ERROR); } if (newnfs_directio_enable && !newnfs_directio_allow_mmap) { mtx_lock(&np->n_mtx); if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { mtx_unlock(&np->n_mtx); ncl_printf("nfs_getpages: called on non-cacheable vnode??\n"); return (VM_PAGER_ERROR); } else mtx_unlock(&np->n_mtx); } mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); /* We'll never get here for v4, because we always have fsinfo */ (void)ncl_fsinfo(nmp, vp, cred, td); } else mtx_unlock(&nmp->nm_mtx); npages = btoc(count); /* * If the requested page is partially valid, just return it and * allow the pager to zero-out the blanks. Partially valid pages * can only occur at the file EOF. */ VM_OBJECT_LOCK(object); if (pages[ap->a_reqpage]->valid != 0) { for (i = 0; i < npages; ++i) { if (i != ap->a_reqpage) { vm_page_lock(pages[i]); vm_page_free(pages[i]); vm_page_unlock(pages[i]); } } VM_OBJECT_UNLOCK(object); return (0); } VM_OBJECT_UNLOCK(object); /* * We use only the kva address for the buffer, but this is extremely * convienient and fast. */ bp = getpbuf(&ncl_pbuf_freecnt); kva = (vm_offset_t) bp->b_data; pmap_qenter(kva, pages, npages); PCPU_INC(cnt.v_vnodein); PCPU_ADD(cnt.v_vnodepgsin, npages); iov.iov_base = (caddr_t) kva; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); uio.uio_resid = count; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_READ; uio.uio_td = td; error = ncl_readrpc(vp, &uio, cred); pmap_qremove(kva, npages); relpbuf(bp, &ncl_pbuf_freecnt); if (error && (uio.uio_resid == count)) { ncl_printf("nfs_getpages: error %d\n", error); VM_OBJECT_LOCK(object); for (i = 0; i < npages; ++i) { if (i != ap->a_reqpage) { vm_page_lock(pages[i]); vm_page_free(pages[i]); vm_page_unlock(pages[i]); } } VM_OBJECT_UNLOCK(object); return (VM_PAGER_ERROR); } /* * Calculate the number of bytes read and validate only that number * of bytes. Note that due to pending writes, size may be 0. This * does not mean that the remaining data is invalid! */ size = count - uio.uio_resid; VM_OBJECT_LOCK(object); for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { vm_page_t m; nextoff = toff + PAGE_SIZE; m = pages[i]; if (nextoff <= size) { /* * Read operation filled an entire page */ m->valid = VM_PAGE_BITS_ALL; KASSERT(m->dirty == 0, ("nfs_getpages: page %p is dirty", m)); } else if (size > toff) { /* * Read operation filled a partial page. */ m->valid = 0; vm_page_set_valid(m, 0, size - toff); KASSERT(m->dirty == 0, ("nfs_getpages: page %p is dirty", m)); } else { /* * Read operation was short. If no error occured * we may have hit a zero-fill section. We simply * leave valid set to 0. */ ; } if (i != ap->a_reqpage) { /* * Whether or not to leave the page activated is up in * the air, but we should put the page on a page queue * somewhere (it already is in the object). Result: * It appears that emperical results show that * deactivating pages is best. */ /* * Just in case someone was asking for this page we * now tell them that it is ok to use. */ if (!error) { if (m->oflags & VPO_WANTED) { vm_page_lock(m); vm_page_activate(m); vm_page_unlock(m); } else { vm_page_lock(m); vm_page_deactivate(m); vm_page_unlock(m); } vm_page_wakeup(m); } else { vm_page_lock(m); vm_page_free(m); vm_page_unlock(m); } } } VM_OBJECT_UNLOCK(object); return (0); } /* * Vnode op for VM putpages. */ int ncl_putpages(struct vop_putpages_args *ap) { struct uio uio; struct iovec iov; vm_offset_t kva; struct buf *bp; int iomode, must_commit, i, error, npages, count; off_t offset; int *rtvals; struct vnode *vp; struct thread *td; struct ucred *cred; struct nfsmount *nmp; struct nfsnode *np; vm_page_t *pages; vp = ap->a_vp; np = VTONFS(vp); td = curthread; /* XXX */ cred = curthread->td_ucred; /* XXX */ nmp = VFSTONFS(vp->v_mount); pages = ap->a_m; count = ap->a_count; rtvals = ap->a_rtvals; npages = btoc(count); offset = IDX_TO_OFF(pages[0]->pindex); mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); (void)ncl_fsinfo(nmp, vp, cred, td); } else mtx_unlock(&nmp->nm_mtx); mtx_lock(&np->n_mtx); if (newnfs_directio_enable && !newnfs_directio_allow_mmap && (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { mtx_unlock(&np->n_mtx); ncl_printf("ncl_putpages: called on noncache-able vnode??\n"); mtx_lock(&np->n_mtx); } for (i = 0; i < npages; i++) - rtvals[i] = VM_PAGER_AGAIN; + rtvals[i] = VM_PAGER_ERROR; /* * When putting pages, do not extend file past EOF. */ if (offset + count > np->n_size) { count = np->n_size - offset; if (count < 0) count = 0; } mtx_unlock(&np->n_mtx); /* * We use only the kva address for the buffer, but this is extremely * convienient and fast. */ bp = getpbuf(&ncl_pbuf_freecnt); kva = (vm_offset_t) bp->b_data; pmap_qenter(kva, pages, npages); PCPU_INC(cnt.v_vnodeout); PCPU_ADD(cnt.v_vnodepgsout, count); iov.iov_base = (caddr_t) kva; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = offset; uio.uio_resid = count; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_WRITE; uio.uio_td = td; if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) iomode = NFSWRITE_UNSTABLE; else iomode = NFSWRITE_FILESYNC; error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0); pmap_qremove(kva, npages); relpbuf(bp, &ncl_pbuf_freecnt); - if (!error) { - int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; - for (i = 0; i < nwritten; i++) { - rtvals[i] = VM_PAGER_OK; - vm_page_undirty(pages[i]); - } - if (must_commit) { - ncl_clearcommit(vp->v_mount); - } - } + vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid); + if (must_commit) + ncl_clearcommit(vp->v_mount); return rtvals[0]; } /* * For nfs, cache consistency can only be maintained approximately. * Although RFC1094 does not specify the criteria, the following is * believed to be compatible with the reference port. * For nfs: * If the file's modify time on the server has changed since the * last read rpc or you have written to the file, * you may have lost data cache consistency with the * server, so flush all of the file's data out of the cache. * Then force a getattr rpc to ensure that you have up to date * attributes. * NB: This implies that cache data can be read when up to * NFS_ATTRTIMEO seconds out of date. If you find that you need current * attributes this could be forced by setting n_attrstamp to 0 before * the VOP_GETATTR() call. */ static inline int nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) { int error = 0; struct vattr vattr; struct nfsnode *np = VTONFS(vp); int old_lock; /* * Grab the exclusive lock before checking whether the cache is * consistent. * XXX - We can make this cheaper later (by acquiring cheaper locks). * But for now, this suffices. */ old_lock = ncl_upgrade_vnlock(vp); if (vp->v_iflag & VI_DOOMED) { ncl_downgrade_vnlock(vp, old_lock); return (EBADF); } mtx_lock(&np->n_mtx); if (np->n_flag & NMODIFIED) { mtx_unlock(&np->n_mtx); if (vp->v_type != VREG) { if (vp->v_type != VDIR) panic("nfs: bioread, not dir"); ncl_invaldir(vp); error = ncl_vinvalbuf(vp, V_SAVE, td, 1); if (error) goto out; } np->n_attrstamp = 0; error = VOP_GETATTR(vp, &vattr, cred); if (error) goto out; mtx_lock(&np->n_mtx); np->n_mtime = vattr.va_mtime; mtx_unlock(&np->n_mtx); } else { mtx_unlock(&np->n_mtx); error = VOP_GETATTR(vp, &vattr, cred); if (error) return (error); mtx_lock(&np->n_mtx); if ((np->n_flag & NSIZECHANGED) || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { mtx_unlock(&np->n_mtx); if (vp->v_type == VDIR) ncl_invaldir(vp); error = ncl_vinvalbuf(vp, V_SAVE, td, 1); if (error) goto out; mtx_lock(&np->n_mtx); np->n_mtime = vattr.va_mtime; np->n_flag &= ~NSIZECHANGED; } mtx_unlock(&np->n_mtx); } out: ncl_downgrade_vnlock(vp, old_lock); return error; } /* * Vnode op for read using bio */ int ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) { struct nfsnode *np = VTONFS(vp); int biosize, i; struct buf *bp, *rabp; struct thread *td; struct nfsmount *nmp = VFSTONFS(vp->v_mount); daddr_t lbn, rabn; int bcount; int seqcount; int nra, error = 0, n = 0, on = 0; off_t tmp_off; KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode")); if (uio->uio_resid == 0) return (0); if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ return (EINVAL); td = uio->uio_td; mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); (void)ncl_fsinfo(nmp, vp, cred, td); mtx_lock(&nmp->nm_mtx); } if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0) (void) newnfs_iosize(nmp); tmp_off = uio->uio_offset + uio->uio_resid; if (vp->v_type != VDIR && (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) { mtx_unlock(&nmp->nm_mtx); return (EFBIG); } mtx_unlock(&nmp->nm_mtx); if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) /* No caching/ no readaheads. Just read data into the user buffer */ return ncl_readrpc(vp, uio, cred); biosize = vp->v_mount->mnt_stat.f_iosize; seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); error = nfs_bioread_check_cons(vp, td, cred); if (error) return error; do { u_quad_t nsize; mtx_lock(&np->n_mtx); nsize = np->n_size; mtx_unlock(&np->n_mtx); switch (vp->v_type) { case VREG: NFSINCRGLOBAL(newnfsstats.biocache_reads); lbn = uio->uio_offset / biosize; on = uio->uio_offset & (biosize - 1); /* * Start the read ahead(s), as required. */ if (nmp->nm_readahead > 0) { for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { rabn = lbn + 1 + nra; if (incore(&vp->v_bufobj, rabn) == NULL) { rabp = nfs_getcacheblk(vp, rabn, biosize, td); if (!rabp) { error = newnfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { rabp->b_flags |= B_ASYNC; rabp->b_iocmd = BIO_READ; vfs_busy_pages(rabp, 0); if (ncl_asyncio(nmp, rabp, cred, td)) { rabp->b_flags |= B_INVAL; rabp->b_ioflags |= BIO_ERROR; vfs_unbusy_pages(rabp); brelse(rabp); break; } } else { brelse(rabp); } } } } /* Note that bcount is *not* DEV_BSIZE aligned. */ bcount = biosize; if ((off_t)lbn * biosize >= nsize) { bcount = 0; } else if ((off_t)(lbn + 1) * biosize > nsize) { bcount = nsize - (off_t)lbn * biosize; } bp = nfs_getcacheblk(vp, lbn, bcount, td); if (!bp) { error = newnfs_sigintr(nmp, td); return (error ? error : EINTR); } /* * If B_CACHE is not set, we must issue the read. If this * fails, we return an error. */ if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = ncl_doio(vp, bp, cred, td, 0); if (error) { brelse(bp); return (error); } } /* * on is the offset into the current bp. Figure out how many * bytes we can copy out of the bp. Note that bcount is * NOT DEV_BSIZE aligned. * * Then figure out how many bytes we can copy into the uio. */ n = 0; if (on < bcount) n = min((unsigned)(bcount - on), uio->uio_resid); break; case VLNK: NFSINCRGLOBAL(newnfsstats.biocache_readlinks); bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); if (!bp) { error = newnfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = ncl_doio(vp, bp, cred, td, 0); if (error) { bp->b_ioflags |= BIO_ERROR; brelse(bp); return (error); } } n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); on = 0; break; case VDIR: NFSINCRGLOBAL(newnfsstats.biocache_readdirs); if (np->n_direofoffset && uio->uio_offset >= np->n_direofoffset) { return (0); } lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); if (!bp) { error = newnfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = ncl_doio(vp, bp, cred, td, 0); if (error) { brelse(bp); } while (error == NFSERR_BAD_COOKIE) { ncl_invaldir(vp); error = ncl_vinvalbuf(vp, 0, td, 1); /* * Yuck! The directory has been modified on the * server. The only way to get the block is by * reading from the beginning to get all the * offset cookies. * * Leave the last bp intact unless there is an error. * Loop back up to the while if the error is another * NFSERR_BAD_COOKIE (double yuch!). */ for (i = 0; i <= lbn && !error; i++) { if (np->n_direofoffset && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) return (0); bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); if (!bp) { error = newnfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = ncl_doio(vp, bp, cred, td, 0); /* * no error + B_INVAL == directory EOF, * use the block. */ if (error == 0 && (bp->b_flags & B_INVAL)) break; } /* * An error will throw away the block and the * for loop will break out. If no error and this * is not the block we want, we throw away the * block and go for the next one via the for loop. */ if (error || i < lbn) brelse(bp); } } /* * The above while is repeated if we hit another cookie * error. If we hit an error and it wasn't a cookie error, * we give up. */ if (error) return (error); } /* * If not eof and read aheads are enabled, start one. * (You need the current block first, so that you have the * directory offset cookie of the next block.) */ if (nmp->nm_readahead > 0 && (bp->b_flags & B_INVAL) == 0 && (np->n_direofoffset == 0 || (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && incore(&vp->v_bufobj, lbn + 1) == NULL) { rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); if (rabp) { if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { rabp->b_flags |= B_ASYNC; rabp->b_iocmd = BIO_READ; vfs_busy_pages(rabp, 0); if (ncl_asyncio(nmp, rabp, cred, td)) { rabp->b_flags |= B_INVAL; rabp->b_ioflags |= BIO_ERROR; vfs_unbusy_pages(rabp); brelse(rabp); } } else { brelse(rabp); } } } /* * Unlike VREG files, whos buffer size ( bp->b_bcount ) is * chopped for the EOF condition, we cannot tell how large * NFS directories are going to be until we hit EOF. So * an NFS directory buffer is *not* chopped to its EOF. Now, * it just so happens that b_resid will effectively chop it * to EOF. *BUT* this information is lost if the buffer goes * away and is reconstituted into a B_CACHE state ( due to * being VMIO ) later. So we keep track of the directory eof * in np->n_direofoffset and chop it off as an extra step * right here. */ n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) n = np->n_direofoffset - uio->uio_offset; break; default: ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type); bp = NULL; break; }; if (n > 0) { error = uiomove(bp->b_data + on, (int)n, uio); } if (vp->v_type == VLNK) n = 0; if (bp != NULL) brelse(bp); } while (error == 0 && uio->uio_resid > 0 && n > 0); return (error); } /* * The NFS write path cannot handle iovecs with len > 1. So we need to * break up iovecs accordingly (restricting them to wsize). * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). * For the ASYNC case, 2 copies are needed. The first a copy from the * user buffer to a staging buffer and then a second copy from the staging * buffer to mbufs. This can be optimized by copying from the user buffer * directly into mbufs and passing the chain down, but that requires a * fair amount of re-working of the relevant codepaths (and can be done * later). */ static int nfs_directio_write(vp, uiop, cred, ioflag) struct vnode *vp; struct uio *uiop; struct ucred *cred; int ioflag; { int error; struct nfsmount *nmp = VFSTONFS(vp->v_mount); struct thread *td = uiop->uio_td; int size; int wsize; mtx_lock(&nmp->nm_mtx); wsize = nmp->nm_wsize; mtx_unlock(&nmp->nm_mtx); if (ioflag & IO_SYNC) { int iomode, must_commit; struct uio uio; struct iovec iov; do_sync: while (uiop->uio_resid > 0) { size = min(uiop->uio_resid, wsize); size = min(uiop->uio_iov->iov_len, size); iov.iov_base = uiop->uio_iov->iov_base; iov.iov_len = size; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = uiop->uio_offset; uio.uio_resid = size; uio.uio_segflg = UIO_USERSPACE; uio.uio_rw = UIO_WRITE; uio.uio_td = td; iomode = NFSWRITE_FILESYNC; error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0); KASSERT((must_commit == 0), ("ncl_directio_write: Did not commit write")); if (error) return (error); uiop->uio_offset += size; uiop->uio_resid -= size; if (uiop->uio_iov->iov_len <= size) { uiop->uio_iovcnt--; uiop->uio_iov++; } else { uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + size; uiop->uio_iov->iov_len -= size; } } } else { struct uio *t_uio; struct iovec *t_iov; struct buf *bp; /* * Break up the write into blocksize chunks and hand these * over to nfsiod's for write back. * Unfortunately, this incurs a copy of the data. Since * the user could modify the buffer before the write is * initiated. * * The obvious optimization here is that one of the 2 copies * in the async write path can be eliminated by copying the * data here directly into mbufs and passing the mbuf chain * down. But that will require a fair amount of re-working * of the code and can be done if there's enough interest * in NFS directio access. */ while (uiop->uio_resid > 0) { size = min(uiop->uio_resid, wsize); size = min(uiop->uio_iov->iov_len, size); bp = getpbuf(&ncl_pbuf_freecnt); t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); t_iov->iov_len = size; t_uio->uio_iov = t_iov; t_uio->uio_iovcnt = 1; t_uio->uio_offset = uiop->uio_offset; t_uio->uio_resid = size; t_uio->uio_segflg = UIO_SYSSPACE; t_uio->uio_rw = UIO_WRITE; t_uio->uio_td = td; bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); bp->b_flags |= B_DIRECT; bp->b_iocmd = BIO_WRITE; if (cred != NOCRED) { crhold(cred); bp->b_wcred = cred; } else bp->b_wcred = NOCRED; bp->b_caller1 = (void *)t_uio; bp->b_vp = vp; error = ncl_asyncio(nmp, bp, NOCRED, td); if (error) { free(t_iov->iov_base, M_NFSDIRECTIO); free(t_iov, M_NFSDIRECTIO); free(t_uio, M_NFSDIRECTIO); bp->b_vp = NULL; relpbuf(bp, &ncl_pbuf_freecnt); if (error == EINTR) return (error); goto do_sync; } uiop->uio_offset += size; uiop->uio_resid -= size; if (uiop->uio_iov->iov_len <= size) { uiop->uio_iovcnt--; uiop->uio_iov++; } else { uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + size; uiop->uio_iov->iov_len -= size; } } } return (0); } /* * Vnode op for write using bio */ int ncl_write(struct vop_write_args *ap) { int biosize; struct uio *uio = ap->a_uio; struct thread *td = uio->uio_td; struct vnode *vp = ap->a_vp; struct nfsnode *np = VTONFS(vp); struct ucred *cred = ap->a_cred; int ioflag = ap->a_ioflag; struct buf *bp; struct vattr vattr; struct nfsmount *nmp = VFSTONFS(vp->v_mount); daddr_t lbn; int bcount; int n, on, error = 0; off_t tmp_off; KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode")); KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, ("ncl_write proc")); if (vp->v_type != VREG) return (EIO); mtx_lock(&np->n_mtx); if (np->n_flag & NWRITEERR) { np->n_flag &= ~NWRITEERR; mtx_unlock(&np->n_mtx); return (np->n_error); } else mtx_unlock(&np->n_mtx); mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); (void)ncl_fsinfo(nmp, vp, cred, td); mtx_lock(&nmp->nm_mtx); } if (nmp->nm_wsize == 0) (void) newnfs_iosize(nmp); mtx_unlock(&nmp->nm_mtx); /* * Synchronously flush pending buffers if we are in synchronous * mode or if we are appending. */ if (ioflag & (IO_APPEND | IO_SYNC)) { mtx_lock(&np->n_mtx); if (np->n_flag & NMODIFIED) { mtx_unlock(&np->n_mtx); #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ /* * Require non-blocking, synchronous writes to * dirty files to inform the program it needs * to fsync(2) explicitly. */ if (ioflag & IO_NDELAY) return (EAGAIN); #endif flush_and_restart: np->n_attrstamp = 0; error = ncl_vinvalbuf(vp, V_SAVE, td, 1); if (error) return (error); } else mtx_unlock(&np->n_mtx); } /* * If IO_APPEND then load uio_offset. We restart here if we cannot * get the append lock. */ if (ioflag & IO_APPEND) { np->n_attrstamp = 0; error = VOP_GETATTR(vp, &vattr, cred); if (error) return (error); mtx_lock(&np->n_mtx); uio->uio_offset = np->n_size; mtx_unlock(&np->n_mtx); } if (uio->uio_offset < 0) return (EINVAL); tmp_off = uio->uio_offset + uio->uio_resid; if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset) return (EFBIG); if (uio->uio_resid == 0) return (0); if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) return nfs_directio_write(vp, uio, cred, ioflag); /* * Maybe this should be above the vnode op call, but so long as * file servers have no limits, i don't think it matters */ if (vn_rlimit_fsize(vp, uio, td)) return (EFBIG); biosize = vp->v_mount->mnt_stat.f_iosize; /* * Find all of this file's B_NEEDCOMMIT buffers. If our writes * would exceed the local maximum per-file write commit size when * combined with those, we must decide whether to flush, * go synchronous, or return error. We don't bother checking * IO_UNIT -- we just make all writes atomic anyway, as there's * no point optimizing for something that really won't ever happen. */ if (!(ioflag & IO_SYNC)) { int nflag; mtx_lock(&np->n_mtx); nflag = np->n_flag; mtx_unlock(&np->n_mtx); int needrestart = 0; if (nmp->nm_wcommitsize < uio->uio_resid) { /* * If this request could not possibly be completed * without exceeding the maximum outstanding write * commit size, see if we can convert it into a * synchronous write operation. */ if (ioflag & IO_NDELAY) return (EAGAIN); ioflag |= IO_SYNC; if (nflag & NMODIFIED) needrestart = 1; } else if (nflag & NMODIFIED) { int wouldcommit = 0; BO_LOCK(&vp->v_bufobj); if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { if (bp->b_flags & B_NEEDCOMMIT) wouldcommit += bp->b_bcount; } } BO_UNLOCK(&vp->v_bufobj); /* * Since we're not operating synchronously and * bypassing the buffer cache, we are in a commit * and holding all of these buffers whether * transmitted or not. If not limited, this * will lead to the buffer cache deadlocking, * as no one else can flush our uncommitted buffers. */ wouldcommit += uio->uio_resid; /* * If we would initially exceed the maximum * outstanding write commit size, flush and restart. */ if (wouldcommit > nmp->nm_wcommitsize) needrestart = 1; } if (needrestart) goto flush_and_restart; } do { NFSINCRGLOBAL(newnfsstats.biocache_writes); lbn = uio->uio_offset / biosize; on = uio->uio_offset & (biosize-1); n = min((unsigned)(biosize - on), uio->uio_resid); again: /* * Handle direct append and file extension cases, calculate * unaligned buffer size. */ mtx_lock(&np->n_mtx); if (uio->uio_offset == np->n_size && n) { mtx_unlock(&np->n_mtx); /* * Get the buffer (in its pre-append state to maintain * B_CACHE if it was previously set). Resize the * nfsnode after we have locked the buffer to prevent * readers from reading garbage. */ bcount = on; bp = nfs_getcacheblk(vp, lbn, bcount, td); if (bp != NULL) { long save; mtx_lock(&np->n_mtx); np->n_size = uio->uio_offset + n; np->n_flag |= NMODIFIED; vnode_pager_setsize(vp, np->n_size); mtx_unlock(&np->n_mtx); save = bp->b_flags & B_CACHE; bcount += n; allocbuf(bp, bcount); bp->b_flags |= save; } } else { /* * Obtain the locked cache block first, and then * adjust the file's size as appropriate. */ bcount = on + n; if ((off_t)lbn * biosize + bcount < np->n_size) { if ((off_t)(lbn + 1) * biosize < np->n_size) bcount = biosize; else bcount = np->n_size - (off_t)lbn * biosize; } mtx_unlock(&np->n_mtx); bp = nfs_getcacheblk(vp, lbn, bcount, td); mtx_lock(&np->n_mtx); if (uio->uio_offset + n > np->n_size) { np->n_size = uio->uio_offset + n; np->n_flag |= NMODIFIED; vnode_pager_setsize(vp, np->n_size); } mtx_unlock(&np->n_mtx); } if (!bp) { error = newnfs_sigintr(nmp, td); if (!error) error = EINTR; break; } /* * Issue a READ if B_CACHE is not set. In special-append * mode, B_CACHE is based on the buffer prior to the write * op and is typically set, avoiding the read. If a read * is required in special append mode, the server will * probably send us a short-read since we extended the file * on our end, resulting in b_resid == 0 and, thusly, * B_CACHE getting set. * * We can also avoid issuing the read if the write covers * the entire buffer. We have to make sure the buffer state * is reasonable in this case since we will not be initiating * I/O. See the comments in kern/vfs_bio.c's getblk() for * more information. * * B_CACHE may also be set due to the buffer being cached * normally. */ if (on == 0 && n == bcount) { bp->b_flags |= B_CACHE; bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = ncl_doio(vp, bp, cred, td, 0); if (error) { brelse(bp); break; } } if (bp->b_wcred == NOCRED) bp->b_wcred = crhold(cred); mtx_lock(&np->n_mtx); np->n_flag |= NMODIFIED; mtx_unlock(&np->n_mtx); /* * If dirtyend exceeds file size, chop it down. This should * not normally occur but there is an append race where it * might occur XXX, so we log it. * * If the chopping creates a reverse-indexed or degenerate * situation with dirtyoff/end, we 0 both of them. */ if (bp->b_dirtyend > bcount) { ncl_printf("NFS append race @%lx:%d\n", (long)bp->b_blkno * DEV_BSIZE, bp->b_dirtyend - bcount); bp->b_dirtyend = bcount; } if (bp->b_dirtyoff >= bp->b_dirtyend) bp->b_dirtyoff = bp->b_dirtyend = 0; /* * If the new write will leave a contiguous dirty * area, just update the b_dirtyoff and b_dirtyend, * otherwise force a write rpc of the old dirty area. * * While it is possible to merge discontiguous writes due to * our having a B_CACHE buffer ( and thus valid read data * for the hole), we don't because it could lead to * significant cache coherency problems with multiple clients, * especially if locking is implemented later on. * * as an optimization we could theoretically maintain * a linked list of discontinuous areas, but we would still * have to commit them separately so there isn't much * advantage to it except perhaps a bit of asynchronization. */ if (bp->b_dirtyend > 0 && (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { if (bwrite(bp) == EINTR) { error = EINTR; break; } goto again; } error = uiomove((char *)bp->b_data + on, n, uio); /* * Since this block is being modified, it must be written * again and not just committed. Since write clustering does * not work for the stage 1 data write, only the stage 2 * commit rpc, we have to clear B_CLUSTEROK as well. */ bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); if (error) { bp->b_ioflags |= BIO_ERROR; brelse(bp); break; } /* * Only update dirtyoff/dirtyend if not a degenerate * condition. */ if (n) { if (bp->b_dirtyend > 0) { bp->b_dirtyoff = min(on, bp->b_dirtyoff); bp->b_dirtyend = max((on + n), bp->b_dirtyend); } else { bp->b_dirtyoff = on; bp->b_dirtyend = on + n; } vfs_bio_set_valid(bp, on, n); } /* * If IO_SYNC do bwrite(). * * IO_INVAL appears to be unused. The idea appears to be * to turn off caching in this case. Very odd. XXX */ if ((ioflag & IO_SYNC)) { if (ioflag & IO_INVAL) bp->b_flags |= B_NOCACHE; error = bwrite(bp); if (error) break; } else if ((n + on) == biosize) { bp->b_flags |= B_ASYNC; (void) ncl_writebp(bp, 0, NULL); } else { bdwrite(bp); } } while (uio->uio_resid > 0 && n > 0); return (error); } /* * Get an nfs cache block. * * Allocate a new one if the block isn't currently in the cache * and return the block marked busy. If the calling process is * interrupted by a signal for an interruptible mount point, return * NULL. * * The caller must carefully deal with the possible B_INVAL state of * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it * indirectly), so synchronous reads can be issued without worrying about * the B_INVAL state. We have to be a little more careful when dealing * with writes (see comments in nfs_write()) when extending a file past * its EOF. */ static struct buf * nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) { struct buf *bp; struct mount *mp; struct nfsmount *nmp; mp = vp->v_mount; nmp = VFSTONFS(mp); if (nmp->nm_flag & NFSMNT_INT) { sigset_t oldset; newnfs_set_sigmask(td, &oldset); bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0); newnfs_restore_sigmask(td, &oldset); while (bp == NULL) { if (newnfs_sigintr(nmp, td)) return (NULL); bp = getblk(vp, bn, size, 0, 2 * hz, 0); } } else { bp = getblk(vp, bn, size, 0, 0, 0); } if (vp->v_type == VREG) { int biosize; biosize = mp->mnt_stat.f_iosize; bp->b_blkno = bn * (biosize / DEV_BSIZE); } return (bp); } /* * Flush and invalidate all dirty buffers. If another process is already * doing the flush, just wait for completion. */ int ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) { struct nfsnode *np = VTONFS(vp); struct nfsmount *nmp = VFSTONFS(vp->v_mount); int error = 0, slpflag, slptimeo; int old_lock = 0; ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf"); if ((nmp->nm_flag & NFSMNT_INT) == 0) intrflg = 0; if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)) intrflg = 1; if (intrflg) { slpflag = NFS_PCATCH; slptimeo = 2 * hz; } else { slpflag = 0; slptimeo = 0; } old_lock = ncl_upgrade_vnlock(vp); if (vp->v_iflag & VI_DOOMED) { /* * Since vgonel() uses the generic vinvalbuf() to flush * dirty buffers and it does not call this function, it * is safe to just return OK when VI_DOOMED is set. */ ncl_downgrade_vnlock(vp, old_lock); return (0); } /* * Now, flush as required. */ if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { VM_OBJECT_LOCK(vp->v_bufobj.bo_object); vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); /* * If the page clean was interrupted, fail the invalidation. * Not doing so, we run the risk of losing dirty pages in the * vinvalbuf() call below. */ if (intrflg && (error = newnfs_sigintr(nmp, td))) goto out; } error = vinvalbuf(vp, flags, slpflag, 0); while (error) { if (intrflg && (error = newnfs_sigintr(nmp, td))) goto out; error = vinvalbuf(vp, flags, 0, slptimeo); } mtx_lock(&np->n_mtx); if (np->n_directio_asyncwr == 0) np->n_flag &= ~NMODIFIED; mtx_unlock(&np->n_mtx); out: ncl_downgrade_vnlock(vp, old_lock); return error; } /* * Initiate asynchronous I/O. Return an error if no nfsiods are available. * This is mainly to avoid queueing async I/O requests when the nfsiods * are all hung on a dead server. * * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp * is eventually dequeued by the async daemon, ncl_doio() *will*. */ int ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) { int iod; int gotiod; int slpflag = 0; int slptimeo = 0; int error, error2; /* * Commits are usually short and sweet so lets save some cpu and * leave the async daemons for more important rpc's (such as reads * and writes). */ mtx_lock(&ncl_iod_mutex); if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && (nmp->nm_bufqiods > ncl_numasync / 2)) { mtx_unlock(&ncl_iod_mutex); return(EIO); } again: if (nmp->nm_flag & NFSMNT_INT) slpflag = NFS_PCATCH; gotiod = FALSE; /* * Find a free iod to process this request. */ for (iod = 0; iod < ncl_numasync; iod++) if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) { gotiod = TRUE; break; } /* * Try to create one if none are free. */ if (!gotiod) ncl_nfsiodnew(); else { /* * Found one, so wake it up and tell it which * mount to process. */ NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n", iod, nmp)); ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE; ncl_iodmount[iod] = nmp; nmp->nm_bufqiods++; wakeup(&ncl_iodwant[iod]); } /* * If none are free, we may already have an iod working on this mount * point. If so, it will process our request. */ if (!gotiod) { if (nmp->nm_bufqiods > 0) { NFS_DPF(ASYNCIO, ("ncl_asyncio: %d iods are already processing mount %p\n", nmp->nm_bufqiods, nmp)); gotiod = TRUE; } } /* * If we have an iod which can process the request, then queue * the buffer. */ if (gotiod) { /* * Ensure that the queue never grows too large. We still want * to asynchronize so we block rather then return EIO. */ while (nmp->nm_bufqlen >= 2*ncl_numasync) { NFS_DPF(ASYNCIO, ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp)); nmp->nm_bufqwant = TRUE; error = newnfs_msleep(td, &nmp->nm_bufq, &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio", slptimeo); if (error) { error2 = newnfs_sigintr(nmp, td); if (error2) { mtx_unlock(&ncl_iod_mutex); return (error2); } if (slpflag == NFS_PCATCH) { slpflag = 0; slptimeo = 2 * hz; } } /* * We might have lost our iod while sleeping, * so check and loop if nescessary. */ goto again; } /* We might have lost our nfsiod */ if (nmp->nm_bufqiods == 0) { NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); goto again; } if (bp->b_iocmd == BIO_READ) { if (bp->b_rcred == NOCRED && cred != NOCRED) bp->b_rcred = crhold(cred); } else { if (bp->b_wcred == NOCRED && cred != NOCRED) bp->b_wcred = crhold(cred); } if (bp->b_flags & B_REMFREE) bremfreef(bp); BUF_KERNPROC(bp); TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); nmp->nm_bufqlen++; if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); VTONFS(bp->b_vp)->n_flag |= NMODIFIED; VTONFS(bp->b_vp)->n_directio_asyncwr++; mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); } mtx_unlock(&ncl_iod_mutex); return (0); } mtx_unlock(&ncl_iod_mutex); /* * All the iods are busy on other mounts, so return EIO to * force the caller to process the i/o synchronously. */ NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n")); return (EIO); } void ncl_doio_directwrite(struct buf *bp) { int iomode, must_commit; struct uio *uiop = (struct uio *)bp->b_caller1; char *iov_base = uiop->uio_iov->iov_base; iomode = NFSWRITE_FILESYNC; uiop->uio_td = NULL; /* NULL since we're in nfsiod */ ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0); KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write")); free(iov_base, M_NFSDIRECTIO); free(uiop->uio_iov, M_NFSDIRECTIO); free(uiop, M_NFSDIRECTIO); if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { struct nfsnode *np = VTONFS(bp->b_vp); mtx_lock(&np->n_mtx); np->n_directio_asyncwr--; if (np->n_directio_asyncwr == 0) { np->n_flag &= ~NMODIFIED; if ((np->n_flag & NFSYNCWAIT)) { np->n_flag &= ~NFSYNCWAIT; wakeup((caddr_t)&np->n_directio_asyncwr); } } mtx_unlock(&np->n_mtx); } bp->b_vp = NULL; relpbuf(bp, &ncl_pbuf_freecnt); } /* * Do an I/O operation to/from a cache block. This may be called * synchronously or from an nfsiod. */ int ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td, int called_from_strategy) { struct uio *uiop; struct nfsnode *np; struct nfsmount *nmp; int error = 0, iomode, must_commit = 0; struct uio uio; struct iovec io; struct proc *p = td ? td->td_proc : NULL; uint8_t iocmd; np = VTONFS(vp); nmp = VFSTONFS(vp->v_mount); uiop = &uio; uiop->uio_iov = &io; uiop->uio_iovcnt = 1; uiop->uio_segflg = UIO_SYSSPACE; uiop->uio_td = td; /* * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We * do this here so we do not have to do it in all the code that * calls us. */ bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp)); iocmd = bp->b_iocmd; if (iocmd == BIO_READ) { io.iov_len = uiop->uio_resid = bp->b_bcount; io.iov_base = bp->b_data; uiop->uio_rw = UIO_READ; switch (vp->v_type) { case VREG: uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; NFSINCRGLOBAL(newnfsstats.read_bios); error = ncl_readrpc(vp, uiop, cr); if (!error) { if (uiop->uio_resid) { /* * If we had a short read with no error, we must have * hit a file hole. We should zero-fill the remainder. * This can also occur if the server hits the file EOF. * * Holes used to be able to occur due to pending * writes, but that is not possible any longer. */ int nread = bp->b_bcount - uiop->uio_resid; int left = uiop->uio_resid; if (left > 0) bzero((char *)bp->b_data + nread, left); uiop->uio_resid = 0; } } /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */ if (p && (vp->v_vflag & VV_TEXT)) { mtx_lock(&np->n_mtx); if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) { mtx_unlock(&np->n_mtx); PROC_LOCK(p); killproc(p, "text file modification"); PROC_UNLOCK(p); } else mtx_unlock(&np->n_mtx); } break; case VLNK: uiop->uio_offset = (off_t)0; NFSINCRGLOBAL(newnfsstats.readlink_bios); error = ncl_readlinkrpc(vp, uiop, cr); break; case VDIR: NFSINCRGLOBAL(newnfsstats.readdir_bios); uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { error = ncl_readdirplusrpc(vp, uiop, cr, td); if (error == NFSERR_NOTSUPP) nmp->nm_flag &= ~NFSMNT_RDIRPLUS; } if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) error = ncl_readdirrpc(vp, uiop, cr, td); /* * end-of-directory sets B_INVAL but does not generate an * error. */ if (error == 0 && uiop->uio_resid == bp->b_bcount) bp->b_flags |= B_INVAL; break; default: ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type); break; }; if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_error = error; } } else { /* * If we only need to commit, try to commit */ if (bp->b_flags & B_NEEDCOMMIT) { int retv; off_t off; off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff, bp->b_wcred, td); if (retv == 0) { bp->b_dirtyoff = bp->b_dirtyend = 0; bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); bp->b_resid = 0; bufdone(bp); return (0); } if (retv == NFSERR_STALEWRITEVERF) { ncl_clearcommit(vp->v_mount); } } /* * Setup for actual write */ mtx_lock(&np->n_mtx); if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; mtx_unlock(&np->n_mtx); if (bp->b_dirtyend > bp->b_dirtyoff) { io.iov_len = uiop->uio_resid = bp->b_dirtyend - bp->b_dirtyoff; uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyoff; io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; uiop->uio_rw = UIO_WRITE; NFSINCRGLOBAL(newnfsstats.write_bios); if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) iomode = NFSWRITE_UNSTABLE; else iomode = NFSWRITE_FILESYNC; error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit, called_from_strategy); /* * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try * to cluster the buffers needing commit. This will allow * the system to submit a single commit rpc for the whole * cluster. We can do this even if the buffer is not 100% * dirty (relative to the NFS blocksize), so we optimize the * append-to-file-case. * * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be * cleared because write clustering only works for commit * rpc's, not for the data portion of the write). */ if (!error && iomode == NFSWRITE_UNSTABLE) { bp->b_flags |= B_NEEDCOMMIT; if (bp->b_dirtyoff == 0 && bp->b_dirtyend == bp->b_bcount) bp->b_flags |= B_CLUSTEROK; } else { bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); } /* * For an interrupted write, the buffer is still valid * and the write hasn't been pushed to the server yet, * so we can't set BIO_ERROR and report the interruption * by setting B_EINTR. For the B_ASYNC case, B_EINTR * is not relevant, so the rpc attempt is essentially * a noop. For the case of a V3 write rpc not being * committed to stable storage, the block is still * dirty and requires either a commit rpc or another * write rpc with iomode == NFSV3WRITE_FILESYNC before * the block is reused. This is indicated by setting * the B_DELWRI and B_NEEDCOMMIT flags. * * EIO is returned by ncl_writerpc() to indicate a recoverable * write error and is handled as above, except that * B_EINTR isn't set. One cause of this is a stale stateid * error for the RPC that indicates recovery is required, * when called with called_from_strategy != 0. * * If the buffer is marked B_PAGING, it does not reside on * the vp's paging queues so we cannot call bdirty(). The * bp in this case is not an NFS cache block so we should * be safe. XXX * * The logic below breaks up errors into recoverable and * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE * and keep the buffer around for potential write retries. * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) * and save the error in the nfsnode. This is less than ideal * but necessary. Keeping such buffers around could potentially * cause buffer exhaustion eventually (they can never be written * out, so will get constantly be re-dirtied). It also causes * all sorts of vfs panics. For non-recoverable write errors, * also invalidate the attrcache, so we'll be forced to go over * the wire for this object, returning an error to user on next * call (most of the time). */ if (error == EINTR || error == EIO || error == ETIMEDOUT || (!error && (bp->b_flags & B_NEEDCOMMIT))) { int s; s = splbio(); bp->b_flags &= ~(B_INVAL|B_NOCACHE); if ((bp->b_flags & B_PAGING) == 0) { bdirty(bp); bp->b_flags &= ~B_DONE; } if ((error == EINTR || error == ETIMEDOUT) && (bp->b_flags & B_ASYNC) == 0) bp->b_flags |= B_EINTR; splx(s); } else { if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_flags |= B_INVAL; bp->b_error = np->n_error = error; mtx_lock(&np->n_mtx); np->n_flag |= NWRITEERR; np->n_attrstamp = 0; mtx_unlock(&np->n_mtx); } bp->b_dirtyoff = bp->b_dirtyend = 0; } } else { bp->b_resid = 0; bufdone(bp); return (0); } } bp->b_resid = uiop->uio_resid; if (must_commit) ncl_clearcommit(vp->v_mount); bufdone(bp); return (error); } /* * Used to aid in handling ftruncate() operations on the NFS client side. * Truncation creates a number of special problems for NFS. We have to * throw away VM pages and buffer cache buffers that are beyond EOF, and * we have to properly handle VM pages or (potentially dirty) buffers * that straddle the truncation point. */ int ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) { struct nfsnode *np = VTONFS(vp); u_quad_t tsize; int biosize = vp->v_mount->mnt_stat.f_iosize; int error = 0; mtx_lock(&np->n_mtx); tsize = np->n_size; np->n_size = nsize; mtx_unlock(&np->n_mtx); if (nsize < tsize) { struct buf *bp; daddr_t lbn; int bufsize; /* * vtruncbuf() doesn't get the buffer overlapping the * truncation point. We may have a B_DELWRI and/or B_CACHE * buffer that now needs to be truncated. */ error = vtruncbuf(vp, cred, td, nsize, biosize); lbn = nsize / biosize; bufsize = nsize & (biosize - 1); bp = nfs_getcacheblk(vp, lbn, bufsize, td); if (!bp) return EINTR; if (bp->b_dirtyoff > bp->b_bcount) bp->b_dirtyoff = bp->b_bcount; if (bp->b_dirtyend > bp->b_bcount) bp->b_dirtyend = bp->b_bcount; bp->b_flags |= B_RELBUF; /* don't leave garbage around */ brelse(bp); } else { vnode_pager_setsize(vp, nsize); } return(error); } Index: head/sys/fs/nwfs/nwfs_io.c =================================================================== --- head/sys/fs/nwfs/nwfs_io.c (revision 222585) +++ head/sys/fs/nwfs/nwfs_io.c (revision 222586) @@ -1,631 +1,626 @@ /*- * Copyright (c) 1999 Boris Popov * 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. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR 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 AUTHOR 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. * * $FreeBSD$ * */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static int nwfs_fastlookup = 1; SYSCTL_DECL(_vfs_nwfs); SYSCTL_INT(_vfs_nwfs, OID_AUTO, fastlookup, CTLFLAG_RW, &nwfs_fastlookup, 0, ""); extern int nwfs_pbuf_freecnt; #define DE_SIZE (sizeof(struct dirent)) #define NWFS_RWCACHE static int nwfs_readvdir(struct vnode *vp, struct uio *uio, struct ucred *cred) { struct nwmount *nmp = VTONWFS(vp); int error, count, i; struct dirent dp; struct nwnode *np = VTONW(vp); struct nw_entry_info fattr; struct vnode *newvp; struct componentname cn; ncpfid fid; np = VTONW(vp); NCPVNDEBUG("dirname='%s'\n",np->n_name); if (uio->uio_resid < DE_SIZE || (uio->uio_offset < 0)) return (EINVAL); error = 0; count = 0; i = uio->uio_offset / DE_SIZE; /* offset in directory */ if (i == 0) { error = ncp_initsearch(vp, uio->uio_td, cred); if (error) { NCPVNDEBUG("cannot initialize search, error=%d",error); return( error ); } } for (; uio->uio_resid >= DE_SIZE; i++) { bzero((char *) &dp, DE_SIZE); dp.d_reclen = DE_SIZE; switch (i) { case 0: /* `.' */ case 1: /* `..' */ dp.d_fileno = (i == 0) ? np->n_fid.f_id : np->n_parent.f_id; if (!dp.d_fileno) dp.d_fileno = NWFS_ROOT_INO; dp.d_namlen = i + 1; dp.d_name[0] = '.'; dp.d_name[1] = '.'; dp.d_name[i + 1] = '\0'; dp.d_type = DT_DIR; break; default: error = ncp_search_for_file_or_subdir(nmp, &np->n_seq, &fattr, uio->uio_td, cred); if (error && error < 0x80) break; dp.d_fileno = fattr.dirEntNum; dp.d_type = (fattr.attributes & aDIR) ? DT_DIR : DT_REG; dp.d_namlen = fattr.nameLen; bcopy(fattr.entryName, dp.d_name, dp.d_namlen); dp.d_name[dp.d_namlen] = '\0'; #if 0 if (error && eofflag) { /* *eofflag = 1;*/ break; } #endif break; } if (nwfs_fastlookup && !error && i > 1) { fid.f_id = fattr.dirEntNum; fid.f_parent = np->n_fid.f_id; error = nwfs_nget(vp->v_mount, fid, &fattr, vp, &newvp); if (!error) { VTONW(newvp)->n_ctime = VTONW(newvp)->n_vattr.va_ctime.tv_sec; cn.cn_nameptr = dp.d_name; cn.cn_namelen = dp.d_namlen; cache_enter(vp, newvp, &cn); vput(newvp); } else error = 0; } if (error >= 0x80) { error = 0; break; } if ((error = uiomove(&dp, DE_SIZE, uio))) break; } uio->uio_offset = i * DE_SIZE; return (error); } int nwfs_readvnode(struct vnode *vp, struct uio *uiop, struct ucred *cred) { struct nwmount *nmp = VFSTONWFS(vp->v_mount); struct nwnode *np = VTONW(vp); struct thread *td; struct vattr vattr; int error, biosize; if (vp->v_type != VREG && vp->v_type != VDIR) { printf("%s: vn types other than VREG or VDIR are unsupported !\n",__func__); return EIO; } if (uiop->uio_resid == 0) return 0; if (uiop->uio_offset < 0) return EINVAL; /* if (uiop->uio_offset + uiop->uio_resid > nmp->nm_maxfilesize) return (EFBIG);*/ td = uiop->uio_td; if (vp->v_type == VDIR) { error = nwfs_readvdir(vp, uiop, cred); return error; } biosize = NWFSTOCONN(nmp)->buffer_size; if (np->n_flag & NMODIFIED) { nwfs_attr_cacheremove(vp); error = VOP_GETATTR(vp, &vattr, cred); if (error) return (error); np->n_mtime = vattr.va_mtime.tv_sec; } else { error = VOP_GETATTR(vp, &vattr, cred); if (error) return (error); if (np->n_mtime != vattr.va_mtime.tv_sec) { error = nwfs_vinvalbuf(vp, td); if (error) return (error); np->n_mtime = vattr.va_mtime.tv_sec; } } error = ncp_read(NWFSTOCONN(nmp), &np->n_fh, uiop, cred); return (error); } int nwfs_writevnode(vp, uiop, cred, ioflag) struct vnode *vp; struct uio *uiop; struct ucred *cred; int ioflag; { struct nwmount *nmp = VTONWFS(vp); struct nwnode *np = VTONW(vp); struct thread *td; /* struct vattr vattr;*/ int error = 0; if (vp->v_type != VREG) { printf("%s: vn types other than VREG unsupported !\n",__func__); return EIO; } NCPVNDEBUG("ofs=%d,resid=%d\n",(int)uiop->uio_offset, uiop->uio_resid); if (uiop->uio_offset < 0) return EINVAL; /* if (uiop->uio_offset + uiop->uio_resid > nmp->nm_maxfilesize) return (EFBIG);*/ td = uiop->uio_td; if (ioflag & (IO_APPEND | IO_SYNC)) { if (np->n_flag & NMODIFIED) { nwfs_attr_cacheremove(vp); error = nwfs_vinvalbuf(vp, td); if (error) return (error); } if (ioflag & IO_APPEND) { /* We can relay only on local information about file size, * because until file is closed NetWare will not return * the correct size. */ #ifdef notyet nwfs_attr_cacheremove(vp); error = VOP_GETATTR(vp, &vattr, cred); if (error) return (error); #endif uiop->uio_offset = np->n_size; } } if (uiop->uio_resid == 0) return 0; if (vn_rlimit_fsize(vp, uiop, td)) return (EFBIG); error = ncp_write(NWFSTOCONN(nmp), &np->n_fh, uiop, cred); NCPVNDEBUG("after: ofs=%d,resid=%d\n",(int)uiop->uio_offset, uiop->uio_resid); if (!error) { if (uiop->uio_offset > np->n_size) { np->n_vattr.va_size = np->n_size = uiop->uio_offset; vnode_pager_setsize(vp, np->n_size); } } return (error); } /* * Do an I/O operation to/from a cache block. */ int nwfs_doio(vp, bp, cr, td) struct vnode *vp; struct buf *bp; struct ucred *cr; struct thread *td; { struct uio *uiop; struct nwnode *np; struct nwmount *nmp; int error = 0; struct uio uio; struct iovec io; np = VTONW(vp); nmp = VFSTONWFS(vp->v_mount); uiop = &uio; uiop->uio_iov = &io; uiop->uio_iovcnt = 1; uiop->uio_segflg = UIO_SYSSPACE; uiop->uio_td = td; if (bp->b_iocmd == BIO_READ) { io.iov_len = uiop->uio_resid = bp->b_bcount; io.iov_base = bp->b_data; uiop->uio_rw = UIO_READ; switch (vp->v_type) { case VREG: uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; error = ncp_read(NWFSTOCONN(nmp), &np->n_fh, uiop, cr); if (error) break; if (uiop->uio_resid) { int left = uiop->uio_resid; int nread = bp->b_bcount - left; if (left > 0) bzero((char *)bp->b_data + nread, left); } break; /* case VDIR: nfsstats.readdir_bios++; uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; if (nmp->nm_flag & NFSMNT_RDIRPLUS) { error = nfs_readdirplusrpc(vp, uiop, cr); if (error == NFSERR_NOTSUPP) nmp->nm_flag &= ~NFSMNT_RDIRPLUS; } if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) error = nfs_readdirrpc(vp, uiop, cr); if (error == 0 && uiop->uio_resid == bp->b_bcount) bp->b_flags |= B_INVAL; break; */ default: printf("nwfs_doio: type %x unexpected\n",vp->v_type); break; }; if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_error = error; } } else { /* write */ if (((bp->b_blkno * DEV_BSIZE) + bp->b_dirtyend) > np->n_size) bp->b_dirtyend = np->n_size - (bp->b_blkno * DEV_BSIZE); if (bp->b_dirtyend > bp->b_dirtyoff) { io.iov_len = uiop->uio_resid = bp->b_dirtyend - bp->b_dirtyoff; uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; uiop->uio_rw = UIO_WRITE; error = ncp_write(NWFSTOCONN(nmp), &np->n_fh, uiop, cr); /* * For an interrupted write, the buffer is still valid * and the write hasn't been pushed to the server yet, * so we can't set BIO_ERROR and report the interruption * by setting B_EINTR. For the B_ASYNC case, B_EINTR * is not relevant, so the rpc attempt is essentially * a noop. For the case of a V3 write rpc not being * committed to stable storage, the block is still * dirty and requires either a commit rpc or another * write rpc with iomode == NFSV3WRITE_FILESYNC before * the block is reused. This is indicated by setting * the B_DELWRI and B_NEEDCOMMIT flags. */ if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) { int s; s = splbio(); bp->b_flags &= ~(B_INVAL|B_NOCACHE); if ((bp->b_flags & B_ASYNC) == 0) bp->b_flags |= B_EINTR; if ((bp->b_flags & B_PAGING) == 0) { bdirty(bp); bp->b_flags &= ~B_DONE; } if ((bp->b_flags & B_ASYNC) == 0) bp->b_flags |= B_EINTR; splx(s); } else { if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_error /*= np->n_error */= error; /* np->n_flag |= NWRITEERR;*/ } bp->b_dirtyoff = bp->b_dirtyend = 0; } } else { bp->b_resid = 0; bufdone(bp); return (0); } } bp->b_resid = uiop->uio_resid; bufdone(bp); return (error); } /* * Vnode op for VM getpages. * Wish wish .... get rid from multiple IO routines */ int nwfs_getpages(ap) struct vop_getpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int a_reqpage; vm_ooffset_t a_offset; } */ *ap; { #ifndef NWFS_RWCACHE return vop_stdgetpages(ap);(ap->a_vp, ap->a_m, ap->a_count, #else int i, error, nextoff, size, toff, npages, count; struct uio uio; struct iovec iov; vm_offset_t kva; struct buf *bp; struct vnode *vp; struct thread *td; struct ucred *cred; struct nwmount *nmp; struct nwnode *np; vm_object_t object; vm_page_t *pages; vp = ap->a_vp; td = curthread; /* XXX */ cred = td->td_ucred; /* XXX */ np = VTONW(vp); nmp = VFSTONWFS(vp->v_mount); pages = ap->a_m; count = ap->a_count; if ((object = vp->v_object) == NULL) { printf("nwfs_getpages: called with non-merged cache vnode??\n"); return VM_PAGER_ERROR; } bp = getpbuf(&nwfs_pbuf_freecnt); npages = btoc(count); kva = (vm_offset_t) bp->b_data; pmap_qenter(kva, pages, npages); iov.iov_base = (caddr_t) kva; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); uio.uio_resid = count; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_READ; uio.uio_td = td; error = ncp_read(NWFSTOCONN(nmp), &np->n_fh, &uio,cred); pmap_qremove(kva, npages); relpbuf(bp, &nwfs_pbuf_freecnt); VM_OBJECT_LOCK(object); if (error && (uio.uio_resid == count)) { printf("nwfs_getpages: error %d\n",error); for (i = 0; i < npages; i++) { if (ap->a_reqpage != i) { vm_page_lock(pages[i]); vm_page_free(pages[i]); vm_page_unlock(pages[i]); } } VM_OBJECT_UNLOCK(object); return VM_PAGER_ERROR; } size = count - uio.uio_resid; for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { vm_page_t m; nextoff = toff + PAGE_SIZE; m = pages[i]; if (nextoff <= size) { m->valid = VM_PAGE_BITS_ALL; KASSERT(m->dirty == 0, ("nwfs_getpages: page %p is dirty", m)); } else { int nvalid = ((size + DEV_BSIZE - 1) - toff) & ~(DEV_BSIZE - 1); vm_page_set_valid(m, 0, nvalid); KASSERT((m->dirty & vm_page_bits(0, nvalid)) == 0, ("nwfs_getpages: page %p is dirty", m)); } if (i != ap->a_reqpage) { /* * Whether or not to leave the page activated is up in * the air, but we should put the page on a page queue * somewhere (it already is in the object). Result: * It appears that emperical results show that * deactivating pages is best. */ /* * Just in case someone was asking for this page we * now tell them that it is ok to use. */ if (!error) { if (m->oflags & VPO_WANTED) { vm_page_lock(m); vm_page_activate(m); vm_page_unlock(m); } else { vm_page_lock(m); vm_page_deactivate(m); vm_page_unlock(m); } vm_page_wakeup(m); } else { vm_page_lock(m); vm_page_free(m); vm_page_unlock(m); } } } VM_OBJECT_UNLOCK(object); return 0; #endif /* NWFS_RWCACHE */ } /* * Vnode op for VM putpages. * possible bug: all IO done in sync mode * Note that vop_close always invalidate pages before close, so it's * not necessary to open vnode. */ int nwfs_putpages(ap) struct vop_putpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int a_sync; int *a_rtvals; vm_ooffset_t a_offset; } */ *ap; { int error; struct vnode *vp = ap->a_vp; struct thread *td; struct ucred *cred; #ifndef NWFS_RWCACHE td = curthread; /* XXX */ cred = td->td_ucred; /* XXX */ VOP_OPEN(vp, FWRITE, cred, td, NULL); error = vop_stdputpages(ap); VOP_CLOSE(vp, FWRITE, cred, td); return error; #else struct uio uio; struct iovec iov; vm_offset_t kva; struct buf *bp; int i, npages, count; int *rtvals; struct nwmount *nmp; struct nwnode *np; vm_page_t *pages; td = curthread; /* XXX */ cred = td->td_ucred; /* XXX */ /* VOP_OPEN(vp, FWRITE, cred, td, NULL);*/ np = VTONW(vp); nmp = VFSTONWFS(vp->v_mount); pages = ap->a_m; count = ap->a_count; rtvals = ap->a_rtvals; npages = btoc(count); for (i = 0; i < npages; i++) { - rtvals[i] = VM_PAGER_AGAIN; + rtvals[i] = VM_PAGER_ERROR; } bp = getpbuf(&nwfs_pbuf_freecnt); kva = (vm_offset_t) bp->b_data; pmap_qenter(kva, pages, npages); iov.iov_base = (caddr_t) kva; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); uio.uio_resid = count; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_WRITE; uio.uio_td = td; NCPVNDEBUG("ofs=%d,resid=%d\n",(int)uio.uio_offset, uio.uio_resid); error = ncp_write(NWFSTOCONN(nmp), &np->n_fh, &uio, cred); /* VOP_CLOSE(vp, FWRITE, cred, td);*/ NCPVNDEBUG("paged write done: %d\n", error); pmap_qremove(kva, npages); relpbuf(bp, &nwfs_pbuf_freecnt); - if (!error) { - int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; - for (i = 0; i < nwritten; i++) { - rtvals[i] = VM_PAGER_OK; - vm_page_undirty(pages[i]); - } - } + if (!error) + vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid); return rtvals[0]; #endif /* NWFS_RWCACHE */ } /* * Flush and invalidate all dirty buffers. If another process is already * doing the flush, just wait for completion. */ int nwfs_vinvalbuf(vp, td) struct vnode *vp; struct thread *td; { struct nwnode *np = VTONW(vp); /* struct nwmount *nmp = VTONWFS(vp);*/ int error = 0; if (vp->v_iflag & VI_DOOMED) return (0); while (np->n_flag & NFLUSHINPROG) { np->n_flag |= NFLUSHWANT; error = tsleep(&np->n_flag, PRIBIO + 2, "nwfsvinv", 2 * hz); error = ncp_chkintr(NWFSTOCONN(VTONWFS(vp)), td); if (error == EINTR) return EINTR; } np->n_flag |= NFLUSHINPROG; if (vp->v_bufobj.bo_object != NULL) { VM_OBJECT_LOCK(vp->v_bufobj.bo_object); vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); } error = vinvalbuf(vp, V_SAVE, PCATCH, 0); while (error) { if (error == ERESTART || error == EINTR) { np->n_flag &= ~NFLUSHINPROG; if (np->n_flag & NFLUSHWANT) { np->n_flag &= ~NFLUSHWANT; wakeup(&np->n_flag); } return EINTR; } error = vinvalbuf(vp, V_SAVE, PCATCH, 0); } np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); if (np->n_flag & NFLUSHWANT) { np->n_flag &= ~NFLUSHWANT; wakeup(&np->n_flag); } return (error); } Index: head/sys/fs/smbfs/smbfs_io.c =================================================================== --- head/sys/fs/smbfs/smbfs_io.c (revision 222585) +++ head/sys/fs/smbfs/smbfs_io.c (revision 222586) @@ -1,699 +1,694 @@ /*- * Copyright (c) 2000-2001 Boris Popov * 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. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR 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 AUTHOR 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. * * $FreeBSD$ * */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* #include */ #include #include #include #include #include #include /*#define SMBFS_RWGENERIC*/ extern int smbfs_pbuf_freecnt; static int smbfs_fastlookup = 1; SYSCTL_DECL(_vfs_smbfs); SYSCTL_INT(_vfs_smbfs, OID_AUTO, fastlookup, CTLFLAG_RW, &smbfs_fastlookup, 0, ""); #define DE_SIZE (sizeof(struct dirent)) static int smbfs_readvdir(struct vnode *vp, struct uio *uio, struct ucred *cred) { struct dirent de; struct componentname cn; struct smb_cred scred; struct smbfs_fctx *ctx; struct vnode *newvp; struct smbnode *np = VTOSMB(vp); int error/*, *eofflag = ap->a_eofflag*/; long offset, limit; np = VTOSMB(vp); SMBVDEBUG("dirname='%s'\n", np->n_name); smb_makescred(&scred, uio->uio_td, cred); offset = uio->uio_offset / DE_SIZE; /* offset in the directory */ limit = uio->uio_resid / DE_SIZE; if (uio->uio_resid < DE_SIZE || uio->uio_offset < 0) return EINVAL; while (limit && offset < 2) { limit--; bzero((caddr_t)&de, DE_SIZE); de.d_reclen = DE_SIZE; de.d_fileno = (offset == 0) ? np->n_ino : (np->n_parent ? VTOSMB(np->n_parent)->n_ino : 2); if (de.d_fileno == 0) de.d_fileno = 0x7ffffffd + offset; de.d_namlen = offset + 1; de.d_name[0] = '.'; de.d_name[1] = '.'; de.d_name[offset + 1] = '\0'; de.d_type = DT_DIR; error = uiomove(&de, DE_SIZE, uio); if (error) return error; offset++; uio->uio_offset += DE_SIZE; } if (limit == 0) return 0; if (offset != np->n_dirofs || np->n_dirseq == NULL) { SMBVDEBUG("Reopening search %ld:%ld\n", offset, np->n_dirofs); if (np->n_dirseq) { smbfs_findclose(np->n_dirseq, &scred); np->n_dirseq = NULL; } np->n_dirofs = 2; error = smbfs_findopen(np, "*", 1, SMB_FA_SYSTEM | SMB_FA_HIDDEN | SMB_FA_DIR, &scred, &ctx); if (error) { SMBVDEBUG("can not open search, error = %d", error); return error; } np->n_dirseq = ctx; } else ctx = np->n_dirseq; while (np->n_dirofs < offset) { error = smbfs_findnext(ctx, offset - np->n_dirofs++, &scred); if (error) { smbfs_findclose(np->n_dirseq, &scred); np->n_dirseq = NULL; return error == ENOENT ? 0 : error; } } error = 0; for (; limit; limit--, offset++) { error = smbfs_findnext(ctx, limit, &scred); if (error) break; np->n_dirofs++; bzero((caddr_t)&de, DE_SIZE); de.d_reclen = DE_SIZE; de.d_fileno = ctx->f_attr.fa_ino; de.d_type = (ctx->f_attr.fa_attr & SMB_FA_DIR) ? DT_DIR : DT_REG; de.d_namlen = ctx->f_nmlen; bcopy(ctx->f_name, de.d_name, de.d_namlen); de.d_name[de.d_namlen] = '\0'; if (smbfs_fastlookup) { error = smbfs_nget(vp->v_mount, vp, ctx->f_name, ctx->f_nmlen, &ctx->f_attr, &newvp); if (!error) { cn.cn_nameptr = de.d_name; cn.cn_namelen = de.d_namlen; cache_enter(vp, newvp, &cn); vput(newvp); } } error = uiomove(&de, DE_SIZE, uio); if (error) break; } if (error == ENOENT) error = 0; uio->uio_offset = offset * DE_SIZE; return error; } int smbfs_readvnode(struct vnode *vp, struct uio *uiop, struct ucred *cred) { struct smbmount *smp = VFSTOSMBFS(vp->v_mount); struct smbnode *np = VTOSMB(vp); struct thread *td; struct vattr vattr; struct smb_cred scred; int error, lks; /* * Protect against method which is not supported for now */ if (uiop->uio_segflg == UIO_NOCOPY) return EOPNOTSUPP; if (vp->v_type != VREG && vp->v_type != VDIR) { SMBFSERR("vn types other than VREG or VDIR are unsupported !\n"); return EIO; } if (uiop->uio_resid == 0) return 0; if (uiop->uio_offset < 0) return EINVAL; /* if (uiop->uio_offset + uiop->uio_resid > smp->nm_maxfilesize) return EFBIG;*/ td = uiop->uio_td; if (vp->v_type == VDIR) { lks = LK_EXCLUSIVE; /* lockstatus(vp->v_vnlock); */ if (lks == LK_SHARED) vn_lock(vp, LK_UPGRADE | LK_RETRY); error = smbfs_readvdir(vp, uiop, cred); if (lks == LK_SHARED) vn_lock(vp, LK_DOWNGRADE | LK_RETRY); return error; } /* biosize = SSTOCN(smp->sm_share)->sc_txmax;*/ if (np->n_flag & NMODIFIED) { smbfs_attr_cacheremove(vp); error = VOP_GETATTR(vp, &vattr, cred); if (error) return error; np->n_mtime.tv_sec = vattr.va_mtime.tv_sec; } else { error = VOP_GETATTR(vp, &vattr, cred); if (error) return error; if (np->n_mtime.tv_sec != vattr.va_mtime.tv_sec) { error = smbfs_vinvalbuf(vp, td); if (error) return error; np->n_mtime.tv_sec = vattr.va_mtime.tv_sec; } } smb_makescred(&scred, td, cred); return smb_read(smp->sm_share, np->n_fid, uiop, &scred); } int smbfs_writevnode(struct vnode *vp, struct uio *uiop, struct ucred *cred, int ioflag) { struct smbmount *smp = VTOSMBFS(vp); struct smbnode *np = VTOSMB(vp); struct smb_cred scred; struct thread *td; int error = 0; if (vp->v_type != VREG) { SMBERROR("vn types other than VREG unsupported !\n"); return EIO; } SMBVDEBUG("ofs=%d,resid=%d\n",(int)uiop->uio_offset, uiop->uio_resid); if (uiop->uio_offset < 0) return EINVAL; /* if (uiop->uio_offset + uiop->uio_resid > smp->nm_maxfilesize) return (EFBIG);*/ td = uiop->uio_td; if (ioflag & (IO_APPEND | IO_SYNC)) { if (np->n_flag & NMODIFIED) { smbfs_attr_cacheremove(vp); error = smbfs_vinvalbuf(vp, td); if (error) return error; } if (ioflag & IO_APPEND) { #ifdef notyet /* * File size can be changed by another client */ smbfs_attr_cacheremove(vp); error = VOP_GETATTR(vp, &vattr, cred); if (error) return (error); #endif uiop->uio_offset = np->n_size; } } if (uiop->uio_resid == 0) return 0; if (vn_rlimit_fsize(vp, uiop, td)) return (EFBIG); smb_makescred(&scred, td, cred); error = smb_write(smp->sm_share, np->n_fid, uiop, &scred); SMBVDEBUG("after: ofs=%d,resid=%d\n",(int)uiop->uio_offset, uiop->uio_resid); if (!error) { if (uiop->uio_offset > np->n_size) { np->n_size = uiop->uio_offset; vnode_pager_setsize(vp, np->n_size); } } return error; } /* * Do an I/O operation to/from a cache block. */ int smbfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) { struct smbmount *smp = VFSTOSMBFS(vp->v_mount); struct smbnode *np = VTOSMB(vp); struct uio uio, *uiop = &uio; struct iovec io; struct smb_cred scred; int error = 0; uiop->uio_iov = &io; uiop->uio_iovcnt = 1; uiop->uio_segflg = UIO_SYSSPACE; uiop->uio_td = td; smb_makescred(&scred, td, cr); if (bp->b_iocmd == BIO_READ) { io.iov_len = uiop->uio_resid = bp->b_bcount; io.iov_base = bp->b_data; uiop->uio_rw = UIO_READ; switch (vp->v_type) { case VREG: uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; error = smb_read(smp->sm_share, np->n_fid, uiop, &scred); if (error) break; if (uiop->uio_resid) { int left = uiop->uio_resid; int nread = bp->b_bcount - left; if (left > 0) bzero((char *)bp->b_data + nread, left); } break; default: printf("smbfs_doio: type %x unexpected\n",vp->v_type); break; }; if (error) { bp->b_error = error; bp->b_ioflags |= BIO_ERROR; } } else { /* write */ if (((bp->b_blkno * DEV_BSIZE) + bp->b_dirtyend) > np->n_size) bp->b_dirtyend = np->n_size - (bp->b_blkno * DEV_BSIZE); if (bp->b_dirtyend > bp->b_dirtyoff) { io.iov_len = uiop->uio_resid = bp->b_dirtyend - bp->b_dirtyoff; uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; uiop->uio_rw = UIO_WRITE; error = smb_write(smp->sm_share, np->n_fid, uiop, &scred); /* * For an interrupted write, the buffer is still valid * and the write hasn't been pushed to the server yet, * so we can't set BIO_ERROR and report the interruption * by setting B_EINTR. For the B_ASYNC case, B_EINTR * is not relevant, so the rpc attempt is essentially * a noop. For the case of a V3 write rpc not being * committed to stable storage, the block is still * dirty and requires either a commit rpc or another * write rpc with iomode == NFSV3WRITE_FILESYNC before * the block is reused. This is indicated by setting * the B_DELWRI and B_NEEDCOMMIT flags. */ if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) { int s; s = splbio(); bp->b_flags &= ~(B_INVAL|B_NOCACHE); if ((bp->b_flags & B_ASYNC) == 0) bp->b_flags |= B_EINTR; if ((bp->b_flags & B_PAGING) == 0) { bdirty(bp); bp->b_flags &= ~B_DONE; } if ((bp->b_flags & B_ASYNC) == 0) bp->b_flags |= B_EINTR; splx(s); } else { if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_error = error; } bp->b_dirtyoff = bp->b_dirtyend = 0; } } else { bp->b_resid = 0; bufdone(bp); return 0; } } bp->b_resid = uiop->uio_resid; bufdone(bp); return error; } /* * Vnode op for VM getpages. * Wish wish .... get rid from multiple IO routines */ int smbfs_getpages(ap) struct vop_getpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int a_reqpage; vm_ooffset_t a_offset; } */ *ap; { #ifdef SMBFS_RWGENERIC return vop_stdgetpages(ap); #else int i, error, nextoff, size, toff, npages, count, reqpage; struct uio uio; struct iovec iov; vm_offset_t kva; struct buf *bp; struct vnode *vp; struct thread *td; struct ucred *cred; struct smbmount *smp; struct smbnode *np; struct smb_cred scred; vm_object_t object; vm_page_t *pages, m; vp = ap->a_vp; if ((object = vp->v_object) == NULL) { printf("smbfs_getpages: called with non-merged cache vnode??\n"); return VM_PAGER_ERROR; } td = curthread; /* XXX */ cred = td->td_ucred; /* XXX */ np = VTOSMB(vp); smp = VFSTOSMBFS(vp->v_mount); pages = ap->a_m; count = ap->a_count; npages = btoc(count); reqpage = ap->a_reqpage; /* * If the requested page is partially valid, just return it and * allow the pager to zero-out the blanks. Partially valid pages * can only occur at the file EOF. */ m = pages[reqpage]; VM_OBJECT_LOCK(object); if (m->valid != 0) { for (i = 0; i < npages; ++i) { if (i != reqpage) { vm_page_lock(pages[i]); vm_page_free(pages[i]); vm_page_unlock(pages[i]); } } VM_OBJECT_UNLOCK(object); return 0; } VM_OBJECT_UNLOCK(object); smb_makescred(&scred, td, cred); bp = getpbuf(&smbfs_pbuf_freecnt); kva = (vm_offset_t) bp->b_data; pmap_qenter(kva, pages, npages); PCPU_INC(cnt.v_vnodein); PCPU_ADD(cnt.v_vnodepgsin, npages); iov.iov_base = (caddr_t) kva; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); uio.uio_resid = count; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_READ; uio.uio_td = td; error = smb_read(smp->sm_share, np->n_fid, &uio, &scred); pmap_qremove(kva, npages); relpbuf(bp, &smbfs_pbuf_freecnt); VM_OBJECT_LOCK(object); if (error && (uio.uio_resid == count)) { printf("smbfs_getpages: error %d\n",error); for (i = 0; i < npages; i++) { if (reqpage != i) { vm_page_lock(pages[i]); vm_page_free(pages[i]); vm_page_unlock(pages[i]); } } VM_OBJECT_UNLOCK(object); return VM_PAGER_ERROR; } size = count - uio.uio_resid; for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { vm_page_t m; nextoff = toff + PAGE_SIZE; m = pages[i]; if (nextoff <= size) { /* * Read operation filled an entire page */ m->valid = VM_PAGE_BITS_ALL; KASSERT(m->dirty == 0, ("smbfs_getpages: page %p is dirty", m)); } else if (size > toff) { /* * Read operation filled a partial page. */ m->valid = 0; vm_page_set_valid(m, 0, size - toff); KASSERT(m->dirty == 0, ("smbfs_getpages: page %p is dirty", m)); } else { /* * Read operation was short. If no error occured * we may have hit a zero-fill section. We simply * leave valid set to 0. */ ; } if (i != reqpage) { /* * Whether or not to leave the page activated is up in * the air, but we should put the page on a page queue * somewhere (it already is in the object). Result: * It appears that emperical results show that * deactivating pages is best. */ /* * Just in case someone was asking for this page we * now tell them that it is ok to use. */ if (!error) { if (m->oflags & VPO_WANTED) { vm_page_lock(m); vm_page_activate(m); vm_page_unlock(m); } else { vm_page_lock(m); vm_page_deactivate(m); vm_page_unlock(m); } vm_page_wakeup(m); } else { vm_page_lock(m); vm_page_free(m); vm_page_unlock(m); } } } VM_OBJECT_UNLOCK(object); return 0; #endif /* SMBFS_RWGENERIC */ } /* * Vnode op for VM putpages. * possible bug: all IO done in sync mode * Note that vop_close always invalidate pages before close, so it's * not necessary to open vnode. */ int smbfs_putpages(ap) struct vop_putpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int a_sync; int *a_rtvals; vm_ooffset_t a_offset; } */ *ap; { int error; struct vnode *vp = ap->a_vp; struct thread *td; struct ucred *cred; #ifdef SMBFS_RWGENERIC td = curthread; /* XXX */ cred = td->td_ucred; /* XXX */ VOP_OPEN(vp, FWRITE, cred, td, NULL); error = vop_stdputpages(ap); VOP_CLOSE(vp, FWRITE, cred, td); return error; #else struct uio uio; struct iovec iov; vm_offset_t kva; struct buf *bp; int i, npages, count; int *rtvals; struct smbmount *smp; struct smbnode *np; struct smb_cred scred; vm_page_t *pages; td = curthread; /* XXX */ cred = td->td_ucred; /* XXX */ /* VOP_OPEN(vp, FWRITE, cred, td, NULL);*/ np = VTOSMB(vp); smp = VFSTOSMBFS(vp->v_mount); pages = ap->a_m; count = ap->a_count; rtvals = ap->a_rtvals; npages = btoc(count); for (i = 0; i < npages; i++) { - rtvals[i] = VM_PAGER_AGAIN; + rtvals[i] = VM_PAGER_ERROR; } bp = getpbuf(&smbfs_pbuf_freecnt); kva = (vm_offset_t) bp->b_data; pmap_qenter(kva, pages, npages); PCPU_INC(cnt.v_vnodeout); PCPU_ADD(cnt.v_vnodepgsout, count); iov.iov_base = (caddr_t) kva; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); uio.uio_resid = count; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_WRITE; uio.uio_td = td; SMBVDEBUG("ofs=%d,resid=%d\n",(int)uio.uio_offset, uio.uio_resid); smb_makescred(&scred, td, cred); error = smb_write(smp->sm_share, np->n_fid, &uio, &scred); /* VOP_CLOSE(vp, FWRITE, cred, td);*/ SMBVDEBUG("paged write done: %d\n", error); pmap_qremove(kva, npages); relpbuf(bp, &smbfs_pbuf_freecnt); - if (!error) { - int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; - for (i = 0; i < nwritten; i++) { - rtvals[i] = VM_PAGER_OK; - vm_page_undirty(pages[i]); - } - } + if (!error) + vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid); return rtvals[0]; #endif /* SMBFS_RWGENERIC */ } /* * Flush and invalidate all dirty buffers. If another process is already * doing the flush, just wait for completion. */ int smbfs_vinvalbuf(struct vnode *vp, struct thread *td) { struct smbnode *np = VTOSMB(vp); int error = 0; if (vp->v_iflag & VI_DOOMED) return 0; while (np->n_flag & NFLUSHINPROG) { np->n_flag |= NFLUSHWANT; error = tsleep(&np->n_flag, PRIBIO + 2, "smfsvinv", 2 * hz); error = smb_td_intr(td); if (error == EINTR) return EINTR; } np->n_flag |= NFLUSHINPROG; if (vp->v_bufobj.bo_object != NULL) { VM_OBJECT_LOCK(vp->v_bufobj.bo_object); vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); } error = vinvalbuf(vp, V_SAVE, PCATCH, 0); while (error) { if (error == ERESTART || error == EINTR) { np->n_flag &= ~NFLUSHINPROG; if (np->n_flag & NFLUSHWANT) { np->n_flag &= ~NFLUSHWANT; wakeup(&np->n_flag); } return EINTR; } error = vinvalbuf(vp, V_SAVE, PCATCH, 0); } np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); if (np->n_flag & NFLUSHWANT) { np->n_flag &= ~NFLUSHWANT; wakeup(&np->n_flag); } return (error); } Index: head/sys/nfsclient/nfs_bio.c =================================================================== --- head/sys/nfsclient/nfs_bio.c (revision 222585) +++ head/sys/nfsclient/nfs_bio.c (revision 222586) @@ -1,1805 +1,1801 @@ /*- * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * Rick Macklem at The University of Guelph. * * 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. * 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. * * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_kdtrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td); static int nfs_directio_write(struct vnode *vp, struct uio *uiop, struct ucred *cred, int ioflag); extern int nfs_directio_enable; extern int nfs_directio_allow_mmap; /* * Vnode op for VM getpages. */ int nfs_getpages(struct vop_getpages_args *ap) { int i, error, nextoff, size, toff, count, npages; struct uio uio; struct iovec iov; vm_offset_t kva; struct buf *bp; struct vnode *vp; struct thread *td; struct ucred *cred; struct nfsmount *nmp; vm_object_t object; vm_page_t *pages; struct nfsnode *np; vp = ap->a_vp; np = VTONFS(vp); td = curthread; /* XXX */ cred = curthread->td_ucred; /* XXX */ nmp = VFSTONFS(vp->v_mount); pages = ap->a_m; count = ap->a_count; if ((object = vp->v_object) == NULL) { nfs_printf("nfs_getpages: called with non-merged cache vnode??\n"); return (VM_PAGER_ERROR); } if (nfs_directio_enable && !nfs_directio_allow_mmap) { mtx_lock(&np->n_mtx); if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { mtx_unlock(&np->n_mtx); nfs_printf("nfs_getpages: called on non-cacheable vnode??\n"); return (VM_PAGER_ERROR); } else mtx_unlock(&np->n_mtx); } mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); /* We'll never get here for v4, because we always have fsinfo */ (void)nfs_fsinfo(nmp, vp, cred, td); } else mtx_unlock(&nmp->nm_mtx); npages = btoc(count); /* * If the requested page is partially valid, just return it and * allow the pager to zero-out the blanks. Partially valid pages * can only occur at the file EOF. */ VM_OBJECT_LOCK(object); if (pages[ap->a_reqpage]->valid != 0) { for (i = 0; i < npages; ++i) { if (i != ap->a_reqpage) { vm_page_lock(pages[i]); vm_page_free(pages[i]); vm_page_unlock(pages[i]); } } VM_OBJECT_UNLOCK(object); return (0); } VM_OBJECT_UNLOCK(object); /* * We use only the kva address for the buffer, but this is extremely * convienient and fast. */ bp = getpbuf(&nfs_pbuf_freecnt); kva = (vm_offset_t) bp->b_data; pmap_qenter(kva, pages, npages); PCPU_INC(cnt.v_vnodein); PCPU_ADD(cnt.v_vnodepgsin, npages); iov.iov_base = (caddr_t) kva; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); uio.uio_resid = count; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_READ; uio.uio_td = td; error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred); pmap_qremove(kva, npages); relpbuf(bp, &nfs_pbuf_freecnt); if (error && (uio.uio_resid == count)) { nfs_printf("nfs_getpages: error %d\n", error); VM_OBJECT_LOCK(object); for (i = 0; i < npages; ++i) { if (i != ap->a_reqpage) { vm_page_lock(pages[i]); vm_page_free(pages[i]); vm_page_unlock(pages[i]); } } VM_OBJECT_UNLOCK(object); return (VM_PAGER_ERROR); } /* * Calculate the number of bytes read and validate only that number * of bytes. Note that due to pending writes, size may be 0. This * does not mean that the remaining data is invalid! */ size = count - uio.uio_resid; VM_OBJECT_LOCK(object); for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { vm_page_t m; nextoff = toff + PAGE_SIZE; m = pages[i]; if (nextoff <= size) { /* * Read operation filled an entire page */ m->valid = VM_PAGE_BITS_ALL; KASSERT(m->dirty == 0, ("nfs_getpages: page %p is dirty", m)); } else if (size > toff) { /* * Read operation filled a partial page. */ m->valid = 0; vm_page_set_valid(m, 0, size - toff); KASSERT(m->dirty == 0, ("nfs_getpages: page %p is dirty", m)); } else { /* * Read operation was short. If no error occured * we may have hit a zero-fill section. We simply * leave valid set to 0. */ ; } if (i != ap->a_reqpage) { /* * Whether or not to leave the page activated is up in * the air, but we should put the page on a page queue * somewhere (it already is in the object). Result: * It appears that emperical results show that * deactivating pages is best. */ /* * Just in case someone was asking for this page we * now tell them that it is ok to use. */ if (!error) { if (m->oflags & VPO_WANTED) { vm_page_lock(m); vm_page_activate(m); vm_page_unlock(m); } else { vm_page_lock(m); vm_page_deactivate(m); vm_page_unlock(m); } vm_page_wakeup(m); } else { vm_page_lock(m); vm_page_free(m); vm_page_unlock(m); } } } VM_OBJECT_UNLOCK(object); return (0); } /* * Vnode op for VM putpages. */ int nfs_putpages(struct vop_putpages_args *ap) { struct uio uio; struct iovec iov; vm_offset_t kva; struct buf *bp; int iomode, must_commit, i, error, npages, count; off_t offset; int *rtvals; struct vnode *vp; struct thread *td; struct ucred *cred; struct nfsmount *nmp; struct nfsnode *np; vm_page_t *pages; vp = ap->a_vp; np = VTONFS(vp); td = curthread; /* XXX */ cred = curthread->td_ucred; /* XXX */ nmp = VFSTONFS(vp->v_mount); pages = ap->a_m; count = ap->a_count; rtvals = ap->a_rtvals; npages = btoc(count); offset = IDX_TO_OFF(pages[0]->pindex); mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); (void)nfs_fsinfo(nmp, vp, cred, td); } else mtx_unlock(&nmp->nm_mtx); mtx_lock(&np->n_mtx); if (nfs_directio_enable && !nfs_directio_allow_mmap && (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { mtx_unlock(&np->n_mtx); nfs_printf("nfs_putpages: called on noncache-able vnode??\n"); mtx_lock(&np->n_mtx); } for (i = 0; i < npages; i++) - rtvals[i] = VM_PAGER_AGAIN; + rtvals[i] = VM_PAGER_ERROR; /* * When putting pages, do not extend file past EOF. */ if (offset + count > np->n_size) { count = np->n_size - offset; if (count < 0) count = 0; } mtx_unlock(&np->n_mtx); /* * We use only the kva address for the buffer, but this is extremely * convienient and fast. */ bp = getpbuf(&nfs_pbuf_freecnt); kva = (vm_offset_t) bp->b_data; pmap_qenter(kva, pages, npages); PCPU_INC(cnt.v_vnodeout); PCPU_ADD(cnt.v_vnodepgsout, count); iov.iov_base = (caddr_t) kva; iov.iov_len = count; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = offset; uio.uio_resid = count; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_WRITE; uio.uio_td = td; if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) iomode = NFSV3WRITE_UNSTABLE; else iomode = NFSV3WRITE_FILESYNC; error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit); pmap_qremove(kva, npages); relpbuf(bp, &nfs_pbuf_freecnt); if (!error) { - int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; - for (i = 0; i < nwritten; i++) { - rtvals[i] = VM_PAGER_OK; - vm_page_undirty(pages[i]); - } + vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid); if (must_commit) { nfs_clearcommit(vp->v_mount); } } return rtvals[0]; } /* * For nfs, cache consistency can only be maintained approximately. * Although RFC1094 does not specify the criteria, the following is * believed to be compatible with the reference port. * For nfs: * If the file's modify time on the server has changed since the * last read rpc or you have written to the file, * you may have lost data cache consistency with the * server, so flush all of the file's data out of the cache. * Then force a getattr rpc to ensure that you have up to date * attributes. * NB: This implies that cache data can be read when up to * NFS_ATTRTIMEO seconds out of date. If you find that you need current * attributes this could be forced by setting n_attrstamp to 0 before * the VOP_GETATTR() call. */ static inline int nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) { int error = 0; struct vattr vattr; struct nfsnode *np = VTONFS(vp); int old_lock; struct nfsmount *nmp = VFSTONFS(vp->v_mount); /* * Grab the exclusive lock before checking whether the cache is * consistent. * XXX - We can make this cheaper later (by acquiring cheaper locks). * But for now, this suffices. */ old_lock = nfs_upgrade_vnlock(vp); if (vp->v_iflag & VI_DOOMED) { nfs_downgrade_vnlock(vp, old_lock); return (EBADF); } mtx_lock(&np->n_mtx); if (np->n_flag & NMODIFIED) { mtx_unlock(&np->n_mtx); if (vp->v_type != VREG) { if (vp->v_type != VDIR) panic("nfs: bioread, not dir"); (nmp->nm_rpcops->nr_invaldir)(vp); error = nfs_vinvalbuf(vp, V_SAVE, td, 1); if (error) goto out; } np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); error = VOP_GETATTR(vp, &vattr, cred); if (error) goto out; mtx_lock(&np->n_mtx); np->n_mtime = vattr.va_mtime; mtx_unlock(&np->n_mtx); } else { mtx_unlock(&np->n_mtx); error = VOP_GETATTR(vp, &vattr, cred); if (error) return (error); mtx_lock(&np->n_mtx); if ((np->n_flag & NSIZECHANGED) || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { mtx_unlock(&np->n_mtx); if (vp->v_type == VDIR) (nmp->nm_rpcops->nr_invaldir)(vp); error = nfs_vinvalbuf(vp, V_SAVE, td, 1); if (error) goto out; mtx_lock(&np->n_mtx); np->n_mtime = vattr.va_mtime; np->n_flag &= ~NSIZECHANGED; } mtx_unlock(&np->n_mtx); } out: nfs_downgrade_vnlock(vp, old_lock); return error; } /* * Vnode op for read using bio */ int nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) { struct nfsnode *np = VTONFS(vp); int biosize, i; struct buf *bp, *rabp; struct thread *td; struct nfsmount *nmp = VFSTONFS(vp->v_mount); daddr_t lbn, rabn; int bcount; int seqcount; int nra, error = 0, n = 0, on = 0; KASSERT(uio->uio_rw == UIO_READ, ("nfs_read mode")); if (uio->uio_resid == 0) return (0); if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ return (EINVAL); td = uio->uio_td; mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); (void)nfs_fsinfo(nmp, vp, cred, td); } else mtx_unlock(&nmp->nm_mtx); if (vp->v_type != VDIR && (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) return (EFBIG); if (nfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) /* No caching/ no readaheads. Just read data into the user buffer */ return nfs_readrpc(vp, uio, cred); biosize = vp->v_mount->mnt_stat.f_iosize; seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); error = nfs_bioread_check_cons(vp, td, cred); if (error) return error; do { u_quad_t nsize; mtx_lock(&np->n_mtx); nsize = np->n_size; mtx_unlock(&np->n_mtx); switch (vp->v_type) { case VREG: nfsstats.biocache_reads++; lbn = uio->uio_offset / biosize; on = uio->uio_offset & (biosize - 1); /* * Start the read ahead(s), as required. */ if (nmp->nm_readahead > 0) { for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { rabn = lbn + 1 + nra; if (incore(&vp->v_bufobj, rabn) == NULL) { rabp = nfs_getcacheblk(vp, rabn, biosize, td); if (!rabp) { error = nfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { rabp->b_flags |= B_ASYNC; rabp->b_iocmd = BIO_READ; vfs_busy_pages(rabp, 0); if (nfs_asyncio(nmp, rabp, cred, td)) { rabp->b_flags |= B_INVAL; rabp->b_ioflags |= BIO_ERROR; vfs_unbusy_pages(rabp); brelse(rabp); break; } } else { brelse(rabp); } } } } /* Note that bcount is *not* DEV_BSIZE aligned. */ bcount = biosize; if ((off_t)lbn * biosize >= nsize) { bcount = 0; } else if ((off_t)(lbn + 1) * biosize > nsize) { bcount = nsize - (off_t)lbn * biosize; } bp = nfs_getcacheblk(vp, lbn, bcount, td); if (!bp) { error = nfs_sigintr(nmp, td); return (error ? error : EINTR); } /* * If B_CACHE is not set, we must issue the read. If this * fails, we return an error. */ if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = nfs_doio(vp, bp, cred, td); if (error) { brelse(bp); return (error); } } /* * on is the offset into the current bp. Figure out how many * bytes we can copy out of the bp. Note that bcount is * NOT DEV_BSIZE aligned. * * Then figure out how many bytes we can copy into the uio. */ n = 0; if (on < bcount) n = min((unsigned)(bcount - on), uio->uio_resid); break; case VLNK: nfsstats.biocache_readlinks++; bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); if (!bp) { error = nfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = nfs_doio(vp, bp, cred, td); if (error) { bp->b_ioflags |= BIO_ERROR; brelse(bp); return (error); } } n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); on = 0; break; case VDIR: nfsstats.biocache_readdirs++; if (np->n_direofoffset && uio->uio_offset >= np->n_direofoffset) { return (0); } lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); if (!bp) { error = nfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = nfs_doio(vp, bp, cred, td); if (error) { brelse(bp); } while (error == NFSERR_BAD_COOKIE) { (nmp->nm_rpcops->nr_invaldir)(vp); error = nfs_vinvalbuf(vp, 0, td, 1); /* * Yuck! The directory has been modified on the * server. The only way to get the block is by * reading from the beginning to get all the * offset cookies. * * Leave the last bp intact unless there is an error. * Loop back up to the while if the error is another * NFSERR_BAD_COOKIE (double yuch!). */ for (i = 0; i <= lbn && !error; i++) { if (np->n_direofoffset && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) return (0); bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); if (!bp) { error = nfs_sigintr(nmp, td); return (error ? error : EINTR); } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = nfs_doio(vp, bp, cred, td); /* * no error + B_INVAL == directory EOF, * use the block. */ if (error == 0 && (bp->b_flags & B_INVAL)) break; } /* * An error will throw away the block and the * for loop will break out. If no error and this * is not the block we want, we throw away the * block and go for the next one via the for loop. */ if (error || i < lbn) brelse(bp); } } /* * The above while is repeated if we hit another cookie * error. If we hit an error and it wasn't a cookie error, * we give up. */ if (error) return (error); } /* * If not eof and read aheads are enabled, start one. * (You need the current block first, so that you have the * directory offset cookie of the next block.) */ if (nmp->nm_readahead > 0 && (bp->b_flags & B_INVAL) == 0 && (np->n_direofoffset == 0 || (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && incore(&vp->v_bufobj, lbn + 1) == NULL) { rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); if (rabp) { if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { rabp->b_flags |= B_ASYNC; rabp->b_iocmd = BIO_READ; vfs_busy_pages(rabp, 0); if (nfs_asyncio(nmp, rabp, cred, td)) { rabp->b_flags |= B_INVAL; rabp->b_ioflags |= BIO_ERROR; vfs_unbusy_pages(rabp); brelse(rabp); } } else { brelse(rabp); } } } /* * Unlike VREG files, whos buffer size ( bp->b_bcount ) is * chopped for the EOF condition, we cannot tell how large * NFS directories are going to be until we hit EOF. So * an NFS directory buffer is *not* chopped to its EOF. Now, * it just so happens that b_resid will effectively chop it * to EOF. *BUT* this information is lost if the buffer goes * away and is reconstituted into a B_CACHE state ( due to * being VMIO ) later. So we keep track of the directory eof * in np->n_direofoffset and chop it off as an extra step * right here. */ n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) n = np->n_direofoffset - uio->uio_offset; break; default: nfs_printf(" nfs_bioread: type %x unexpected\n", vp->v_type); bp = NULL; break; }; if (n > 0) { error = uiomove(bp->b_data + on, (int)n, uio); } if (vp->v_type == VLNK) n = 0; if (bp != NULL) brelse(bp); } while (error == 0 && uio->uio_resid > 0 && n > 0); return (error); } /* * The NFS write path cannot handle iovecs with len > 1. So we need to * break up iovecs accordingly (restricting them to wsize). * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). * For the ASYNC case, 2 copies are needed. The first a copy from the * user buffer to a staging buffer and then a second copy from the staging * buffer to mbufs. This can be optimized by copying from the user buffer * directly into mbufs and passing the chain down, but that requires a * fair amount of re-working of the relevant codepaths (and can be done * later). */ static int nfs_directio_write(vp, uiop, cred, ioflag) struct vnode *vp; struct uio *uiop; struct ucred *cred; int ioflag; { int error; struct nfsmount *nmp = VFSTONFS(vp->v_mount); struct thread *td = uiop->uio_td; int size; int wsize; mtx_lock(&nmp->nm_mtx); wsize = nmp->nm_wsize; mtx_unlock(&nmp->nm_mtx); if (ioflag & IO_SYNC) { int iomode, must_commit; struct uio uio; struct iovec iov; do_sync: while (uiop->uio_resid > 0) { size = min(uiop->uio_resid, wsize); size = min(uiop->uio_iov->iov_len, size); iov.iov_base = uiop->uio_iov->iov_base; iov.iov_len = size; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = uiop->uio_offset; uio.uio_resid = size; uio.uio_segflg = UIO_USERSPACE; uio.uio_rw = UIO_WRITE; uio.uio_td = td; iomode = NFSV3WRITE_FILESYNC; error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit); KASSERT((must_commit == 0), ("nfs_directio_write: Did not commit write")); if (error) return (error); uiop->uio_offset += size; uiop->uio_resid -= size; if (uiop->uio_iov->iov_len <= size) { uiop->uio_iovcnt--; uiop->uio_iov++; } else { uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + size; uiop->uio_iov->iov_len -= size; } } } else { struct uio *t_uio; struct iovec *t_iov; struct buf *bp; /* * Break up the write into blocksize chunks and hand these * over to nfsiod's for write back. * Unfortunately, this incurs a copy of the data. Since * the user could modify the buffer before the write is * initiated. * * The obvious optimization here is that one of the 2 copies * in the async write path can be eliminated by copying the * data here directly into mbufs and passing the mbuf chain * down. But that will require a fair amount of re-working * of the code and can be done if there's enough interest * in NFS directio access. */ while (uiop->uio_resid > 0) { size = min(uiop->uio_resid, wsize); size = min(uiop->uio_iov->iov_len, size); bp = getpbuf(&nfs_pbuf_freecnt); t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); t_iov->iov_len = size; t_uio->uio_iov = t_iov; t_uio->uio_iovcnt = 1; t_uio->uio_offset = uiop->uio_offset; t_uio->uio_resid = size; t_uio->uio_segflg = UIO_SYSSPACE; t_uio->uio_rw = UIO_WRITE; t_uio->uio_td = td; bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); bp->b_flags |= B_DIRECT; bp->b_iocmd = BIO_WRITE; if (cred != NOCRED) { crhold(cred); bp->b_wcred = cred; } else bp->b_wcred = NOCRED; bp->b_caller1 = (void *)t_uio; bp->b_vp = vp; error = nfs_asyncio(nmp, bp, NOCRED, td); if (error) { free(t_iov->iov_base, M_NFSDIRECTIO); free(t_iov, M_NFSDIRECTIO); free(t_uio, M_NFSDIRECTIO); bp->b_vp = NULL; relpbuf(bp, &nfs_pbuf_freecnt); if (error == EINTR) return (error); goto do_sync; } uiop->uio_offset += size; uiop->uio_resid -= size; if (uiop->uio_iov->iov_len <= size) { uiop->uio_iovcnt--; uiop->uio_iov++; } else { uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base + size; uiop->uio_iov->iov_len -= size; } } } return (0); } /* * Vnode op for write using bio */ int nfs_write(struct vop_write_args *ap) { int biosize; struct uio *uio = ap->a_uio; struct thread *td = uio->uio_td; struct vnode *vp = ap->a_vp; struct nfsnode *np = VTONFS(vp); struct ucred *cred = ap->a_cred; int ioflag = ap->a_ioflag; struct buf *bp; struct vattr vattr; struct nfsmount *nmp = VFSTONFS(vp->v_mount); daddr_t lbn; int bcount; int n, on, error = 0; KASSERT(uio->uio_rw == UIO_WRITE, ("nfs_write mode")); KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, ("nfs_write proc")); if (vp->v_type != VREG) return (EIO); mtx_lock(&np->n_mtx); if (np->n_flag & NWRITEERR) { np->n_flag &= ~NWRITEERR; mtx_unlock(&np->n_mtx); return (np->n_error); } else mtx_unlock(&np->n_mtx); mtx_lock(&nmp->nm_mtx); if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { mtx_unlock(&nmp->nm_mtx); (void)nfs_fsinfo(nmp, vp, cred, td); } else mtx_unlock(&nmp->nm_mtx); /* * Synchronously flush pending buffers if we are in synchronous * mode or if we are appending. */ if (ioflag & (IO_APPEND | IO_SYNC)) { mtx_lock(&np->n_mtx); if (np->n_flag & NMODIFIED) { mtx_unlock(&np->n_mtx); #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ /* * Require non-blocking, synchronous writes to * dirty files to inform the program it needs * to fsync(2) explicitly. */ if (ioflag & IO_NDELAY) return (EAGAIN); #endif flush_and_restart: np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); error = nfs_vinvalbuf(vp, V_SAVE, td, 1); if (error) return (error); } else mtx_unlock(&np->n_mtx); } /* * If IO_APPEND then load uio_offset. We restart here if we cannot * get the append lock. */ if (ioflag & IO_APPEND) { np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); error = VOP_GETATTR(vp, &vattr, cred); if (error) return (error); mtx_lock(&np->n_mtx); uio->uio_offset = np->n_size; mtx_unlock(&np->n_mtx); } if (uio->uio_offset < 0) return (EINVAL); if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) return (EFBIG); if (uio->uio_resid == 0) return (0); if (nfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) return nfs_directio_write(vp, uio, cred, ioflag); /* * Maybe this should be above the vnode op call, but so long as * file servers have no limits, i don't think it matters */ if (vn_rlimit_fsize(vp, uio, td)) return (EFBIG); biosize = vp->v_mount->mnt_stat.f_iosize; /* * Find all of this file's B_NEEDCOMMIT buffers. If our writes * would exceed the local maximum per-file write commit size when * combined with those, we must decide whether to flush, * go synchronous, or return error. We don't bother checking * IO_UNIT -- we just make all writes atomic anyway, as there's * no point optimizing for something that really won't ever happen. */ if (!(ioflag & IO_SYNC)) { int nflag; mtx_lock(&np->n_mtx); nflag = np->n_flag; mtx_unlock(&np->n_mtx); int needrestart = 0; if (nmp->nm_wcommitsize < uio->uio_resid) { /* * If this request could not possibly be completed * without exceeding the maximum outstanding write * commit size, see if we can convert it into a * synchronous write operation. */ if (ioflag & IO_NDELAY) return (EAGAIN); ioflag |= IO_SYNC; if (nflag & NMODIFIED) needrestart = 1; } else if (nflag & NMODIFIED) { int wouldcommit = 0; BO_LOCK(&vp->v_bufobj); if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { if (bp->b_flags & B_NEEDCOMMIT) wouldcommit += bp->b_bcount; } } BO_UNLOCK(&vp->v_bufobj); /* * Since we're not operating synchronously and * bypassing the buffer cache, we are in a commit * and holding all of these buffers whether * transmitted or not. If not limited, this * will lead to the buffer cache deadlocking, * as no one else can flush our uncommitted buffers. */ wouldcommit += uio->uio_resid; /* * If we would initially exceed the maximum * outstanding write commit size, flush and restart. */ if (wouldcommit > nmp->nm_wcommitsize) needrestart = 1; } if (needrestart) goto flush_and_restart; } do { nfsstats.biocache_writes++; lbn = uio->uio_offset / biosize; on = uio->uio_offset & (biosize-1); n = min((unsigned)(biosize - on), uio->uio_resid); again: /* * Handle direct append and file extension cases, calculate * unaligned buffer size. */ mtx_lock(&np->n_mtx); if (uio->uio_offset == np->n_size && n) { mtx_unlock(&np->n_mtx); /* * Get the buffer (in its pre-append state to maintain * B_CACHE if it was previously set). Resize the * nfsnode after we have locked the buffer to prevent * readers from reading garbage. */ bcount = on; bp = nfs_getcacheblk(vp, lbn, bcount, td); if (bp != NULL) { long save; mtx_lock(&np->n_mtx); np->n_size = uio->uio_offset + n; np->n_flag |= NMODIFIED; vnode_pager_setsize(vp, np->n_size); mtx_unlock(&np->n_mtx); save = bp->b_flags & B_CACHE; bcount += n; allocbuf(bp, bcount); bp->b_flags |= save; } } else { /* * Obtain the locked cache block first, and then * adjust the file's size as appropriate. */ bcount = on + n; if ((off_t)lbn * biosize + bcount < np->n_size) { if ((off_t)(lbn + 1) * biosize < np->n_size) bcount = biosize; else bcount = np->n_size - (off_t)lbn * biosize; } mtx_unlock(&np->n_mtx); bp = nfs_getcacheblk(vp, lbn, bcount, td); mtx_lock(&np->n_mtx); if (uio->uio_offset + n > np->n_size) { np->n_size = uio->uio_offset + n; np->n_flag |= NMODIFIED; vnode_pager_setsize(vp, np->n_size); } mtx_unlock(&np->n_mtx); } if (!bp) { error = nfs_sigintr(nmp, td); if (!error) error = EINTR; break; } /* * Issue a READ if B_CACHE is not set. In special-append * mode, B_CACHE is based on the buffer prior to the write * op and is typically set, avoiding the read. If a read * is required in special append mode, the server will * probably send us a short-read since we extended the file * on our end, resulting in b_resid == 0 and, thusly, * B_CACHE getting set. * * We can also avoid issuing the read if the write covers * the entire buffer. We have to make sure the buffer state * is reasonable in this case since we will not be initiating * I/O. See the comments in kern/vfs_bio.c's getblk() for * more information. * * B_CACHE may also be set due to the buffer being cached * normally. */ if (on == 0 && n == bcount) { bp->b_flags |= B_CACHE; bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); error = nfs_doio(vp, bp, cred, td); if (error) { brelse(bp); break; } } if (bp->b_wcred == NOCRED) bp->b_wcred = crhold(cred); mtx_lock(&np->n_mtx); np->n_flag |= NMODIFIED; mtx_unlock(&np->n_mtx); /* * If dirtyend exceeds file size, chop it down. This should * not normally occur but there is an append race where it * might occur XXX, so we log it. * * If the chopping creates a reverse-indexed or degenerate * situation with dirtyoff/end, we 0 both of them. */ if (bp->b_dirtyend > bcount) { nfs_printf("NFS append race @%lx:%d\n", (long)bp->b_blkno * DEV_BSIZE, bp->b_dirtyend - bcount); bp->b_dirtyend = bcount; } if (bp->b_dirtyoff >= bp->b_dirtyend) bp->b_dirtyoff = bp->b_dirtyend = 0; /* * If the new write will leave a contiguous dirty * area, just update the b_dirtyoff and b_dirtyend, * otherwise force a write rpc of the old dirty area. * * While it is possible to merge discontiguous writes due to * our having a B_CACHE buffer ( and thus valid read data * for the hole), we don't because it could lead to * significant cache coherency problems with multiple clients, * especially if locking is implemented later on. * * as an optimization we could theoretically maintain * a linked list of discontinuous areas, but we would still * have to commit them separately so there isn't much * advantage to it except perhaps a bit of asynchronization. */ if (bp->b_dirtyend > 0 && (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { if (bwrite(bp) == EINTR) { error = EINTR; break; } goto again; } error = uiomove((char *)bp->b_data + on, n, uio); /* * Since this block is being modified, it must be written * again and not just committed. Since write clustering does * not work for the stage 1 data write, only the stage 2 * commit rpc, we have to clear B_CLUSTEROK as well. */ bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); if (error) { bp->b_ioflags |= BIO_ERROR; brelse(bp); break; } /* * Only update dirtyoff/dirtyend if not a degenerate * condition. */ if (n) { if (bp->b_dirtyend > 0) { bp->b_dirtyoff = min(on, bp->b_dirtyoff); bp->b_dirtyend = max((on + n), bp->b_dirtyend); } else { bp->b_dirtyoff = on; bp->b_dirtyend = on + n; } vfs_bio_set_valid(bp, on, n); } /* * If IO_SYNC do bwrite(). * * IO_INVAL appears to be unused. The idea appears to be * to turn off caching in this case. Very odd. XXX */ if ((ioflag & IO_SYNC)) { if (ioflag & IO_INVAL) bp->b_flags |= B_NOCACHE; error = bwrite(bp); if (error) break; } else if ((n + on) == biosize) { bp->b_flags |= B_ASYNC; (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, NULL); } else { bdwrite(bp); } } while (uio->uio_resid > 0 && n > 0); return (error); } /* * Get an nfs cache block. * * Allocate a new one if the block isn't currently in the cache * and return the block marked busy. If the calling process is * interrupted by a signal for an interruptible mount point, return * NULL. * * The caller must carefully deal with the possible B_INVAL state of * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it * indirectly), so synchronous reads can be issued without worrying about * the B_INVAL state. We have to be a little more careful when dealing * with writes (see comments in nfs_write()) when extending a file past * its EOF. */ static struct buf * nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) { struct buf *bp; struct mount *mp; struct nfsmount *nmp; mp = vp->v_mount; nmp = VFSTONFS(mp); if (nmp->nm_flag & NFSMNT_INT) { sigset_t oldset; nfs_set_sigmask(td, &oldset); bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0); nfs_restore_sigmask(td, &oldset); while (bp == NULL) { if (nfs_sigintr(nmp, td)) return (NULL); bp = getblk(vp, bn, size, 0, 2 * hz, 0); } } else { bp = getblk(vp, bn, size, 0, 0, 0); } if (vp->v_type == VREG) { int biosize; biosize = mp->mnt_stat.f_iosize; bp->b_blkno = bn * (biosize / DEV_BSIZE); } return (bp); } /* * Flush and invalidate all dirty buffers. If another process is already * doing the flush, just wait for completion. */ int nfs_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) { struct nfsnode *np = VTONFS(vp); struct nfsmount *nmp = VFSTONFS(vp->v_mount); int error = 0, slpflag, slptimeo; int old_lock = 0; ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf"); if ((nmp->nm_flag & NFSMNT_INT) == 0) intrflg = 0; if (intrflg) { slpflag = NFS_PCATCH; slptimeo = 2 * hz; } else { slpflag = 0; slptimeo = 0; } old_lock = nfs_upgrade_vnlock(vp); if (vp->v_iflag & VI_DOOMED) { /* * Since vgonel() uses the generic vinvalbuf() to flush * dirty buffers and it does not call this function, it * is safe to just return OK when VI_DOOMED is set. */ nfs_downgrade_vnlock(vp, old_lock); return (0); } /* * Now, flush as required. */ if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { VM_OBJECT_LOCK(vp->v_bufobj.bo_object); vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); /* * If the page clean was interrupted, fail the invalidation. * Not doing so, we run the risk of losing dirty pages in the * vinvalbuf() call below. */ if (intrflg && (error = nfs_sigintr(nmp, td))) goto out; } error = vinvalbuf(vp, flags, slpflag, 0); while (error) { if (intrflg && (error = nfs_sigintr(nmp, td))) goto out; error = vinvalbuf(vp, flags, 0, slptimeo); } mtx_lock(&np->n_mtx); if (np->n_directio_asyncwr == 0) np->n_flag &= ~NMODIFIED; mtx_unlock(&np->n_mtx); out: nfs_downgrade_vnlock(vp, old_lock); return error; } /* * Initiate asynchronous I/O. Return an error if no nfsiods are available. * This is mainly to avoid queueing async I/O requests when the nfsiods * are all hung on a dead server. * * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp * is eventually dequeued by the async daemon, nfs_doio() *will*. */ int nfs_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) { int iod; int gotiod; int slpflag = 0; int slptimeo = 0; int error, error2; /* * Commits are usually short and sweet so lets save some cpu and * leave the async daemons for more important rpc's (such as reads * and writes). */ mtx_lock(&nfs_iod_mtx); if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && (nmp->nm_bufqiods > nfs_numasync / 2)) { mtx_unlock(&nfs_iod_mtx); return(EIO); } again: if (nmp->nm_flag & NFSMNT_INT) slpflag = NFS_PCATCH; gotiod = FALSE; /* * Find a free iod to process this request. */ for (iod = 0; iod < nfs_numasync; iod++) if (nfs_iodwant[iod] == NFSIOD_AVAILABLE) { gotiod = TRUE; break; } /* * Try to create one if none are free. */ if (!gotiod) nfs_nfsiodnew(); else { /* * Found one, so wake it up and tell it which * mount to process. */ NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n", iod, nmp)); nfs_iodwant[iod] = NFSIOD_NOT_AVAILABLE; nfs_iodmount[iod] = nmp; nmp->nm_bufqiods++; wakeup(&nfs_iodwant[iod]); } /* * If none are free, we may already have an iod working on this mount * point. If so, it will process our request. */ if (!gotiod) { if (nmp->nm_bufqiods > 0) { NFS_DPF(ASYNCIO, ("nfs_asyncio: %d iods are already processing mount %p\n", nmp->nm_bufqiods, nmp)); gotiod = TRUE; } } /* * If we have an iod which can process the request, then queue * the buffer. */ if (gotiod) { /* * Ensure that the queue never grows too large. We still want * to asynchronize so we block rather then return EIO. */ while (nmp->nm_bufqlen >= 2 * nfs_numasync) { NFS_DPF(ASYNCIO, ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); nmp->nm_bufqwant = TRUE; error = nfs_msleep(td, &nmp->nm_bufq, &nfs_iod_mtx, slpflag | PRIBIO, "nfsaio", slptimeo); if (error) { error2 = nfs_sigintr(nmp, td); if (error2) { mtx_unlock(&nfs_iod_mtx); return (error2); } if (slpflag == NFS_PCATCH) { slpflag = 0; slptimeo = 2 * hz; } } /* * We might have lost our iod while sleeping, * so check and loop if nescessary. */ goto again; } /* We might have lost our nfsiod */ if (nmp->nm_bufqiods == 0) { NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); goto again; } if (bp->b_iocmd == BIO_READ) { if (bp->b_rcred == NOCRED && cred != NOCRED) bp->b_rcred = crhold(cred); } else { if (bp->b_wcred == NOCRED && cred != NOCRED) bp->b_wcred = crhold(cred); } if (bp->b_flags & B_REMFREE) bremfreef(bp); BUF_KERNPROC(bp); TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); nmp->nm_bufqlen++; if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); VTONFS(bp->b_vp)->n_flag |= NMODIFIED; VTONFS(bp->b_vp)->n_directio_asyncwr++; mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); } mtx_unlock(&nfs_iod_mtx); return (0); } mtx_unlock(&nfs_iod_mtx); /* * All the iods are busy on other mounts, so return EIO to * force the caller to process the i/o synchronously. */ NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); return (EIO); } void nfs_doio_directwrite(struct buf *bp) { int iomode, must_commit; struct uio *uiop = (struct uio *)bp->b_caller1; char *iov_base = uiop->uio_iov->iov_base; struct nfsmount *nmp = VFSTONFS(bp->b_vp->v_mount); iomode = NFSV3WRITE_FILESYNC; uiop->uio_td = NULL; /* NULL since we're in nfsiod */ (nmp->nm_rpcops->nr_writerpc)(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); KASSERT((must_commit == 0), ("nfs_doio_directwrite: Did not commit write")); free(iov_base, M_NFSDIRECTIO); free(uiop->uio_iov, M_NFSDIRECTIO); free(uiop, M_NFSDIRECTIO); if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { struct nfsnode *np = VTONFS(bp->b_vp); mtx_lock(&np->n_mtx); np->n_directio_asyncwr--; if (np->n_directio_asyncwr == 0) { VTONFS(bp->b_vp)->n_flag &= ~NMODIFIED; if ((np->n_flag & NFSYNCWAIT)) { np->n_flag &= ~NFSYNCWAIT; wakeup((caddr_t)&np->n_directio_asyncwr); } } mtx_unlock(&np->n_mtx); } bp->b_vp = NULL; relpbuf(bp, &nfs_pbuf_freecnt); } /* * Do an I/O operation to/from a cache block. This may be called * synchronously or from an nfsiod. */ int nfs_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) { struct uio *uiop; struct nfsnode *np; struct nfsmount *nmp; int error = 0, iomode, must_commit = 0; struct uio uio; struct iovec io; struct proc *p = td ? td->td_proc : NULL; uint8_t iocmd; np = VTONFS(vp); nmp = VFSTONFS(vp->v_mount); uiop = &uio; uiop->uio_iov = &io; uiop->uio_iovcnt = 1; uiop->uio_segflg = UIO_SYSSPACE; uiop->uio_td = td; /* * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We * do this here so we do not have to do it in all the code that * calls us. */ bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); iocmd = bp->b_iocmd; if (iocmd == BIO_READ) { io.iov_len = uiop->uio_resid = bp->b_bcount; io.iov_base = bp->b_data; uiop->uio_rw = UIO_READ; switch (vp->v_type) { case VREG: uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; nfsstats.read_bios++; error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr); if (!error) { if (uiop->uio_resid) { /* * If we had a short read with no error, we must have * hit a file hole. We should zero-fill the remainder. * This can also occur if the server hits the file EOF. * * Holes used to be able to occur due to pending * writes, but that is not possible any longer. */ int nread = bp->b_bcount - uiop->uio_resid; int left = uiop->uio_resid; if (left > 0) bzero((char *)bp->b_data + nread, left); uiop->uio_resid = 0; } } /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */ if (p && (vp->v_vflag & VV_TEXT)) { mtx_lock(&np->n_mtx); if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.va_mtime)) { mtx_unlock(&np->n_mtx); PROC_LOCK(p); killproc(p, "text file modification"); PROC_UNLOCK(p); } else mtx_unlock(&np->n_mtx); } break; case VLNK: uiop->uio_offset = (off_t)0; nfsstats.readlink_bios++; error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr); break; case VDIR: nfsstats.readdir_bios++; uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { error = nfs_readdirplusrpc(vp, uiop, cr); if (error == NFSERR_NOTSUPP) nmp->nm_flag &= ~NFSMNT_RDIRPLUS; } if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) error = nfs_readdirrpc(vp, uiop, cr); /* * end-of-directory sets B_INVAL but does not generate an * error. */ if (error == 0 && uiop->uio_resid == bp->b_bcount) bp->b_flags |= B_INVAL; break; default: nfs_printf("nfs_doio: type %x unexpected\n", vp->v_type); break; }; if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_error = error; } } else { /* * If we only need to commit, try to commit */ if (bp->b_flags & B_NEEDCOMMIT) { int retv; off_t off; off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; retv = (nmp->nm_rpcops->nr_commit)( vp, off, bp->b_dirtyend-bp->b_dirtyoff, bp->b_wcred, td); if (retv == 0) { bp->b_dirtyoff = bp->b_dirtyend = 0; bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); bp->b_resid = 0; bufdone(bp); return (0); } if (retv == NFSERR_STALEWRITEVERF) { nfs_clearcommit(vp->v_mount); } } /* * Setup for actual write */ mtx_lock(&np->n_mtx); if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; mtx_unlock(&np->n_mtx); if (bp->b_dirtyend > bp->b_dirtyoff) { io.iov_len = uiop->uio_resid = bp->b_dirtyend - bp->b_dirtyoff; uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyoff; io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; uiop->uio_rw = UIO_WRITE; nfsstats.write_bios++; if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) iomode = NFSV3WRITE_UNSTABLE; else iomode = NFSV3WRITE_FILESYNC; error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit); /* * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try * to cluster the buffers needing commit. This will allow * the system to submit a single commit rpc for the whole * cluster. We can do this even if the buffer is not 100% * dirty (relative to the NFS blocksize), so we optimize the * append-to-file-case. * * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be * cleared because write clustering only works for commit * rpc's, not for the data portion of the write). */ if (!error && iomode == NFSV3WRITE_UNSTABLE) { bp->b_flags |= B_NEEDCOMMIT; if (bp->b_dirtyoff == 0 && bp->b_dirtyend == bp->b_bcount) bp->b_flags |= B_CLUSTEROK; } else { bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); } /* * For an interrupted write, the buffer is still valid * and the write hasn't been pushed to the server yet, * so we can't set BIO_ERROR and report the interruption * by setting B_EINTR. For the B_ASYNC case, B_EINTR * is not relevant, so the rpc attempt is essentially * a noop. For the case of a V3 write rpc not being * committed to stable storage, the block is still * dirty and requires either a commit rpc or another * write rpc with iomode == NFSV3WRITE_FILESYNC before * the block is reused. This is indicated by setting * the B_DELWRI and B_NEEDCOMMIT flags. * * If the buffer is marked B_PAGING, it does not reside on * the vp's paging queues so we cannot call bdirty(). The * bp in this case is not an NFS cache block so we should * be safe. XXX * * The logic below breaks up errors into recoverable and * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE * and keep the buffer around for potential write retries. * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) * and save the error in the nfsnode. This is less than ideal * but necessary. Keeping such buffers around could potentially * cause buffer exhaustion eventually (they can never be written * out, so will get constantly be re-dirtied). It also causes * all sorts of vfs panics. For non-recoverable write errors, * also invalidate the attrcache, so we'll be forced to go over * the wire for this object, returning an error to user on next * call (most of the time). */ if (error == EINTR || error == EIO || error == ETIMEDOUT || (!error && (bp->b_flags & B_NEEDCOMMIT))) { int s; s = splbio(); bp->b_flags &= ~(B_INVAL|B_NOCACHE); if ((bp->b_flags & B_PAGING) == 0) { bdirty(bp); bp->b_flags &= ~B_DONE; } if (error && (bp->b_flags & B_ASYNC) == 0) bp->b_flags |= B_EINTR; splx(s); } else { if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_flags |= B_INVAL; bp->b_error = np->n_error = error; mtx_lock(&np->n_mtx); np->n_flag |= NWRITEERR; np->n_attrstamp = 0; KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); mtx_unlock(&np->n_mtx); } bp->b_dirtyoff = bp->b_dirtyend = 0; } } else { bp->b_resid = 0; bufdone(bp); return (0); } } bp->b_resid = uiop->uio_resid; if (must_commit) nfs_clearcommit(vp->v_mount); bufdone(bp); return (error); } /* * Used to aid in handling ftruncate() operations on the NFS client side. * Truncation creates a number of special problems for NFS. We have to * throw away VM pages and buffer cache buffers that are beyond EOF, and * we have to properly handle VM pages or (potentially dirty) buffers * that straddle the truncation point. */ int nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) { struct nfsnode *np = VTONFS(vp); u_quad_t tsize; int biosize = vp->v_mount->mnt_stat.f_iosize; int error = 0; mtx_lock(&np->n_mtx); tsize = np->n_size; np->n_size = nsize; mtx_unlock(&np->n_mtx); if (nsize < tsize) { struct buf *bp; daddr_t lbn; int bufsize; /* * vtruncbuf() doesn't get the buffer overlapping the * truncation point. We may have a B_DELWRI and/or B_CACHE * buffer that now needs to be truncated. */ error = vtruncbuf(vp, cred, td, nsize, biosize); lbn = nsize / biosize; bufsize = nsize & (biosize - 1); bp = nfs_getcacheblk(vp, lbn, bufsize, td); if (!bp) return EINTR; if (bp->b_dirtyoff > bp->b_bcount) bp->b_dirtyoff = bp->b_bcount; if (bp->b_dirtyend > bp->b_bcount) bp->b_dirtyend = bp->b_bcount; bp->b_flags |= B_RELBUF; /* don't leave garbage around */ brelse(bp); } else { vnode_pager_setsize(vp, nsize); } return(error); } Index: head/sys/vm/vm_object.c =================================================================== --- head/sys/vm/vm_object.c (revision 222585) +++ head/sys/vm/vm_object.c (revision 222586) @@ -1,2259 +1,2274 @@ /*- * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * 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. * 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. * * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Virtual memory object module. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include /* for curproc, pageproc */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static int old_msync; SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, "Use old (insecure) msync behavior"); static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, int flags, int *clearobjflags); static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, int *clearobjflags); static void vm_object_qcollapse(vm_object_t object); static void vm_object_vndeallocate(vm_object_t object); /* * Virtual memory objects maintain the actual data * associated with allocated virtual memory. A given * page of memory exists within exactly one object. * * An object is only deallocated when all "references" * are given up. Only one "reference" to a given * region of an object should be writeable. * * Associated with each object is a list of all resident * memory pages belonging to that object; this list is * maintained by the "vm_page" module, and locked by the object's * lock. * * Each object also records a "pager" routine which is * used to retrieve (and store) pages to the proper backing * storage. In addition, objects may be backed by other * objects from which they were virtual-copied. * * The only items within the object structure which are * modified after time of creation are: * reference count locked by object's lock * pager routine locked by object's lock * */ struct object_q vm_object_list; struct mtx vm_object_list_mtx; /* lock for object list and count */ struct vm_object kernel_object_store; struct vm_object kmem_object_store; SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats"); static long object_collapses; SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, &object_collapses, 0, "VM object collapses"); static long object_bypasses; SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, &object_bypasses, 0, "VM object bypasses"); static uma_zone_t obj_zone; static int vm_object_zinit(void *mem, int size, int flags); #ifdef INVARIANTS static void vm_object_zdtor(void *mem, int size, void *arg); static void vm_object_zdtor(void *mem, int size, void *arg) { vm_object_t object; object = (vm_object_t)mem; KASSERT(TAILQ_EMPTY(&object->memq), ("object %p has resident pages", object)); #if VM_NRESERVLEVEL > 0 KASSERT(LIST_EMPTY(&object->rvq), ("object %p has reservations", object)); #endif KASSERT(object->cache == NULL, ("object %p has cached pages", object)); KASSERT(object->paging_in_progress == 0, ("object %p paging_in_progress = %d", object, object->paging_in_progress)); KASSERT(object->resident_page_count == 0, ("object %p resident_page_count = %d", object, object->resident_page_count)); KASSERT(object->shadow_count == 0, ("object %p shadow_count = %d", object, object->shadow_count)); } #endif static int vm_object_zinit(void *mem, int size, int flags) { vm_object_t object; object = (vm_object_t)mem; bzero(&object->mtx, sizeof(object->mtx)); VM_OBJECT_LOCK_INIT(object, "standard object"); /* These are true for any object that has been freed */ object->paging_in_progress = 0; object->resident_page_count = 0; object->shadow_count = 0; return (0); } void _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) { TAILQ_INIT(&object->memq); LIST_INIT(&object->shadow_head); object->root = NULL; object->type = type; object->size = size; object->generation = 1; object->ref_count = 1; object->memattr = VM_MEMATTR_DEFAULT; object->flags = 0; object->cred = NULL; object->charge = 0; if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP)) object->flags = OBJ_ONEMAPPING; object->pg_color = 0; object->handle = NULL; object->backing_object = NULL; object->backing_object_offset = (vm_ooffset_t) 0; #if VM_NRESERVLEVEL > 0 LIST_INIT(&object->rvq); #endif object->cache = NULL; mtx_lock(&vm_object_list_mtx); TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); mtx_unlock(&vm_object_list_mtx); } /* * vm_object_init: * * Initialize the VM objects module. */ void vm_object_init(void) { TAILQ_INIT(&vm_object_list); mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); VM_OBJECT_LOCK_INIT(kernel_object, "kernel object"); _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), kernel_object); #if VM_NRESERVLEVEL > 0 kernel_object->flags |= OBJ_COLORED; kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); #endif VM_OBJECT_LOCK_INIT(kmem_object, "kmem object"); _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), kmem_object); #if VM_NRESERVLEVEL > 0 kmem_object->flags |= OBJ_COLORED; kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); #endif /* * The lock portion of struct vm_object must be type stable due * to vm_pageout_fallback_object_lock locking a vm object * without holding any references to it. */ obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, #ifdef INVARIANTS vm_object_zdtor, #else NULL, #endif vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE); } void vm_object_clear_flag(vm_object_t object, u_short bits) { VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); object->flags &= ~bits; } /* * Sets the default memory attribute for the specified object. Pages * that are allocated to this object are by default assigned this memory * attribute. * * Presently, this function must be called before any pages are allocated * to the object. In the future, this requirement may be relaxed for * "default" and "swap" objects. */ int vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) { VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); switch (object->type) { case OBJT_DEFAULT: case OBJT_DEVICE: case OBJT_PHYS: case OBJT_SG: case OBJT_SWAP: case OBJT_VNODE: if (!TAILQ_EMPTY(&object->memq)) return (KERN_FAILURE); break; case OBJT_DEAD: return (KERN_INVALID_ARGUMENT); } object->memattr = memattr; return (KERN_SUCCESS); } void vm_object_pip_add(vm_object_t object, short i) { VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); object->paging_in_progress += i; } void vm_object_pip_subtract(vm_object_t object, short i) { VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); object->paging_in_progress -= i; } void vm_object_pip_wakeup(vm_object_t object) { VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); object->paging_in_progress--; if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { vm_object_clear_flag(object, OBJ_PIPWNT); wakeup(object); } } void vm_object_pip_wakeupn(vm_object_t object, short i) { VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); if (i) object->paging_in_progress -= i; if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) { vm_object_clear_flag(object, OBJ_PIPWNT); wakeup(object); } } void vm_object_pip_wait(vm_object_t object, char *waitid) { VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); while (object->paging_in_progress) { object->flags |= OBJ_PIPWNT; msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0); } } /* * vm_object_allocate: * * Returns a new object with the given size. */ vm_object_t vm_object_allocate(objtype_t type, vm_pindex_t size) { vm_object_t object; object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); _vm_object_allocate(type, size, object); return (object); } /* * vm_object_reference: * * Gets another reference to the given object. Note: OBJ_DEAD * objects can be referenced during final cleaning. */ void vm_object_reference(vm_object_t object) { if (object == NULL) return; VM_OBJECT_LOCK(object); vm_object_reference_locked(object); VM_OBJECT_UNLOCK(object); } /* * vm_object_reference_locked: * * Gets another reference to the given object. * * The object must be locked. */ void vm_object_reference_locked(vm_object_t object) { struct vnode *vp; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); object->ref_count++; if (object->type == OBJT_VNODE) { vp = object->handle; vref(vp); } } /* * Handle deallocating an object of type OBJT_VNODE. */ static void vm_object_vndeallocate(vm_object_t object) { struct vnode *vp = (struct vnode *) object->handle; VFS_ASSERT_GIANT(vp->v_mount); VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); KASSERT(object->type == OBJT_VNODE, ("vm_object_vndeallocate: not a vnode object")); KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); #ifdef INVARIANTS if (object->ref_count == 0) { vprint("vm_object_vndeallocate", vp); panic("vm_object_vndeallocate: bad object reference count"); } #endif if (object->ref_count > 1) { object->ref_count--; VM_OBJECT_UNLOCK(object); /* vrele may need the vnode lock. */ vrele(vp); } else { vhold(vp); VM_OBJECT_UNLOCK(object); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); vdrop(vp); VM_OBJECT_LOCK(object); object->ref_count--; if (object->type == OBJT_DEAD) { VM_OBJECT_UNLOCK(object); VOP_UNLOCK(vp, 0); } else { if (object->ref_count == 0) vp->v_vflag &= ~VV_TEXT; VM_OBJECT_UNLOCK(object); vput(vp); } } } /* * vm_object_deallocate: * * Release a reference to the specified object, * gained either through a vm_object_allocate * or a vm_object_reference call. When all references * are gone, storage associated with this object * may be relinquished. * * No object may be locked. */ void vm_object_deallocate(vm_object_t object) { vm_object_t temp; while (object != NULL) { int vfslocked; vfslocked = 0; restart: VM_OBJECT_LOCK(object); if (object->type == OBJT_VNODE) { struct vnode *vp = (struct vnode *) object->handle; /* * Conditionally acquire Giant for a vnode-backed * object. We have to be careful since the type of * a vnode object can change while the object is * unlocked. */ if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) { vfslocked = 1; if (!mtx_trylock(&Giant)) { VM_OBJECT_UNLOCK(object); mtx_lock(&Giant); goto restart; } } vm_object_vndeallocate(object); VFS_UNLOCK_GIANT(vfslocked); return; } else /* * This is to handle the case that the object * changed type while we dropped its lock to * obtain Giant. */ VFS_UNLOCK_GIANT(vfslocked); KASSERT(object->ref_count != 0, ("vm_object_deallocate: object deallocated too many times: %d", object->type)); /* * If the reference count goes to 0 we start calling * vm_object_terminate() on the object chain. * A ref count of 1 may be a special case depending on the * shadow count being 0 or 1. */ object->ref_count--; if (object->ref_count > 1) { VM_OBJECT_UNLOCK(object); return; } else if (object->ref_count == 1) { if (object->shadow_count == 0 && object->handle == NULL && (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { vm_object_set_flag(object, OBJ_ONEMAPPING); } else if ((object->shadow_count == 1) && (object->handle == NULL) && (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { vm_object_t robject; robject = LIST_FIRST(&object->shadow_head); KASSERT(robject != NULL, ("vm_object_deallocate: ref_count: %d, shadow_count: %d", object->ref_count, object->shadow_count)); if (!VM_OBJECT_TRYLOCK(robject)) { /* * Avoid a potential deadlock. */ object->ref_count++; VM_OBJECT_UNLOCK(object); /* * More likely than not the thread * holding robject's lock has lower * priority than the current thread. * Let the lower priority thread run. */ pause("vmo_de", 1); continue; } /* * Collapse object into its shadow unless its * shadow is dead. In that case, object will * be deallocated by the thread that is * deallocating its shadow. */ if ((robject->flags & OBJ_DEAD) == 0 && (robject->handle == NULL) && (robject->type == OBJT_DEFAULT || robject->type == OBJT_SWAP)) { robject->ref_count++; retry: if (robject->paging_in_progress) { VM_OBJECT_UNLOCK(object); vm_object_pip_wait(robject, "objde1"); temp = robject->backing_object; if (object == temp) { VM_OBJECT_LOCK(object); goto retry; } } else if (object->paging_in_progress) { VM_OBJECT_UNLOCK(robject); object->flags |= OBJ_PIPWNT; msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "objde2", 0); VM_OBJECT_LOCK(robject); temp = robject->backing_object; if (object == temp) { VM_OBJECT_LOCK(object); goto retry; } } else VM_OBJECT_UNLOCK(object); if (robject->ref_count == 1) { robject->ref_count--; object = robject; goto doterm; } object = robject; vm_object_collapse(object); VM_OBJECT_UNLOCK(object); continue; } VM_OBJECT_UNLOCK(robject); } VM_OBJECT_UNLOCK(object); return; } doterm: temp = object->backing_object; if (temp != NULL) { VM_OBJECT_LOCK(temp); LIST_REMOVE(object, shadow_list); temp->shadow_count--; VM_OBJECT_UNLOCK(temp); object->backing_object = NULL; } /* * Don't double-terminate, we could be in a termination * recursion due to the terminate having to sync data * to disk. */ if ((object->flags & OBJ_DEAD) == 0) vm_object_terminate(object); else VM_OBJECT_UNLOCK(object); object = temp; } } /* * vm_object_destroy removes the object from the global object list * and frees the space for the object. */ void vm_object_destroy(vm_object_t object) { /* * Remove the object from the global object list. */ mtx_lock(&vm_object_list_mtx); TAILQ_REMOVE(&vm_object_list, object, object_list); mtx_unlock(&vm_object_list_mtx); /* * Release the allocation charge. */ if (object->cred != NULL) { KASSERT(object->type == OBJT_DEFAULT || object->type == OBJT_SWAP, ("vm_object_terminate: non-swap obj %p has cred", object)); swap_release_by_cred(object->charge, object->cred); object->charge = 0; crfree(object->cred); object->cred = NULL; } /* * Free the space for the object. */ uma_zfree(obj_zone, object); } /* * vm_object_terminate actually destroys the specified object, freeing * up all previously used resources. * * The object must be locked. * This routine may block. */ void vm_object_terminate(vm_object_t object) { vm_page_t p, p_next; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); /* * Make sure no one uses us. */ vm_object_set_flag(object, OBJ_DEAD); /* * wait for the pageout daemon to be done with the object */ vm_object_pip_wait(object, "objtrm"); KASSERT(!object->paging_in_progress, ("vm_object_terminate: pageout in progress")); /* * Clean and free the pages, as appropriate. All references to the * object are gone, so we don't need to lock it. */ if (object->type == OBJT_VNODE) { struct vnode *vp = (struct vnode *)object->handle; /* * Clean pages and flush buffers. */ vm_object_page_clean(object, 0, 0, OBJPC_SYNC); VM_OBJECT_UNLOCK(object); vinvalbuf(vp, V_SAVE, 0, 0); VM_OBJECT_LOCK(object); } KASSERT(object->ref_count == 0, ("vm_object_terminate: object with references, ref_count=%d", object->ref_count)); /* * Free any remaining pageable pages. This also removes them from the * paging queues. However, don't free wired pages, just remove them * from the object. Rather than incrementally removing each page from * the object, the page and object are reset to any empty state. */ TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0, ("vm_object_terminate: freeing busy page %p", p)); vm_page_lock(p); /* * Optimize the page's removal from the object by resetting * its "object" field. Specifically, if the page is not * wired, then the effect of this assignment is that * vm_page_free()'s call to vm_page_remove() will return * immediately without modifying the page or the object. */ p->object = NULL; if (p->wire_count == 0) { vm_page_free(p); PCPU_INC(cnt.v_pfree); } vm_page_unlock(p); } /* * If the object contained any pages, then reset it to an empty state. * None of the object's fields, including "resident_page_count", were * modified by the preceding loop. */ if (object->resident_page_count != 0) { object->root = NULL; TAILQ_INIT(&object->memq); object->resident_page_count = 0; if (object->type == OBJT_VNODE) vdrop(object->handle); } #if VM_NRESERVLEVEL > 0 if (__predict_false(!LIST_EMPTY(&object->rvq))) vm_reserv_break_all(object); #endif if (__predict_false(object->cache != NULL)) vm_page_cache_free(object, 0, 0); /* * Let the pager know object is dead. */ vm_pager_deallocate(object); VM_OBJECT_UNLOCK(object); vm_object_destroy(object); } /* * Make the page read-only so that we can clear the object flags. However, if * this is a nosync mmap then the object is likely to stay dirty so do not * mess with the page and do not clear the object flags. Returns TRUE if the * page should be flushed, and FALSE otherwise. */ static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, int *clearobjflags) { /* * If we have been asked to skip nosync pages and this is a * nosync page, skip it. Note that the object flags were not * cleared in this case so we do not have to set them. */ if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) { *clearobjflags = 0; return (FALSE); } else { pmap_remove_write(p); return (p->dirty != 0); } } /* * vm_object_page_clean * * Clean all dirty pages in the specified range of object. Leaves page * on whatever queue it is currently on. If NOSYNC is set then do not * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC), * leaving the object dirty. * * When stuffing pages asynchronously, allow clustering. XXX we need a * synchronous clustering mode implementation. * * Odd semantics: if start == end, we clean everything. * * The object must be locked. */ void vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, int flags) { vm_page_t np, p; vm_pindex_t pi, tend, tstart; int clearobjflags, curgeneration, n, pagerflags; mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); KASSERT(object->type == OBJT_VNODE, ("Not a vnode object")); if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 || object->resident_page_count == 0) return; pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; tstart = OFF_TO_IDX(start); tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); clearobjflags = tstart == 0 && tend >= object->size; rescan: curgeneration = object->generation; for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { pi = p->pindex; if (pi >= tend) break; np = TAILQ_NEXT(p, listq); if (p->valid == 0) continue; if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) { if (object->generation != curgeneration) goto rescan; np = vm_page_find_least(object, pi); continue; } if (!vm_object_page_remove_write(p, flags, &clearobjflags)) continue; n = vm_object_page_collect_flush(object, p, pagerflags, flags, &clearobjflags); if (object->generation != curgeneration) goto rescan; + + /* + * If the VOP_PUTPAGES() did a truncated write, so + * that even the first page of the run is not fully + * written, vm_pageout_flush() returns 0 as the run + * length. Since the condition that caused truncated + * write may be permanent, e.g. exhausted free space, + * accepting n == 0 would cause an infinite loop. + * + * Forwarding the iterator leaves the unwritten page + * behind, but there is not much we can do there if + * filesystem refuses to write it. + */ + if (n == 0) + n = 1; np = vm_page_find_least(object, pi + n); } #if 0 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); #endif if (clearobjflags) vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY); } static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, int flags, int *clearobjflags) { vm_page_t ma[vm_pageout_page_count], p_first, tp; int count, i, mreq, runlen; mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED); vm_page_lock_assert(p, MA_NOTOWNED); VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); count = 1; mreq = 0; for (tp = p; count < vm_pageout_page_count; count++) { tp = vm_page_next(tp); if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0) break; if (!vm_object_page_remove_write(tp, flags, clearobjflags)) break; } for (p_first = p; count < vm_pageout_page_count; count++) { tp = vm_page_prev(p_first); if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0) break; if (!vm_object_page_remove_write(tp, flags, clearobjflags)) break; p_first = tp; mreq++; } for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) ma[i] = tp; vm_pageout_flush(ma, count, pagerflags, mreq, &runlen); return (runlen); } /* * Note that there is absolutely no sense in writing out * anonymous objects, so we track down the vnode object * to write out. * We invalidate (remove) all pages from the address space * for semantic correctness. * * Note: certain anonymous maps, such as MAP_NOSYNC maps, * may start out with a NULL object. */ void vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, boolean_t syncio, boolean_t invalidate) { vm_object_t backing_object; struct vnode *vp; struct mount *mp; int flags; if (object == NULL) return; VM_OBJECT_LOCK(object); while ((backing_object = object->backing_object) != NULL) { VM_OBJECT_LOCK(backing_object); offset += object->backing_object_offset; VM_OBJECT_UNLOCK(object); object = backing_object; if (object->size < OFF_TO_IDX(offset + size)) size = IDX_TO_OFF(object->size) - offset; } /* * Flush pages if writing is allowed, invalidate them * if invalidation requested. Pages undergoing I/O * will be ignored by vm_object_page_remove(). * * We cannot lock the vnode and then wait for paging * to complete without deadlocking against vm_fault. * Instead we simply call vm_object_page_remove() and * allow it to block internally on a page-by-page * basis when it encounters pages undergoing async * I/O. */ if (object->type == OBJT_VNODE && (object->flags & OBJ_MIGHTBEDIRTY) != 0) { int vfslocked; vp = object->handle; VM_OBJECT_UNLOCK(object); (void) vn_start_write(vp, &mp, V_WAIT); vfslocked = VFS_LOCK_GIANT(vp->v_mount); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); flags = (syncio || invalidate) ? OBJPC_SYNC : 0; flags |= invalidate ? OBJPC_INVAL : 0; VM_OBJECT_LOCK(object); vm_object_page_clean(object, offset, offset + size, flags); VM_OBJECT_UNLOCK(object); VOP_UNLOCK(vp, 0); VFS_UNLOCK_GIANT(vfslocked); vn_finished_write(mp); VM_OBJECT_LOCK(object); } if ((object->type == OBJT_VNODE || object->type == OBJT_DEVICE) && invalidate) { boolean_t purge; purge = old_msync || (object->type == OBJT_DEVICE); vm_object_page_remove(object, OFF_TO_IDX(offset), OFF_TO_IDX(offset + size + PAGE_MASK), purge ? FALSE : TRUE); } VM_OBJECT_UNLOCK(object); } /* * vm_object_madvise: * * Implements the madvise function at the object/page level. * * MADV_WILLNEED (any object) * * Activate the specified pages if they are resident. * * MADV_DONTNEED (any object) * * Deactivate the specified pages if they are resident. * * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, * OBJ_ONEMAPPING only) * * Deactivate and clean the specified pages if they are * resident. This permits the process to reuse the pages * without faulting or the kernel to reclaim the pages * without I/O. */ void vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise) { vm_pindex_t end, tpindex; vm_object_t backing_object, tobject; vm_page_t m; if (object == NULL) return; VM_OBJECT_LOCK(object); end = pindex + count; /* * Locate and adjust resident pages */ for (; pindex < end; pindex += 1) { relookup: tobject = object; tpindex = pindex; shadowlookup: /* * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages * and those pages must be OBJ_ONEMAPPING. */ if (advise == MADV_FREE) { if ((tobject->type != OBJT_DEFAULT && tobject->type != OBJT_SWAP) || (tobject->flags & OBJ_ONEMAPPING) == 0) { goto unlock_tobject; } } else if (tobject->type == OBJT_PHYS) goto unlock_tobject; m = vm_page_lookup(tobject, tpindex); if (m == NULL && advise == MADV_WILLNEED) { /* * If the page is cached, reactivate it. */ m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED | VM_ALLOC_NOBUSY); } if (m == NULL) { /* * There may be swap even if there is no backing page */ if (advise == MADV_FREE && tobject->type == OBJT_SWAP) swap_pager_freespace(tobject, tpindex, 1); /* * next object */ backing_object = tobject->backing_object; if (backing_object == NULL) goto unlock_tobject; VM_OBJECT_LOCK(backing_object); tpindex += OFF_TO_IDX(tobject->backing_object_offset); if (tobject != object) VM_OBJECT_UNLOCK(tobject); tobject = backing_object; goto shadowlookup; } else if (m->valid != VM_PAGE_BITS_ALL) goto unlock_tobject; /* * If the page is not in a normal state, skip it. */ vm_page_lock(m); if (m->hold_count != 0 || m->wire_count != 0) { vm_page_unlock(m); goto unlock_tobject; } KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0, ("vm_object_madvise: page %p is not managed", m)); if ((m->oflags & VPO_BUSY) || m->busy) { if (advise == MADV_WILLNEED) { /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ vm_page_lock_queues(); vm_page_flag_set(m, PG_REFERENCED); vm_page_unlock_queues(); } vm_page_unlock(m); if (object != tobject) VM_OBJECT_UNLOCK(object); m->oflags |= VPO_WANTED; msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo", 0); VM_OBJECT_LOCK(object); goto relookup; } if (advise == MADV_WILLNEED) { vm_page_activate(m); } else if (advise == MADV_DONTNEED) { vm_page_dontneed(m); } else if (advise == MADV_FREE) { /* * Mark the page clean. This will allow the page * to be freed up by the system. However, such pages * are often reused quickly by malloc()/free() * so we do not do anything that would cause * a page fault if we can help it. * * Specifically, we do not try to actually free * the page now nor do we try to put it in the * cache (which would cause a page fault on reuse). * * But we do make the page is freeable as we * can without actually taking the step of unmapping * it. */ pmap_clear_modify(m); m->dirty = 0; m->act_count = 0; vm_page_dontneed(m); } vm_page_unlock(m); if (advise == MADV_FREE && tobject->type == OBJT_SWAP) swap_pager_freespace(tobject, tpindex, 1); unlock_tobject: if (tobject != object) VM_OBJECT_UNLOCK(tobject); } VM_OBJECT_UNLOCK(object); } /* * vm_object_shadow: * * Create a new object which is backed by the * specified existing object range. The source * object reference is deallocated. * * The new object and offset into that object * are returned in the source parameters. */ void vm_object_shadow( vm_object_t *object, /* IN/OUT */ vm_ooffset_t *offset, /* IN/OUT */ vm_size_t length) { vm_object_t source; vm_object_t result; source = *object; /* * Don't create the new object if the old object isn't shared. */ if (source != NULL) { VM_OBJECT_LOCK(source); if (source->ref_count == 1 && source->handle == NULL && (source->type == OBJT_DEFAULT || source->type == OBJT_SWAP)) { VM_OBJECT_UNLOCK(source); return; } VM_OBJECT_UNLOCK(source); } /* * Allocate a new object with the given length. */ result = vm_object_allocate(OBJT_DEFAULT, atop(length)); /* * The new object shadows the source object, adding a reference to it. * Our caller changes his reference to point to the new object, * removing a reference to the source object. Net result: no change * of reference count. * * Try to optimize the result object's page color when shadowing * in order to maintain page coloring consistency in the combined * shadowed object. */ result->backing_object = source; /* * Store the offset into the source object, and fix up the offset into * the new object. */ result->backing_object_offset = *offset; if (source != NULL) { VM_OBJECT_LOCK(source); LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); source->shadow_count++; #if VM_NRESERVLEVEL > 0 result->flags |= source->flags & OBJ_COLORED; result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER - 1)) - 1); #endif VM_OBJECT_UNLOCK(source); } /* * Return the new things */ *offset = 0; *object = result; } /* * vm_object_split: * * Split the pages in a map entry into a new object. This affords * easier removal of unused pages, and keeps object inheritance from * being a negative impact on memory usage. */ void vm_object_split(vm_map_entry_t entry) { vm_page_t m, m_next; vm_object_t orig_object, new_object, source; vm_pindex_t idx, offidxstart; vm_size_t size; orig_object = entry->object.vm_object; if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) return; if (orig_object->ref_count <= 1) return; VM_OBJECT_UNLOCK(orig_object); offidxstart = OFF_TO_IDX(entry->offset); size = atop(entry->end - entry->start); /* * If swap_pager_copy() is later called, it will convert new_object * into a swap object. */ new_object = vm_object_allocate(OBJT_DEFAULT, size); /* * At this point, the new object is still private, so the order in * which the original and new objects are locked does not matter. */ VM_OBJECT_LOCK(new_object); VM_OBJECT_LOCK(orig_object); source = orig_object->backing_object; if (source != NULL) { VM_OBJECT_LOCK(source); if ((source->flags & OBJ_DEAD) != 0) { VM_OBJECT_UNLOCK(source); VM_OBJECT_UNLOCK(orig_object); VM_OBJECT_UNLOCK(new_object); vm_object_deallocate(new_object); VM_OBJECT_LOCK(orig_object); return; } LIST_INSERT_HEAD(&source->shadow_head, new_object, shadow_list); source->shadow_count++; vm_object_reference_locked(source); /* for new_object */ vm_object_clear_flag(source, OBJ_ONEMAPPING); VM_OBJECT_UNLOCK(source); new_object->backing_object_offset = orig_object->backing_object_offset + entry->offset; new_object->backing_object = source; } if (orig_object->cred != NULL) { new_object->cred = orig_object->cred; crhold(orig_object->cred); new_object->charge = ptoa(size); KASSERT(orig_object->charge >= ptoa(size), ("orig_object->charge < 0")); orig_object->charge -= ptoa(size); } retry: m = vm_page_find_least(orig_object, offidxstart); for (; m != NULL && (idx = m->pindex - offidxstart) < size; m = m_next) { m_next = TAILQ_NEXT(m, listq); /* * We must wait for pending I/O to complete before we can * rename the page. * * We do not have to VM_PROT_NONE the page as mappings should * not be changed by this operation. */ if ((m->oflags & VPO_BUSY) || m->busy) { VM_OBJECT_UNLOCK(new_object); m->oflags |= VPO_WANTED; msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0); VM_OBJECT_LOCK(new_object); goto retry; } vm_page_lock(m); vm_page_rename(m, new_object, idx); vm_page_unlock(m); /* page automatically made dirty by rename and cache handled */ vm_page_busy(m); } if (orig_object->type == OBJT_SWAP) { /* * swap_pager_copy() can sleep, in which case the orig_object's * and new_object's locks are released and reacquired. */ swap_pager_copy(orig_object, new_object, offidxstart, 0); /* * Transfer any cached pages from orig_object to new_object. */ if (__predict_false(orig_object->cache != NULL)) vm_page_cache_transfer(orig_object, offidxstart, new_object); } VM_OBJECT_UNLOCK(orig_object); TAILQ_FOREACH(m, &new_object->memq, listq) vm_page_wakeup(m); VM_OBJECT_UNLOCK(new_object); entry->object.vm_object = new_object; entry->offset = 0LL; vm_object_deallocate(orig_object); VM_OBJECT_LOCK(new_object); } #define OBSC_TEST_ALL_SHADOWED 0x0001 #define OBSC_COLLAPSE_NOWAIT 0x0002 #define OBSC_COLLAPSE_WAIT 0x0004 static int vm_object_backing_scan(vm_object_t object, int op) { int r = 1; vm_page_t p; vm_object_t backing_object; vm_pindex_t backing_offset_index; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED); backing_object = object->backing_object; backing_offset_index = OFF_TO_IDX(object->backing_object_offset); /* * Initial conditions */ if (op & OBSC_TEST_ALL_SHADOWED) { /* * We do not want to have to test for the existence of cache * or swap pages in the backing object. XXX but with the * new swapper this would be pretty easy to do. * * XXX what about anonymous MAP_SHARED memory that hasn't * been ZFOD faulted yet? If we do not test for this, the * shadow test may succeed! XXX */ if (backing_object->type != OBJT_DEFAULT) { return (0); } } if (op & OBSC_COLLAPSE_WAIT) { vm_object_set_flag(backing_object, OBJ_DEAD); } /* * Our scan */ p = TAILQ_FIRST(&backing_object->memq); while (p) { vm_page_t next = TAILQ_NEXT(p, listq); vm_pindex_t new_pindex = p->pindex - backing_offset_index; if (op & OBSC_TEST_ALL_SHADOWED) { vm_page_t pp; /* * Ignore pages outside the parent object's range * and outside the parent object's mapping of the * backing object. * * note that we do not busy the backing object's * page. */ if ( p->pindex < backing_offset_index || new_pindex >= object->size ) { p = next; continue; } /* * See if the parent has the page or if the parent's * object pager has the page. If the parent has the * page but the page is not valid, the parent's * object pager must have the page. * * If this fails, the parent does not completely shadow * the object and we might as well give up now. */ pp = vm_page_lookup(object, new_pindex); if ( (pp == NULL || pp->valid == 0) && !vm_pager_has_page(object, new_pindex, NULL, NULL) ) { r = 0; break; } } /* * Check for busy page */ if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) { vm_page_t pp; if (op & OBSC_COLLAPSE_NOWAIT) { if ((p->oflags & VPO_BUSY) || !p->valid || p->busy) { p = next; continue; } } else if (op & OBSC_COLLAPSE_WAIT) { if ((p->oflags & VPO_BUSY) || p->busy) { VM_OBJECT_UNLOCK(object); p->oflags |= VPO_WANTED; msleep(p, VM_OBJECT_MTX(backing_object), PDROP | PVM, "vmocol", 0); VM_OBJECT_LOCK(object); VM_OBJECT_LOCK(backing_object); /* * If we slept, anything could have * happened. Since the object is * marked dead, the backing offset * should not have changed so we * just restart our scan. */ p = TAILQ_FIRST(&backing_object->memq); continue; } } KASSERT( p->object == backing_object, ("vm_object_backing_scan: object mismatch") ); /* * Destroy any associated swap */ if (backing_object->type == OBJT_SWAP) { swap_pager_freespace( backing_object, p->pindex, 1 ); } if ( p->pindex < backing_offset_index || new_pindex >= object->size ) { /* * Page is out of the parent object's range, we * can simply destroy it. */ vm_page_lock(p); KASSERT(!pmap_page_is_mapped(p), ("freeing mapped page %p", p)); if (p->wire_count == 0) vm_page_free(p); else vm_page_remove(p); vm_page_unlock(p); p = next; continue; } pp = vm_page_lookup(object, new_pindex); if ( (op & OBSC_COLLAPSE_NOWAIT) != 0 && (pp != NULL && pp->valid == 0) ) { /* * The page in the parent is not (yet) valid. * We don't know anything about the state of * the original page. It might be mapped, * so we must avoid the next if here. * * This is due to a race in vm_fault() where * we must unbusy the original (backing_obj) * page before we can (re)lock the parent. * Hence we can get here. */ p = next; continue; } if ( pp != NULL || vm_pager_has_page(object, new_pindex, NULL, NULL) ) { /* * page already exists in parent OR swap exists * for this location in the parent. Destroy * the original page from the backing object. * * Leave the parent's page alone */ vm_page_lock(p); KASSERT(!pmap_page_is_mapped(p), ("freeing mapped page %p", p)); if (p->wire_count == 0) vm_page_free(p); else vm_page_remove(p); vm_page_unlock(p); p = next; continue; } #if VM_NRESERVLEVEL > 0 /* * Rename the reservation. */ vm_reserv_rename(p, object, backing_object, backing_offset_index); #endif /* * Page does not exist in parent, rename the * page from the backing object to the main object. * * If the page was mapped to a process, it can remain * mapped through the rename. */ vm_page_lock(p); vm_page_rename(p, object, new_pindex); vm_page_unlock(p); /* page automatically made dirty by rename */ } p = next; } return (r); } /* * this version of collapse allows the operation to occur earlier and * when paging_in_progress is true for an object... This is not a complete * operation, but should plug 99.9% of the rest of the leaks. */ static void vm_object_qcollapse(vm_object_t object) { vm_object_t backing_object = object->backing_object; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED); if (backing_object->ref_count != 1) return; vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT); } /* * vm_object_collapse: * * Collapse an object with the object backing it. * Pages in the backing object are moved into the * parent, and the backing object is deallocated. */ void vm_object_collapse(vm_object_t object) { VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); while (TRUE) { vm_object_t backing_object; /* * Verify that the conditions are right for collapse: * * The object exists and the backing object exists. */ if ((backing_object = object->backing_object) == NULL) break; /* * we check the backing object first, because it is most likely * not collapsable. */ VM_OBJECT_LOCK(backing_object); if (backing_object->handle != NULL || (backing_object->type != OBJT_DEFAULT && backing_object->type != OBJT_SWAP) || (backing_object->flags & OBJ_DEAD) || object->handle != NULL || (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) || (object->flags & OBJ_DEAD)) { VM_OBJECT_UNLOCK(backing_object); break; } if ( object->paging_in_progress != 0 || backing_object->paging_in_progress != 0 ) { vm_object_qcollapse(object); VM_OBJECT_UNLOCK(backing_object); break; } /* * We know that we can either collapse the backing object (if * the parent is the only reference to it) or (perhaps) have * the parent bypass the object if the parent happens to shadow * all the resident pages in the entire backing object. * * This is ignoring pager-backed pages such as swap pages. * vm_object_backing_scan fails the shadowing test in this * case. */ if (backing_object->ref_count == 1) { /* * If there is exactly one reference to the backing * object, we can collapse it into the parent. */ vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); #if VM_NRESERVLEVEL > 0 /* * Break any reservations from backing_object. */ if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) vm_reserv_break_all(backing_object); #endif /* * Move the pager from backing_object to object. */ if (backing_object->type == OBJT_SWAP) { /* * swap_pager_copy() can sleep, in which case * the backing_object's and object's locks are * released and reacquired. */ swap_pager_copy( backing_object, object, OFF_TO_IDX(object->backing_object_offset), TRUE); /* * Free any cached pages from backing_object. */ if (__predict_false(backing_object->cache != NULL)) vm_page_cache_free(backing_object, 0, 0); } /* * Object now shadows whatever backing_object did. * Note that the reference to * backing_object->backing_object moves from within * backing_object to within object. */ LIST_REMOVE(object, shadow_list); backing_object->shadow_count--; if (backing_object->backing_object) { VM_OBJECT_LOCK(backing_object->backing_object); LIST_REMOVE(backing_object, shadow_list); LIST_INSERT_HEAD( &backing_object->backing_object->shadow_head, object, shadow_list); /* * The shadow_count has not changed. */ VM_OBJECT_UNLOCK(backing_object->backing_object); } object->backing_object = backing_object->backing_object; object->backing_object_offset += backing_object->backing_object_offset; /* * Discard backing_object. * * Since the backing object has no pages, no pager left, * and no object references within it, all that is * necessary is to dispose of it. */ KASSERT(backing_object->ref_count == 1, ( "backing_object %p was somehow re-referenced during collapse!", backing_object)); VM_OBJECT_UNLOCK(backing_object); vm_object_destroy(backing_object); object_collapses++; } else { vm_object_t new_backing_object; /* * If we do not entirely shadow the backing object, * there is nothing we can do so we give up. */ if (object->resident_page_count != object->size && vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) { VM_OBJECT_UNLOCK(backing_object); break; } /* * Make the parent shadow the next object in the * chain. Deallocating backing_object will not remove * it, since its reference count is at least 2. */ LIST_REMOVE(object, shadow_list); backing_object->shadow_count--; new_backing_object = backing_object->backing_object; if ((object->backing_object = new_backing_object) != NULL) { VM_OBJECT_LOCK(new_backing_object); LIST_INSERT_HEAD( &new_backing_object->shadow_head, object, shadow_list ); new_backing_object->shadow_count++; vm_object_reference_locked(new_backing_object); VM_OBJECT_UNLOCK(new_backing_object); object->backing_object_offset += backing_object->backing_object_offset; } /* * Drop the reference count on backing_object. Since * its ref_count was at least 2, it will not vanish. */ backing_object->ref_count--; VM_OBJECT_UNLOCK(backing_object); object_bypasses++; } /* * Try again with this object's new backing object. */ } } /* * vm_object_page_remove: * * For the given object, either frees or invalidates each of the * specified pages. In general, a page is freed. However, if a * page is wired for any reason other than the existence of a * managed, wired mapping, then it may be invalidated but not * removed from the object. Pages are specified by the given * range ["start", "end") and Boolean "clean_only". As a * special case, if "end" is zero, then the range extends from * "start" to the end of the object. If "clean_only" is TRUE, * then only the non-dirty pages within the specified range are * affected. * * In general, this operation should only be performed on objects * that contain managed pages. There are two exceptions. First, * it may be performed on the kernel and kmem objects. Second, * it may be used by msync(..., MS_INVALIDATE) to invalidate * device-backed pages. In both of these cases, "clean_only" * must be FALSE. * * The object must be locked. */ void vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, boolean_t clean_only) { vm_page_t p, next; int wirings; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); if (object->resident_page_count == 0) goto skipmemq; /* * Since physically-backed objects do not use managed pages, we can't * remove pages from the object (we must instead remove the page * references, and then destroy the object). */ KASSERT(object->type != OBJT_PHYS || object == kernel_object || object == kmem_object, ("attempt to remove pages from a physical object")); vm_object_pip_add(object, 1); again: p = vm_page_find_least(object, start); /* * Assert: the variable p is either (1) the page with the * least pindex greater than or equal to the parameter pindex * or (2) NULL. */ for (; p != NULL && (p->pindex < end || end == 0); p = next) { next = TAILQ_NEXT(p, listq); /* * If the page is wired for any reason besides the * existence of managed, wired mappings, then it cannot * be freed. For example, fictitious pages, which * represent device memory, are inherently wired and * cannot be freed. They can, however, be invalidated * if "clean_only" is FALSE. */ vm_page_lock(p); if ((wirings = p->wire_count) != 0 && (wirings = pmap_page_wired_mappings(p)) != p->wire_count) { /* Fictitious pages do not have managed mappings. */ if ((p->flags & PG_FICTITIOUS) == 0) pmap_remove_all(p); /* Account for removal of managed, wired mappings. */ p->wire_count -= wirings; if (!clean_only) { p->valid = 0; vm_page_undirty(p); } vm_page_unlock(p); continue; } if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) goto again; KASSERT((p->flags & PG_FICTITIOUS) == 0, ("vm_object_page_remove: page %p is fictitious", p)); if (clean_only && p->valid) { pmap_remove_write(p); if (p->dirty) { vm_page_unlock(p); continue; } } pmap_remove_all(p); /* Account for removal of managed, wired mappings. */ if (wirings != 0) p->wire_count -= wirings; vm_page_free(p); vm_page_unlock(p); } vm_object_pip_wakeup(object); skipmemq: if (__predict_false(object->cache != NULL)) vm_page_cache_free(object, start, end); } /* * Populate the specified range of the object with valid pages. Returns * TRUE if the range is successfully populated and FALSE otherwise. * * Note: This function should be optimized to pass a larger array of * pages to vm_pager_get_pages() before it is applied to a non- * OBJT_DEVICE object. * * The object must be locked. */ boolean_t vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) { vm_page_t m, ma[1]; vm_pindex_t pindex; int rv; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); for (pindex = start; pindex < end; pindex++) { m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); if (m->valid != VM_PAGE_BITS_ALL) { ma[0] = m; rv = vm_pager_get_pages(object, ma, 1, 0); m = vm_page_lookup(object, pindex); if (m == NULL) break; if (rv != VM_PAGER_OK) { vm_page_lock(m); vm_page_free(m); vm_page_unlock(m); break; } } /* * Keep "m" busy because a subsequent iteration may unlock * the object. */ } if (pindex > start) { m = vm_page_lookup(object, start); while (m != NULL && m->pindex < pindex) { vm_page_wakeup(m); m = TAILQ_NEXT(m, listq); } } return (pindex == end); } /* * Routine: vm_object_coalesce * Function: Coalesces two objects backing up adjoining * regions of memory into a single object. * * returns TRUE if objects were combined. * * NOTE: Only works at the moment if the second object is NULL - * if it's not, which object do we lock first? * * Parameters: * prev_object First object to coalesce * prev_offset Offset into prev_object * prev_size Size of reference to prev_object * next_size Size of reference to the second object * reserved Indicator that extension region has * swap accounted for * * Conditions: * The object must *not* be locked. */ boolean_t vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) { vm_pindex_t next_pindex; if (prev_object == NULL) return (TRUE); VM_OBJECT_LOCK(prev_object); if (prev_object->type != OBJT_DEFAULT && prev_object->type != OBJT_SWAP) { VM_OBJECT_UNLOCK(prev_object); return (FALSE); } /* * Try to collapse the object first */ vm_object_collapse(prev_object); /* * Can't coalesce if: . more than one reference . paged out . shadows * another object . has a copy elsewhere (any of which mean that the * pages not mapped to prev_entry may be in use anyway) */ if (prev_object->backing_object != NULL) { VM_OBJECT_UNLOCK(prev_object); return (FALSE); } prev_size >>= PAGE_SHIFT; next_size >>= PAGE_SHIFT; next_pindex = OFF_TO_IDX(prev_offset) + prev_size; if ((prev_object->ref_count > 1) && (prev_object->size != next_pindex)) { VM_OBJECT_UNLOCK(prev_object); return (FALSE); } /* * Account for the charge. */ if (prev_object->cred != NULL) { /* * If prev_object was charged, then this mapping, * althought not charged now, may become writable * later. Non-NULL cred in the object would prevent * swap reservation during enabling of the write * access, so reserve swap now. Failed reservation * cause allocation of the separate object for the map * entry, and swap reservation for this entry is * managed in appropriate time. */ if (!reserved && !swap_reserve_by_cred(ptoa(next_size), prev_object->cred)) { return (FALSE); } prev_object->charge += ptoa(next_size); } /* * Remove any pages that may still be in the object from a previous * deallocation. */ if (next_pindex < prev_object->size) { vm_object_page_remove(prev_object, next_pindex, next_pindex + next_size, FALSE); if (prev_object->type == OBJT_SWAP) swap_pager_freespace(prev_object, next_pindex, next_size); #if 0 if (prev_object->cred != NULL) { KASSERT(prev_object->charge >= ptoa(prev_object->size - next_pindex), ("object %p overcharged 1 %jx %jx", prev_object, (uintmax_t)next_pindex, (uintmax_t)next_size)); prev_object->charge -= ptoa(prev_object->size - next_pindex); } #endif } /* * Extend the object if necessary. */ if (next_pindex + next_size > prev_object->size) prev_object->size = next_pindex + next_size; VM_OBJECT_UNLOCK(prev_object); return (TRUE); } void vm_object_set_writeable_dirty(vm_object_t object) { VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); if (object->type != OBJT_VNODE) return; object->generation++; if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) return; vm_object_set_flag(object, OBJ_MIGHTBEDIRTY); } #include "opt_ddb.h" #ifdef DDB #include #include #include static int _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) { vm_map_t tmpm; vm_map_entry_t tmpe; vm_object_t obj; int entcount; if (map == 0) return 0; if (entry == 0) { tmpe = map->header.next; entcount = map->nentries; while (entcount-- && (tmpe != &map->header)) { if (_vm_object_in_map(map, object, tmpe)) { return 1; } tmpe = tmpe->next; } } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { tmpm = entry->object.sub_map; tmpe = tmpm->header.next; entcount = tmpm->nentries; while (entcount-- && tmpe != &tmpm->header) { if (_vm_object_in_map(tmpm, object, tmpe)) { return 1; } tmpe = tmpe->next; } } else if ((obj = entry->object.vm_object) != NULL) { for (; obj; obj = obj->backing_object) if (obj == object) { return 1; } } return 0; } static int vm_object_in_map(vm_object_t object) { struct proc *p; /* sx_slock(&allproc_lock); */ FOREACH_PROC_IN_SYSTEM(p) { if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) continue; if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { /* sx_sunlock(&allproc_lock); */ return 1; } } /* sx_sunlock(&allproc_lock); */ if (_vm_object_in_map(kernel_map, object, 0)) return 1; if (_vm_object_in_map(kmem_map, object, 0)) return 1; if (_vm_object_in_map(pager_map, object, 0)) return 1; if (_vm_object_in_map(buffer_map, object, 0)) return 1; return 0; } DB_SHOW_COMMAND(vmochk, vm_object_check) { vm_object_t object; /* * make sure that internal objs are in a map somewhere * and none have zero ref counts. */ TAILQ_FOREACH(object, &vm_object_list, object_list) { if (object->handle == NULL && (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { if (object->ref_count == 0) { db_printf("vmochk: internal obj has zero ref count: %ld\n", (long)object->size); } if (!vm_object_in_map(object)) { db_printf( "vmochk: internal obj is not in a map: " "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", object->ref_count, (u_long)object->size, (u_long)object->size, (void *)object->backing_object); } } } } /* * vm_object_print: [ debug ] */ DB_SHOW_COMMAND(object, vm_object_print_static) { /* XXX convert args. */ vm_object_t object = (vm_object_t)addr; boolean_t full = have_addr; vm_page_t p; /* XXX count is an (unused) arg. Avoid shadowing it. */ #define count was_count int count; if (object == NULL) return; db_iprintf( "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", object, (int)object->type, (uintmax_t)object->size, object->resident_page_count, object->ref_count, object->flags, object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", object->shadow_count, object->backing_object ? object->backing_object->ref_count : 0, object->backing_object, (uintmax_t)object->backing_object_offset); if (!full) return; db_indent += 2; count = 0; TAILQ_FOREACH(p, &object->memq, listq) { if (count == 0) db_iprintf("memory:="); else if (count == 6) { db_printf("\n"); db_iprintf(" ..."); count = 0; } else db_printf(","); count++; db_printf("(off=0x%jx,page=0x%jx)", (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); } if (count != 0) db_printf("\n"); db_indent -= 2; } /* XXX. */ #undef count /* XXX need this non-static entry for calling from vm_map_print. */ void vm_object_print( /* db_expr_t */ long addr, boolean_t have_addr, /* db_expr_t */ long count, char *modif) { vm_object_print_static(addr, have_addr, count, modif); } DB_SHOW_COMMAND(vmopag, vm_object_print_pages) { vm_object_t object; vm_pindex_t fidx; vm_paddr_t pa; vm_page_t m, prev_m; int rcount, nl, c; nl = 0; TAILQ_FOREACH(object, &vm_object_list, object_list) { db_printf("new object: %p\n", (void *)object); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; rcount = 0; fidx = 0; pa = -1; TAILQ_FOREACH(m, &object->memq, listq) { if (m->pindex > 128) break; if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && prev_m->pindex + 1 != m->pindex) { if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; rcount = 0; } } if (rcount && (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { ++rcount; continue; } if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; } fidx = m->pindex; pa = VM_PAGE_TO_PHYS(m); rcount = 1; } if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; } } } #endif /* DDB */ Index: head/sys/vm/vnode_pager.c =================================================================== --- head/sys/vm/vnode_pager.c (revision 222585) +++ head/sys/vm/vnode_pager.c (revision 222586) @@ -1,1193 +1,1210 @@ /*- * Copyright (c) 1990 University of Utah. * Copyright (c) 1991 The Regents of the University of California. * All rights reserved. * Copyright (c) 1993, 1994 John S. Dyson * Copyright (c) 1995, David Greenman * * 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 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. * * from: @(#)vnode_pager.c 7.5 (Berkeley) 4/20/91 */ /* * Page to/from files (vnodes). */ /* * TODO: * Implement VOP_GETPAGES/PUTPAGES interface for filesystems. Will * greatly re-simplify the vnode_pager. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, int *run); static int vnode_pager_input_smlfs(vm_object_t object, vm_page_t m); static int vnode_pager_input_old(vm_object_t object, vm_page_t m); static void vnode_pager_dealloc(vm_object_t); static int vnode_pager_getpages(vm_object_t, vm_page_t *, int, int); static void vnode_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); static boolean_t vnode_pager_haspage(vm_object_t, vm_pindex_t, int *, int *); static vm_object_t vnode_pager_alloc(void *, vm_ooffset_t, vm_prot_t, vm_ooffset_t, struct ucred *cred); struct pagerops vnodepagerops = { .pgo_alloc = vnode_pager_alloc, .pgo_dealloc = vnode_pager_dealloc, .pgo_getpages = vnode_pager_getpages, .pgo_putpages = vnode_pager_putpages, .pgo_haspage = vnode_pager_haspage, }; int vnode_pbuf_freecnt; /* Create the VM system backing object for this vnode */ int vnode_create_vobject(struct vnode *vp, off_t isize, struct thread *td) { vm_object_t object; vm_ooffset_t size = isize; struct vattr va; if (!vn_isdisk(vp, NULL) && vn_canvmio(vp) == FALSE) return (0); while ((object = vp->v_object) != NULL) { VM_OBJECT_LOCK(object); if (!(object->flags & OBJ_DEAD)) { VM_OBJECT_UNLOCK(object); return (0); } VOP_UNLOCK(vp, 0); vm_object_set_flag(object, OBJ_DISCONNECTWNT); msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vodead", 0); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); } if (size == 0) { if (vn_isdisk(vp, NULL)) { size = IDX_TO_OFF(INT_MAX); } else { if (VOP_GETATTR(vp, &va, td->td_ucred)) return (0); size = va.va_size; } } object = vnode_pager_alloc(vp, size, 0, 0, td->td_ucred); /* * Dereference the reference we just created. This assumes * that the object is associated with the vp. */ VM_OBJECT_LOCK(object); object->ref_count--; VM_OBJECT_UNLOCK(object); vrele(vp); KASSERT(vp->v_object != NULL, ("vnode_create_vobject: NULL object")); return (0); } void vnode_destroy_vobject(struct vnode *vp) { struct vm_object *obj; obj = vp->v_object; if (obj == NULL) return; ASSERT_VOP_ELOCKED(vp, "vnode_destroy_vobject"); VM_OBJECT_LOCK(obj); if (obj->ref_count == 0) { /* * vclean() may be called twice. The first time * removes the primary reference to the object, * the second time goes one further and is a * special-case to terminate the object. * * don't double-terminate the object */ if ((obj->flags & OBJ_DEAD) == 0) vm_object_terminate(obj); else VM_OBJECT_UNLOCK(obj); } else { /* * Woe to the process that tries to page now :-). */ vm_pager_deallocate(obj); VM_OBJECT_UNLOCK(obj); } vp->v_object = NULL; } /* * Allocate (or lookup) pager for a vnode. * Handle is a vnode pointer. * * MPSAFE */ vm_object_t vnode_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t offset, struct ucred *cred) { vm_object_t object; struct vnode *vp; /* * Pageout to vnode, no can do yet. */ if (handle == NULL) return (NULL); vp = (struct vnode *) handle; /* * If the object is being terminated, wait for it to * go away. */ retry: while ((object = vp->v_object) != NULL) { VM_OBJECT_LOCK(object); if ((object->flags & OBJ_DEAD) == 0) break; vm_object_set_flag(object, OBJ_DISCONNECTWNT); msleep(object, VM_OBJECT_MTX(object), PDROP | PVM, "vadead", 0); } if (vp->v_usecount == 0) panic("vnode_pager_alloc: no vnode reference"); if (object == NULL) { /* * Add an object of the appropriate size */ object = vm_object_allocate(OBJT_VNODE, OFF_TO_IDX(round_page(size))); object->un_pager.vnp.vnp_size = size; object->handle = handle; VI_LOCK(vp); if (vp->v_object != NULL) { /* * Object has been created while we were sleeping */ VI_UNLOCK(vp); vm_object_destroy(object); goto retry; } vp->v_object = object; VI_UNLOCK(vp); } else { object->ref_count++; VM_OBJECT_UNLOCK(object); } vref(vp); return (object); } /* * The object must be locked. */ static void vnode_pager_dealloc(object) vm_object_t object; { struct vnode *vp; int refs; vp = object->handle; if (vp == NULL) panic("vnode_pager_dealloc: pager already dealloced"); VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); vm_object_pip_wait(object, "vnpdea"); refs = object->ref_count; object->handle = NULL; object->type = OBJT_DEAD; if (object->flags & OBJ_DISCONNECTWNT) { vm_object_clear_flag(object, OBJ_DISCONNECTWNT); wakeup(object); } ASSERT_VOP_ELOCKED(vp, "vnode_pager_dealloc"); vp->v_object = NULL; vp->v_vflag &= ~VV_TEXT; while (refs-- > 0) vunref(vp); } static boolean_t vnode_pager_haspage(object, pindex, before, after) vm_object_t object; vm_pindex_t pindex; int *before; int *after; { struct vnode *vp = object->handle; daddr_t bn; int err; daddr_t reqblock; int poff; int bsize; int pagesperblock, blocksperpage; int vfslocked; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); /* * If no vp or vp is doomed or marked transparent to VM, we do not * have the page. */ if (vp == NULL || vp->v_iflag & VI_DOOMED) return FALSE; /* * If the offset is beyond end of file we do * not have the page. */ if (IDX_TO_OFF(pindex) >= object->un_pager.vnp.vnp_size) return FALSE; bsize = vp->v_mount->mnt_stat.f_iosize; pagesperblock = bsize / PAGE_SIZE; blocksperpage = 0; if (pagesperblock > 0) { reqblock = pindex / pagesperblock; } else { blocksperpage = (PAGE_SIZE / bsize); reqblock = pindex * blocksperpage; } VM_OBJECT_UNLOCK(object); vfslocked = VFS_LOCK_GIANT(vp->v_mount); err = VOP_BMAP(vp, reqblock, NULL, &bn, after, before); VFS_UNLOCK_GIANT(vfslocked); VM_OBJECT_LOCK(object); if (err) return TRUE; if (bn == -1) return FALSE; if (pagesperblock > 0) { poff = pindex - (reqblock * pagesperblock); if (before) { *before *= pagesperblock; *before += poff; } if (after) { int numafter; *after *= pagesperblock; numafter = pagesperblock - (poff + 1); if (IDX_TO_OFF(pindex + numafter) > object->un_pager.vnp.vnp_size) { numafter = OFF_TO_IDX(object->un_pager.vnp.vnp_size) - pindex; } *after += numafter; } } else { if (before) { *before /= blocksperpage; } if (after) { *after /= blocksperpage; } } return TRUE; } /* * Lets the VM system know about a change in size for a file. * We adjust our own internal size and flush any cached pages in * the associated object that are affected by the size change. * * Note: this routine may be invoked as a result of a pager put * operation (possibly at object termination time), so we must be careful. */ void vnode_pager_setsize(vp, nsize) struct vnode *vp; vm_ooffset_t nsize; { vm_object_t object; vm_page_t m; vm_pindex_t nobjsize; if ((object = vp->v_object) == NULL) return; /* ASSERT_VOP_ELOCKED(vp, "vnode_pager_setsize and not locked vnode"); */ VM_OBJECT_LOCK(object); if (nsize == object->un_pager.vnp.vnp_size) { /* * Hasn't changed size */ VM_OBJECT_UNLOCK(object); return; } nobjsize = OFF_TO_IDX(nsize + PAGE_MASK); if (nsize < object->un_pager.vnp.vnp_size) { /* * File has shrunk. Toss any cached pages beyond the new EOF. */ if (nobjsize < object->size) vm_object_page_remove(object, nobjsize, object->size, FALSE); /* * this gets rid of garbage at the end of a page that is now * only partially backed by the vnode. * * XXX for some reason (I don't know yet), if we take a * completely invalid page and mark it partially valid * it can screw up NFS reads, so we don't allow the case. */ if ((nsize & PAGE_MASK) && (m = vm_page_lookup(object, OFF_TO_IDX(nsize))) != NULL && m->valid != 0) { int base = (int)nsize & PAGE_MASK; int size = PAGE_SIZE - base; /* * Clear out partial-page garbage in case * the page has been mapped. */ pmap_zero_page_area(m, base, size); /* * Update the valid bits to reflect the blocks that * have been zeroed. Some of these valid bits may * have already been set. */ vm_page_set_valid(m, base, size); /* * Round "base" to the next block boundary so that the * dirty bit for a partially zeroed block is not * cleared. */ base = roundup2(base, DEV_BSIZE); /* * Clear out partial-page dirty bits. * * note that we do not clear out the valid * bits. This would prevent bogus_page * replacement from working properly. */ vm_page_clear_dirty(m, base, PAGE_SIZE - base); } else if ((nsize & PAGE_MASK) && __predict_false(object->cache != NULL)) { vm_page_cache_free(object, OFF_TO_IDX(nsize), nobjsize); } } object->un_pager.vnp.vnp_size = nsize; object->size = nobjsize; VM_OBJECT_UNLOCK(object); } /* * calculate the linear (byte) disk address of specified virtual * file address */ static int vnode_pager_addr(struct vnode *vp, vm_ooffset_t address, daddr_t *rtaddress, int *run) { int bsize; int err; daddr_t vblock; daddr_t voffset; if (address < 0) return -1; if (vp->v_iflag & VI_DOOMED) return -1; bsize = vp->v_mount->mnt_stat.f_iosize; vblock = address / bsize; voffset = address % bsize; err = VOP_BMAP(vp, vblock, NULL, rtaddress, run, NULL); if (err == 0) { if (*rtaddress != -1) *rtaddress += voffset / DEV_BSIZE; if (run) { *run += 1; *run *= bsize/PAGE_SIZE; *run -= voffset/PAGE_SIZE; } } return (err); } /* * small block filesystem vnode pager input */ static int vnode_pager_input_smlfs(object, m) vm_object_t object; vm_page_t m; { int bits, i; struct vnode *vp; struct bufobj *bo; struct buf *bp; struct sf_buf *sf; daddr_t fileaddr; vm_offset_t bsize; int error = 0; vp = object->handle; if (vp->v_iflag & VI_DOOMED) return VM_PAGER_BAD; bsize = vp->v_mount->mnt_stat.f_iosize; VOP_BMAP(vp, 0, &bo, 0, NULL, NULL); sf = sf_buf_alloc(m, 0); for (i = 0; i < PAGE_SIZE / bsize; i++) { vm_ooffset_t address; bits = vm_page_bits(i * bsize, bsize); if (m->valid & bits) continue; address = IDX_TO_OFF(m->pindex) + i * bsize; if (address >= object->un_pager.vnp.vnp_size) { fileaddr = -1; } else { error = vnode_pager_addr(vp, address, &fileaddr, NULL); if (error) break; } if (fileaddr != -1) { bp = getpbuf(&vnode_pbuf_freecnt); /* build a minimal buffer header */ bp->b_iocmd = BIO_READ; bp->b_iodone = bdone; KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); bp->b_rcred = crhold(curthread->td_ucred); bp->b_wcred = crhold(curthread->td_ucred); bp->b_data = (caddr_t)sf_buf_kva(sf) + i * bsize; bp->b_blkno = fileaddr; pbgetbo(bo, bp); bp->b_bcount = bsize; bp->b_bufsize = bsize; bp->b_runningbufspace = bp->b_bufsize; atomic_add_long(&runningbufspace, bp->b_runningbufspace); /* do the input */ bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); bwait(bp, PVM, "vnsrd"); if ((bp->b_ioflags & BIO_ERROR) != 0) error = EIO; /* * free the buffer header back to the swap buffer pool */ pbrelbo(bp); relpbuf(bp, &vnode_pbuf_freecnt); if (error) break; } else bzero((caddr_t)sf_buf_kva(sf) + i * bsize, bsize); KASSERT((m->dirty & bits) == 0, ("vnode_pager_input_smlfs: page %p is dirty", m)); VM_OBJECT_LOCK(object); m->valid |= bits; VM_OBJECT_UNLOCK(object); } sf_buf_free(sf); if (error) { return VM_PAGER_ERROR; } return VM_PAGER_OK; } /* * old style vnode pager input routine */ static int vnode_pager_input_old(object, m) vm_object_t object; vm_page_t m; { struct uio auio; struct iovec aiov; int error; int size; struct sf_buf *sf; struct vnode *vp; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); error = 0; /* * Return failure if beyond current EOF */ if (IDX_TO_OFF(m->pindex) >= object->un_pager.vnp.vnp_size) { return VM_PAGER_BAD; } else { size = PAGE_SIZE; if (IDX_TO_OFF(m->pindex) + size > object->un_pager.vnp.vnp_size) size = object->un_pager.vnp.vnp_size - IDX_TO_OFF(m->pindex); vp = object->handle; VM_OBJECT_UNLOCK(object); /* * Allocate a kernel virtual address and initialize so that * we can use VOP_READ/WRITE routines. */ sf = sf_buf_alloc(m, 0); aiov.iov_base = (caddr_t)sf_buf_kva(sf); aiov.iov_len = size; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_offset = IDX_TO_OFF(m->pindex); auio.uio_segflg = UIO_SYSSPACE; auio.uio_rw = UIO_READ; auio.uio_resid = size; auio.uio_td = curthread; error = VOP_READ(vp, &auio, 0, curthread->td_ucred); if (!error) { int count = size - auio.uio_resid; if (count == 0) error = EINVAL; else if (count != PAGE_SIZE) bzero((caddr_t)sf_buf_kva(sf) + count, PAGE_SIZE - count); } sf_buf_free(sf); VM_OBJECT_LOCK(object); } KASSERT(m->dirty == 0, ("vnode_pager_input_old: page %p is dirty", m)); if (!error) m->valid = VM_PAGE_BITS_ALL; return error ? VM_PAGER_ERROR : VM_PAGER_OK; } /* * generic vnode pager input routine */ /* * Local media VFS's that do not implement their own VOP_GETPAGES * should have their VOP_GETPAGES call to vnode_pager_generic_getpages() * to implement the previous behaviour. * * All other FS's should use the bypass to get to the local media * backing vp's VOP_GETPAGES. */ static int vnode_pager_getpages(object, m, count, reqpage) vm_object_t object; vm_page_t *m; int count; int reqpage; { int rtval; struct vnode *vp; int bytes = count * PAGE_SIZE; int vfslocked; vp = object->handle; VM_OBJECT_UNLOCK(object); vfslocked = VFS_LOCK_GIANT(vp->v_mount); rtval = VOP_GETPAGES(vp, m, bytes, reqpage, 0); KASSERT(rtval != EOPNOTSUPP, ("vnode_pager: FS getpages not implemented\n")); VFS_UNLOCK_GIANT(vfslocked); VM_OBJECT_LOCK(object); return rtval; } /* * This is now called from local media FS's to operate against their * own vnodes if they fail to implement VOP_GETPAGES. */ int vnode_pager_generic_getpages(vp, m, bytecount, reqpage) struct vnode *vp; vm_page_t *m; int bytecount; int reqpage; { vm_object_t object; vm_offset_t kva; off_t foff, tfoff, nextoff; int i, j, size, bsize, first; daddr_t firstaddr, reqblock; struct bufobj *bo; int runpg; int runend; struct buf *bp; int count; int error; object = vp->v_object; count = bytecount / PAGE_SIZE; KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, ("vnode_pager_generic_getpages does not support devices")); if (vp->v_iflag & VI_DOOMED) return VM_PAGER_BAD; bsize = vp->v_mount->mnt_stat.f_iosize; /* get the UNDERLYING device for the file with VOP_BMAP() */ /* * originally, we did not check for an error return value -- assuming * an fs always has a bmap entry point -- that assumption is wrong!!! */ foff = IDX_TO_OFF(m[reqpage]->pindex); /* * if we can't bmap, use old VOP code */ error = VOP_BMAP(vp, foff / bsize, &bo, &reqblock, NULL, NULL); if (error == EOPNOTSUPP) { VM_OBJECT_LOCK(object); for (i = 0; i < count; i++) if (i != reqpage) { vm_page_lock(m[i]); vm_page_free(m[i]); vm_page_unlock(m[i]); } PCPU_INC(cnt.v_vnodein); PCPU_INC(cnt.v_vnodepgsin); error = vnode_pager_input_old(object, m[reqpage]); VM_OBJECT_UNLOCK(object); return (error); } else if (error != 0) { VM_OBJECT_LOCK(object); for (i = 0; i < count; i++) if (i != reqpage) { vm_page_lock(m[i]); vm_page_free(m[i]); vm_page_unlock(m[i]); } VM_OBJECT_UNLOCK(object); return (VM_PAGER_ERROR); /* * if the blocksize is smaller than a page size, then use * special small filesystem code. NFS sometimes has a small * blocksize, but it can handle large reads itself. */ } else if ((PAGE_SIZE / bsize) > 1 && (vp->v_mount->mnt_stat.f_type != nfs_mount_type)) { VM_OBJECT_LOCK(object); for (i = 0; i < count; i++) if (i != reqpage) { vm_page_lock(m[i]); vm_page_free(m[i]); vm_page_unlock(m[i]); } VM_OBJECT_UNLOCK(object); PCPU_INC(cnt.v_vnodein); PCPU_INC(cnt.v_vnodepgsin); return vnode_pager_input_smlfs(object, m[reqpage]); } /* * If we have a completely valid page available to us, we can * clean up and return. Otherwise we have to re-read the * media. */ VM_OBJECT_LOCK(object); if (m[reqpage]->valid == VM_PAGE_BITS_ALL) { for (i = 0; i < count; i++) if (i != reqpage) { vm_page_lock(m[i]); vm_page_free(m[i]); vm_page_unlock(m[i]); } VM_OBJECT_UNLOCK(object); return VM_PAGER_OK; } else if (reqblock == -1) { pmap_zero_page(m[reqpage]); KASSERT(m[reqpage]->dirty == 0, ("vnode_pager_generic_getpages: page %p is dirty", m)); m[reqpage]->valid = VM_PAGE_BITS_ALL; for (i = 0; i < count; i++) if (i != reqpage) { vm_page_lock(m[i]); vm_page_free(m[i]); vm_page_unlock(m[i]); } VM_OBJECT_UNLOCK(object); return (VM_PAGER_OK); } m[reqpage]->valid = 0; VM_OBJECT_UNLOCK(object); /* * here on direct device I/O */ firstaddr = -1; /* * calculate the run that includes the required page */ for (first = 0, i = 0; i < count; i = runend) { if (vnode_pager_addr(vp, IDX_TO_OFF(m[i]->pindex), &firstaddr, &runpg) != 0) { VM_OBJECT_LOCK(object); for (; i < count; i++) if (i != reqpage) { vm_page_lock(m[i]); vm_page_free(m[i]); vm_page_unlock(m[i]); } VM_OBJECT_UNLOCK(object); return (VM_PAGER_ERROR); } if (firstaddr == -1) { VM_OBJECT_LOCK(object); if (i == reqpage && foff < object->un_pager.vnp.vnp_size) { panic("vnode_pager_getpages: unexpected missing page: firstaddr: %jd, foff: 0x%jx%08jx, vnp_size: 0x%jx%08jx", (intmax_t)firstaddr, (uintmax_t)(foff >> 32), (uintmax_t)foff, (uintmax_t) (object->un_pager.vnp.vnp_size >> 32), (uintmax_t)object->un_pager.vnp.vnp_size); } vm_page_lock(m[i]); vm_page_free(m[i]); vm_page_unlock(m[i]); VM_OBJECT_UNLOCK(object); runend = i + 1; first = runend; continue; } runend = i + runpg; if (runend <= reqpage) { VM_OBJECT_LOCK(object); for (j = i; j < runend; j++) { vm_page_lock(m[j]); vm_page_free(m[j]); vm_page_unlock(m[j]); } VM_OBJECT_UNLOCK(object); } else { if (runpg < (count - first)) { VM_OBJECT_LOCK(object); for (i = first + runpg; i < count; i++) { vm_page_lock(m[i]); vm_page_free(m[i]); vm_page_unlock(m[i]); } VM_OBJECT_UNLOCK(object); count = first + runpg; } break; } first = runend; } /* * the first and last page have been calculated now, move input pages * to be zero based... */ if (first != 0) { m += first; count -= first; reqpage -= first; } /* * calculate the file virtual address for the transfer */ foff = IDX_TO_OFF(m[0]->pindex); /* * calculate the size of the transfer */ size = count * PAGE_SIZE; KASSERT(count > 0, ("zero count")); if ((foff + size) > object->un_pager.vnp.vnp_size) size = object->un_pager.vnp.vnp_size - foff; KASSERT(size > 0, ("zero size")); /* * round up physical size for real devices. */ if (1) { int secmask = bo->bo_bsize - 1; KASSERT(secmask < PAGE_SIZE && secmask > 0, ("vnode_pager_generic_getpages: sector size %d too large", secmask + 1)); size = (size + secmask) & ~secmask; } bp = getpbuf(&vnode_pbuf_freecnt); kva = (vm_offset_t) bp->b_data; /* * and map the pages to be read into the kva */ pmap_qenter(kva, m, count); /* build a minimal buffer header */ bp->b_iocmd = BIO_READ; bp->b_iodone = bdone; KASSERT(bp->b_rcred == NOCRED, ("leaking read ucred")); KASSERT(bp->b_wcred == NOCRED, ("leaking write ucred")); bp->b_rcred = crhold(curthread->td_ucred); bp->b_wcred = crhold(curthread->td_ucred); bp->b_blkno = firstaddr; pbgetbo(bo, bp); bp->b_bcount = size; bp->b_bufsize = size; bp->b_runningbufspace = bp->b_bufsize; atomic_add_long(&runningbufspace, bp->b_runningbufspace); PCPU_INC(cnt.v_vnodein); PCPU_ADD(cnt.v_vnodepgsin, count); /* do the input */ bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); bwait(bp, PVM, "vnread"); if ((bp->b_ioflags & BIO_ERROR) != 0) error = EIO; if (!error) { if (size != count * PAGE_SIZE) bzero((caddr_t) kva + size, PAGE_SIZE * count - size); } pmap_qremove(kva, count); /* * free the buffer header back to the swap buffer pool */ pbrelbo(bp); relpbuf(bp, &vnode_pbuf_freecnt); VM_OBJECT_LOCK(object); for (i = 0, tfoff = foff; i < count; i++, tfoff = nextoff) { vm_page_t mt; nextoff = tfoff + PAGE_SIZE; mt = m[i]; if (nextoff <= object->un_pager.vnp.vnp_size) { /* * Read filled up entire page. */ mt->valid = VM_PAGE_BITS_ALL; KASSERT(mt->dirty == 0, ("vnode_pager_generic_getpages: page %p is dirty", mt)); KASSERT(!pmap_page_is_mapped(mt), ("vnode_pager_generic_getpages: page %p is mapped", mt)); } else { /* * Read did not fill up entire page. * * Currently we do not set the entire page valid, * we just try to clear the piece that we couldn't * read. */ vm_page_set_valid(mt, 0, object->un_pager.vnp.vnp_size - tfoff); KASSERT((mt->dirty & vm_page_bits(0, object->un_pager.vnp.vnp_size - tfoff)) == 0, ("vnode_pager_generic_getpages: page %p is dirty", mt)); } if (i != reqpage) { /* * whether or not to leave the page activated is up in * the air, but we should put the page on a page queue * somewhere. (it already is in the object). Result: * It appears that empirical results show that * deactivating pages is best. */ /* * just in case someone was asking for this page we * now tell them that it is ok to use */ if (!error) { if (mt->oflags & VPO_WANTED) { vm_page_lock(mt); vm_page_activate(mt); vm_page_unlock(mt); } else { vm_page_lock(mt); vm_page_deactivate(mt); vm_page_unlock(mt); } vm_page_wakeup(mt); } else { vm_page_lock(mt); vm_page_free(mt); vm_page_unlock(mt); } } } VM_OBJECT_UNLOCK(object); if (error) { printf("vnode_pager_getpages: I/O read error\n"); } return (error ? VM_PAGER_ERROR : VM_PAGER_OK); } /* * EOPNOTSUPP is no longer legal. For local media VFS's that do not * implement their own VOP_PUTPAGES, their VOP_PUTPAGES should call to * vnode_pager_generic_putpages() to implement the previous behaviour. * * All other FS's should use the bypass to get to the local media * backing vp's VOP_PUTPAGES. */ static void vnode_pager_putpages(object, m, count, sync, rtvals) vm_object_t object; vm_page_t *m; int count; boolean_t sync; int *rtvals; { int rtval; struct vnode *vp; int bytes = count * PAGE_SIZE; /* * Force synchronous operation if we are extremely low on memory * to prevent a low-memory deadlock. VOP operations often need to * allocate more memory to initiate the I/O ( i.e. do a BMAP * operation ). The swapper handles the case by limiting the amount * of asynchronous I/O, but that sort of solution doesn't scale well * for the vnode pager without a lot of work. * * Also, the backing vnode's iodone routine may not wake the pageout * daemon up. This should be probably be addressed XXX. */ if ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min) sync |= OBJPC_SYNC; /* * Call device-specific putpages function */ vp = object->handle; VM_OBJECT_UNLOCK(object); rtval = VOP_PUTPAGES(vp, m, bytes, sync, rtvals, 0); KASSERT(rtval != EOPNOTSUPP, ("vnode_pager: stale FS putpages\n")); VM_OBJECT_LOCK(object); } /* * This is now called from local media FS's to operate against their * own vnodes if they fail to implement VOP_PUTPAGES. * * This is typically called indirectly via the pageout daemon and * clustering has already typically occured, so in general we ask the * underlying filesystem to write the data out asynchronously rather * then delayed. */ int vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *ma, int bytecount, int flags, int *rtvals) { int i; vm_object_t object; vm_page_t m; int count; int maxsize, ncount; vm_ooffset_t poffset; struct uio auio; struct iovec aiov; int error; int ioflags; int ppscheck = 0; static struct timeval lastfail; static int curfail; object = vp->v_object; count = bytecount / PAGE_SIZE; for (i = 0; i < count; i++) - rtvals[i] = VM_PAGER_AGAIN; + rtvals[i] = VM_PAGER_ERROR; if ((int64_t)ma[0]->pindex < 0) { printf("vnode_pager_putpages: attempt to write meta-data!!! -- 0x%lx(%lx)\n", (long)ma[0]->pindex, (u_long)ma[0]->dirty); rtvals[0] = VM_PAGER_BAD; return VM_PAGER_BAD; } maxsize = count * PAGE_SIZE; ncount = count; poffset = IDX_TO_OFF(ma[0]->pindex); /* * If the page-aligned write is larger then the actual file we * have to invalidate pages occuring beyond the file EOF. However, * there is an edge case where a file may not be page-aligned where * the last page is partially invalid. In this case the filesystem * may not properly clear the dirty bits for the entire page (which * could be VM_PAGE_BITS_ALL due to the page having been mmap()d). * With the page locked we are free to fix-up the dirty bits here. * * We do not under any circumstances truncate the valid bits, as * this will screw up bogus page replacement. */ VM_OBJECT_LOCK(object); if (maxsize + poffset > object->un_pager.vnp.vnp_size) { if (object->un_pager.vnp.vnp_size > poffset) { int pgoff; maxsize = object->un_pager.vnp.vnp_size - poffset; ncount = btoc(maxsize); if ((pgoff = (int)maxsize & PAGE_MASK) != 0) { /* * If the object is locked and the following * conditions hold, then the page's dirty * field cannot be concurrently changed by a * pmap operation. */ m = ma[ncount - 1]; KASSERT(m->busy > 0, ("vnode_pager_generic_putpages: page %p is not busy", m)); KASSERT((m->flags & PG_WRITEABLE) == 0, ("vnode_pager_generic_putpages: page %p is not read-only", m)); vm_page_clear_dirty(m, pgoff, PAGE_SIZE - pgoff); } } else { maxsize = 0; ncount = 0; } if (ncount < count) { for (i = ncount; i < count; i++) { rtvals[i] = VM_PAGER_BAD; } } } VM_OBJECT_UNLOCK(object); /* * pageouts are already clustered, use IO_ASYNC t o force a bawrite() * rather then a bdwrite() to prevent paging I/O from saturating * the buffer cache. Dummy-up the sequential heuristic to cause * large ranges to cluster. If neither IO_SYNC or IO_ASYNC is set, * the system decides how to cluster. */ ioflags = IO_VMIO; if (flags & (VM_PAGER_PUT_SYNC | VM_PAGER_PUT_INVAL)) ioflags |= IO_SYNC; else if ((flags & VM_PAGER_CLUSTER_OK) == 0) ioflags |= IO_ASYNC; ioflags |= (flags & VM_PAGER_PUT_INVAL) ? IO_INVAL: 0; ioflags |= IO_SEQMAX << IO_SEQSHIFT; aiov.iov_base = (caddr_t) 0; aiov.iov_len = maxsize; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_offset = poffset; auio.uio_segflg = UIO_NOCOPY; auio.uio_rw = UIO_WRITE; auio.uio_resid = maxsize; auio.uio_td = (struct thread *) 0; error = VOP_WRITE(vp, &auio, ioflags, curthread->td_ucred); PCPU_INC(cnt.v_vnodeout); PCPU_ADD(cnt.v_vnodepgsout, ncount); if (error) { if ((ppscheck = ppsratecheck(&lastfail, &curfail, 1))) printf("vnode_pager_putpages: I/O error %d\n", error); } if (auio.uio_resid) { if (ppscheck || ppsratecheck(&lastfail, &curfail, 1)) printf("vnode_pager_putpages: residual I/O %zd at %lu\n", auio.uio_resid, (u_long)ma[0]->pindex); } for (i = 0; i < ncount; i++) { rtvals[i] = VM_PAGER_OK; } return rtvals[0]; +} + +void +vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written) +{ + int i, pos; + + for (i = 0, pos = 0; pos < written; i++, pos += PAGE_SIZE) { + if (pos < trunc_page(written)) { + rtvals[i] = VM_PAGER_OK; + vm_page_undirty(ma[i]); + } else { + /* Partially written page. */ + rtvals[i] = VM_PAGER_AGAIN; + vm_page_clear_dirty(ma[i], 0, written & PAGE_MASK); + } + } } Index: head/sys/vm/vnode_pager.h =================================================================== --- head/sys/vm/vnode_pager.h (revision 222585) +++ head/sys/vm/vnode_pager.h (revision 222586) @@ -1,53 +1,56 @@ /*- * Copyright (c) 1990 University of Utah. * Copyright (c) 1991, 1993 * The Regents of the University of California. 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. * 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.h 8.1 (Berkeley) 6/11/93 * $FreeBSD$ */ #ifndef _VNODE_PAGER_ #define _VNODE_PAGER_ 1 #ifdef _KERNEL /* * XXX Generic routines; currently called by badly written FS code; these * XXX should go away soon. */ int vnode_pager_generic_getpages(struct vnode *vp, vm_page_t *m, int count, int reqpage); int vnode_pager_generic_putpages(struct vnode *vp, vm_page_t *m, int count, boolean_t sync, int *rtvals); + +void vnode_pager_undirty_pages(vm_page_t *ma, int *rtvals, int written); + #endif /* _KERNEL */ #endif /* _VNODE_PAGER_ */