nfs_clbio.c revision 1.2.10.3
1 1.2.10.2 tls /* $NetBSD: nfs_clbio.c,v 1.2.10.3 2017/12/03 11:38:42 jdolecek Exp $ */ 2 1.2.10.2 tls /*- 3 1.2.10.2 tls * Copyright (c) 1989, 1993 4 1.2.10.2 tls * The Regents of the University of California. All rights reserved. 5 1.2.10.2 tls * 6 1.2.10.2 tls * This code is derived from software contributed to Berkeley by 7 1.2.10.2 tls * Rick Macklem at The University of Guelph. 8 1.2.10.2 tls * 9 1.2.10.2 tls * Redistribution and use in source and binary forms, with or without 10 1.2.10.2 tls * modification, are permitted provided that the following conditions 11 1.2.10.2 tls * are met: 12 1.2.10.2 tls * 1. Redistributions of source code must retain the above copyright 13 1.2.10.2 tls * notice, this list of conditions and the following disclaimer. 14 1.2.10.2 tls * 2. Redistributions in binary form must reproduce the above copyright 15 1.2.10.2 tls * notice, this list of conditions and the following disclaimer in the 16 1.2.10.2 tls * documentation and/or other materials provided with the distribution. 17 1.2.10.2 tls * 4. Neither the name of the University nor the names of its contributors 18 1.2.10.2 tls * may be used to endorse or promote products derived from this software 19 1.2.10.2 tls * without specific prior written permission. 20 1.2.10.2 tls * 21 1.2.10.2 tls * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 1.2.10.2 tls * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 1.2.10.2 tls * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 1.2.10.2 tls * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 1.2.10.2 tls * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 1.2.10.2 tls * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 1.2.10.2 tls * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 1.2.10.2 tls * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 1.2.10.2 tls * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 1.2.10.2 tls * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 1.2.10.2 tls * SUCH DAMAGE. 32 1.2.10.2 tls * 33 1.2.10.2 tls * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 34 1.2.10.2 tls */ 35 1.2.10.2 tls 36 1.2.10.2 tls #include <sys/cdefs.h> 37 1.2.10.3 jdolecek /* __FBSDID("FreeBSD: head/sys/fs/nfsclient/nfs_clbio.c 304026 2016-08-12 22:44:59Z rmacklem "); */ 38 1.2.10.2 tls __RCSID("$NetBSD: nfs_clbio.c,v 1.2.10.3 2017/12/03 11:38:42 jdolecek Exp $"); 39 1.2.10.2 tls 40 1.2.10.2 tls #include <sys/param.h> 41 1.2.10.2 tls #include <sys/systm.h> 42 1.2.10.2 tls #include <sys/buf.h> 43 1.2.10.2 tls #include <sys/kernel.h> 44 1.2.10.2 tls #include <sys/mount.h> 45 1.2.10.2 tls #include <sys/rwlock.h> 46 1.2.10.2 tls #include <sys/vmmeter.h> 47 1.2.10.2 tls #include <sys/vnode.h> 48 1.2.10.2 tls 49 1.2.10.3 jdolecek #include <fs/nfs/common/nfsport.h> 50 1.2.10.3 jdolecek #include <fs/nfs/client/nfsmount.h> 51 1.2.10.3 jdolecek #include <fs/nfs/client/nfs.h> 52 1.2.10.3 jdolecek #include <fs/nfs/client/nfsnode.h> 53 1.2.10.3 jdolecek #include <fs/nfs/client/nfs_kdtrace.h> 54 1.2.10.2 tls 55 1.2.10.2 tls extern int newnfs_directio_allow_mmap; 56 1.2.10.3 jdolecek extern struct nfsstatsv1 nfsstatsv1; 57 1.2.10.2 tls extern struct mtx ncl_iod_mutex; 58 1.2.10.2 tls extern int ncl_numasync; 59 1.2.10.2 tls extern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON]; 60 1.2.10.2 tls extern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON]; 61 1.2.10.2 tls extern int newnfs_directio_enable; 62 1.2.10.2 tls extern int nfs_keep_dirty_on_error; 63 1.2.10.2 tls 64 1.2.10.2 tls int ncl_pbuf_freecnt = -1; /* start out unlimited */ 65 1.2.10.2 tls 66 1.2.10.2 tls static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 67 1.2.10.2 tls struct thread *td); 68 1.2.10.2 tls static int nfs_directio_write(struct vnode *vp, struct uio *uiop, 69 1.2.10.2 tls struct ucred *cred, int ioflag); 70 1.2.10.2 tls 71 1.2.10.2 tls /* 72 1.2.10.2 tls * Vnode op for VM getpages. 73 1.2.10.2 tls */ 74 1.2.10.2 tls int 75 1.2.10.2 tls ncl_getpages(struct vop_getpages_args *ap) 76 1.2.10.2 tls { 77 1.2.10.2 tls int i, error, nextoff, size, toff, count, npages; 78 1.2.10.2 tls struct uio uio; 79 1.2.10.2 tls struct iovec iov; 80 1.2.10.2 tls vm_offset_t kva; 81 1.2.10.2 tls struct buf *bp; 82 1.2.10.2 tls struct vnode *vp; 83 1.2.10.2 tls struct thread *td; 84 1.2.10.2 tls struct ucred *cred; 85 1.2.10.2 tls struct nfsmount *nmp; 86 1.2.10.2 tls vm_object_t object; 87 1.2.10.2 tls vm_page_t *pages; 88 1.2.10.2 tls struct nfsnode *np; 89 1.2.10.2 tls 90 1.2.10.2 tls vp = ap->a_vp; 91 1.2.10.2 tls np = VTONFS(vp); 92 1.2.10.2 tls td = curthread; /* XXX */ 93 1.2.10.2 tls cred = curthread->td_ucred; /* XXX */ 94 1.2.10.2 tls nmp = VFSTONFS(vp->v_mount); 95 1.2.10.2 tls pages = ap->a_m; 96 1.2.10.3 jdolecek npages = ap->a_count; 97 1.2.10.2 tls 98 1.2.10.2 tls if ((object = vp->v_object) == NULL) { 99 1.2.10.3 jdolecek printf("ncl_getpages: called with non-merged cache vnode\n"); 100 1.2.10.2 tls return (VM_PAGER_ERROR); 101 1.2.10.2 tls } 102 1.2.10.2 tls 103 1.2.10.2 tls if (newnfs_directio_enable && !newnfs_directio_allow_mmap) { 104 1.2.10.2 tls mtx_lock(&np->n_mtx); 105 1.2.10.2 tls if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 106 1.2.10.2 tls mtx_unlock(&np->n_mtx); 107 1.2.10.3 jdolecek printf("ncl_getpages: called on non-cacheable vnode\n"); 108 1.2.10.2 tls return (VM_PAGER_ERROR); 109 1.2.10.2 tls } else 110 1.2.10.2 tls mtx_unlock(&np->n_mtx); 111 1.2.10.2 tls } 112 1.2.10.2 tls 113 1.2.10.2 tls mtx_lock(&nmp->nm_mtx); 114 1.2.10.2 tls if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 115 1.2.10.2 tls (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 116 1.2.10.2 tls mtx_unlock(&nmp->nm_mtx); 117 1.2.10.2 tls /* We'll never get here for v4, because we always have fsinfo */ 118 1.2.10.2 tls (void)ncl_fsinfo(nmp, vp, cred, td); 119 1.2.10.2 tls } else 120 1.2.10.2 tls mtx_unlock(&nmp->nm_mtx); 121 1.2.10.2 tls 122 1.2.10.2 tls /* 123 1.2.10.2 tls * If the requested page is partially valid, just return it and 124 1.2.10.2 tls * allow the pager to zero-out the blanks. Partially valid pages 125 1.2.10.2 tls * can only occur at the file EOF. 126 1.2.10.3 jdolecek * 127 1.2.10.3 jdolecek * XXXGL: is that true for NFS, where short read can occur??? 128 1.2.10.2 tls */ 129 1.2.10.2 tls VM_OBJECT_WLOCK(object); 130 1.2.10.3 jdolecek if (pages[npages - 1]->valid != 0 && --npages == 0) 131 1.2.10.3 jdolecek goto out; 132 1.2.10.2 tls VM_OBJECT_WUNLOCK(object); 133 1.2.10.2 tls 134 1.2.10.2 tls /* 135 1.2.10.2 tls * We use only the kva address for the buffer, but this is extremely 136 1.2.10.3 jdolecek * convenient and fast. 137 1.2.10.2 tls */ 138 1.2.10.2 tls bp = getpbuf(&ncl_pbuf_freecnt); 139 1.2.10.2 tls 140 1.2.10.2 tls kva = (vm_offset_t) bp->b_data; 141 1.2.10.2 tls pmap_qenter(kva, pages, npages); 142 1.2.10.2 tls PCPU_INC(cnt.v_vnodein); 143 1.2.10.2 tls PCPU_ADD(cnt.v_vnodepgsin, npages); 144 1.2.10.2 tls 145 1.2.10.3 jdolecek count = npages << PAGE_SHIFT; 146 1.2.10.2 tls iov.iov_base = (caddr_t) kva; 147 1.2.10.2 tls iov.iov_len = count; 148 1.2.10.2 tls uio.uio_iov = &iov; 149 1.2.10.2 tls uio.uio_iovcnt = 1; 150 1.2.10.2 tls uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 151 1.2.10.2 tls uio.uio_resid = count; 152 1.2.10.2 tls uio.uio_segflg = UIO_SYSSPACE; 153 1.2.10.2 tls uio.uio_rw = UIO_READ; 154 1.2.10.2 tls uio.uio_td = td; 155 1.2.10.2 tls 156 1.2.10.2 tls error = ncl_readrpc(vp, &uio, cred); 157 1.2.10.2 tls pmap_qremove(kva, npages); 158 1.2.10.2 tls 159 1.2.10.2 tls relpbuf(bp, &ncl_pbuf_freecnt); 160 1.2.10.2 tls 161 1.2.10.2 tls if (error && (uio.uio_resid == count)) { 162 1.2.10.3 jdolecek printf("ncl_getpages: error %d\n", error); 163 1.2.10.2 tls return (VM_PAGER_ERROR); 164 1.2.10.2 tls } 165 1.2.10.2 tls 166 1.2.10.2 tls /* 167 1.2.10.2 tls * Calculate the number of bytes read and validate only that number 168 1.2.10.2 tls * of bytes. Note that due to pending writes, size may be 0. This 169 1.2.10.2 tls * does not mean that the remaining data is invalid! 170 1.2.10.2 tls */ 171 1.2.10.2 tls 172 1.2.10.2 tls size = count - uio.uio_resid; 173 1.2.10.2 tls VM_OBJECT_WLOCK(object); 174 1.2.10.2 tls for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 175 1.2.10.2 tls vm_page_t m; 176 1.2.10.2 tls nextoff = toff + PAGE_SIZE; 177 1.2.10.2 tls m = pages[i]; 178 1.2.10.2 tls 179 1.2.10.2 tls if (nextoff <= size) { 180 1.2.10.2 tls /* 181 1.2.10.2 tls * Read operation filled an entire page 182 1.2.10.2 tls */ 183 1.2.10.2 tls m->valid = VM_PAGE_BITS_ALL; 184 1.2.10.2 tls KASSERT(m->dirty == 0, 185 1.2.10.2 tls ("nfs_getpages: page %p is dirty", m)); 186 1.2.10.2 tls } else if (size > toff) { 187 1.2.10.2 tls /* 188 1.2.10.2 tls * Read operation filled a partial page. 189 1.2.10.2 tls */ 190 1.2.10.2 tls m->valid = 0; 191 1.2.10.2 tls vm_page_set_valid_range(m, 0, size - toff); 192 1.2.10.2 tls KASSERT(m->dirty == 0, 193 1.2.10.2 tls ("nfs_getpages: page %p is dirty", m)); 194 1.2.10.2 tls } else { 195 1.2.10.2 tls /* 196 1.2.10.2 tls * Read operation was short. If no error 197 1.2.10.3 jdolecek * occurred we may have hit a zero-fill 198 1.2.10.2 tls * section. We leave valid set to 0, and page 199 1.2.10.2 tls * is freed by vm_page_readahead_finish() if 200 1.2.10.2 tls * its index is not equal to requested, or 201 1.2.10.2 tls * page is zeroed and set valid by 202 1.2.10.2 tls * vm_pager_get_pages() for requested page. 203 1.2.10.2 tls */ 204 1.2.10.2 tls ; 205 1.2.10.2 tls } 206 1.2.10.2 tls } 207 1.2.10.3 jdolecek out: 208 1.2.10.2 tls VM_OBJECT_WUNLOCK(object); 209 1.2.10.3 jdolecek if (ap->a_rbehind) 210 1.2.10.3 jdolecek *ap->a_rbehind = 0; 211 1.2.10.3 jdolecek if (ap->a_rahead) 212 1.2.10.3 jdolecek *ap->a_rahead = 0; 213 1.2.10.3 jdolecek return (VM_PAGER_OK); 214 1.2.10.2 tls } 215 1.2.10.2 tls 216 1.2.10.2 tls /* 217 1.2.10.2 tls * Vnode op for VM putpages. 218 1.2.10.2 tls */ 219 1.2.10.2 tls int 220 1.2.10.2 tls ncl_putpages(struct vop_putpages_args *ap) 221 1.2.10.2 tls { 222 1.2.10.2 tls struct uio uio; 223 1.2.10.2 tls struct iovec iov; 224 1.2.10.2 tls vm_offset_t kva; 225 1.2.10.2 tls struct buf *bp; 226 1.2.10.2 tls int iomode, must_commit, i, error, npages, count; 227 1.2.10.2 tls off_t offset; 228 1.2.10.2 tls int *rtvals; 229 1.2.10.2 tls struct vnode *vp; 230 1.2.10.2 tls struct thread *td; 231 1.2.10.2 tls struct ucred *cred; 232 1.2.10.2 tls struct nfsmount *nmp; 233 1.2.10.2 tls struct nfsnode *np; 234 1.2.10.2 tls vm_page_t *pages; 235 1.2.10.2 tls 236 1.2.10.2 tls vp = ap->a_vp; 237 1.2.10.2 tls np = VTONFS(vp); 238 1.2.10.2 tls td = curthread; /* XXX */ 239 1.2.10.2 tls /* Set the cred to n_writecred for the write rpcs. */ 240 1.2.10.2 tls if (np->n_writecred != NULL) 241 1.2.10.2 tls cred = crhold(np->n_writecred); 242 1.2.10.2 tls else 243 1.2.10.2 tls cred = crhold(curthread->td_ucred); /* XXX */ 244 1.2.10.2 tls nmp = VFSTONFS(vp->v_mount); 245 1.2.10.2 tls pages = ap->a_m; 246 1.2.10.2 tls count = ap->a_count; 247 1.2.10.2 tls rtvals = ap->a_rtvals; 248 1.2.10.2 tls npages = btoc(count); 249 1.2.10.2 tls offset = IDX_TO_OFF(pages[0]->pindex); 250 1.2.10.2 tls 251 1.2.10.2 tls mtx_lock(&nmp->nm_mtx); 252 1.2.10.2 tls if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 253 1.2.10.2 tls (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 254 1.2.10.2 tls mtx_unlock(&nmp->nm_mtx); 255 1.2.10.2 tls (void)ncl_fsinfo(nmp, vp, cred, td); 256 1.2.10.2 tls } else 257 1.2.10.2 tls mtx_unlock(&nmp->nm_mtx); 258 1.2.10.2 tls 259 1.2.10.2 tls mtx_lock(&np->n_mtx); 260 1.2.10.2 tls if (newnfs_directio_enable && !newnfs_directio_allow_mmap && 261 1.2.10.2 tls (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 262 1.2.10.2 tls mtx_unlock(&np->n_mtx); 263 1.2.10.3 jdolecek printf("ncl_putpages: called on noncache-able vnode\n"); 264 1.2.10.2 tls mtx_lock(&np->n_mtx); 265 1.2.10.2 tls } 266 1.2.10.2 tls 267 1.2.10.2 tls for (i = 0; i < npages; i++) 268 1.2.10.2 tls rtvals[i] = VM_PAGER_ERROR; 269 1.2.10.2 tls 270 1.2.10.2 tls /* 271 1.2.10.2 tls * When putting pages, do not extend file past EOF. 272 1.2.10.2 tls */ 273 1.2.10.2 tls if (offset + count > np->n_size) { 274 1.2.10.2 tls count = np->n_size - offset; 275 1.2.10.2 tls if (count < 0) 276 1.2.10.2 tls count = 0; 277 1.2.10.2 tls } 278 1.2.10.2 tls mtx_unlock(&np->n_mtx); 279 1.2.10.2 tls 280 1.2.10.2 tls /* 281 1.2.10.2 tls * We use only the kva address for the buffer, but this is extremely 282 1.2.10.3 jdolecek * convenient and fast. 283 1.2.10.2 tls */ 284 1.2.10.2 tls bp = getpbuf(&ncl_pbuf_freecnt); 285 1.2.10.2 tls 286 1.2.10.2 tls kva = (vm_offset_t) bp->b_data; 287 1.2.10.2 tls pmap_qenter(kva, pages, npages); 288 1.2.10.2 tls PCPU_INC(cnt.v_vnodeout); 289 1.2.10.2 tls PCPU_ADD(cnt.v_vnodepgsout, count); 290 1.2.10.2 tls 291 1.2.10.2 tls iov.iov_base = (caddr_t) kva; 292 1.2.10.2 tls iov.iov_len = count; 293 1.2.10.2 tls uio.uio_iov = &iov; 294 1.2.10.2 tls uio.uio_iovcnt = 1; 295 1.2.10.2 tls uio.uio_offset = offset; 296 1.2.10.2 tls uio.uio_resid = count; 297 1.2.10.2 tls uio.uio_segflg = UIO_SYSSPACE; 298 1.2.10.2 tls uio.uio_rw = UIO_WRITE; 299 1.2.10.2 tls uio.uio_td = td; 300 1.2.10.2 tls 301 1.2.10.2 tls if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 302 1.2.10.2 tls iomode = NFSWRITE_UNSTABLE; 303 1.2.10.2 tls else 304 1.2.10.2 tls iomode = NFSWRITE_FILESYNC; 305 1.2.10.2 tls 306 1.2.10.2 tls error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0); 307 1.2.10.2 tls crfree(cred); 308 1.2.10.2 tls 309 1.2.10.2 tls pmap_qremove(kva, npages); 310 1.2.10.2 tls relpbuf(bp, &ncl_pbuf_freecnt); 311 1.2.10.2 tls 312 1.2.10.2 tls if (error == 0 || !nfs_keep_dirty_on_error) { 313 1.2.10.2 tls vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid); 314 1.2.10.2 tls if (must_commit) 315 1.2.10.2 tls ncl_clearcommit(vp->v_mount); 316 1.2.10.2 tls } 317 1.2.10.2 tls return rtvals[0]; 318 1.2.10.2 tls } 319 1.2.10.2 tls 320 1.2.10.2 tls /* 321 1.2.10.2 tls * For nfs, cache consistency can only be maintained approximately. 322 1.2.10.2 tls * Although RFC1094 does not specify the criteria, the following is 323 1.2.10.2 tls * believed to be compatible with the reference port. 324 1.2.10.2 tls * For nfs: 325 1.2.10.2 tls * If the file's modify time on the server has changed since the 326 1.2.10.2 tls * last read rpc or you have written to the file, 327 1.2.10.2 tls * you may have lost data cache consistency with the 328 1.2.10.2 tls * server, so flush all of the file's data out of the cache. 329 1.2.10.2 tls * Then force a getattr rpc to ensure that you have up to date 330 1.2.10.2 tls * attributes. 331 1.2.10.2 tls * NB: This implies that cache data can be read when up to 332 1.2.10.2 tls * NFS_ATTRTIMEO seconds out of date. If you find that you need current 333 1.2.10.2 tls * attributes this could be forced by setting n_attrstamp to 0 before 334 1.2.10.2 tls * the VOP_GETATTR() call. 335 1.2.10.2 tls */ 336 1.2.10.2 tls static inline int 337 1.2.10.2 tls nfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 338 1.2.10.2 tls { 339 1.2.10.2 tls int error = 0; 340 1.2.10.2 tls struct vattr vattr; 341 1.2.10.2 tls struct nfsnode *np = VTONFS(vp); 342 1.2.10.2 tls int old_lock; 343 1.2.10.2 tls 344 1.2.10.2 tls /* 345 1.2.10.2 tls * Grab the exclusive lock before checking whether the cache is 346 1.2.10.2 tls * consistent. 347 1.2.10.2 tls * XXX - We can make this cheaper later (by acquiring cheaper locks). 348 1.2.10.2 tls * But for now, this suffices. 349 1.2.10.2 tls */ 350 1.2.10.2 tls old_lock = ncl_upgrade_vnlock(vp); 351 1.2.10.2 tls if (vp->v_iflag & VI_DOOMED) { 352 1.2.10.2 tls ncl_downgrade_vnlock(vp, old_lock); 353 1.2.10.2 tls return (EBADF); 354 1.2.10.2 tls } 355 1.2.10.2 tls 356 1.2.10.2 tls mtx_lock(&np->n_mtx); 357 1.2.10.2 tls if (np->n_flag & NMODIFIED) { 358 1.2.10.2 tls mtx_unlock(&np->n_mtx); 359 1.2.10.2 tls if (vp->v_type != VREG) { 360 1.2.10.2 tls if (vp->v_type != VDIR) 361 1.2.10.2 tls panic("nfs: bioread, not dir"); 362 1.2.10.2 tls ncl_invaldir(vp); 363 1.2.10.2 tls error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 364 1.2.10.2 tls if (error) 365 1.2.10.2 tls goto out; 366 1.2.10.2 tls } 367 1.2.10.2 tls np->n_attrstamp = 0; 368 1.2.10.2 tls KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 369 1.2.10.2 tls error = VOP_GETATTR(vp, &vattr, cred); 370 1.2.10.2 tls if (error) 371 1.2.10.2 tls goto out; 372 1.2.10.2 tls mtx_lock(&np->n_mtx); 373 1.2.10.2 tls np->n_mtime = vattr.va_mtime; 374 1.2.10.2 tls mtx_unlock(&np->n_mtx); 375 1.2.10.2 tls } else { 376 1.2.10.2 tls mtx_unlock(&np->n_mtx); 377 1.2.10.2 tls error = VOP_GETATTR(vp, &vattr, cred); 378 1.2.10.2 tls if (error) 379 1.2.10.2 tls return (error); 380 1.2.10.2 tls mtx_lock(&np->n_mtx); 381 1.2.10.2 tls if ((np->n_flag & NSIZECHANGED) 382 1.2.10.2 tls || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 383 1.2.10.2 tls mtx_unlock(&np->n_mtx); 384 1.2.10.2 tls if (vp->v_type == VDIR) 385 1.2.10.2 tls ncl_invaldir(vp); 386 1.2.10.2 tls error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 387 1.2.10.2 tls if (error) 388 1.2.10.2 tls goto out; 389 1.2.10.2 tls mtx_lock(&np->n_mtx); 390 1.2.10.2 tls np->n_mtime = vattr.va_mtime; 391 1.2.10.2 tls np->n_flag &= ~NSIZECHANGED; 392 1.2.10.2 tls } 393 1.2.10.2 tls mtx_unlock(&np->n_mtx); 394 1.2.10.2 tls } 395 1.2.10.2 tls out: 396 1.2.10.2 tls ncl_downgrade_vnlock(vp, old_lock); 397 1.2.10.2 tls return error; 398 1.2.10.2 tls } 399 1.2.10.2 tls 400 1.2.10.2 tls /* 401 1.2.10.2 tls * Vnode op for read using bio 402 1.2.10.2 tls */ 403 1.2.10.2 tls int 404 1.2.10.2 tls ncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 405 1.2.10.2 tls { 406 1.2.10.2 tls struct nfsnode *np = VTONFS(vp); 407 1.2.10.2 tls int biosize, i; 408 1.2.10.2 tls struct buf *bp, *rabp; 409 1.2.10.2 tls struct thread *td; 410 1.2.10.2 tls struct nfsmount *nmp = VFSTONFS(vp->v_mount); 411 1.2.10.2 tls daddr_t lbn, rabn; 412 1.2.10.2 tls int bcount; 413 1.2.10.2 tls int seqcount; 414 1.2.10.2 tls int nra, error = 0, n = 0, on = 0; 415 1.2.10.2 tls off_t tmp_off; 416 1.2.10.2 tls 417 1.2.10.2 tls KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode")); 418 1.2.10.2 tls if (uio->uio_resid == 0) 419 1.2.10.2 tls return (0); 420 1.2.10.2 tls if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 421 1.2.10.2 tls return (EINVAL); 422 1.2.10.2 tls td = uio->uio_td; 423 1.2.10.2 tls 424 1.2.10.2 tls mtx_lock(&nmp->nm_mtx); 425 1.2.10.2 tls if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 426 1.2.10.2 tls (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 427 1.2.10.2 tls mtx_unlock(&nmp->nm_mtx); 428 1.2.10.2 tls (void)ncl_fsinfo(nmp, vp, cred, td); 429 1.2.10.2 tls mtx_lock(&nmp->nm_mtx); 430 1.2.10.2 tls } 431 1.2.10.2 tls if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0) 432 1.2.10.2 tls (void) newnfs_iosize(nmp); 433 1.2.10.2 tls 434 1.2.10.2 tls tmp_off = uio->uio_offset + uio->uio_resid; 435 1.2.10.2 tls if (vp->v_type != VDIR && 436 1.2.10.2 tls (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) { 437 1.2.10.2 tls mtx_unlock(&nmp->nm_mtx); 438 1.2.10.2 tls return (EFBIG); 439 1.2.10.2 tls } 440 1.2.10.2 tls mtx_unlock(&nmp->nm_mtx); 441 1.2.10.2 tls 442 1.2.10.2 tls if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 443 1.2.10.2 tls /* No caching/ no readaheads. Just read data into the user buffer */ 444 1.2.10.2 tls return ncl_readrpc(vp, uio, cred); 445 1.2.10.2 tls 446 1.2.10.2 tls biosize = vp->v_bufobj.bo_bsize; 447 1.2.10.2 tls seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 448 1.2.10.2 tls 449 1.2.10.2 tls error = nfs_bioread_check_cons(vp, td, cred); 450 1.2.10.2 tls if (error) 451 1.2.10.2 tls return error; 452 1.2.10.2 tls 453 1.2.10.2 tls do { 454 1.2.10.2 tls u_quad_t nsize; 455 1.2.10.2 tls 456 1.2.10.2 tls mtx_lock(&np->n_mtx); 457 1.2.10.2 tls nsize = np->n_size; 458 1.2.10.2 tls mtx_unlock(&np->n_mtx); 459 1.2.10.2 tls 460 1.2.10.2 tls switch (vp->v_type) { 461 1.2.10.2 tls case VREG: 462 1.2.10.3 jdolecek NFSINCRGLOBAL(nfsstatsv1.biocache_reads); 463 1.2.10.2 tls lbn = uio->uio_offset / biosize; 464 1.2.10.2 tls on = uio->uio_offset - (lbn * biosize); 465 1.2.10.2 tls 466 1.2.10.2 tls /* 467 1.2.10.2 tls * Start the read ahead(s), as required. 468 1.2.10.2 tls */ 469 1.2.10.2 tls if (nmp->nm_readahead > 0) { 470 1.2.10.2 tls for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 471 1.2.10.2 tls (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 472 1.2.10.2 tls rabn = lbn + 1 + nra; 473 1.2.10.2 tls if (incore(&vp->v_bufobj, rabn) == NULL) { 474 1.2.10.2 tls rabp = nfs_getcacheblk(vp, rabn, biosize, td); 475 1.2.10.2 tls if (!rabp) { 476 1.2.10.2 tls error = newnfs_sigintr(nmp, td); 477 1.2.10.2 tls return (error ? error : EINTR); 478 1.2.10.2 tls } 479 1.2.10.2 tls if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 480 1.2.10.2 tls rabp->b_flags |= B_ASYNC; 481 1.2.10.2 tls rabp->b_iocmd = BIO_READ; 482 1.2.10.2 tls vfs_busy_pages(rabp, 0); 483 1.2.10.2 tls if (ncl_asyncio(nmp, rabp, cred, td)) { 484 1.2.10.2 tls rabp->b_flags |= B_INVAL; 485 1.2.10.2 tls rabp->b_ioflags |= BIO_ERROR; 486 1.2.10.2 tls vfs_unbusy_pages(rabp); 487 1.2.10.2 tls brelse(rabp); 488 1.2.10.2 tls break; 489 1.2.10.2 tls } 490 1.2.10.2 tls } else { 491 1.2.10.2 tls brelse(rabp); 492 1.2.10.2 tls } 493 1.2.10.2 tls } 494 1.2.10.2 tls } 495 1.2.10.2 tls } 496 1.2.10.2 tls 497 1.2.10.2 tls /* Note that bcount is *not* DEV_BSIZE aligned. */ 498 1.2.10.2 tls bcount = biosize; 499 1.2.10.2 tls if ((off_t)lbn * biosize >= nsize) { 500 1.2.10.2 tls bcount = 0; 501 1.2.10.2 tls } else if ((off_t)(lbn + 1) * biosize > nsize) { 502 1.2.10.2 tls bcount = nsize - (off_t)lbn * biosize; 503 1.2.10.2 tls } 504 1.2.10.2 tls bp = nfs_getcacheblk(vp, lbn, bcount, td); 505 1.2.10.2 tls 506 1.2.10.2 tls if (!bp) { 507 1.2.10.2 tls error = newnfs_sigintr(nmp, td); 508 1.2.10.2 tls return (error ? error : EINTR); 509 1.2.10.2 tls } 510 1.2.10.2 tls 511 1.2.10.2 tls /* 512 1.2.10.2 tls * If B_CACHE is not set, we must issue the read. If this 513 1.2.10.2 tls * fails, we return an error. 514 1.2.10.2 tls */ 515 1.2.10.2 tls 516 1.2.10.2 tls if ((bp->b_flags & B_CACHE) == 0) { 517 1.2.10.2 tls bp->b_iocmd = BIO_READ; 518 1.2.10.2 tls vfs_busy_pages(bp, 0); 519 1.2.10.2 tls error = ncl_doio(vp, bp, cred, td, 0); 520 1.2.10.2 tls if (error) { 521 1.2.10.2 tls brelse(bp); 522 1.2.10.2 tls return (error); 523 1.2.10.2 tls } 524 1.2.10.2 tls } 525 1.2.10.2 tls 526 1.2.10.2 tls /* 527 1.2.10.2 tls * on is the offset into the current bp. Figure out how many 528 1.2.10.2 tls * bytes we can copy out of the bp. Note that bcount is 529 1.2.10.2 tls * NOT DEV_BSIZE aligned. 530 1.2.10.2 tls * 531 1.2.10.2 tls * Then figure out how many bytes we can copy into the uio. 532 1.2.10.2 tls */ 533 1.2.10.2 tls 534 1.2.10.2 tls n = 0; 535 1.2.10.2 tls if (on < bcount) 536 1.2.10.2 tls n = MIN((unsigned)(bcount - on), uio->uio_resid); 537 1.2.10.2 tls break; 538 1.2.10.2 tls case VLNK: 539 1.2.10.3 jdolecek NFSINCRGLOBAL(nfsstatsv1.biocache_readlinks); 540 1.2.10.2 tls bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 541 1.2.10.2 tls if (!bp) { 542 1.2.10.2 tls error = newnfs_sigintr(nmp, td); 543 1.2.10.2 tls return (error ? error : EINTR); 544 1.2.10.2 tls } 545 1.2.10.2 tls if ((bp->b_flags & B_CACHE) == 0) { 546 1.2.10.2 tls bp->b_iocmd = BIO_READ; 547 1.2.10.2 tls vfs_busy_pages(bp, 0); 548 1.2.10.2 tls error = ncl_doio(vp, bp, cred, td, 0); 549 1.2.10.2 tls if (error) { 550 1.2.10.2 tls bp->b_ioflags |= BIO_ERROR; 551 1.2.10.2 tls brelse(bp); 552 1.2.10.2 tls return (error); 553 1.2.10.2 tls } 554 1.2.10.2 tls } 555 1.2.10.2 tls n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 556 1.2.10.2 tls on = 0; 557 1.2.10.2 tls break; 558 1.2.10.2 tls case VDIR: 559 1.2.10.3 jdolecek NFSINCRGLOBAL(nfsstatsv1.biocache_readdirs); 560 1.2.10.2 tls if (np->n_direofoffset 561 1.2.10.2 tls && uio->uio_offset >= np->n_direofoffset) { 562 1.2.10.2 tls return (0); 563 1.2.10.2 tls } 564 1.2.10.2 tls lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 565 1.2.10.2 tls on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 566 1.2.10.2 tls bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 567 1.2.10.2 tls if (!bp) { 568 1.2.10.2 tls error = newnfs_sigintr(nmp, td); 569 1.2.10.2 tls return (error ? error : EINTR); 570 1.2.10.2 tls } 571 1.2.10.2 tls if ((bp->b_flags & B_CACHE) == 0) { 572 1.2.10.2 tls bp->b_iocmd = BIO_READ; 573 1.2.10.2 tls vfs_busy_pages(bp, 0); 574 1.2.10.2 tls error = ncl_doio(vp, bp, cred, td, 0); 575 1.2.10.2 tls if (error) { 576 1.2.10.2 tls brelse(bp); 577 1.2.10.2 tls } 578 1.2.10.2 tls while (error == NFSERR_BAD_COOKIE) { 579 1.2.10.2 tls ncl_invaldir(vp); 580 1.2.10.2 tls error = ncl_vinvalbuf(vp, 0, td, 1); 581 1.2.10.2 tls /* 582 1.2.10.2 tls * Yuck! The directory has been modified on the 583 1.2.10.2 tls * server. The only way to get the block is by 584 1.2.10.2 tls * reading from the beginning to get all the 585 1.2.10.2 tls * offset cookies. 586 1.2.10.2 tls * 587 1.2.10.2 tls * Leave the last bp intact unless there is an error. 588 1.2.10.2 tls * Loop back up to the while if the error is another 589 1.2.10.2 tls * NFSERR_BAD_COOKIE (double yuch!). 590 1.2.10.2 tls */ 591 1.2.10.2 tls for (i = 0; i <= lbn && !error; i++) { 592 1.2.10.2 tls if (np->n_direofoffset 593 1.2.10.2 tls && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 594 1.2.10.2 tls return (0); 595 1.2.10.2 tls bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 596 1.2.10.2 tls if (!bp) { 597 1.2.10.2 tls error = newnfs_sigintr(nmp, td); 598 1.2.10.2 tls return (error ? error : EINTR); 599 1.2.10.2 tls } 600 1.2.10.2 tls if ((bp->b_flags & B_CACHE) == 0) { 601 1.2.10.2 tls bp->b_iocmd = BIO_READ; 602 1.2.10.2 tls vfs_busy_pages(bp, 0); 603 1.2.10.2 tls error = ncl_doio(vp, bp, cred, td, 0); 604 1.2.10.2 tls /* 605 1.2.10.2 tls * no error + B_INVAL == directory EOF, 606 1.2.10.2 tls * use the block. 607 1.2.10.2 tls */ 608 1.2.10.2 tls if (error == 0 && (bp->b_flags & B_INVAL)) 609 1.2.10.2 tls break; 610 1.2.10.2 tls } 611 1.2.10.2 tls /* 612 1.2.10.2 tls * An error will throw away the block and the 613 1.2.10.2 tls * for loop will break out. If no error and this 614 1.2.10.2 tls * is not the block we want, we throw away the 615 1.2.10.2 tls * block and go for the next one via the for loop. 616 1.2.10.2 tls */ 617 1.2.10.2 tls if (error || i < lbn) 618 1.2.10.2 tls brelse(bp); 619 1.2.10.2 tls } 620 1.2.10.2 tls } 621 1.2.10.2 tls /* 622 1.2.10.2 tls * The above while is repeated if we hit another cookie 623 1.2.10.2 tls * error. If we hit an error and it wasn't a cookie error, 624 1.2.10.2 tls * we give up. 625 1.2.10.2 tls */ 626 1.2.10.2 tls if (error) 627 1.2.10.2 tls return (error); 628 1.2.10.2 tls } 629 1.2.10.2 tls 630 1.2.10.2 tls /* 631 1.2.10.2 tls * If not eof and read aheads are enabled, start one. 632 1.2.10.2 tls * (You need the current block first, so that you have the 633 1.2.10.2 tls * directory offset cookie of the next block.) 634 1.2.10.2 tls */ 635 1.2.10.2 tls if (nmp->nm_readahead > 0 && 636 1.2.10.2 tls (bp->b_flags & B_INVAL) == 0 && 637 1.2.10.2 tls (np->n_direofoffset == 0 || 638 1.2.10.2 tls (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 639 1.2.10.2 tls incore(&vp->v_bufobj, lbn + 1) == NULL) { 640 1.2.10.2 tls rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 641 1.2.10.2 tls if (rabp) { 642 1.2.10.2 tls if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 643 1.2.10.2 tls rabp->b_flags |= B_ASYNC; 644 1.2.10.2 tls rabp->b_iocmd = BIO_READ; 645 1.2.10.2 tls vfs_busy_pages(rabp, 0); 646 1.2.10.2 tls if (ncl_asyncio(nmp, rabp, cred, td)) { 647 1.2.10.2 tls rabp->b_flags |= B_INVAL; 648 1.2.10.2 tls rabp->b_ioflags |= BIO_ERROR; 649 1.2.10.2 tls vfs_unbusy_pages(rabp); 650 1.2.10.2 tls brelse(rabp); 651 1.2.10.2 tls } 652 1.2.10.2 tls } else { 653 1.2.10.2 tls brelse(rabp); 654 1.2.10.2 tls } 655 1.2.10.2 tls } 656 1.2.10.2 tls } 657 1.2.10.2 tls /* 658 1.2.10.2 tls * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 659 1.2.10.2 tls * chopped for the EOF condition, we cannot tell how large 660 1.2.10.2 tls * NFS directories are going to be until we hit EOF. So 661 1.2.10.2 tls * an NFS directory buffer is *not* chopped to its EOF. Now, 662 1.2.10.2 tls * it just so happens that b_resid will effectively chop it 663 1.2.10.2 tls * to EOF. *BUT* this information is lost if the buffer goes 664 1.2.10.2 tls * away and is reconstituted into a B_CACHE state ( due to 665 1.2.10.2 tls * being VMIO ) later. So we keep track of the directory eof 666 1.2.10.2 tls * in np->n_direofoffset and chop it off as an extra step 667 1.2.10.2 tls * right here. 668 1.2.10.2 tls */ 669 1.2.10.2 tls n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 670 1.2.10.2 tls if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 671 1.2.10.2 tls n = np->n_direofoffset - uio->uio_offset; 672 1.2.10.2 tls break; 673 1.2.10.2 tls default: 674 1.2.10.3 jdolecek printf(" ncl_bioread: type %x unexpected\n", vp->v_type); 675 1.2.10.2 tls bp = NULL; 676 1.2.10.2 tls break; 677 1.2.10.3 jdolecek } 678 1.2.10.2 tls 679 1.2.10.2 tls if (n > 0) { 680 1.2.10.2 tls error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio); 681 1.2.10.2 tls } 682 1.2.10.2 tls if (vp->v_type == VLNK) 683 1.2.10.2 tls n = 0; 684 1.2.10.2 tls if (bp != NULL) 685 1.2.10.2 tls brelse(bp); 686 1.2.10.2 tls } while (error == 0 && uio->uio_resid > 0 && n > 0); 687 1.2.10.2 tls return (error); 688 1.2.10.2 tls } 689 1.2.10.2 tls 690 1.2.10.2 tls /* 691 1.2.10.2 tls * The NFS write path cannot handle iovecs with len > 1. So we need to 692 1.2.10.2 tls * break up iovecs accordingly (restricting them to wsize). 693 1.2.10.2 tls * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 694 1.2.10.2 tls * For the ASYNC case, 2 copies are needed. The first a copy from the 695 1.2.10.2 tls * user buffer to a staging buffer and then a second copy from the staging 696 1.2.10.2 tls * buffer to mbufs. This can be optimized by copying from the user buffer 697 1.2.10.2 tls * directly into mbufs and passing the chain down, but that requires a 698 1.2.10.2 tls * fair amount of re-working of the relevant codepaths (and can be done 699 1.2.10.2 tls * later). 700 1.2.10.2 tls */ 701 1.2.10.2 tls static int 702 1.2.10.2 tls nfs_directio_write(vp, uiop, cred, ioflag) 703 1.2.10.2 tls struct vnode *vp; 704 1.2.10.2 tls struct uio *uiop; 705 1.2.10.2 tls struct ucred *cred; 706 1.2.10.2 tls int ioflag; 707 1.2.10.2 tls { 708 1.2.10.2 tls int error; 709 1.2.10.2 tls struct nfsmount *nmp = VFSTONFS(vp->v_mount); 710 1.2.10.2 tls struct thread *td = uiop->uio_td; 711 1.2.10.2 tls int size; 712 1.2.10.2 tls int wsize; 713 1.2.10.2 tls 714 1.2.10.2 tls mtx_lock(&nmp->nm_mtx); 715 1.2.10.2 tls wsize = nmp->nm_wsize; 716 1.2.10.2 tls mtx_unlock(&nmp->nm_mtx); 717 1.2.10.2 tls if (ioflag & IO_SYNC) { 718 1.2.10.2 tls int iomode, must_commit; 719 1.2.10.2 tls struct uio uio; 720 1.2.10.2 tls struct iovec iov; 721 1.2.10.2 tls do_sync: 722 1.2.10.2 tls while (uiop->uio_resid > 0) { 723 1.2.10.2 tls size = MIN(uiop->uio_resid, wsize); 724 1.2.10.2 tls size = MIN(uiop->uio_iov->iov_len, size); 725 1.2.10.2 tls iov.iov_base = uiop->uio_iov->iov_base; 726 1.2.10.2 tls iov.iov_len = size; 727 1.2.10.2 tls uio.uio_iov = &iov; 728 1.2.10.2 tls uio.uio_iovcnt = 1; 729 1.2.10.2 tls uio.uio_offset = uiop->uio_offset; 730 1.2.10.2 tls uio.uio_resid = size; 731 1.2.10.2 tls uio.uio_segflg = UIO_USERSPACE; 732 1.2.10.2 tls uio.uio_rw = UIO_WRITE; 733 1.2.10.2 tls uio.uio_td = td; 734 1.2.10.2 tls iomode = NFSWRITE_FILESYNC; 735 1.2.10.2 tls error = ncl_writerpc(vp, &uio, cred, &iomode, 736 1.2.10.2 tls &must_commit, 0); 737 1.2.10.2 tls KASSERT((must_commit == 0), 738 1.2.10.2 tls ("ncl_directio_write: Did not commit write")); 739 1.2.10.2 tls if (error) 740 1.2.10.2 tls return (error); 741 1.2.10.2 tls uiop->uio_offset += size; 742 1.2.10.2 tls uiop->uio_resid -= size; 743 1.2.10.2 tls if (uiop->uio_iov->iov_len <= size) { 744 1.2.10.2 tls uiop->uio_iovcnt--; 745 1.2.10.2 tls uiop->uio_iov++; 746 1.2.10.2 tls } else { 747 1.2.10.2 tls uiop->uio_iov->iov_base = 748 1.2.10.2 tls (char *)uiop->uio_iov->iov_base + size; 749 1.2.10.2 tls uiop->uio_iov->iov_len -= size; 750 1.2.10.2 tls } 751 1.2.10.2 tls } 752 1.2.10.2 tls } else { 753 1.2.10.2 tls struct uio *t_uio; 754 1.2.10.2 tls struct iovec *t_iov; 755 1.2.10.2 tls struct buf *bp; 756 1.2.10.2 tls 757 1.2.10.2 tls /* 758 1.2.10.2 tls * Break up the write into blocksize chunks and hand these 759 1.2.10.2 tls * over to nfsiod's for write back. 760 1.2.10.2 tls * Unfortunately, this incurs a copy of the data. Since 761 1.2.10.2 tls * the user could modify the buffer before the write is 762 1.2.10.2 tls * initiated. 763 1.2.10.2 tls * 764 1.2.10.2 tls * The obvious optimization here is that one of the 2 copies 765 1.2.10.2 tls * in the async write path can be eliminated by copying the 766 1.2.10.2 tls * data here directly into mbufs and passing the mbuf chain 767 1.2.10.2 tls * down. But that will require a fair amount of re-working 768 1.2.10.2 tls * of the code and can be done if there's enough interest 769 1.2.10.2 tls * in NFS directio access. 770 1.2.10.2 tls */ 771 1.2.10.2 tls while (uiop->uio_resid > 0) { 772 1.2.10.2 tls size = MIN(uiop->uio_resid, wsize); 773 1.2.10.2 tls size = MIN(uiop->uio_iov->iov_len, size); 774 1.2.10.2 tls bp = getpbuf(&ncl_pbuf_freecnt); 775 1.2.10.2 tls t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 776 1.2.10.2 tls t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 777 1.2.10.2 tls t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 778 1.2.10.2 tls t_iov->iov_len = size; 779 1.2.10.2 tls t_uio->uio_iov = t_iov; 780 1.2.10.2 tls t_uio->uio_iovcnt = 1; 781 1.2.10.2 tls t_uio->uio_offset = uiop->uio_offset; 782 1.2.10.2 tls t_uio->uio_resid = size; 783 1.2.10.2 tls t_uio->uio_segflg = UIO_SYSSPACE; 784 1.2.10.2 tls t_uio->uio_rw = UIO_WRITE; 785 1.2.10.2 tls t_uio->uio_td = td; 786 1.2.10.2 tls KASSERT(uiop->uio_segflg == UIO_USERSPACE || 787 1.2.10.2 tls uiop->uio_segflg == UIO_SYSSPACE, 788 1.2.10.2 tls ("nfs_directio_write: Bad uio_segflg")); 789 1.2.10.2 tls if (uiop->uio_segflg == UIO_USERSPACE) { 790 1.2.10.2 tls error = copyin(uiop->uio_iov->iov_base, 791 1.2.10.2 tls t_iov->iov_base, size); 792 1.2.10.2 tls if (error != 0) 793 1.2.10.2 tls goto err_free; 794 1.2.10.2 tls } else 795 1.2.10.2 tls /* 796 1.2.10.2 tls * UIO_SYSSPACE may never happen, but handle 797 1.2.10.2 tls * it just in case it does. 798 1.2.10.2 tls */ 799 1.2.10.2 tls bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, 800 1.2.10.2 tls size); 801 1.2.10.2 tls bp->b_flags |= B_DIRECT; 802 1.2.10.2 tls bp->b_iocmd = BIO_WRITE; 803 1.2.10.2 tls if (cred != NOCRED) { 804 1.2.10.2 tls crhold(cred); 805 1.2.10.2 tls bp->b_wcred = cred; 806 1.2.10.2 tls } else 807 1.2.10.2 tls bp->b_wcred = NOCRED; 808 1.2.10.2 tls bp->b_caller1 = (void *)t_uio; 809 1.2.10.2 tls bp->b_vp = vp; 810 1.2.10.2 tls error = ncl_asyncio(nmp, bp, NOCRED, td); 811 1.2.10.2 tls err_free: 812 1.2.10.2 tls if (error) { 813 1.2.10.2 tls free(t_iov->iov_base, M_NFSDIRECTIO); 814 1.2.10.2 tls free(t_iov, M_NFSDIRECTIO); 815 1.2.10.2 tls free(t_uio, M_NFSDIRECTIO); 816 1.2.10.2 tls bp->b_vp = NULL; 817 1.2.10.2 tls relpbuf(bp, &ncl_pbuf_freecnt); 818 1.2.10.2 tls if (error == EINTR) 819 1.2.10.2 tls return (error); 820 1.2.10.2 tls goto do_sync; 821 1.2.10.2 tls } 822 1.2.10.2 tls uiop->uio_offset += size; 823 1.2.10.2 tls uiop->uio_resid -= size; 824 1.2.10.2 tls if (uiop->uio_iov->iov_len <= size) { 825 1.2.10.2 tls uiop->uio_iovcnt--; 826 1.2.10.2 tls uiop->uio_iov++; 827 1.2.10.2 tls } else { 828 1.2.10.2 tls uiop->uio_iov->iov_base = 829 1.2.10.2 tls (char *)uiop->uio_iov->iov_base + size; 830 1.2.10.2 tls uiop->uio_iov->iov_len -= size; 831 1.2.10.2 tls } 832 1.2.10.2 tls } 833 1.2.10.2 tls } 834 1.2.10.2 tls return (0); 835 1.2.10.2 tls } 836 1.2.10.2 tls 837 1.2.10.2 tls /* 838 1.2.10.2 tls * Vnode op for write using bio 839 1.2.10.2 tls */ 840 1.2.10.2 tls int 841 1.2.10.2 tls ncl_write(struct vop_write_args *ap) 842 1.2.10.2 tls { 843 1.2.10.2 tls int biosize; 844 1.2.10.2 tls struct uio *uio = ap->a_uio; 845 1.2.10.2 tls struct thread *td = uio->uio_td; 846 1.2.10.2 tls struct vnode *vp = ap->a_vp; 847 1.2.10.2 tls struct nfsnode *np = VTONFS(vp); 848 1.2.10.2 tls struct ucred *cred = ap->a_cred; 849 1.2.10.2 tls int ioflag = ap->a_ioflag; 850 1.2.10.2 tls struct buf *bp; 851 1.2.10.2 tls struct vattr vattr; 852 1.2.10.2 tls struct nfsmount *nmp = VFSTONFS(vp->v_mount); 853 1.2.10.2 tls daddr_t lbn; 854 1.2.10.3 jdolecek int bcount, noncontig_write, obcount; 855 1.2.10.3 jdolecek int bp_cached, n, on, error = 0, error1, wouldcommit; 856 1.2.10.2 tls size_t orig_resid, local_resid; 857 1.2.10.2 tls off_t orig_size, tmp_off; 858 1.2.10.2 tls 859 1.2.10.2 tls KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode")); 860 1.2.10.2 tls KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, 861 1.2.10.2 tls ("ncl_write proc")); 862 1.2.10.2 tls if (vp->v_type != VREG) 863 1.2.10.2 tls return (EIO); 864 1.2.10.2 tls mtx_lock(&np->n_mtx); 865 1.2.10.2 tls if (np->n_flag & NWRITEERR) { 866 1.2.10.2 tls np->n_flag &= ~NWRITEERR; 867 1.2.10.2 tls mtx_unlock(&np->n_mtx); 868 1.2.10.2 tls return (np->n_error); 869 1.2.10.2 tls } else 870 1.2.10.2 tls mtx_unlock(&np->n_mtx); 871 1.2.10.2 tls mtx_lock(&nmp->nm_mtx); 872 1.2.10.2 tls if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 873 1.2.10.2 tls (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 874 1.2.10.2 tls mtx_unlock(&nmp->nm_mtx); 875 1.2.10.2 tls (void)ncl_fsinfo(nmp, vp, cred, td); 876 1.2.10.2 tls mtx_lock(&nmp->nm_mtx); 877 1.2.10.2 tls } 878 1.2.10.2 tls if (nmp->nm_wsize == 0) 879 1.2.10.2 tls (void) newnfs_iosize(nmp); 880 1.2.10.2 tls mtx_unlock(&nmp->nm_mtx); 881 1.2.10.2 tls 882 1.2.10.2 tls /* 883 1.2.10.2 tls * Synchronously flush pending buffers if we are in synchronous 884 1.2.10.2 tls * mode or if we are appending. 885 1.2.10.2 tls */ 886 1.2.10.2 tls if (ioflag & (IO_APPEND | IO_SYNC)) { 887 1.2.10.2 tls mtx_lock(&np->n_mtx); 888 1.2.10.2 tls if (np->n_flag & NMODIFIED) { 889 1.2.10.2 tls mtx_unlock(&np->n_mtx); 890 1.2.10.2 tls #ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 891 1.2.10.2 tls /* 892 1.2.10.2 tls * Require non-blocking, synchronous writes to 893 1.2.10.2 tls * dirty files to inform the program it needs 894 1.2.10.2 tls * to fsync(2) explicitly. 895 1.2.10.2 tls */ 896 1.2.10.2 tls if (ioflag & IO_NDELAY) 897 1.2.10.2 tls return (EAGAIN); 898 1.2.10.2 tls #endif 899 1.2.10.2 tls np->n_attrstamp = 0; 900 1.2.10.2 tls KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 901 1.2.10.2 tls error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 902 1.2.10.2 tls if (error) 903 1.2.10.2 tls return (error); 904 1.2.10.2 tls } else 905 1.2.10.2 tls mtx_unlock(&np->n_mtx); 906 1.2.10.2 tls } 907 1.2.10.2 tls 908 1.2.10.2 tls orig_resid = uio->uio_resid; 909 1.2.10.2 tls mtx_lock(&np->n_mtx); 910 1.2.10.2 tls orig_size = np->n_size; 911 1.2.10.2 tls mtx_unlock(&np->n_mtx); 912 1.2.10.2 tls 913 1.2.10.2 tls /* 914 1.2.10.2 tls * If IO_APPEND then load uio_offset. We restart here if we cannot 915 1.2.10.2 tls * get the append lock. 916 1.2.10.2 tls */ 917 1.2.10.2 tls if (ioflag & IO_APPEND) { 918 1.2.10.2 tls np->n_attrstamp = 0; 919 1.2.10.2 tls KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 920 1.2.10.2 tls error = VOP_GETATTR(vp, &vattr, cred); 921 1.2.10.2 tls if (error) 922 1.2.10.2 tls return (error); 923 1.2.10.2 tls mtx_lock(&np->n_mtx); 924 1.2.10.2 tls uio->uio_offset = np->n_size; 925 1.2.10.2 tls mtx_unlock(&np->n_mtx); 926 1.2.10.2 tls } 927 1.2.10.2 tls 928 1.2.10.2 tls if (uio->uio_offset < 0) 929 1.2.10.2 tls return (EINVAL); 930 1.2.10.2 tls tmp_off = uio->uio_offset + uio->uio_resid; 931 1.2.10.2 tls if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset) 932 1.2.10.2 tls return (EFBIG); 933 1.2.10.2 tls if (uio->uio_resid == 0) 934 1.2.10.2 tls return (0); 935 1.2.10.2 tls 936 1.2.10.2 tls if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 937 1.2.10.2 tls return nfs_directio_write(vp, uio, cred, ioflag); 938 1.2.10.2 tls 939 1.2.10.2 tls /* 940 1.2.10.2 tls * Maybe this should be above the vnode op call, but so long as 941 1.2.10.2 tls * file servers have no limits, i don't think it matters 942 1.2.10.2 tls */ 943 1.2.10.2 tls if (vn_rlimit_fsize(vp, uio, td)) 944 1.2.10.2 tls return (EFBIG); 945 1.2.10.2 tls 946 1.2.10.2 tls biosize = vp->v_bufobj.bo_bsize; 947 1.2.10.2 tls /* 948 1.2.10.2 tls * Find all of this file's B_NEEDCOMMIT buffers. If our writes 949 1.2.10.2 tls * would exceed the local maximum per-file write commit size when 950 1.2.10.2 tls * combined with those, we must decide whether to flush, 951 1.2.10.2 tls * go synchronous, or return error. We don't bother checking 952 1.2.10.2 tls * IO_UNIT -- we just make all writes atomic anyway, as there's 953 1.2.10.2 tls * no point optimizing for something that really won't ever happen. 954 1.2.10.2 tls */ 955 1.2.10.3 jdolecek wouldcommit = 0; 956 1.2.10.2 tls if (!(ioflag & IO_SYNC)) { 957 1.2.10.2 tls int nflag; 958 1.2.10.2 tls 959 1.2.10.2 tls mtx_lock(&np->n_mtx); 960 1.2.10.2 tls nflag = np->n_flag; 961 1.2.10.2 tls mtx_unlock(&np->n_mtx); 962 1.2.10.3 jdolecek if (nflag & NMODIFIED) { 963 1.2.10.2 tls BO_LOCK(&vp->v_bufobj); 964 1.2.10.2 tls if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 965 1.2.10.2 tls TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 966 1.2.10.2 tls b_bobufs) { 967 1.2.10.2 tls if (bp->b_flags & B_NEEDCOMMIT) 968 1.2.10.2 tls wouldcommit += bp->b_bcount; 969 1.2.10.2 tls } 970 1.2.10.2 tls } 971 1.2.10.2 tls BO_UNLOCK(&vp->v_bufobj); 972 1.2.10.2 tls } 973 1.2.10.2 tls } 974 1.2.10.2 tls 975 1.2.10.2 tls do { 976 1.2.10.3 jdolecek if (!(ioflag & IO_SYNC)) { 977 1.2.10.3 jdolecek wouldcommit += biosize; 978 1.2.10.3 jdolecek if (wouldcommit > nmp->nm_wcommitsize) { 979 1.2.10.3 jdolecek np->n_attrstamp = 0; 980 1.2.10.3 jdolecek KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 981 1.2.10.3 jdolecek error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 982 1.2.10.3 jdolecek if (error) 983 1.2.10.3 jdolecek return (error); 984 1.2.10.3 jdolecek wouldcommit = biosize; 985 1.2.10.3 jdolecek } 986 1.2.10.3 jdolecek } 987 1.2.10.3 jdolecek 988 1.2.10.3 jdolecek NFSINCRGLOBAL(nfsstatsv1.biocache_writes); 989 1.2.10.2 tls lbn = uio->uio_offset / biosize; 990 1.2.10.2 tls on = uio->uio_offset - (lbn * biosize); 991 1.2.10.2 tls n = MIN((unsigned)(biosize - on), uio->uio_resid); 992 1.2.10.2 tls again: 993 1.2.10.2 tls /* 994 1.2.10.2 tls * Handle direct append and file extension cases, calculate 995 1.2.10.2 tls * unaligned buffer size. 996 1.2.10.2 tls */ 997 1.2.10.2 tls mtx_lock(&np->n_mtx); 998 1.2.10.3 jdolecek if ((np->n_flag & NHASBEENLOCKED) == 0 && 999 1.2.10.3 jdolecek (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0) 1000 1.2.10.3 jdolecek noncontig_write = 1; 1001 1.2.10.3 jdolecek else 1002 1.2.10.3 jdolecek noncontig_write = 0; 1003 1.2.10.3 jdolecek if ((uio->uio_offset == np->n_size || 1004 1.2.10.3 jdolecek (noncontig_write != 0 && 1005 1.2.10.3 jdolecek lbn == (np->n_size / biosize) && 1006 1.2.10.3 jdolecek uio->uio_offset + n > np->n_size)) && n) { 1007 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1008 1.2.10.2 tls /* 1009 1.2.10.2 tls * Get the buffer (in its pre-append state to maintain 1010 1.2.10.2 tls * B_CACHE if it was previously set). Resize the 1011 1.2.10.2 tls * nfsnode after we have locked the buffer to prevent 1012 1.2.10.2 tls * readers from reading garbage. 1013 1.2.10.2 tls */ 1014 1.2.10.3 jdolecek obcount = np->n_size - (lbn * biosize); 1015 1.2.10.3 jdolecek bp = nfs_getcacheblk(vp, lbn, obcount, td); 1016 1.2.10.2 tls 1017 1.2.10.2 tls if (bp != NULL) { 1018 1.2.10.2 tls long save; 1019 1.2.10.2 tls 1020 1.2.10.2 tls mtx_lock(&np->n_mtx); 1021 1.2.10.2 tls np->n_size = uio->uio_offset + n; 1022 1.2.10.2 tls np->n_flag |= NMODIFIED; 1023 1.2.10.2 tls vnode_pager_setsize(vp, np->n_size); 1024 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1025 1.2.10.2 tls 1026 1.2.10.2 tls save = bp->b_flags & B_CACHE; 1027 1.2.10.3 jdolecek bcount = on + n; 1028 1.2.10.2 tls allocbuf(bp, bcount); 1029 1.2.10.2 tls bp->b_flags |= save; 1030 1.2.10.3 jdolecek if (noncontig_write != 0 && on > obcount) 1031 1.2.10.3 jdolecek vfs_bio_bzero_buf(bp, obcount, on - 1032 1.2.10.3 jdolecek obcount); 1033 1.2.10.2 tls } 1034 1.2.10.2 tls } else { 1035 1.2.10.2 tls /* 1036 1.2.10.2 tls * Obtain the locked cache block first, and then 1037 1.2.10.2 tls * adjust the file's size as appropriate. 1038 1.2.10.2 tls */ 1039 1.2.10.2 tls bcount = on + n; 1040 1.2.10.2 tls if ((off_t)lbn * biosize + bcount < np->n_size) { 1041 1.2.10.2 tls if ((off_t)(lbn + 1) * biosize < np->n_size) 1042 1.2.10.2 tls bcount = biosize; 1043 1.2.10.2 tls else 1044 1.2.10.2 tls bcount = np->n_size - (off_t)lbn * biosize; 1045 1.2.10.2 tls } 1046 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1047 1.2.10.2 tls bp = nfs_getcacheblk(vp, lbn, bcount, td); 1048 1.2.10.2 tls mtx_lock(&np->n_mtx); 1049 1.2.10.2 tls if (uio->uio_offset + n > np->n_size) { 1050 1.2.10.2 tls np->n_size = uio->uio_offset + n; 1051 1.2.10.2 tls np->n_flag |= NMODIFIED; 1052 1.2.10.2 tls vnode_pager_setsize(vp, np->n_size); 1053 1.2.10.2 tls } 1054 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1055 1.2.10.2 tls } 1056 1.2.10.2 tls 1057 1.2.10.2 tls if (!bp) { 1058 1.2.10.2 tls error = newnfs_sigintr(nmp, td); 1059 1.2.10.2 tls if (!error) 1060 1.2.10.2 tls error = EINTR; 1061 1.2.10.2 tls break; 1062 1.2.10.2 tls } 1063 1.2.10.2 tls 1064 1.2.10.2 tls /* 1065 1.2.10.2 tls * Issue a READ if B_CACHE is not set. In special-append 1066 1.2.10.2 tls * mode, B_CACHE is based on the buffer prior to the write 1067 1.2.10.2 tls * op and is typically set, avoiding the read. If a read 1068 1.2.10.2 tls * is required in special append mode, the server will 1069 1.2.10.2 tls * probably send us a short-read since we extended the file 1070 1.2.10.2 tls * on our end, resulting in b_resid == 0 and, thusly, 1071 1.2.10.2 tls * B_CACHE getting set. 1072 1.2.10.2 tls * 1073 1.2.10.2 tls * We can also avoid issuing the read if the write covers 1074 1.2.10.2 tls * the entire buffer. We have to make sure the buffer state 1075 1.2.10.2 tls * is reasonable in this case since we will not be initiating 1076 1.2.10.2 tls * I/O. See the comments in kern/vfs_bio.c's getblk() for 1077 1.2.10.2 tls * more information. 1078 1.2.10.2 tls * 1079 1.2.10.2 tls * B_CACHE may also be set due to the buffer being cached 1080 1.2.10.2 tls * normally. 1081 1.2.10.2 tls */ 1082 1.2.10.2 tls 1083 1.2.10.2 tls bp_cached = 1; 1084 1.2.10.2 tls if (on == 0 && n == bcount) { 1085 1.2.10.2 tls if ((bp->b_flags & B_CACHE) == 0) 1086 1.2.10.2 tls bp_cached = 0; 1087 1.2.10.2 tls bp->b_flags |= B_CACHE; 1088 1.2.10.2 tls bp->b_flags &= ~B_INVAL; 1089 1.2.10.2 tls bp->b_ioflags &= ~BIO_ERROR; 1090 1.2.10.2 tls } 1091 1.2.10.2 tls 1092 1.2.10.2 tls if ((bp->b_flags & B_CACHE) == 0) { 1093 1.2.10.2 tls bp->b_iocmd = BIO_READ; 1094 1.2.10.2 tls vfs_busy_pages(bp, 0); 1095 1.2.10.2 tls error = ncl_doio(vp, bp, cred, td, 0); 1096 1.2.10.2 tls if (error) { 1097 1.2.10.2 tls brelse(bp); 1098 1.2.10.2 tls break; 1099 1.2.10.2 tls } 1100 1.2.10.2 tls } 1101 1.2.10.2 tls if (bp->b_wcred == NOCRED) 1102 1.2.10.2 tls bp->b_wcred = crhold(cred); 1103 1.2.10.2 tls mtx_lock(&np->n_mtx); 1104 1.2.10.2 tls np->n_flag |= NMODIFIED; 1105 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1106 1.2.10.2 tls 1107 1.2.10.2 tls /* 1108 1.2.10.2 tls * If dirtyend exceeds file size, chop it down. This should 1109 1.2.10.2 tls * not normally occur but there is an append race where it 1110 1.2.10.2 tls * might occur XXX, so we log it. 1111 1.2.10.2 tls * 1112 1.2.10.2 tls * If the chopping creates a reverse-indexed or degenerate 1113 1.2.10.2 tls * situation with dirtyoff/end, we 0 both of them. 1114 1.2.10.2 tls */ 1115 1.2.10.2 tls 1116 1.2.10.2 tls if (bp->b_dirtyend > bcount) { 1117 1.2.10.3 jdolecek printf("NFS append race @%lx:%d\n", 1118 1.2.10.2 tls (long)bp->b_blkno * DEV_BSIZE, 1119 1.2.10.2 tls bp->b_dirtyend - bcount); 1120 1.2.10.2 tls bp->b_dirtyend = bcount; 1121 1.2.10.2 tls } 1122 1.2.10.2 tls 1123 1.2.10.2 tls if (bp->b_dirtyoff >= bp->b_dirtyend) 1124 1.2.10.2 tls bp->b_dirtyoff = bp->b_dirtyend = 0; 1125 1.2.10.2 tls 1126 1.2.10.2 tls /* 1127 1.2.10.2 tls * If the new write will leave a contiguous dirty 1128 1.2.10.2 tls * area, just update the b_dirtyoff and b_dirtyend, 1129 1.2.10.2 tls * otherwise force a write rpc of the old dirty area. 1130 1.2.10.2 tls * 1131 1.2.10.3 jdolecek * If there has been a file lock applied to this file 1132 1.2.10.3 jdolecek * or vfs.nfs.old_noncontig_writing is set, do the following: 1133 1.2.10.2 tls * While it is possible to merge discontiguous writes due to 1134 1.2.10.2 tls * our having a B_CACHE buffer ( and thus valid read data 1135 1.2.10.2 tls * for the hole), we don't because it could lead to 1136 1.2.10.2 tls * significant cache coherency problems with multiple clients, 1137 1.2.10.2 tls * especially if locking is implemented later on. 1138 1.2.10.2 tls * 1139 1.2.10.3 jdolecek * If vfs.nfs.old_noncontig_writing is not set and there has 1140 1.2.10.3 jdolecek * not been file locking done on this file: 1141 1.2.10.3 jdolecek * Relax coherency a bit for the sake of performance and 1142 1.2.10.3 jdolecek * expand the current dirty region to contain the new 1143 1.2.10.3 jdolecek * write even if it means we mark some non-dirty data as 1144 1.2.10.3 jdolecek * dirty. 1145 1.2.10.2 tls */ 1146 1.2.10.2 tls 1147 1.2.10.3 jdolecek if (noncontig_write == 0 && bp->b_dirtyend > 0 && 1148 1.2.10.2 tls (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1149 1.2.10.2 tls if (bwrite(bp) == EINTR) { 1150 1.2.10.2 tls error = EINTR; 1151 1.2.10.2 tls break; 1152 1.2.10.2 tls } 1153 1.2.10.2 tls goto again; 1154 1.2.10.2 tls } 1155 1.2.10.2 tls 1156 1.2.10.2 tls local_resid = uio->uio_resid; 1157 1.2.10.2 tls error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio); 1158 1.2.10.2 tls 1159 1.2.10.2 tls if (error != 0 && !bp_cached) { 1160 1.2.10.2 tls /* 1161 1.2.10.2 tls * This block has no other content than what 1162 1.2.10.2 tls * possibly was written by the faulty uiomove. 1163 1.2.10.2 tls * Release it, forgetting the data pages, to 1164 1.2.10.2 tls * prevent the leak of uninitialized data to 1165 1.2.10.2 tls * usermode. 1166 1.2.10.2 tls */ 1167 1.2.10.2 tls bp->b_ioflags |= BIO_ERROR; 1168 1.2.10.2 tls brelse(bp); 1169 1.2.10.2 tls uio->uio_offset -= local_resid - uio->uio_resid; 1170 1.2.10.2 tls uio->uio_resid = local_resid; 1171 1.2.10.2 tls break; 1172 1.2.10.2 tls } 1173 1.2.10.2 tls 1174 1.2.10.2 tls /* 1175 1.2.10.2 tls * Since this block is being modified, it must be written 1176 1.2.10.2 tls * again and not just committed. Since write clustering does 1177 1.2.10.2 tls * not work for the stage 1 data write, only the stage 2 1178 1.2.10.2 tls * commit rpc, we have to clear B_CLUSTEROK as well. 1179 1.2.10.2 tls */ 1180 1.2.10.2 tls bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1181 1.2.10.2 tls 1182 1.2.10.2 tls /* 1183 1.2.10.2 tls * Get the partial update on the progress made from 1184 1.2.10.3 jdolecek * uiomove, if an error occurred. 1185 1.2.10.2 tls */ 1186 1.2.10.2 tls if (error != 0) 1187 1.2.10.2 tls n = local_resid - uio->uio_resid; 1188 1.2.10.2 tls 1189 1.2.10.2 tls /* 1190 1.2.10.2 tls * Only update dirtyoff/dirtyend if not a degenerate 1191 1.2.10.2 tls * condition. 1192 1.2.10.2 tls */ 1193 1.2.10.2 tls if (n > 0) { 1194 1.2.10.2 tls if (bp->b_dirtyend > 0) { 1195 1.2.10.2 tls bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1196 1.2.10.2 tls bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1197 1.2.10.2 tls } else { 1198 1.2.10.2 tls bp->b_dirtyoff = on; 1199 1.2.10.2 tls bp->b_dirtyend = on + n; 1200 1.2.10.2 tls } 1201 1.2.10.2 tls vfs_bio_set_valid(bp, on, n); 1202 1.2.10.2 tls } 1203 1.2.10.2 tls 1204 1.2.10.2 tls /* 1205 1.2.10.2 tls * If IO_SYNC do bwrite(). 1206 1.2.10.2 tls * 1207 1.2.10.2 tls * IO_INVAL appears to be unused. The idea appears to be 1208 1.2.10.2 tls * to turn off caching in this case. Very odd. XXX 1209 1.2.10.2 tls */ 1210 1.2.10.2 tls if ((ioflag & IO_SYNC)) { 1211 1.2.10.2 tls if (ioflag & IO_INVAL) 1212 1.2.10.2 tls bp->b_flags |= B_NOCACHE; 1213 1.2.10.2 tls error1 = bwrite(bp); 1214 1.2.10.2 tls if (error1 != 0) { 1215 1.2.10.2 tls if (error == 0) 1216 1.2.10.2 tls error = error1; 1217 1.2.10.2 tls break; 1218 1.2.10.2 tls } 1219 1.2.10.2 tls } else if ((n + on) == biosize) { 1220 1.2.10.2 tls bp->b_flags |= B_ASYNC; 1221 1.2.10.2 tls (void) ncl_writebp(bp, 0, NULL); 1222 1.2.10.2 tls } else { 1223 1.2.10.2 tls bdwrite(bp); 1224 1.2.10.2 tls } 1225 1.2.10.2 tls 1226 1.2.10.2 tls if (error != 0) 1227 1.2.10.2 tls break; 1228 1.2.10.2 tls } while (uio->uio_resid > 0 && n > 0); 1229 1.2.10.2 tls 1230 1.2.10.2 tls if (error != 0) { 1231 1.2.10.2 tls if (ioflag & IO_UNIT) { 1232 1.2.10.2 tls VATTR_NULL(&vattr); 1233 1.2.10.2 tls vattr.va_size = orig_size; 1234 1.2.10.2 tls /* IO_SYNC is handled implicitely */ 1235 1.2.10.2 tls (void)VOP_SETATTR(vp, &vattr, cred); 1236 1.2.10.2 tls uio->uio_offset -= orig_resid - uio->uio_resid; 1237 1.2.10.2 tls uio->uio_resid = orig_resid; 1238 1.2.10.2 tls } 1239 1.2.10.2 tls } 1240 1.2.10.2 tls 1241 1.2.10.2 tls return (error); 1242 1.2.10.2 tls } 1243 1.2.10.2 tls 1244 1.2.10.2 tls /* 1245 1.2.10.2 tls * Get an nfs cache block. 1246 1.2.10.2 tls * 1247 1.2.10.2 tls * Allocate a new one if the block isn't currently in the cache 1248 1.2.10.2 tls * and return the block marked busy. If the calling process is 1249 1.2.10.2 tls * interrupted by a signal for an interruptible mount point, return 1250 1.2.10.2 tls * NULL. 1251 1.2.10.2 tls * 1252 1.2.10.2 tls * The caller must carefully deal with the possible B_INVAL state of 1253 1.2.10.2 tls * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it 1254 1.2.10.2 tls * indirectly), so synchronous reads can be issued without worrying about 1255 1.2.10.2 tls * the B_INVAL state. We have to be a little more careful when dealing 1256 1.2.10.2 tls * with writes (see comments in nfs_write()) when extending a file past 1257 1.2.10.2 tls * its EOF. 1258 1.2.10.2 tls */ 1259 1.2.10.2 tls static struct buf * 1260 1.2.10.2 tls nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1261 1.2.10.2 tls { 1262 1.2.10.2 tls struct buf *bp; 1263 1.2.10.2 tls struct mount *mp; 1264 1.2.10.2 tls struct nfsmount *nmp; 1265 1.2.10.2 tls 1266 1.2.10.2 tls mp = vp->v_mount; 1267 1.2.10.2 tls nmp = VFSTONFS(mp); 1268 1.2.10.2 tls 1269 1.2.10.2 tls if (nmp->nm_flag & NFSMNT_INT) { 1270 1.2.10.2 tls sigset_t oldset; 1271 1.2.10.2 tls 1272 1.2.10.2 tls newnfs_set_sigmask(td, &oldset); 1273 1.2.10.2 tls bp = getblk(vp, bn, size, PCATCH, 0, 0); 1274 1.2.10.2 tls newnfs_restore_sigmask(td, &oldset); 1275 1.2.10.2 tls while (bp == NULL) { 1276 1.2.10.2 tls if (newnfs_sigintr(nmp, td)) 1277 1.2.10.2 tls return (NULL); 1278 1.2.10.2 tls bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1279 1.2.10.2 tls } 1280 1.2.10.2 tls } else { 1281 1.2.10.2 tls bp = getblk(vp, bn, size, 0, 0, 0); 1282 1.2.10.2 tls } 1283 1.2.10.2 tls 1284 1.2.10.2 tls if (vp->v_type == VREG) 1285 1.2.10.2 tls bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE); 1286 1.2.10.2 tls return (bp); 1287 1.2.10.2 tls } 1288 1.2.10.2 tls 1289 1.2.10.2 tls /* 1290 1.2.10.2 tls * Flush and invalidate all dirty buffers. If another process is already 1291 1.2.10.2 tls * doing the flush, just wait for completion. 1292 1.2.10.2 tls */ 1293 1.2.10.2 tls int 1294 1.2.10.2 tls ncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1295 1.2.10.2 tls { 1296 1.2.10.2 tls struct nfsnode *np = VTONFS(vp); 1297 1.2.10.2 tls struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1298 1.2.10.2 tls int error = 0, slpflag, slptimeo; 1299 1.2.10.2 tls int old_lock = 0; 1300 1.2.10.2 tls 1301 1.2.10.2 tls ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf"); 1302 1.2.10.2 tls 1303 1.2.10.2 tls if ((nmp->nm_flag & NFSMNT_INT) == 0) 1304 1.2.10.2 tls intrflg = 0; 1305 1.2.10.2 tls if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)) 1306 1.2.10.2 tls intrflg = 1; 1307 1.2.10.2 tls if (intrflg) { 1308 1.2.10.2 tls slpflag = PCATCH; 1309 1.2.10.2 tls slptimeo = 2 * hz; 1310 1.2.10.2 tls } else { 1311 1.2.10.2 tls slpflag = 0; 1312 1.2.10.2 tls slptimeo = 0; 1313 1.2.10.2 tls } 1314 1.2.10.2 tls 1315 1.2.10.2 tls old_lock = ncl_upgrade_vnlock(vp); 1316 1.2.10.2 tls if (vp->v_iflag & VI_DOOMED) { 1317 1.2.10.2 tls /* 1318 1.2.10.2 tls * Since vgonel() uses the generic vinvalbuf() to flush 1319 1.2.10.2 tls * dirty buffers and it does not call this function, it 1320 1.2.10.2 tls * is safe to just return OK when VI_DOOMED is set. 1321 1.2.10.2 tls */ 1322 1.2.10.2 tls ncl_downgrade_vnlock(vp, old_lock); 1323 1.2.10.2 tls return (0); 1324 1.2.10.2 tls } 1325 1.2.10.2 tls 1326 1.2.10.2 tls /* 1327 1.2.10.2 tls * Now, flush as required. 1328 1.2.10.2 tls */ 1329 1.2.10.2 tls if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1330 1.2.10.2 tls VM_OBJECT_WLOCK(vp->v_bufobj.bo_object); 1331 1.2.10.2 tls vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1332 1.2.10.2 tls VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object); 1333 1.2.10.2 tls /* 1334 1.2.10.2 tls * If the page clean was interrupted, fail the invalidation. 1335 1.2.10.2 tls * Not doing so, we run the risk of losing dirty pages in the 1336 1.2.10.2 tls * vinvalbuf() call below. 1337 1.2.10.2 tls */ 1338 1.2.10.2 tls if (intrflg && (error = newnfs_sigintr(nmp, td))) 1339 1.2.10.2 tls goto out; 1340 1.2.10.2 tls } 1341 1.2.10.2 tls 1342 1.2.10.2 tls error = vinvalbuf(vp, flags, slpflag, 0); 1343 1.2.10.2 tls while (error) { 1344 1.2.10.2 tls if (intrflg && (error = newnfs_sigintr(nmp, td))) 1345 1.2.10.2 tls goto out; 1346 1.2.10.2 tls error = vinvalbuf(vp, flags, 0, slptimeo); 1347 1.2.10.2 tls } 1348 1.2.10.2 tls if (NFSHASPNFS(nmp)) { 1349 1.2.10.2 tls nfscl_layoutcommit(vp, td); 1350 1.2.10.2 tls /* 1351 1.2.10.2 tls * Invalidate the attribute cache, since writes to a DS 1352 1.2.10.2 tls * won't update the size attribute. 1353 1.2.10.2 tls */ 1354 1.2.10.2 tls mtx_lock(&np->n_mtx); 1355 1.2.10.2 tls np->n_attrstamp = 0; 1356 1.2.10.2 tls } else 1357 1.2.10.2 tls mtx_lock(&np->n_mtx); 1358 1.2.10.2 tls if (np->n_directio_asyncwr == 0) 1359 1.2.10.2 tls np->n_flag &= ~NMODIFIED; 1360 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1361 1.2.10.2 tls out: 1362 1.2.10.2 tls ncl_downgrade_vnlock(vp, old_lock); 1363 1.2.10.2 tls return error; 1364 1.2.10.2 tls } 1365 1.2.10.2 tls 1366 1.2.10.2 tls /* 1367 1.2.10.2 tls * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1368 1.2.10.2 tls * This is mainly to avoid queueing async I/O requests when the nfsiods 1369 1.2.10.2 tls * are all hung on a dead server. 1370 1.2.10.2 tls * 1371 1.2.10.2 tls * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1372 1.2.10.2 tls * is eventually dequeued by the async daemon, ncl_doio() *will*. 1373 1.2.10.2 tls */ 1374 1.2.10.2 tls int 1375 1.2.10.2 tls ncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1376 1.2.10.2 tls { 1377 1.2.10.2 tls int iod; 1378 1.2.10.2 tls int gotiod; 1379 1.2.10.2 tls int slpflag = 0; 1380 1.2.10.2 tls int slptimeo = 0; 1381 1.2.10.2 tls int error, error2; 1382 1.2.10.2 tls 1383 1.2.10.2 tls /* 1384 1.2.10.2 tls * Commits are usually short and sweet so lets save some cpu and 1385 1.2.10.2 tls * leave the async daemons for more important rpc's (such as reads 1386 1.2.10.2 tls * and writes). 1387 1.2.10.2 tls * 1388 1.2.10.2 tls * Readdirplus RPCs do vget()s to acquire the vnodes for entries 1389 1.2.10.2 tls * in the directory in order to update attributes. This can deadlock 1390 1.2.10.2 tls * with another thread that is waiting for async I/O to be done by 1391 1.2.10.2 tls * an nfsiod thread while holding a lock on one of these vnodes. 1392 1.2.10.2 tls * To avoid this deadlock, don't allow the async nfsiod threads to 1393 1.2.10.2 tls * perform Readdirplus RPCs. 1394 1.2.10.2 tls */ 1395 1.2.10.2 tls mtx_lock(&ncl_iod_mutex); 1396 1.2.10.2 tls if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1397 1.2.10.2 tls (nmp->nm_bufqiods > ncl_numasync / 2)) || 1398 1.2.10.2 tls (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) { 1399 1.2.10.2 tls mtx_unlock(&ncl_iod_mutex); 1400 1.2.10.2 tls return(EIO); 1401 1.2.10.2 tls } 1402 1.2.10.2 tls again: 1403 1.2.10.2 tls if (nmp->nm_flag & NFSMNT_INT) 1404 1.2.10.2 tls slpflag = PCATCH; 1405 1.2.10.2 tls gotiod = FALSE; 1406 1.2.10.2 tls 1407 1.2.10.2 tls /* 1408 1.2.10.2 tls * Find a free iod to process this request. 1409 1.2.10.2 tls */ 1410 1.2.10.2 tls for (iod = 0; iod < ncl_numasync; iod++) 1411 1.2.10.2 tls if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) { 1412 1.2.10.2 tls gotiod = TRUE; 1413 1.2.10.2 tls break; 1414 1.2.10.2 tls } 1415 1.2.10.2 tls 1416 1.2.10.2 tls /* 1417 1.2.10.2 tls * Try to create one if none are free. 1418 1.2.10.2 tls */ 1419 1.2.10.2 tls if (!gotiod) 1420 1.2.10.2 tls ncl_nfsiodnew(); 1421 1.2.10.2 tls else { 1422 1.2.10.2 tls /* 1423 1.2.10.2 tls * Found one, so wake it up and tell it which 1424 1.2.10.2 tls * mount to process. 1425 1.2.10.2 tls */ 1426 1.2.10.2 tls NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n", 1427 1.2.10.2 tls iod, nmp)); 1428 1.2.10.2 tls ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1429 1.2.10.2 tls ncl_iodmount[iod] = nmp; 1430 1.2.10.2 tls nmp->nm_bufqiods++; 1431 1.2.10.2 tls wakeup(&ncl_iodwant[iod]); 1432 1.2.10.2 tls } 1433 1.2.10.2 tls 1434 1.2.10.2 tls /* 1435 1.2.10.2 tls * If none are free, we may already have an iod working on this mount 1436 1.2.10.2 tls * point. If so, it will process our request. 1437 1.2.10.2 tls */ 1438 1.2.10.2 tls if (!gotiod) { 1439 1.2.10.2 tls if (nmp->nm_bufqiods > 0) { 1440 1.2.10.2 tls NFS_DPF(ASYNCIO, 1441 1.2.10.2 tls ("ncl_asyncio: %d iods are already processing mount %p\n", 1442 1.2.10.2 tls nmp->nm_bufqiods, nmp)); 1443 1.2.10.2 tls gotiod = TRUE; 1444 1.2.10.2 tls } 1445 1.2.10.2 tls } 1446 1.2.10.2 tls 1447 1.2.10.2 tls /* 1448 1.2.10.2 tls * If we have an iod which can process the request, then queue 1449 1.2.10.2 tls * the buffer. 1450 1.2.10.2 tls */ 1451 1.2.10.2 tls if (gotiod) { 1452 1.2.10.2 tls /* 1453 1.2.10.2 tls * Ensure that the queue never grows too large. We still want 1454 1.2.10.2 tls * to asynchronize so we block rather than return EIO. 1455 1.2.10.2 tls */ 1456 1.2.10.2 tls while (nmp->nm_bufqlen >= 2*ncl_numasync) { 1457 1.2.10.2 tls NFS_DPF(ASYNCIO, 1458 1.2.10.2 tls ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp)); 1459 1.2.10.2 tls nmp->nm_bufqwant = TRUE; 1460 1.2.10.2 tls error = newnfs_msleep(td, &nmp->nm_bufq, 1461 1.2.10.2 tls &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio", 1462 1.2.10.2 tls slptimeo); 1463 1.2.10.2 tls if (error) { 1464 1.2.10.2 tls error2 = newnfs_sigintr(nmp, td); 1465 1.2.10.2 tls if (error2) { 1466 1.2.10.2 tls mtx_unlock(&ncl_iod_mutex); 1467 1.2.10.2 tls return (error2); 1468 1.2.10.2 tls } 1469 1.2.10.2 tls if (slpflag == PCATCH) { 1470 1.2.10.2 tls slpflag = 0; 1471 1.2.10.2 tls slptimeo = 2 * hz; 1472 1.2.10.2 tls } 1473 1.2.10.2 tls } 1474 1.2.10.2 tls /* 1475 1.2.10.2 tls * We might have lost our iod while sleeping, 1476 1.2.10.3 jdolecek * so check and loop if necessary. 1477 1.2.10.2 tls */ 1478 1.2.10.2 tls goto again; 1479 1.2.10.2 tls } 1480 1.2.10.2 tls 1481 1.2.10.2 tls /* We might have lost our nfsiod */ 1482 1.2.10.2 tls if (nmp->nm_bufqiods == 0) { 1483 1.2.10.2 tls NFS_DPF(ASYNCIO, 1484 1.2.10.2 tls ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1485 1.2.10.2 tls goto again; 1486 1.2.10.2 tls } 1487 1.2.10.2 tls 1488 1.2.10.2 tls if (bp->b_iocmd == BIO_READ) { 1489 1.2.10.2 tls if (bp->b_rcred == NOCRED && cred != NOCRED) 1490 1.2.10.2 tls bp->b_rcred = crhold(cred); 1491 1.2.10.2 tls } else { 1492 1.2.10.2 tls if (bp->b_wcred == NOCRED && cred != NOCRED) 1493 1.2.10.2 tls bp->b_wcred = crhold(cred); 1494 1.2.10.2 tls } 1495 1.2.10.2 tls 1496 1.2.10.2 tls if (bp->b_flags & B_REMFREE) 1497 1.2.10.2 tls bremfreef(bp); 1498 1.2.10.2 tls BUF_KERNPROC(bp); 1499 1.2.10.2 tls TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1500 1.2.10.2 tls nmp->nm_bufqlen++; 1501 1.2.10.2 tls if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1502 1.2.10.2 tls mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1503 1.2.10.2 tls VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1504 1.2.10.2 tls VTONFS(bp->b_vp)->n_directio_asyncwr++; 1505 1.2.10.2 tls mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1506 1.2.10.2 tls } 1507 1.2.10.2 tls mtx_unlock(&ncl_iod_mutex); 1508 1.2.10.2 tls return (0); 1509 1.2.10.2 tls } 1510 1.2.10.2 tls 1511 1.2.10.2 tls mtx_unlock(&ncl_iod_mutex); 1512 1.2.10.2 tls 1513 1.2.10.2 tls /* 1514 1.2.10.2 tls * All the iods are busy on other mounts, so return EIO to 1515 1.2.10.2 tls * force the caller to process the i/o synchronously. 1516 1.2.10.2 tls */ 1517 1.2.10.2 tls NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n")); 1518 1.2.10.2 tls return (EIO); 1519 1.2.10.2 tls } 1520 1.2.10.2 tls 1521 1.2.10.2 tls void 1522 1.2.10.2 tls ncl_doio_directwrite(struct buf *bp) 1523 1.2.10.2 tls { 1524 1.2.10.2 tls int iomode, must_commit; 1525 1.2.10.2 tls struct uio *uiop = (struct uio *)bp->b_caller1; 1526 1.2.10.2 tls char *iov_base = uiop->uio_iov->iov_base; 1527 1.2.10.2 tls 1528 1.2.10.2 tls iomode = NFSWRITE_FILESYNC; 1529 1.2.10.2 tls uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1530 1.2.10.2 tls ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0); 1531 1.2.10.2 tls KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write")); 1532 1.2.10.2 tls free(iov_base, M_NFSDIRECTIO); 1533 1.2.10.2 tls free(uiop->uio_iov, M_NFSDIRECTIO); 1534 1.2.10.2 tls free(uiop, M_NFSDIRECTIO); 1535 1.2.10.2 tls if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1536 1.2.10.2 tls struct nfsnode *np = VTONFS(bp->b_vp); 1537 1.2.10.2 tls mtx_lock(&np->n_mtx); 1538 1.2.10.2 tls if (NFSHASPNFS(VFSTONFS(vnode_mount(bp->b_vp)))) { 1539 1.2.10.2 tls /* 1540 1.2.10.2 tls * Invalidate the attribute cache, since writes to a DS 1541 1.2.10.2 tls * won't update the size attribute. 1542 1.2.10.2 tls */ 1543 1.2.10.2 tls np->n_attrstamp = 0; 1544 1.2.10.2 tls } 1545 1.2.10.2 tls np->n_directio_asyncwr--; 1546 1.2.10.2 tls if (np->n_directio_asyncwr == 0) { 1547 1.2.10.2 tls np->n_flag &= ~NMODIFIED; 1548 1.2.10.2 tls if ((np->n_flag & NFSYNCWAIT)) { 1549 1.2.10.2 tls np->n_flag &= ~NFSYNCWAIT; 1550 1.2.10.2 tls wakeup((caddr_t)&np->n_directio_asyncwr); 1551 1.2.10.2 tls } 1552 1.2.10.2 tls } 1553 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1554 1.2.10.2 tls } 1555 1.2.10.2 tls bp->b_vp = NULL; 1556 1.2.10.2 tls relpbuf(bp, &ncl_pbuf_freecnt); 1557 1.2.10.2 tls } 1558 1.2.10.2 tls 1559 1.2.10.2 tls /* 1560 1.2.10.2 tls * Do an I/O operation to/from a cache block. This may be called 1561 1.2.10.2 tls * synchronously or from an nfsiod. 1562 1.2.10.2 tls */ 1563 1.2.10.2 tls int 1564 1.2.10.2 tls ncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td, 1565 1.2.10.2 tls int called_from_strategy) 1566 1.2.10.2 tls { 1567 1.2.10.2 tls struct uio *uiop; 1568 1.2.10.2 tls struct nfsnode *np; 1569 1.2.10.2 tls struct nfsmount *nmp; 1570 1.2.10.2 tls int error = 0, iomode, must_commit = 0; 1571 1.2.10.2 tls struct uio uio; 1572 1.2.10.2 tls struct iovec io; 1573 1.2.10.2 tls struct proc *p = td ? td->td_proc : NULL; 1574 1.2.10.2 tls uint8_t iocmd; 1575 1.2.10.2 tls 1576 1.2.10.2 tls np = VTONFS(vp); 1577 1.2.10.2 tls nmp = VFSTONFS(vp->v_mount); 1578 1.2.10.2 tls uiop = &uio; 1579 1.2.10.2 tls uiop->uio_iov = &io; 1580 1.2.10.2 tls uiop->uio_iovcnt = 1; 1581 1.2.10.2 tls uiop->uio_segflg = UIO_SYSSPACE; 1582 1.2.10.2 tls uiop->uio_td = td; 1583 1.2.10.2 tls 1584 1.2.10.2 tls /* 1585 1.2.10.2 tls * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1586 1.2.10.2 tls * do this here so we do not have to do it in all the code that 1587 1.2.10.2 tls * calls us. 1588 1.2.10.2 tls */ 1589 1.2.10.2 tls bp->b_flags &= ~B_INVAL; 1590 1.2.10.2 tls bp->b_ioflags &= ~BIO_ERROR; 1591 1.2.10.2 tls 1592 1.2.10.2 tls KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp)); 1593 1.2.10.2 tls iocmd = bp->b_iocmd; 1594 1.2.10.2 tls if (iocmd == BIO_READ) { 1595 1.2.10.2 tls io.iov_len = uiop->uio_resid = bp->b_bcount; 1596 1.2.10.2 tls io.iov_base = bp->b_data; 1597 1.2.10.2 tls uiop->uio_rw = UIO_READ; 1598 1.2.10.2 tls 1599 1.2.10.2 tls switch (vp->v_type) { 1600 1.2.10.2 tls case VREG: 1601 1.2.10.2 tls uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1602 1.2.10.3 jdolecek NFSINCRGLOBAL(nfsstatsv1.read_bios); 1603 1.2.10.2 tls error = ncl_readrpc(vp, uiop, cr); 1604 1.2.10.2 tls 1605 1.2.10.2 tls if (!error) { 1606 1.2.10.2 tls if (uiop->uio_resid) { 1607 1.2.10.2 tls /* 1608 1.2.10.2 tls * If we had a short read with no error, we must have 1609 1.2.10.2 tls * hit a file hole. We should zero-fill the remainder. 1610 1.2.10.2 tls * This can also occur if the server hits the file EOF. 1611 1.2.10.2 tls * 1612 1.2.10.2 tls * Holes used to be able to occur due to pending 1613 1.2.10.2 tls * writes, but that is not possible any longer. 1614 1.2.10.2 tls */ 1615 1.2.10.2 tls int nread = bp->b_bcount - uiop->uio_resid; 1616 1.2.10.2 tls ssize_t left = uiop->uio_resid; 1617 1.2.10.2 tls 1618 1.2.10.2 tls if (left > 0) 1619 1.2.10.2 tls bzero((char *)bp->b_data + nread, left); 1620 1.2.10.2 tls uiop->uio_resid = 0; 1621 1.2.10.2 tls } 1622 1.2.10.2 tls } 1623 1.2.10.2 tls /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */ 1624 1.2.10.2 tls if (p && (vp->v_vflag & VV_TEXT)) { 1625 1.2.10.2 tls mtx_lock(&np->n_mtx); 1626 1.2.10.2 tls if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) { 1627 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1628 1.2.10.2 tls PROC_LOCK(p); 1629 1.2.10.2 tls killproc(p, "text file modification"); 1630 1.2.10.2 tls PROC_UNLOCK(p); 1631 1.2.10.2 tls } else 1632 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1633 1.2.10.2 tls } 1634 1.2.10.2 tls break; 1635 1.2.10.2 tls case VLNK: 1636 1.2.10.2 tls uiop->uio_offset = (off_t)0; 1637 1.2.10.3 jdolecek NFSINCRGLOBAL(nfsstatsv1.readlink_bios); 1638 1.2.10.2 tls error = ncl_readlinkrpc(vp, uiop, cr); 1639 1.2.10.2 tls break; 1640 1.2.10.2 tls case VDIR: 1641 1.2.10.3 jdolecek NFSINCRGLOBAL(nfsstatsv1.readdir_bios); 1642 1.2.10.2 tls uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1643 1.2.10.2 tls if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1644 1.2.10.2 tls error = ncl_readdirplusrpc(vp, uiop, cr, td); 1645 1.2.10.2 tls if (error == NFSERR_NOTSUPP) 1646 1.2.10.2 tls nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1647 1.2.10.2 tls } 1648 1.2.10.2 tls if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1649 1.2.10.2 tls error = ncl_readdirrpc(vp, uiop, cr, td); 1650 1.2.10.2 tls /* 1651 1.2.10.2 tls * end-of-directory sets B_INVAL but does not generate an 1652 1.2.10.2 tls * error. 1653 1.2.10.2 tls */ 1654 1.2.10.2 tls if (error == 0 && uiop->uio_resid == bp->b_bcount) 1655 1.2.10.2 tls bp->b_flags |= B_INVAL; 1656 1.2.10.2 tls break; 1657 1.2.10.2 tls default: 1658 1.2.10.3 jdolecek printf("ncl_doio: type %x unexpected\n", vp->v_type); 1659 1.2.10.2 tls break; 1660 1.2.10.3 jdolecek } 1661 1.2.10.2 tls if (error) { 1662 1.2.10.2 tls bp->b_ioflags |= BIO_ERROR; 1663 1.2.10.2 tls bp->b_error = error; 1664 1.2.10.2 tls } 1665 1.2.10.2 tls } else { 1666 1.2.10.2 tls /* 1667 1.2.10.2 tls * If we only need to commit, try to commit 1668 1.2.10.2 tls */ 1669 1.2.10.2 tls if (bp->b_flags & B_NEEDCOMMIT) { 1670 1.2.10.2 tls int retv; 1671 1.2.10.2 tls off_t off; 1672 1.2.10.2 tls 1673 1.2.10.2 tls off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1674 1.2.10.2 tls retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1675 1.2.10.2 tls bp->b_wcred, td); 1676 1.2.10.2 tls if (retv == 0) { 1677 1.2.10.2 tls bp->b_dirtyoff = bp->b_dirtyend = 0; 1678 1.2.10.2 tls bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1679 1.2.10.2 tls bp->b_resid = 0; 1680 1.2.10.2 tls bufdone(bp); 1681 1.2.10.2 tls return (0); 1682 1.2.10.2 tls } 1683 1.2.10.2 tls if (retv == NFSERR_STALEWRITEVERF) { 1684 1.2.10.2 tls ncl_clearcommit(vp->v_mount); 1685 1.2.10.2 tls } 1686 1.2.10.2 tls } 1687 1.2.10.2 tls 1688 1.2.10.2 tls /* 1689 1.2.10.2 tls * Setup for actual write 1690 1.2.10.2 tls */ 1691 1.2.10.2 tls mtx_lock(&np->n_mtx); 1692 1.2.10.2 tls if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1693 1.2.10.2 tls bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1694 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1695 1.2.10.2 tls 1696 1.2.10.2 tls if (bp->b_dirtyend > bp->b_dirtyoff) { 1697 1.2.10.2 tls io.iov_len = uiop->uio_resid = bp->b_dirtyend 1698 1.2.10.2 tls - bp->b_dirtyoff; 1699 1.2.10.2 tls uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1700 1.2.10.2 tls + bp->b_dirtyoff; 1701 1.2.10.2 tls io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1702 1.2.10.2 tls uiop->uio_rw = UIO_WRITE; 1703 1.2.10.3 jdolecek NFSINCRGLOBAL(nfsstatsv1.write_bios); 1704 1.2.10.2 tls 1705 1.2.10.2 tls if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1706 1.2.10.2 tls iomode = NFSWRITE_UNSTABLE; 1707 1.2.10.2 tls else 1708 1.2.10.2 tls iomode = NFSWRITE_FILESYNC; 1709 1.2.10.2 tls 1710 1.2.10.2 tls error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit, 1711 1.2.10.2 tls called_from_strategy); 1712 1.2.10.2 tls 1713 1.2.10.2 tls /* 1714 1.2.10.2 tls * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1715 1.2.10.2 tls * to cluster the buffers needing commit. This will allow 1716 1.2.10.2 tls * the system to submit a single commit rpc for the whole 1717 1.2.10.2 tls * cluster. We can do this even if the buffer is not 100% 1718 1.2.10.2 tls * dirty (relative to the NFS blocksize), so we optimize the 1719 1.2.10.2 tls * append-to-file-case. 1720 1.2.10.2 tls * 1721 1.2.10.2 tls * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1722 1.2.10.2 tls * cleared because write clustering only works for commit 1723 1.2.10.2 tls * rpc's, not for the data portion of the write). 1724 1.2.10.2 tls */ 1725 1.2.10.2 tls 1726 1.2.10.2 tls if (!error && iomode == NFSWRITE_UNSTABLE) { 1727 1.2.10.2 tls bp->b_flags |= B_NEEDCOMMIT; 1728 1.2.10.2 tls if (bp->b_dirtyoff == 0 1729 1.2.10.2 tls && bp->b_dirtyend == bp->b_bcount) 1730 1.2.10.2 tls bp->b_flags |= B_CLUSTEROK; 1731 1.2.10.2 tls } else { 1732 1.2.10.2 tls bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1733 1.2.10.2 tls } 1734 1.2.10.2 tls 1735 1.2.10.2 tls /* 1736 1.2.10.2 tls * For an interrupted write, the buffer is still valid 1737 1.2.10.2 tls * and the write hasn't been pushed to the server yet, 1738 1.2.10.2 tls * so we can't set BIO_ERROR and report the interruption 1739 1.2.10.2 tls * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1740 1.2.10.2 tls * is not relevant, so the rpc attempt is essentially 1741 1.2.10.2 tls * a noop. For the case of a V3 write rpc not being 1742 1.2.10.2 tls * committed to stable storage, the block is still 1743 1.2.10.2 tls * dirty and requires either a commit rpc or another 1744 1.2.10.2 tls * write rpc with iomode == NFSV3WRITE_FILESYNC before 1745 1.2.10.2 tls * the block is reused. This is indicated by setting 1746 1.2.10.2 tls * the B_DELWRI and B_NEEDCOMMIT flags. 1747 1.2.10.2 tls * 1748 1.2.10.2 tls * EIO is returned by ncl_writerpc() to indicate a recoverable 1749 1.2.10.2 tls * write error and is handled as above, except that 1750 1.2.10.2 tls * B_EINTR isn't set. One cause of this is a stale stateid 1751 1.2.10.2 tls * error for the RPC that indicates recovery is required, 1752 1.2.10.2 tls * when called with called_from_strategy != 0. 1753 1.2.10.2 tls * 1754 1.2.10.2 tls * If the buffer is marked B_PAGING, it does not reside on 1755 1.2.10.2 tls * the vp's paging queues so we cannot call bdirty(). The 1756 1.2.10.2 tls * bp in this case is not an NFS cache block so we should 1757 1.2.10.2 tls * be safe. XXX 1758 1.2.10.2 tls * 1759 1.2.10.2 tls * The logic below breaks up errors into recoverable and 1760 1.2.10.2 tls * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1761 1.2.10.2 tls * and keep the buffer around for potential write retries. 1762 1.2.10.2 tls * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1763 1.2.10.2 tls * and save the error in the nfsnode. This is less than ideal 1764 1.2.10.2 tls * but necessary. Keeping such buffers around could potentially 1765 1.2.10.2 tls * cause buffer exhaustion eventually (they can never be written 1766 1.2.10.2 tls * out, so will get constantly be re-dirtied). It also causes 1767 1.2.10.2 tls * all sorts of vfs panics. For non-recoverable write errors, 1768 1.2.10.2 tls * also invalidate the attrcache, so we'll be forced to go over 1769 1.2.10.2 tls * the wire for this object, returning an error to user on next 1770 1.2.10.2 tls * call (most of the time). 1771 1.2.10.2 tls */ 1772 1.2.10.2 tls if (error == EINTR || error == EIO || error == ETIMEDOUT 1773 1.2.10.2 tls || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1774 1.2.10.2 tls int s; 1775 1.2.10.2 tls 1776 1.2.10.2 tls s = splbio(); 1777 1.2.10.2 tls bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1778 1.2.10.2 tls if ((bp->b_flags & B_PAGING) == 0) { 1779 1.2.10.2 tls bdirty(bp); 1780 1.2.10.2 tls bp->b_flags &= ~B_DONE; 1781 1.2.10.2 tls } 1782 1.2.10.2 tls if ((error == EINTR || error == ETIMEDOUT) && 1783 1.2.10.2 tls (bp->b_flags & B_ASYNC) == 0) 1784 1.2.10.2 tls bp->b_flags |= B_EINTR; 1785 1.2.10.2 tls splx(s); 1786 1.2.10.2 tls } else { 1787 1.2.10.2 tls if (error) { 1788 1.2.10.2 tls bp->b_ioflags |= BIO_ERROR; 1789 1.2.10.2 tls bp->b_flags |= B_INVAL; 1790 1.2.10.2 tls bp->b_error = np->n_error = error; 1791 1.2.10.2 tls mtx_lock(&np->n_mtx); 1792 1.2.10.2 tls np->n_flag |= NWRITEERR; 1793 1.2.10.2 tls np->n_attrstamp = 0; 1794 1.2.10.2 tls KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1795 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1796 1.2.10.2 tls } 1797 1.2.10.2 tls bp->b_dirtyoff = bp->b_dirtyend = 0; 1798 1.2.10.2 tls } 1799 1.2.10.2 tls } else { 1800 1.2.10.2 tls bp->b_resid = 0; 1801 1.2.10.2 tls bufdone(bp); 1802 1.2.10.2 tls return (0); 1803 1.2.10.2 tls } 1804 1.2.10.2 tls } 1805 1.2.10.2 tls bp->b_resid = uiop->uio_resid; 1806 1.2.10.2 tls if (must_commit) 1807 1.2.10.2 tls ncl_clearcommit(vp->v_mount); 1808 1.2.10.2 tls bufdone(bp); 1809 1.2.10.2 tls return (error); 1810 1.2.10.2 tls } 1811 1.2.10.2 tls 1812 1.2.10.2 tls /* 1813 1.2.10.2 tls * Used to aid in handling ftruncate() operations on the NFS client side. 1814 1.2.10.2 tls * Truncation creates a number of special problems for NFS. We have to 1815 1.2.10.2 tls * throw away VM pages and buffer cache buffers that are beyond EOF, and 1816 1.2.10.2 tls * we have to properly handle VM pages or (potentially dirty) buffers 1817 1.2.10.2 tls * that straddle the truncation point. 1818 1.2.10.2 tls */ 1819 1.2.10.2 tls 1820 1.2.10.2 tls int 1821 1.2.10.2 tls ncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1822 1.2.10.2 tls { 1823 1.2.10.2 tls struct nfsnode *np = VTONFS(vp); 1824 1.2.10.2 tls u_quad_t tsize; 1825 1.2.10.2 tls int biosize = vp->v_bufobj.bo_bsize; 1826 1.2.10.2 tls int error = 0; 1827 1.2.10.2 tls 1828 1.2.10.2 tls mtx_lock(&np->n_mtx); 1829 1.2.10.2 tls tsize = np->n_size; 1830 1.2.10.2 tls np->n_size = nsize; 1831 1.2.10.2 tls mtx_unlock(&np->n_mtx); 1832 1.2.10.2 tls 1833 1.2.10.2 tls if (nsize < tsize) { 1834 1.2.10.2 tls struct buf *bp; 1835 1.2.10.2 tls daddr_t lbn; 1836 1.2.10.2 tls int bufsize; 1837 1.2.10.2 tls 1838 1.2.10.2 tls /* 1839 1.2.10.2 tls * vtruncbuf() doesn't get the buffer overlapping the 1840 1.2.10.2 tls * truncation point. We may have a B_DELWRI and/or B_CACHE 1841 1.2.10.2 tls * buffer that now needs to be truncated. 1842 1.2.10.2 tls */ 1843 1.2.10.2 tls error = vtruncbuf(vp, cred, nsize, biosize); 1844 1.2.10.2 tls lbn = nsize / biosize; 1845 1.2.10.2 tls bufsize = nsize - (lbn * biosize); 1846 1.2.10.2 tls bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1847 1.2.10.2 tls if (!bp) 1848 1.2.10.2 tls return EINTR; 1849 1.2.10.2 tls if (bp->b_dirtyoff > bp->b_bcount) 1850 1.2.10.2 tls bp->b_dirtyoff = bp->b_bcount; 1851 1.2.10.2 tls if (bp->b_dirtyend > bp->b_bcount) 1852 1.2.10.2 tls bp->b_dirtyend = bp->b_bcount; 1853 1.2.10.2 tls bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1854 1.2.10.2 tls brelse(bp); 1855 1.2.10.2 tls } else { 1856 1.2.10.2 tls vnode_pager_setsize(vp, nsize); 1857 1.2.10.2 tls } 1858 1.2.10.2 tls return(error); 1859 1.2.10.2 tls } 1860 1.2.10.2 tls 1861
Indexes created Fri Oct 03 02:09:57 GMT 2025