nfs_clbio.c revision 194425
1191783Srmacklem/*- 2191783Srmacklem * Copyright (c) 1989, 1993 3191783Srmacklem * The Regents of the University of California. All rights reserved. 4191783Srmacklem * 5191783Srmacklem * This code is derived from software contributed to Berkeley by 6191783Srmacklem * Rick Macklem at The University of Guelph. 7191783Srmacklem * 8191783Srmacklem * Redistribution and use in source and binary forms, with or without 9191783Srmacklem * modification, are permitted provided that the following conditions 10191783Srmacklem * are met: 11191783Srmacklem * 1. Redistributions of source code must retain the above copyright 12191783Srmacklem * notice, this list of conditions and the following disclaimer. 13191783Srmacklem * 2. Redistributions in binary form must reproduce the above copyright 14191783Srmacklem * notice, this list of conditions and the following disclaimer in the 15191783Srmacklem * documentation and/or other materials provided with the distribution. 16191783Srmacklem * 4. Neither the name of the University nor the names of its contributors 17191783Srmacklem * may be used to endorse or promote products derived from this software 18191783Srmacklem * without specific prior written permission. 19191783Srmacklem * 20191783Srmacklem * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21191783Srmacklem * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22191783Srmacklem * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23191783Srmacklem * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24191783Srmacklem * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25191783Srmacklem * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26191783Srmacklem * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27191783Srmacklem * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28191783Srmacklem * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29191783Srmacklem * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30191783Srmacklem * SUCH DAMAGE. 31191783Srmacklem * 32191783Srmacklem * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 33191783Srmacklem */ 34191783Srmacklem 35191783Srmacklem#include <sys/cdefs.h> 36191783Srmacklem__FBSDID("$FreeBSD: head/sys/fs/nfsclient/nfs_clbio.c 194425 2009-06-18 05:56:24Z alc $"); 37191783Srmacklem 38191783Srmacklem#include <sys/param.h> 39191783Srmacklem#include <sys/systm.h> 40191783Srmacklem#include <sys/bio.h> 41191783Srmacklem#include <sys/buf.h> 42191783Srmacklem#include <sys/kernel.h> 43191783Srmacklem#include <sys/mount.h> 44191783Srmacklem#include <sys/proc.h> 45191783Srmacklem#include <sys/resourcevar.h> 46191783Srmacklem#include <sys/signalvar.h> 47191783Srmacklem#include <sys/vmmeter.h> 48191783Srmacklem#include <sys/vnode.h> 49191783Srmacklem 50191783Srmacklem#include <vm/vm.h> 51191783Srmacklem#include <vm/vm_extern.h> 52191783Srmacklem#include <vm/vm_page.h> 53191783Srmacklem#include <vm/vm_object.h> 54191783Srmacklem#include <vm/vm_pager.h> 55191783Srmacklem#include <vm/vnode_pager.h> 56191783Srmacklem 57191783Srmacklem#include <fs/nfs/nfsport.h> 58191783Srmacklem#include <fs/nfsclient/nfsmount.h> 59191783Srmacklem#include <fs/nfsclient/nfs.h> 60191783Srmacklem#include <fs/nfsclient/nfsnode.h> 61191783Srmacklem 62191783Srmacklemextern int newnfs_directio_allow_mmap; 63191783Srmacklemextern struct nfsstats newnfsstats; 64191783Srmacklemextern struct mtx ncl_iod_mutex; 65191783Srmacklemextern int ncl_numasync; 66191783Srmacklemextern struct proc *ncl_iodwant[NFS_MAXRAHEAD]; 67191783Srmacklemextern struct nfsmount *ncl_iodmount[NFS_MAXRAHEAD]; 68191783Srmacklemextern int newnfs_directio_enable; 69191783Srmacklem 70191783Srmacklemint ncl_pbuf_freecnt = -1; /* start out unlimited */ 71191783Srmacklem 72191783Srmacklemstatic struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 73191783Srmacklem struct thread *td); 74191783Srmacklemstatic int nfs_directio_write(struct vnode *vp, struct uio *uiop, 75191783Srmacklem struct ucred *cred, int ioflag); 76191783Srmacklem 77191783Srmacklem/* 78191783Srmacklem * Any signal that can interrupt an NFS operation in an intr mount 79191783Srmacklem * should be added to this set. SIGSTOP and SIGKILL cannot be masked. 80191783Srmacklem */ 81191783Srmacklemstatic int nfs_sig_set[] = { 82191783Srmacklem SIGINT, 83191783Srmacklem SIGTERM, 84191783Srmacklem SIGHUP, 85191783Srmacklem SIGKILL, 86191783Srmacklem SIGSTOP, 87191783Srmacklem SIGQUIT 88191783Srmacklem}; 89191783Srmacklem 90191783Srmacklem#ifdef notnow 91191783Srmacklem/* 92191783Srmacklem * Check to see if one of the signals in our subset is pending on 93191783Srmacklem * the process (in an intr mount). 94191783Srmacklem */ 95191783Srmacklemint 96191783Srmacklemncl_sig_pending(sigset_t set) 97191783Srmacklem{ 98191783Srmacklem int i; 99191783Srmacklem 100191783Srmacklem for (i = 0 ; i < sizeof(nfs_sig_set)/sizeof(int) ; i++) 101191783Srmacklem if (SIGISMEMBER(set, nfs_sig_set[i])) 102191783Srmacklem return (1); 103191783Srmacklem return (0); 104191783Srmacklem} 105191783Srmacklem#endif 106191783Srmacklem 107191783Srmacklem/* 108191783Srmacklem * The set/restore sigmask functions are used to (temporarily) overwrite 109191783Srmacklem * the process p_sigmask during an RPC call (for example). These are also 110191783Srmacklem * used in other places in the NFS client that might tsleep(). 111191783Srmacklem */ 112191783Srmacklemstatic void 113191783Srmacklemncl_set_sigmask(struct thread *td, sigset_t *oldset) 114191783Srmacklem{ 115191783Srmacklem sigset_t newset; 116191783Srmacklem int i; 117191783Srmacklem struct proc *p; 118191783Srmacklem 119191783Srmacklem SIGFILLSET(newset); 120191783Srmacklem if (td == NULL) 121191783Srmacklem td = curthread; /* XXX */ 122191783Srmacklem p = td->td_proc; 123191783Srmacklem /* Remove the NFS set of signals from newset */ 124191783Srmacklem PROC_LOCK(p); 125191783Srmacklem mtx_lock(&p->p_sigacts->ps_mtx); 126191783Srmacklem for (i = 0 ; i < sizeof(nfs_sig_set)/sizeof(int) ; i++) { 127191783Srmacklem /* 128191783Srmacklem * But make sure we leave the ones already masked 129191783Srmacklem * by the process, ie. remove the signal from the 130191783Srmacklem * temporary signalmask only if it wasn't already 131191783Srmacklem * in p_sigmask. 132191783Srmacklem */ 133191783Srmacklem if (!SIGISMEMBER(td->td_sigmask, nfs_sig_set[i]) && 134191783Srmacklem !SIGISMEMBER(p->p_sigacts->ps_sigignore, nfs_sig_set[i])) 135191783Srmacklem SIGDELSET(newset, nfs_sig_set[i]); 136191783Srmacklem } 137191783Srmacklem mtx_unlock(&p->p_sigacts->ps_mtx); 138191783Srmacklem PROC_UNLOCK(p); 139191783Srmacklem kern_sigprocmask(td, SIG_SETMASK, &newset, oldset, 0); 140191783Srmacklem} 141191783Srmacklem 142191783Srmacklemstatic void 143191783Srmacklemncl_restore_sigmask(struct thread *td, sigset_t *set) 144191783Srmacklem{ 145191783Srmacklem if (td == NULL) 146191783Srmacklem td = curthread; /* XXX */ 147191783Srmacklem kern_sigprocmask(td, SIG_SETMASK, set, NULL, 0); 148191783Srmacklem} 149191783Srmacklem 150191783Srmacklem/* 151191783Srmacklem * NFS wrapper to msleep(), that shoves a new p_sigmask and restores the 152191783Srmacklem * old one after msleep() returns. 153191783Srmacklem */ 154191783Srmacklemint 155191783Srmacklemncl_msleep(struct thread *td, void *ident, struct mtx *mtx, int priority, char *wmesg, int timo) 156191783Srmacklem{ 157191783Srmacklem sigset_t oldset; 158191783Srmacklem int error; 159191783Srmacklem struct proc *p; 160191783Srmacklem 161191783Srmacklem if ((priority & PCATCH) == 0) 162191783Srmacklem return msleep(ident, mtx, priority, wmesg, timo); 163191783Srmacklem if (td == NULL) 164191783Srmacklem td = curthread; /* XXX */ 165191783Srmacklem ncl_set_sigmask(td, &oldset); 166191783Srmacklem error = msleep(ident, mtx, priority, wmesg, timo); 167191783Srmacklem ncl_restore_sigmask(td, &oldset); 168191783Srmacklem p = td->td_proc; 169191783Srmacklem return (error); 170191783Srmacklem} 171191783Srmacklem 172191783Srmacklem/* 173191783Srmacklem * Vnode op for VM getpages. 174191783Srmacklem */ 175191783Srmacklemint 176191783Srmacklemncl_getpages(struct vop_getpages_args *ap) 177191783Srmacklem{ 178191783Srmacklem int i, error, nextoff, size, toff, count, npages; 179191783Srmacklem struct uio uio; 180191783Srmacklem struct iovec iov; 181191783Srmacklem vm_offset_t kva; 182191783Srmacklem struct buf *bp; 183191783Srmacklem struct vnode *vp; 184191783Srmacklem struct thread *td; 185191783Srmacklem struct ucred *cred; 186191783Srmacklem struct nfsmount *nmp; 187191783Srmacklem vm_object_t object; 188191783Srmacklem vm_page_t *pages; 189191783Srmacklem struct nfsnode *np; 190191783Srmacklem 191191783Srmacklem vp = ap->a_vp; 192191783Srmacklem np = VTONFS(vp); 193191783Srmacklem td = curthread; /* XXX */ 194191783Srmacklem cred = curthread->td_ucred; /* XXX */ 195191783Srmacklem nmp = VFSTONFS(vp->v_mount); 196191783Srmacklem pages = ap->a_m; 197191783Srmacklem count = ap->a_count; 198191783Srmacklem 199191783Srmacklem if ((object = vp->v_object) == NULL) { 200191783Srmacklem ncl_printf("nfs_getpages: called with non-merged cache vnode??\n"); 201194425Salc return (VM_PAGER_ERROR); 202191783Srmacklem } 203191783Srmacklem 204191783Srmacklem if (newnfs_directio_enable && !newnfs_directio_allow_mmap) { 205191783Srmacklem mtx_lock(&np->n_mtx); 206191783Srmacklem if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 207191783Srmacklem mtx_unlock(&np->n_mtx); 208191783Srmacklem ncl_printf("nfs_getpages: called on non-cacheable vnode??\n"); 209194425Salc return (VM_PAGER_ERROR); 210191783Srmacklem } else 211191783Srmacklem mtx_unlock(&np->n_mtx); 212191783Srmacklem } 213191783Srmacklem 214191783Srmacklem mtx_lock(&nmp->nm_mtx); 215191783Srmacklem if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 216191783Srmacklem (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 217191783Srmacklem mtx_unlock(&nmp->nm_mtx); 218191783Srmacklem /* We'll never get here for v4, because we always have fsinfo */ 219191783Srmacklem (void)ncl_fsinfo(nmp, vp, cred, td); 220191783Srmacklem } else 221191783Srmacklem mtx_unlock(&nmp->nm_mtx); 222191783Srmacklem 223191783Srmacklem npages = btoc(count); 224191783Srmacklem 225191783Srmacklem /* 226191783Srmacklem * If the requested page is partially valid, just return it and 227191783Srmacklem * allow the pager to zero-out the blanks. Partially valid pages 228191783Srmacklem * can only occur at the file EOF. 229191783Srmacklem */ 230194425Salc VM_OBJECT_LOCK(object); 231194425Salc if (pages[ap->a_reqpage]->valid != 0) { 232194425Salc vm_page_lock_queues(); 233194425Salc for (i = 0; i < npages; ++i) { 234194425Salc if (i != ap->a_reqpage) 235194425Salc vm_page_free(pages[i]); 236191783Srmacklem } 237194425Salc vm_page_unlock_queues(); 238191783Srmacklem VM_OBJECT_UNLOCK(object); 239194425Salc return (0); 240191783Srmacklem } 241194425Salc VM_OBJECT_UNLOCK(object); 242191783Srmacklem 243191783Srmacklem /* 244191783Srmacklem * We use only the kva address for the buffer, but this is extremely 245191783Srmacklem * convienient and fast. 246191783Srmacklem */ 247191783Srmacklem bp = getpbuf(&ncl_pbuf_freecnt); 248191783Srmacklem 249191783Srmacklem kva = (vm_offset_t) bp->b_data; 250191783Srmacklem pmap_qenter(kva, pages, npages); 251191783Srmacklem PCPU_INC(cnt.v_vnodein); 252191783Srmacklem PCPU_ADD(cnt.v_vnodepgsin, npages); 253191783Srmacklem 254191783Srmacklem iov.iov_base = (caddr_t) kva; 255191783Srmacklem iov.iov_len = count; 256191783Srmacklem uio.uio_iov = &iov; 257191783Srmacklem uio.uio_iovcnt = 1; 258191783Srmacklem uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 259191783Srmacklem uio.uio_resid = count; 260191783Srmacklem uio.uio_segflg = UIO_SYSSPACE; 261191783Srmacklem uio.uio_rw = UIO_READ; 262191783Srmacklem uio.uio_td = td; 263191783Srmacklem 264191783Srmacklem error = ncl_readrpc(vp, &uio, cred); 265191783Srmacklem pmap_qremove(kva, npages); 266191783Srmacklem 267191783Srmacklem relpbuf(bp, &ncl_pbuf_freecnt); 268191783Srmacklem 269191783Srmacklem if (error && (uio.uio_resid == count)) { 270191783Srmacklem ncl_printf("nfs_getpages: error %d\n", error); 271191783Srmacklem VM_OBJECT_LOCK(object); 272191783Srmacklem vm_page_lock_queues(); 273191783Srmacklem for (i = 0; i < npages; ++i) { 274191783Srmacklem if (i != ap->a_reqpage) 275191783Srmacklem vm_page_free(pages[i]); 276191783Srmacklem } 277191783Srmacklem vm_page_unlock_queues(); 278191783Srmacklem VM_OBJECT_UNLOCK(object); 279194425Salc return (VM_PAGER_ERROR); 280191783Srmacklem } 281191783Srmacklem 282191783Srmacklem /* 283191783Srmacklem * Calculate the number of bytes read and validate only that number 284191783Srmacklem * of bytes. Note that due to pending writes, size may be 0. This 285191783Srmacklem * does not mean that the remaining data is invalid! 286191783Srmacklem */ 287191783Srmacklem 288191783Srmacklem size = count - uio.uio_resid; 289191783Srmacklem VM_OBJECT_LOCK(object); 290191783Srmacklem vm_page_lock_queues(); 291191783Srmacklem for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 292191783Srmacklem vm_page_t m; 293191783Srmacklem nextoff = toff + PAGE_SIZE; 294191783Srmacklem m = pages[i]; 295191783Srmacklem 296191783Srmacklem if (nextoff <= size) { 297191783Srmacklem /* 298191783Srmacklem * Read operation filled an entire page 299191783Srmacklem */ 300191783Srmacklem m->valid = VM_PAGE_BITS_ALL; 301192065Srmacklem KASSERT(m->dirty == 0, 302192065Srmacklem ("nfs_getpages: page %p is dirty", m)); 303191783Srmacklem } else if (size > toff) { 304191783Srmacklem /* 305191783Srmacklem * Read operation filled a partial page. 306191783Srmacklem */ 307191783Srmacklem m->valid = 0; 308192231Srmacklem vm_page_set_valid(m, 0, size - toff); 309192986Salc KASSERT(m->dirty == 0, 310192231Srmacklem ("nfs_getpages: page %p is dirty", m)); 311191783Srmacklem } else { 312191783Srmacklem /* 313191783Srmacklem * Read operation was short. If no error occured 314191783Srmacklem * we may have hit a zero-fill section. We simply 315191783Srmacklem * leave valid set to 0. 316191783Srmacklem */ 317191783Srmacklem ; 318191783Srmacklem } 319191783Srmacklem if (i != ap->a_reqpage) { 320191783Srmacklem /* 321191783Srmacklem * Whether or not to leave the page activated is up in 322191783Srmacklem * the air, but we should put the page on a page queue 323191783Srmacklem * somewhere (it already is in the object). Result: 324191783Srmacklem * It appears that emperical results show that 325191783Srmacklem * deactivating pages is best. 326191783Srmacklem */ 327191783Srmacklem 328191783Srmacklem /* 329191783Srmacklem * Just in case someone was asking for this page we 330191783Srmacklem * now tell them that it is ok to use. 331191783Srmacklem */ 332191783Srmacklem if (!error) { 333191783Srmacklem if (m->oflags & VPO_WANTED) 334191783Srmacklem vm_page_activate(m); 335191783Srmacklem else 336191783Srmacklem vm_page_deactivate(m); 337191783Srmacklem vm_page_wakeup(m); 338191783Srmacklem } else { 339191783Srmacklem vm_page_free(m); 340191783Srmacklem } 341191783Srmacklem } 342191783Srmacklem } 343191783Srmacklem vm_page_unlock_queues(); 344191783Srmacklem VM_OBJECT_UNLOCK(object); 345194425Salc return (0); 346191783Srmacklem} 347191783Srmacklem 348191783Srmacklem/* 349191783Srmacklem * Vnode op for VM putpages. 350191783Srmacklem */ 351191783Srmacklemint 352191783Srmacklemncl_putpages(struct vop_putpages_args *ap) 353191783Srmacklem{ 354191783Srmacklem struct uio uio; 355191783Srmacklem struct iovec iov; 356191783Srmacklem vm_offset_t kva; 357191783Srmacklem struct buf *bp; 358191783Srmacklem int iomode, must_commit, i, error, npages, count; 359191783Srmacklem off_t offset; 360191783Srmacklem int *rtvals; 361191783Srmacklem struct vnode *vp; 362191783Srmacklem struct thread *td; 363191783Srmacklem struct ucred *cred; 364191783Srmacklem struct nfsmount *nmp; 365191783Srmacklem struct nfsnode *np; 366191783Srmacklem vm_page_t *pages; 367191783Srmacklem 368191783Srmacklem vp = ap->a_vp; 369191783Srmacklem np = VTONFS(vp); 370191783Srmacklem td = curthread; /* XXX */ 371191783Srmacklem cred = curthread->td_ucred; /* XXX */ 372191783Srmacklem nmp = VFSTONFS(vp->v_mount); 373191783Srmacklem pages = ap->a_m; 374191783Srmacklem count = ap->a_count; 375191783Srmacklem rtvals = ap->a_rtvals; 376191783Srmacklem npages = btoc(count); 377191783Srmacklem offset = IDX_TO_OFF(pages[0]->pindex); 378191783Srmacklem 379191783Srmacklem mtx_lock(&nmp->nm_mtx); 380191783Srmacklem if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 381191783Srmacklem (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 382191783Srmacklem mtx_unlock(&nmp->nm_mtx); 383191783Srmacklem (void)ncl_fsinfo(nmp, vp, cred, td); 384191783Srmacklem } else 385191783Srmacklem mtx_unlock(&nmp->nm_mtx); 386191783Srmacklem 387191783Srmacklem mtx_lock(&np->n_mtx); 388191783Srmacklem if (newnfs_directio_enable && !newnfs_directio_allow_mmap && 389191783Srmacklem (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 390191783Srmacklem mtx_unlock(&np->n_mtx); 391191783Srmacklem ncl_printf("ncl_putpages: called on noncache-able vnode??\n"); 392191783Srmacklem mtx_lock(&np->n_mtx); 393191783Srmacklem } 394191783Srmacklem 395191783Srmacklem for (i = 0; i < npages; i++) 396191783Srmacklem rtvals[i] = VM_PAGER_AGAIN; 397191783Srmacklem 398191783Srmacklem /* 399191783Srmacklem * When putting pages, do not extend file past EOF. 400191783Srmacklem */ 401191783Srmacklem if (offset + count > np->n_size) { 402191783Srmacklem count = np->n_size - offset; 403191783Srmacklem if (count < 0) 404191783Srmacklem count = 0; 405191783Srmacklem } 406191783Srmacklem mtx_unlock(&np->n_mtx); 407191783Srmacklem 408191783Srmacklem /* 409191783Srmacklem * We use only the kva address for the buffer, but this is extremely 410191783Srmacklem * convienient and fast. 411191783Srmacklem */ 412191783Srmacklem bp = getpbuf(&ncl_pbuf_freecnt); 413191783Srmacklem 414191783Srmacklem kva = (vm_offset_t) bp->b_data; 415191783Srmacklem pmap_qenter(kva, pages, npages); 416191783Srmacklem PCPU_INC(cnt.v_vnodeout); 417191783Srmacklem PCPU_ADD(cnt.v_vnodepgsout, count); 418191783Srmacklem 419191783Srmacklem iov.iov_base = (caddr_t) kva; 420191783Srmacklem iov.iov_len = count; 421191783Srmacklem uio.uio_iov = &iov; 422191783Srmacklem uio.uio_iovcnt = 1; 423191783Srmacklem uio.uio_offset = offset; 424191783Srmacklem uio.uio_resid = count; 425191783Srmacklem uio.uio_segflg = UIO_SYSSPACE; 426191783Srmacklem uio.uio_rw = UIO_WRITE; 427191783Srmacklem uio.uio_td = td; 428191783Srmacklem 429191783Srmacklem if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 430191783Srmacklem iomode = NFSWRITE_UNSTABLE; 431191783Srmacklem else 432191783Srmacklem iomode = NFSWRITE_FILESYNC; 433191783Srmacklem 434191783Srmacklem error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit); 435191783Srmacklem 436191783Srmacklem pmap_qremove(kva, npages); 437191783Srmacklem relpbuf(bp, &ncl_pbuf_freecnt); 438191783Srmacklem 439191783Srmacklem if (!error) { 440191783Srmacklem int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 441191783Srmacklem for (i = 0; i < nwritten; i++) { 442191783Srmacklem rtvals[i] = VM_PAGER_OK; 443191783Srmacklem vm_page_undirty(pages[i]); 444191783Srmacklem } 445191783Srmacklem if (must_commit) { 446191783Srmacklem ncl_clearcommit(vp->v_mount); 447191783Srmacklem } 448191783Srmacklem } 449191783Srmacklem return rtvals[0]; 450191783Srmacklem} 451191783Srmacklem 452191783Srmacklem/* 453191783Srmacklem * For nfs, cache consistency can only be maintained approximately. 454191783Srmacklem * Although RFC1094 does not specify the criteria, the following is 455191783Srmacklem * believed to be compatible with the reference port. 456191783Srmacklem * For nfs: 457191783Srmacklem * If the file's modify time on the server has changed since the 458191783Srmacklem * last read rpc or you have written to the file, 459191783Srmacklem * you may have lost data cache consistency with the 460191783Srmacklem * server, so flush all of the file's data out of the cache. 461191783Srmacklem * Then force a getattr rpc to ensure that you have up to date 462191783Srmacklem * attributes. 463191783Srmacklem * NB: This implies that cache data can be read when up to 464191783Srmacklem * NFS_ATTRTIMEO seconds out of date. If you find that you need current 465191783Srmacklem * attributes this could be forced by setting n_attrstamp to 0 before 466191783Srmacklem * the VOP_GETATTR() call. 467191783Srmacklem */ 468191783Srmacklemstatic inline int 469191783Srmacklemnfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 470191783Srmacklem{ 471191783Srmacklem int error = 0; 472191783Srmacklem struct vattr vattr; 473191783Srmacklem struct nfsnode *np = VTONFS(vp); 474191783Srmacklem int old_lock; 475191783Srmacklem 476191783Srmacklem /* 477191783Srmacklem * Grab the exclusive lock before checking whether the cache is 478191783Srmacklem * consistent. 479191783Srmacklem * XXX - We can make this cheaper later (by acquiring cheaper locks). 480191783Srmacklem * But for now, this suffices. 481191783Srmacklem */ 482191783Srmacklem old_lock = ncl_upgrade_vnlock(vp); 483193955Srmacklem if (vp->v_iflag & VI_DOOMED) { 484193955Srmacklem ncl_downgrade_vnlock(vp, old_lock); 485193955Srmacklem return (EBADF); 486193955Srmacklem } 487193955Srmacklem 488191783Srmacklem mtx_lock(&np->n_mtx); 489191783Srmacklem if (np->n_flag & NMODIFIED) { 490191783Srmacklem mtx_unlock(&np->n_mtx); 491191783Srmacklem if (vp->v_type != VREG) { 492191783Srmacklem if (vp->v_type != VDIR) 493191783Srmacklem panic("nfs: bioread, not dir"); 494191783Srmacklem ncl_invaldir(vp); 495191783Srmacklem error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 496191783Srmacklem if (error) 497191783Srmacklem goto out; 498191783Srmacklem } 499191783Srmacklem np->n_attrstamp = 0; 500191783Srmacklem error = VOP_GETATTR(vp, &vattr, cred); 501191783Srmacklem if (error) 502191783Srmacklem goto out; 503191783Srmacklem mtx_lock(&np->n_mtx); 504191783Srmacklem np->n_mtime = vattr.va_mtime; 505191783Srmacklem mtx_unlock(&np->n_mtx); 506191783Srmacklem } else { 507191783Srmacklem mtx_unlock(&np->n_mtx); 508191783Srmacklem error = VOP_GETATTR(vp, &vattr, cred); 509191783Srmacklem if (error) 510191783Srmacklem return (error); 511191783Srmacklem mtx_lock(&np->n_mtx); 512191783Srmacklem if ((np->n_flag & NSIZECHANGED) 513191783Srmacklem || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 514191783Srmacklem mtx_unlock(&np->n_mtx); 515191783Srmacklem if (vp->v_type == VDIR) 516191783Srmacklem ncl_invaldir(vp); 517191783Srmacklem error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 518191783Srmacklem if (error) 519191783Srmacklem goto out; 520191783Srmacklem mtx_lock(&np->n_mtx); 521191783Srmacklem np->n_mtime = vattr.va_mtime; 522191783Srmacklem np->n_flag &= ~NSIZECHANGED; 523191783Srmacklem } 524191783Srmacklem mtx_unlock(&np->n_mtx); 525191783Srmacklem } 526191783Srmacklemout: 527191783Srmacklem ncl_downgrade_vnlock(vp, old_lock); 528191783Srmacklem return error; 529191783Srmacklem} 530191783Srmacklem 531191783Srmacklem/* 532191783Srmacklem * Vnode op for read using bio 533191783Srmacklem */ 534191783Srmacklemint 535191783Srmacklemncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 536191783Srmacklem{ 537191783Srmacklem struct nfsnode *np = VTONFS(vp); 538191783Srmacklem int biosize, i; 539191783Srmacklem struct buf *bp, *rabp; 540191783Srmacklem struct thread *td; 541191783Srmacklem struct nfsmount *nmp = VFSTONFS(vp->v_mount); 542191783Srmacklem daddr_t lbn, rabn; 543191783Srmacklem int bcount; 544191783Srmacklem int seqcount; 545191783Srmacklem int nra, error = 0, n = 0, on = 0; 546191783Srmacklem 547191783Srmacklem#ifdef DIAGNOSTIC 548191783Srmacklem if (uio->uio_rw != UIO_READ) 549191783Srmacklem panic("ncl_read mode"); 550191783Srmacklem#endif 551191783Srmacklem if (uio->uio_resid == 0) 552191783Srmacklem return (0); 553191783Srmacklem if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 554191783Srmacklem return (EINVAL); 555191783Srmacklem td = uio->uio_td; 556191783Srmacklem 557191783Srmacklem mtx_lock(&nmp->nm_mtx); 558191783Srmacklem if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 559191783Srmacklem (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 560191783Srmacklem mtx_unlock(&nmp->nm_mtx); 561191783Srmacklem (void)ncl_fsinfo(nmp, vp, cred, td); 562191783Srmacklem mtx_lock(&nmp->nm_mtx); 563191783Srmacklem } 564191783Srmacklem if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0) 565191783Srmacklem (void) newnfs_iosize(nmp); 566191783Srmacklem mtx_unlock(&nmp->nm_mtx); 567191783Srmacklem 568191783Srmacklem if (vp->v_type != VDIR && 569191783Srmacklem (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 570191783Srmacklem return (EFBIG); 571191783Srmacklem 572191783Srmacklem if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 573191783Srmacklem /* No caching/ no readaheads. Just read data into the user buffer */ 574191783Srmacklem return ncl_readrpc(vp, uio, cred); 575191783Srmacklem 576191783Srmacklem biosize = vp->v_mount->mnt_stat.f_iosize; 577191783Srmacklem seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 578191783Srmacklem 579191783Srmacklem error = nfs_bioread_check_cons(vp, td, cred); 580191783Srmacklem if (error) 581191783Srmacklem return error; 582191783Srmacklem 583191783Srmacklem do { 584191783Srmacklem u_quad_t nsize; 585191783Srmacklem 586191783Srmacklem mtx_lock(&np->n_mtx); 587191783Srmacklem nsize = np->n_size; 588191783Srmacklem mtx_unlock(&np->n_mtx); 589191783Srmacklem 590191783Srmacklem switch (vp->v_type) { 591191783Srmacklem case VREG: 592191783Srmacklem NFSINCRGLOBAL(newnfsstats.biocache_reads); 593191783Srmacklem lbn = uio->uio_offset / biosize; 594191783Srmacklem on = uio->uio_offset & (biosize - 1); 595191783Srmacklem 596191783Srmacklem /* 597191783Srmacklem * Start the read ahead(s), as required. 598191783Srmacklem */ 599191783Srmacklem if (nmp->nm_readahead > 0) { 600191783Srmacklem for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 601191783Srmacklem (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 602191783Srmacklem rabn = lbn + 1 + nra; 603191783Srmacklem if (incore(&vp->v_bufobj, rabn) == NULL) { 604191783Srmacklem rabp = nfs_getcacheblk(vp, rabn, biosize, td); 605191783Srmacklem if (!rabp) { 606191783Srmacklem error = newnfs_sigintr(nmp, td); 607191783Srmacklem if (error) 608191783Srmacklem return (error); 609191783Srmacklem else 610191783Srmacklem break; 611191783Srmacklem } 612191783Srmacklem if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 613191783Srmacklem rabp->b_flags |= B_ASYNC; 614191783Srmacklem rabp->b_iocmd = BIO_READ; 615191783Srmacklem vfs_busy_pages(rabp, 0); 616191783Srmacklem if (ncl_asyncio(nmp, rabp, cred, td)) { 617191783Srmacklem rabp->b_flags |= B_INVAL; 618191783Srmacklem rabp->b_ioflags |= BIO_ERROR; 619191783Srmacklem vfs_unbusy_pages(rabp); 620191783Srmacklem brelse(rabp); 621191783Srmacklem break; 622191783Srmacklem } 623191783Srmacklem } else { 624191783Srmacklem brelse(rabp); 625191783Srmacklem } 626191783Srmacklem } 627191783Srmacklem } 628191783Srmacklem } 629191783Srmacklem 630191783Srmacklem /* Note that bcount is *not* DEV_BSIZE aligned. */ 631191783Srmacklem bcount = biosize; 632191783Srmacklem if ((off_t)lbn * biosize >= nsize) { 633191783Srmacklem bcount = 0; 634191783Srmacklem } else if ((off_t)(lbn + 1) * biosize > nsize) { 635191783Srmacklem bcount = nsize - (off_t)lbn * biosize; 636191783Srmacklem } 637191783Srmacklem bp = nfs_getcacheblk(vp, lbn, bcount, td); 638191783Srmacklem 639191783Srmacklem if (!bp) { 640191783Srmacklem error = newnfs_sigintr(nmp, td); 641191783Srmacklem return (error ? error : EINTR); 642191783Srmacklem } 643191783Srmacklem 644191783Srmacklem /* 645191783Srmacklem * If B_CACHE is not set, we must issue the read. If this 646191783Srmacklem * fails, we return an error. 647191783Srmacklem */ 648191783Srmacklem 649191783Srmacklem if ((bp->b_flags & B_CACHE) == 0) { 650191783Srmacklem bp->b_iocmd = BIO_READ; 651191783Srmacklem vfs_busy_pages(bp, 0); 652191783Srmacklem error = ncl_doio(vp, bp, cred, td); 653191783Srmacklem if (error) { 654191783Srmacklem brelse(bp); 655191783Srmacklem return (error); 656191783Srmacklem } 657191783Srmacklem } 658191783Srmacklem 659191783Srmacklem /* 660191783Srmacklem * on is the offset into the current bp. Figure out how many 661191783Srmacklem * bytes we can copy out of the bp. Note that bcount is 662191783Srmacklem * NOT DEV_BSIZE aligned. 663191783Srmacklem * 664191783Srmacklem * Then figure out how many bytes we can copy into the uio. 665191783Srmacklem */ 666191783Srmacklem 667191783Srmacklem n = 0; 668191783Srmacklem if (on < bcount) 669191783Srmacklem n = min((unsigned)(bcount - on), uio->uio_resid); 670191783Srmacklem break; 671191783Srmacklem case VLNK: 672191783Srmacklem NFSINCRGLOBAL(newnfsstats.biocache_readlinks); 673191783Srmacklem bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 674191783Srmacklem if (!bp) { 675191783Srmacklem error = newnfs_sigintr(nmp, td); 676191783Srmacklem return (error ? error : EINTR); 677191783Srmacklem } 678191783Srmacklem if ((bp->b_flags & B_CACHE) == 0) { 679191783Srmacklem bp->b_iocmd = BIO_READ; 680191783Srmacklem vfs_busy_pages(bp, 0); 681191783Srmacklem error = ncl_doio(vp, bp, cred, td); 682191783Srmacklem if (error) { 683191783Srmacklem bp->b_ioflags |= BIO_ERROR; 684191783Srmacklem brelse(bp); 685191783Srmacklem return (error); 686191783Srmacklem } 687191783Srmacklem } 688191783Srmacklem n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 689191783Srmacklem on = 0; 690191783Srmacklem break; 691191783Srmacklem case VDIR: 692191783Srmacklem NFSINCRGLOBAL(newnfsstats.biocache_readdirs); 693191783Srmacklem if (np->n_direofoffset 694191783Srmacklem && uio->uio_offset >= np->n_direofoffset) { 695191783Srmacklem return (0); 696191783Srmacklem } 697191783Srmacklem lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 698191783Srmacklem on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 699191783Srmacklem bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 700191783Srmacklem if (!bp) { 701191783Srmacklem error = newnfs_sigintr(nmp, td); 702191783Srmacklem return (error ? error : EINTR); 703191783Srmacklem } 704191783Srmacklem if ((bp->b_flags & B_CACHE) == 0) { 705191783Srmacklem bp->b_iocmd = BIO_READ; 706191783Srmacklem vfs_busy_pages(bp, 0); 707191783Srmacklem error = ncl_doio(vp, bp, cred, td); 708191783Srmacklem if (error) { 709191783Srmacklem brelse(bp); 710191783Srmacklem } 711191783Srmacklem while (error == NFSERR_BAD_COOKIE) { 712191783Srmacklem ncl_invaldir(vp); 713191783Srmacklem error = ncl_vinvalbuf(vp, 0, td, 1); 714191783Srmacklem /* 715191783Srmacklem * Yuck! The directory has been modified on the 716191783Srmacklem * server. The only way to get the block is by 717191783Srmacklem * reading from the beginning to get all the 718191783Srmacklem * offset cookies. 719191783Srmacklem * 720191783Srmacklem * Leave the last bp intact unless there is an error. 721191783Srmacklem * Loop back up to the while if the error is another 722191783Srmacklem * NFSERR_BAD_COOKIE (double yuch!). 723191783Srmacklem */ 724191783Srmacklem for (i = 0; i <= lbn && !error; i++) { 725191783Srmacklem if (np->n_direofoffset 726191783Srmacklem && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 727191783Srmacklem return (0); 728191783Srmacklem bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 729191783Srmacklem if (!bp) { 730191783Srmacklem error = newnfs_sigintr(nmp, td); 731191783Srmacklem return (error ? error : EINTR); 732191783Srmacklem } 733191783Srmacklem if ((bp->b_flags & B_CACHE) == 0) { 734191783Srmacklem bp->b_iocmd = BIO_READ; 735191783Srmacklem vfs_busy_pages(bp, 0); 736191783Srmacklem error = ncl_doio(vp, bp, cred, td); 737191783Srmacklem /* 738191783Srmacklem * no error + B_INVAL == directory EOF, 739191783Srmacklem * use the block. 740191783Srmacklem */ 741191783Srmacklem if (error == 0 && (bp->b_flags & B_INVAL)) 742191783Srmacklem break; 743191783Srmacklem } 744191783Srmacklem /* 745191783Srmacklem * An error will throw away the block and the 746191783Srmacklem * for loop will break out. If no error and this 747191783Srmacklem * is not the block we want, we throw away the 748191783Srmacklem * block and go for the next one via the for loop. 749191783Srmacklem */ 750191783Srmacklem if (error || i < lbn) 751191783Srmacklem brelse(bp); 752191783Srmacklem } 753191783Srmacklem } 754191783Srmacklem /* 755191783Srmacklem * The above while is repeated if we hit another cookie 756191783Srmacklem * error. If we hit an error and it wasn't a cookie error, 757191783Srmacklem * we give up. 758191783Srmacklem */ 759191783Srmacklem if (error) 760191783Srmacklem return (error); 761191783Srmacklem } 762191783Srmacklem 763191783Srmacklem /* 764191783Srmacklem * If not eof and read aheads are enabled, start one. 765191783Srmacklem * (You need the current block first, so that you have the 766191783Srmacklem * directory offset cookie of the next block.) 767191783Srmacklem */ 768191783Srmacklem if (nmp->nm_readahead > 0 && 769191783Srmacklem (bp->b_flags & B_INVAL) == 0 && 770191783Srmacklem (np->n_direofoffset == 0 || 771191783Srmacklem (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 772191783Srmacklem incore(&vp->v_bufobj, lbn + 1) == NULL) { 773191783Srmacklem rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 774191783Srmacklem if (rabp) { 775191783Srmacklem if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 776191783Srmacklem rabp->b_flags |= B_ASYNC; 777191783Srmacklem rabp->b_iocmd = BIO_READ; 778191783Srmacklem vfs_busy_pages(rabp, 0); 779191783Srmacklem if (ncl_asyncio(nmp, rabp, cred, td)) { 780191783Srmacklem rabp->b_flags |= B_INVAL; 781191783Srmacklem rabp->b_ioflags |= BIO_ERROR; 782191783Srmacklem vfs_unbusy_pages(rabp); 783191783Srmacklem brelse(rabp); 784191783Srmacklem } 785191783Srmacklem } else { 786191783Srmacklem brelse(rabp); 787191783Srmacklem } 788191783Srmacklem } 789191783Srmacklem } 790191783Srmacklem /* 791191783Srmacklem * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 792191783Srmacklem * chopped for the EOF condition, we cannot tell how large 793191783Srmacklem * NFS directories are going to be until we hit EOF. So 794191783Srmacklem * an NFS directory buffer is *not* chopped to its EOF. Now, 795191783Srmacklem * it just so happens that b_resid will effectively chop it 796191783Srmacklem * to EOF. *BUT* this information is lost if the buffer goes 797191783Srmacklem * away and is reconstituted into a B_CACHE state ( due to 798191783Srmacklem * being VMIO ) later. So we keep track of the directory eof 799191783Srmacklem * in np->n_direofoffset and chop it off as an extra step 800191783Srmacklem * right here. 801191783Srmacklem */ 802191783Srmacklem n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 803191783Srmacklem if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 804191783Srmacklem n = np->n_direofoffset - uio->uio_offset; 805191783Srmacklem break; 806191783Srmacklem default: 807191783Srmacklem ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type); 808191783Srmacklem bp = NULL; 809191783Srmacklem break; 810191783Srmacklem }; 811191783Srmacklem 812191783Srmacklem if (n > 0) { 813191783Srmacklem error = uiomove(bp->b_data + on, (int)n, uio); 814191783Srmacklem } 815191783Srmacklem if (vp->v_type == VLNK) 816191783Srmacklem n = 0; 817191783Srmacklem if (bp != NULL) 818191783Srmacklem brelse(bp); 819191783Srmacklem } while (error == 0 && uio->uio_resid > 0 && n > 0); 820191783Srmacklem return (error); 821191783Srmacklem} 822191783Srmacklem 823191783Srmacklem/* 824191783Srmacklem * The NFS write path cannot handle iovecs with len > 1. So we need to 825191783Srmacklem * break up iovecs accordingly (restricting them to wsize). 826191783Srmacklem * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 827191783Srmacklem * For the ASYNC case, 2 copies are needed. The first a copy from the 828191783Srmacklem * user buffer to a staging buffer and then a second copy from the staging 829191783Srmacklem * buffer to mbufs. This can be optimized by copying from the user buffer 830191783Srmacklem * directly into mbufs and passing the chain down, but that requires a 831191783Srmacklem * fair amount of re-working of the relevant codepaths (and can be done 832191783Srmacklem * later). 833191783Srmacklem */ 834191783Srmacklemstatic int 835191783Srmacklemnfs_directio_write(vp, uiop, cred, ioflag) 836191783Srmacklem struct vnode *vp; 837191783Srmacklem struct uio *uiop; 838191783Srmacklem struct ucred *cred; 839191783Srmacklem int ioflag; 840191783Srmacklem{ 841191783Srmacklem int error; 842191783Srmacklem struct nfsmount *nmp = VFSTONFS(vp->v_mount); 843191783Srmacklem struct thread *td = uiop->uio_td; 844191783Srmacklem int size; 845191783Srmacklem int wsize; 846191783Srmacklem 847191783Srmacklem mtx_lock(&nmp->nm_mtx); 848191783Srmacklem wsize = nmp->nm_wsize; 849191783Srmacklem mtx_unlock(&nmp->nm_mtx); 850191783Srmacklem if (ioflag & IO_SYNC) { 851191783Srmacklem int iomode, must_commit; 852191783Srmacklem struct uio uio; 853191783Srmacklem struct iovec iov; 854191783Srmacklemdo_sync: 855191783Srmacklem while (uiop->uio_resid > 0) { 856191783Srmacklem size = min(uiop->uio_resid, wsize); 857191783Srmacklem size = min(uiop->uio_iov->iov_len, size); 858191783Srmacklem iov.iov_base = uiop->uio_iov->iov_base; 859191783Srmacklem iov.iov_len = size; 860191783Srmacklem uio.uio_iov = &iov; 861191783Srmacklem uio.uio_iovcnt = 1; 862191783Srmacklem uio.uio_offset = uiop->uio_offset; 863191783Srmacklem uio.uio_resid = size; 864191783Srmacklem uio.uio_segflg = UIO_USERSPACE; 865191783Srmacklem uio.uio_rw = UIO_WRITE; 866191783Srmacklem uio.uio_td = td; 867191783Srmacklem iomode = NFSWRITE_FILESYNC; 868191783Srmacklem error = ncl_writerpc(vp, &uio, cred, &iomode, 869191783Srmacklem &must_commit); 870191783Srmacklem KASSERT((must_commit == 0), 871191783Srmacklem ("ncl_directio_write: Did not commit write")); 872191783Srmacklem if (error) 873191783Srmacklem return (error); 874191783Srmacklem uiop->uio_offset += size; 875191783Srmacklem uiop->uio_resid -= size; 876191783Srmacklem if (uiop->uio_iov->iov_len <= size) { 877191783Srmacklem uiop->uio_iovcnt--; 878191783Srmacklem uiop->uio_iov++; 879191783Srmacklem } else { 880191783Srmacklem uiop->uio_iov->iov_base = 881191783Srmacklem (char *)uiop->uio_iov->iov_base + size; 882191783Srmacklem uiop->uio_iov->iov_len -= size; 883191783Srmacklem } 884191783Srmacklem } 885191783Srmacklem } else { 886191783Srmacklem struct uio *t_uio; 887191783Srmacklem struct iovec *t_iov; 888191783Srmacklem struct buf *bp; 889191783Srmacklem 890191783Srmacklem /* 891191783Srmacklem * Break up the write into blocksize chunks and hand these 892191783Srmacklem * over to nfsiod's for write back. 893191783Srmacklem * Unfortunately, this incurs a copy of the data. Since 894191783Srmacklem * the user could modify the buffer before the write is 895191783Srmacklem * initiated. 896191783Srmacklem * 897191783Srmacklem * The obvious optimization here is that one of the 2 copies 898191783Srmacklem * in the async write path can be eliminated by copying the 899191783Srmacklem * data here directly into mbufs and passing the mbuf chain 900191783Srmacklem * down. But that will require a fair amount of re-working 901191783Srmacklem * of the code and can be done if there's enough interest 902191783Srmacklem * in NFS directio access. 903191783Srmacklem */ 904191783Srmacklem while (uiop->uio_resid > 0) { 905191783Srmacklem size = min(uiop->uio_resid, wsize); 906191783Srmacklem size = min(uiop->uio_iov->iov_len, size); 907191783Srmacklem bp = getpbuf(&ncl_pbuf_freecnt); 908191783Srmacklem t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 909191783Srmacklem t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 910191783Srmacklem t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 911191783Srmacklem t_iov->iov_len = size; 912191783Srmacklem t_uio->uio_iov = t_iov; 913191783Srmacklem t_uio->uio_iovcnt = 1; 914191783Srmacklem t_uio->uio_offset = uiop->uio_offset; 915191783Srmacklem t_uio->uio_resid = size; 916191783Srmacklem t_uio->uio_segflg = UIO_SYSSPACE; 917191783Srmacklem t_uio->uio_rw = UIO_WRITE; 918191783Srmacklem t_uio->uio_td = td; 919191783Srmacklem bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, size); 920191783Srmacklem bp->b_flags |= B_DIRECT; 921191783Srmacklem bp->b_iocmd = BIO_WRITE; 922191783Srmacklem if (cred != NOCRED) { 923191783Srmacklem crhold(cred); 924191783Srmacklem bp->b_wcred = cred; 925191783Srmacklem } else 926191783Srmacklem bp->b_wcred = NOCRED; 927191783Srmacklem bp->b_caller1 = (void *)t_uio; 928191783Srmacklem bp->b_vp = vp; 929191783Srmacklem error = ncl_asyncio(nmp, bp, NOCRED, td); 930191783Srmacklem if (error) { 931191783Srmacklem free(t_iov->iov_base, M_NFSDIRECTIO); 932191783Srmacklem free(t_iov, M_NFSDIRECTIO); 933191783Srmacklem free(t_uio, M_NFSDIRECTIO); 934191783Srmacklem bp->b_vp = NULL; 935191783Srmacklem relpbuf(bp, &ncl_pbuf_freecnt); 936191783Srmacklem if (error == EINTR) 937191783Srmacklem return (error); 938191783Srmacklem goto do_sync; 939191783Srmacklem } 940191783Srmacklem uiop->uio_offset += size; 941191783Srmacklem uiop->uio_resid -= size; 942191783Srmacklem if (uiop->uio_iov->iov_len <= size) { 943191783Srmacklem uiop->uio_iovcnt--; 944191783Srmacklem uiop->uio_iov++; 945191783Srmacklem } else { 946191783Srmacklem uiop->uio_iov->iov_base = 947191783Srmacklem (char *)uiop->uio_iov->iov_base + size; 948191783Srmacklem uiop->uio_iov->iov_len -= size; 949191783Srmacklem } 950191783Srmacklem } 951191783Srmacklem } 952191783Srmacklem return (0); 953191783Srmacklem} 954191783Srmacklem 955191783Srmacklem/* 956191783Srmacklem * Vnode op for write using bio 957191783Srmacklem */ 958191783Srmacklemint 959191783Srmacklemncl_write(struct vop_write_args *ap) 960191783Srmacklem{ 961191783Srmacklem int biosize; 962191783Srmacklem struct uio *uio = ap->a_uio; 963191783Srmacklem struct thread *td = uio->uio_td; 964191783Srmacklem struct vnode *vp = ap->a_vp; 965191783Srmacklem struct nfsnode *np = VTONFS(vp); 966191783Srmacklem struct ucred *cred = ap->a_cred; 967191783Srmacklem int ioflag = ap->a_ioflag; 968191783Srmacklem struct buf *bp; 969191783Srmacklem struct vattr vattr; 970191783Srmacklem struct nfsmount *nmp = VFSTONFS(vp->v_mount); 971191783Srmacklem daddr_t lbn; 972191783Srmacklem int bcount; 973191783Srmacklem int n, on, error = 0; 974191783Srmacklem struct proc *p = td?td->td_proc:NULL; 975191783Srmacklem 976191783Srmacklem#ifdef DIAGNOSTIC 977191783Srmacklem if (uio->uio_rw != UIO_WRITE) 978191783Srmacklem panic("ncl_write mode"); 979191783Srmacklem if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 980191783Srmacklem panic("ncl_write proc"); 981191783Srmacklem#endif 982191783Srmacklem if (vp->v_type != VREG) 983191783Srmacklem return (EIO); 984191783Srmacklem mtx_lock(&np->n_mtx); 985191783Srmacklem if (np->n_flag & NWRITEERR) { 986191783Srmacklem np->n_flag &= ~NWRITEERR; 987191783Srmacklem mtx_unlock(&np->n_mtx); 988191783Srmacklem return (np->n_error); 989191783Srmacklem } else 990191783Srmacklem mtx_unlock(&np->n_mtx); 991191783Srmacklem mtx_lock(&nmp->nm_mtx); 992191783Srmacklem if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 993191783Srmacklem (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 994191783Srmacklem mtx_unlock(&nmp->nm_mtx); 995191783Srmacklem (void)ncl_fsinfo(nmp, vp, cred, td); 996191783Srmacklem mtx_lock(&nmp->nm_mtx); 997191783Srmacklem } 998191783Srmacklem if (nmp->nm_wsize == 0) 999191783Srmacklem (void) newnfs_iosize(nmp); 1000191783Srmacklem mtx_unlock(&nmp->nm_mtx); 1001191783Srmacklem 1002191783Srmacklem /* 1003191783Srmacklem * Synchronously flush pending buffers if we are in synchronous 1004191783Srmacklem * mode or if we are appending. 1005191783Srmacklem */ 1006191783Srmacklem if (ioflag & (IO_APPEND | IO_SYNC)) { 1007191783Srmacklem mtx_lock(&np->n_mtx); 1008191783Srmacklem if (np->n_flag & NMODIFIED) { 1009191783Srmacklem mtx_unlock(&np->n_mtx); 1010191783Srmacklem#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 1011191783Srmacklem /* 1012191783Srmacklem * Require non-blocking, synchronous writes to 1013191783Srmacklem * dirty files to inform the program it needs 1014191783Srmacklem * to fsync(2) explicitly. 1015191783Srmacklem */ 1016191783Srmacklem if (ioflag & IO_NDELAY) 1017191783Srmacklem return (EAGAIN); 1018191783Srmacklem#endif 1019191783Srmacklemflush_and_restart: 1020191783Srmacklem np->n_attrstamp = 0; 1021191783Srmacklem error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 1022191783Srmacklem if (error) 1023191783Srmacklem return (error); 1024191783Srmacklem } else 1025191783Srmacklem mtx_unlock(&np->n_mtx); 1026191783Srmacklem } 1027191783Srmacklem 1028191783Srmacklem /* 1029191783Srmacklem * If IO_APPEND then load uio_offset. We restart here if we cannot 1030191783Srmacklem * get the append lock. 1031191783Srmacklem */ 1032191783Srmacklem if (ioflag & IO_APPEND) { 1033191783Srmacklem np->n_attrstamp = 0; 1034191783Srmacklem error = VOP_GETATTR(vp, &vattr, cred); 1035191783Srmacklem if (error) 1036191783Srmacklem return (error); 1037191783Srmacklem mtx_lock(&np->n_mtx); 1038191783Srmacklem uio->uio_offset = np->n_size; 1039191783Srmacklem mtx_unlock(&np->n_mtx); 1040191783Srmacklem } 1041191783Srmacklem 1042191783Srmacklem if (uio->uio_offset < 0) 1043191783Srmacklem return (EINVAL); 1044191783Srmacklem if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 1045191783Srmacklem return (EFBIG); 1046191783Srmacklem if (uio->uio_resid == 0) 1047191783Srmacklem return (0); 1048191783Srmacklem 1049191783Srmacklem if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 1050191783Srmacklem return nfs_directio_write(vp, uio, cred, ioflag); 1051191783Srmacklem 1052191783Srmacklem /* 1053191783Srmacklem * Maybe this should be above the vnode op call, but so long as 1054191783Srmacklem * file servers have no limits, i don't think it matters 1055191783Srmacklem */ 1056191783Srmacklem if (p != NULL) { 1057191783Srmacklem PROC_LOCK(p); 1058191783Srmacklem if (uio->uio_offset + uio->uio_resid > 1059191783Srmacklem lim_cur(p, RLIMIT_FSIZE)) { 1060191783Srmacklem psignal(p, SIGXFSZ); 1061191783Srmacklem PROC_UNLOCK(p); 1062191783Srmacklem return (EFBIG); 1063191783Srmacklem } 1064191783Srmacklem PROC_UNLOCK(p); 1065191783Srmacklem } 1066191783Srmacklem 1067191783Srmacklem biosize = vp->v_mount->mnt_stat.f_iosize; 1068191783Srmacklem /* 1069191783Srmacklem * Find all of this file's B_NEEDCOMMIT buffers. If our writes 1070191783Srmacklem * would exceed the local maximum per-file write commit size when 1071191783Srmacklem * combined with those, we must decide whether to flush, 1072191783Srmacklem * go synchronous, or return error. We don't bother checking 1073191783Srmacklem * IO_UNIT -- we just make all writes atomic anyway, as there's 1074191783Srmacklem * no point optimizing for something that really won't ever happen. 1075191783Srmacklem */ 1076191783Srmacklem if (!(ioflag & IO_SYNC)) { 1077191783Srmacklem int nflag; 1078191783Srmacklem 1079191783Srmacklem mtx_lock(&np->n_mtx); 1080191783Srmacklem nflag = np->n_flag; 1081191783Srmacklem mtx_unlock(&np->n_mtx); 1082191783Srmacklem int needrestart = 0; 1083191783Srmacklem if (nmp->nm_wcommitsize < uio->uio_resid) { 1084191783Srmacklem /* 1085191783Srmacklem * If this request could not possibly be completed 1086191783Srmacklem * without exceeding the maximum outstanding write 1087191783Srmacklem * commit size, see if we can convert it into a 1088191783Srmacklem * synchronous write operation. 1089191783Srmacklem */ 1090191783Srmacklem if (ioflag & IO_NDELAY) 1091191783Srmacklem return (EAGAIN); 1092191783Srmacklem ioflag |= IO_SYNC; 1093191783Srmacklem if (nflag & NMODIFIED) 1094191783Srmacklem needrestart = 1; 1095191783Srmacklem } else if (nflag & NMODIFIED) { 1096191783Srmacklem int wouldcommit = 0; 1097191783Srmacklem BO_LOCK(&vp->v_bufobj); 1098191783Srmacklem if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 1099191783Srmacklem TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 1100191783Srmacklem b_bobufs) { 1101191783Srmacklem if (bp->b_flags & B_NEEDCOMMIT) 1102191783Srmacklem wouldcommit += bp->b_bcount; 1103191783Srmacklem } 1104191783Srmacklem } 1105191783Srmacklem BO_UNLOCK(&vp->v_bufobj); 1106191783Srmacklem /* 1107191783Srmacklem * Since we're not operating synchronously and 1108191783Srmacklem * bypassing the buffer cache, we are in a commit 1109191783Srmacklem * and holding all of these buffers whether 1110191783Srmacklem * transmitted or not. If not limited, this 1111191783Srmacklem * will lead to the buffer cache deadlocking, 1112191783Srmacklem * as no one else can flush our uncommitted buffers. 1113191783Srmacklem */ 1114191783Srmacklem wouldcommit += uio->uio_resid; 1115191783Srmacklem /* 1116191783Srmacklem * If we would initially exceed the maximum 1117191783Srmacklem * outstanding write commit size, flush and restart. 1118191783Srmacklem */ 1119191783Srmacklem if (wouldcommit > nmp->nm_wcommitsize) 1120191783Srmacklem needrestart = 1; 1121191783Srmacklem } 1122191783Srmacklem if (needrestart) 1123191783Srmacklem goto flush_and_restart; 1124191783Srmacklem } 1125191783Srmacklem 1126191783Srmacklem do { 1127191783Srmacklem NFSINCRGLOBAL(newnfsstats.biocache_writes); 1128191783Srmacklem lbn = uio->uio_offset / biosize; 1129191783Srmacklem on = uio->uio_offset & (biosize-1); 1130191783Srmacklem n = min((unsigned)(biosize - on), uio->uio_resid); 1131191783Srmacklemagain: 1132191783Srmacklem /* 1133191783Srmacklem * Handle direct append and file extension cases, calculate 1134191783Srmacklem * unaligned buffer size. 1135191783Srmacklem */ 1136191783Srmacklem mtx_lock(&np->n_mtx); 1137191783Srmacklem if (uio->uio_offset == np->n_size && n) { 1138191783Srmacklem mtx_unlock(&np->n_mtx); 1139191783Srmacklem /* 1140191783Srmacklem * Get the buffer (in its pre-append state to maintain 1141191783Srmacklem * B_CACHE if it was previously set). Resize the 1142191783Srmacklem * nfsnode after we have locked the buffer to prevent 1143191783Srmacklem * readers from reading garbage. 1144191783Srmacklem */ 1145191783Srmacklem bcount = on; 1146191783Srmacklem bp = nfs_getcacheblk(vp, lbn, bcount, td); 1147191783Srmacklem 1148191783Srmacklem if (bp != NULL) { 1149191783Srmacklem long save; 1150191783Srmacklem 1151191783Srmacklem mtx_lock(&np->n_mtx); 1152191783Srmacklem np->n_size = uio->uio_offset + n; 1153191783Srmacklem np->n_flag |= NMODIFIED; 1154191783Srmacklem vnode_pager_setsize(vp, np->n_size); 1155191783Srmacklem mtx_unlock(&np->n_mtx); 1156191783Srmacklem 1157191783Srmacklem save = bp->b_flags & B_CACHE; 1158191783Srmacklem bcount += n; 1159191783Srmacklem allocbuf(bp, bcount); 1160191783Srmacklem bp->b_flags |= save; 1161191783Srmacklem } 1162191783Srmacklem } else { 1163191783Srmacklem /* 1164191783Srmacklem * Obtain the locked cache block first, and then 1165191783Srmacklem * adjust the file's size as appropriate. 1166191783Srmacklem */ 1167191783Srmacklem bcount = on + n; 1168191783Srmacklem if ((off_t)lbn * biosize + bcount < np->n_size) { 1169191783Srmacklem if ((off_t)(lbn + 1) * biosize < np->n_size) 1170191783Srmacklem bcount = biosize; 1171191783Srmacklem else 1172191783Srmacklem bcount = np->n_size - (off_t)lbn * biosize; 1173191783Srmacklem } 1174191783Srmacklem mtx_unlock(&np->n_mtx); 1175191783Srmacklem bp = nfs_getcacheblk(vp, lbn, bcount, td); 1176191783Srmacklem mtx_lock(&np->n_mtx); 1177191783Srmacklem if (uio->uio_offset + n > np->n_size) { 1178191783Srmacklem np->n_size = uio->uio_offset + n; 1179191783Srmacklem np->n_flag |= NMODIFIED; 1180191783Srmacklem vnode_pager_setsize(vp, np->n_size); 1181191783Srmacklem } 1182191783Srmacklem mtx_unlock(&np->n_mtx); 1183191783Srmacklem } 1184191783Srmacklem 1185191783Srmacklem if (!bp) { 1186191783Srmacklem error = newnfs_sigintr(nmp, td); 1187191783Srmacklem if (!error) 1188191783Srmacklem error = EINTR; 1189191783Srmacklem break; 1190191783Srmacklem } 1191191783Srmacklem 1192191783Srmacklem /* 1193191783Srmacklem * Issue a READ if B_CACHE is not set. In special-append 1194191783Srmacklem * mode, B_CACHE is based on the buffer prior to the write 1195191783Srmacklem * op and is typically set, avoiding the read. If a read 1196191783Srmacklem * is required in special append mode, the server will 1197191783Srmacklem * probably send us a short-read since we extended the file 1198191783Srmacklem * on our end, resulting in b_resid == 0 and, thusly, 1199191783Srmacklem * B_CACHE getting set. 1200191783Srmacklem * 1201191783Srmacklem * We can also avoid issuing the read if the write covers 1202191783Srmacklem * the entire buffer. We have to make sure the buffer state 1203191783Srmacklem * is reasonable in this case since we will not be initiating 1204191783Srmacklem * I/O. See the comments in kern/vfs_bio.c's getblk() for 1205191783Srmacklem * more information. 1206191783Srmacklem * 1207191783Srmacklem * B_CACHE may also be set due to the buffer being cached 1208191783Srmacklem * normally. 1209191783Srmacklem */ 1210191783Srmacklem 1211191783Srmacklem if (on == 0 && n == bcount) { 1212191783Srmacklem bp->b_flags |= B_CACHE; 1213191783Srmacklem bp->b_flags &= ~B_INVAL; 1214191783Srmacklem bp->b_ioflags &= ~BIO_ERROR; 1215191783Srmacklem } 1216191783Srmacklem 1217191783Srmacklem if ((bp->b_flags & B_CACHE) == 0) { 1218191783Srmacklem bp->b_iocmd = BIO_READ; 1219191783Srmacklem vfs_busy_pages(bp, 0); 1220191783Srmacklem error = ncl_doio(vp, bp, cred, td); 1221191783Srmacklem if (error) { 1222191783Srmacklem brelse(bp); 1223191783Srmacklem break; 1224191783Srmacklem } 1225191783Srmacklem } 1226191783Srmacklem if (bp->b_wcred == NOCRED) 1227191783Srmacklem bp->b_wcred = crhold(cred); 1228191783Srmacklem mtx_lock(&np->n_mtx); 1229191783Srmacklem np->n_flag |= NMODIFIED; 1230191783Srmacklem mtx_unlock(&np->n_mtx); 1231191783Srmacklem 1232191783Srmacklem /* 1233191783Srmacklem * If dirtyend exceeds file size, chop it down. This should 1234191783Srmacklem * not normally occur but there is an append race where it 1235191783Srmacklem * might occur XXX, so we log it. 1236191783Srmacklem * 1237191783Srmacklem * If the chopping creates a reverse-indexed or degenerate 1238191783Srmacklem * situation with dirtyoff/end, we 0 both of them. 1239191783Srmacklem */ 1240191783Srmacklem 1241191783Srmacklem if (bp->b_dirtyend > bcount) { 1242191783Srmacklem ncl_printf("NFS append race @%lx:%d\n", 1243191783Srmacklem (long)bp->b_blkno * DEV_BSIZE, 1244191783Srmacklem bp->b_dirtyend - bcount); 1245191783Srmacklem bp->b_dirtyend = bcount; 1246191783Srmacklem } 1247191783Srmacklem 1248191783Srmacklem if (bp->b_dirtyoff >= bp->b_dirtyend) 1249191783Srmacklem bp->b_dirtyoff = bp->b_dirtyend = 0; 1250191783Srmacklem 1251191783Srmacklem /* 1252191783Srmacklem * If the new write will leave a contiguous dirty 1253191783Srmacklem * area, just update the b_dirtyoff and b_dirtyend, 1254191783Srmacklem * otherwise force a write rpc of the old dirty area. 1255191783Srmacklem * 1256191783Srmacklem * While it is possible to merge discontiguous writes due to 1257191783Srmacklem * our having a B_CACHE buffer ( and thus valid read data 1258191783Srmacklem * for the hole), we don't because it could lead to 1259191783Srmacklem * significant cache coherency problems with multiple clients, 1260191783Srmacklem * especially if locking is implemented later on. 1261191783Srmacklem * 1262191783Srmacklem * as an optimization we could theoretically maintain 1263191783Srmacklem * a linked list of discontinuous areas, but we would still 1264191783Srmacklem * have to commit them separately so there isn't much 1265191783Srmacklem * advantage to it except perhaps a bit of asynchronization. 1266191783Srmacklem */ 1267191783Srmacklem 1268191783Srmacklem if (bp->b_dirtyend > 0 && 1269191783Srmacklem (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1270191783Srmacklem if (bwrite(bp) == EINTR) { 1271191783Srmacklem error = EINTR; 1272191783Srmacklem break; 1273191783Srmacklem } 1274191783Srmacklem goto again; 1275191783Srmacklem } 1276191783Srmacklem 1277191783Srmacklem error = uiomove((char *)bp->b_data + on, n, uio); 1278191783Srmacklem 1279191783Srmacklem /* 1280191783Srmacklem * Since this block is being modified, it must be written 1281191783Srmacklem * again and not just committed. Since write clustering does 1282191783Srmacklem * not work for the stage 1 data write, only the stage 2 1283191783Srmacklem * commit rpc, we have to clear B_CLUSTEROK as well. 1284191783Srmacklem */ 1285191783Srmacklem bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1286191783Srmacklem 1287191783Srmacklem if (error) { 1288191783Srmacklem bp->b_ioflags |= BIO_ERROR; 1289191783Srmacklem brelse(bp); 1290191783Srmacklem break; 1291191783Srmacklem } 1292191783Srmacklem 1293191783Srmacklem /* 1294191783Srmacklem * Only update dirtyoff/dirtyend if not a degenerate 1295191783Srmacklem * condition. 1296191783Srmacklem */ 1297191783Srmacklem if (n) { 1298191783Srmacklem if (bp->b_dirtyend > 0) { 1299191783Srmacklem bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1300191783Srmacklem bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1301191783Srmacklem } else { 1302191783Srmacklem bp->b_dirtyoff = on; 1303191783Srmacklem bp->b_dirtyend = on + n; 1304191783Srmacklem } 1305193187Salc vfs_bio_set_valid(bp, on, n); 1306191783Srmacklem } 1307191783Srmacklem 1308191783Srmacklem /* 1309191783Srmacklem * If IO_SYNC do bwrite(). 1310191783Srmacklem * 1311191783Srmacklem * IO_INVAL appears to be unused. The idea appears to be 1312191783Srmacklem * to turn off caching in this case. Very odd. XXX 1313191783Srmacklem */ 1314191783Srmacklem if ((ioflag & IO_SYNC)) { 1315191783Srmacklem if (ioflag & IO_INVAL) 1316191783Srmacklem bp->b_flags |= B_NOCACHE; 1317191783Srmacklem error = bwrite(bp); 1318191783Srmacklem if (error) 1319191783Srmacklem break; 1320191783Srmacklem } else if ((n + on) == biosize) { 1321191783Srmacklem bp->b_flags |= B_ASYNC; 1322191783Srmacklem (void) ncl_writebp(bp, 0, NULL); 1323191783Srmacklem } else { 1324191783Srmacklem bdwrite(bp); 1325191783Srmacklem } 1326191783Srmacklem } while (uio->uio_resid > 0 && n > 0); 1327191783Srmacklem 1328191783Srmacklem return (error); 1329191783Srmacklem} 1330191783Srmacklem 1331191783Srmacklem/* 1332191783Srmacklem * Get an nfs cache block. 1333191783Srmacklem * 1334191783Srmacklem * Allocate a new one if the block isn't currently in the cache 1335191783Srmacklem * and return the block marked busy. If the calling process is 1336191783Srmacklem * interrupted by a signal for an interruptible mount point, return 1337191783Srmacklem * NULL. 1338191783Srmacklem * 1339191783Srmacklem * The caller must carefully deal with the possible B_INVAL state of 1340191783Srmacklem * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it 1341191783Srmacklem * indirectly), so synchronous reads can be issued without worrying about 1342191783Srmacklem * the B_INVAL state. We have to be a little more careful when dealing 1343191783Srmacklem * with writes (see comments in nfs_write()) when extending a file past 1344191783Srmacklem * its EOF. 1345191783Srmacklem */ 1346191783Srmacklemstatic struct buf * 1347191783Srmacklemnfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1348191783Srmacklem{ 1349191783Srmacklem struct buf *bp; 1350191783Srmacklem struct mount *mp; 1351191783Srmacklem struct nfsmount *nmp; 1352191783Srmacklem 1353191783Srmacklem mp = vp->v_mount; 1354191783Srmacklem nmp = VFSTONFS(mp); 1355191783Srmacklem 1356191783Srmacklem if (nmp->nm_flag & NFSMNT_INT) { 1357191783Srmacklem sigset_t oldset; 1358191783Srmacklem 1359191783Srmacklem ncl_set_sigmask(td, &oldset); 1360191783Srmacklem bp = getblk(vp, bn, size, PCATCH, 0, 0); 1361191783Srmacklem ncl_restore_sigmask(td, &oldset); 1362191783Srmacklem while (bp == NULL) { 1363191783Srmacklem if (newnfs_sigintr(nmp, td)) 1364191783Srmacklem return (NULL); 1365191783Srmacklem bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1366191783Srmacklem } 1367191783Srmacklem } else { 1368191783Srmacklem bp = getblk(vp, bn, size, 0, 0, 0); 1369191783Srmacklem } 1370191783Srmacklem 1371191783Srmacklem if (vp->v_type == VREG) { 1372191783Srmacklem int biosize; 1373191783Srmacklem 1374191783Srmacklem biosize = mp->mnt_stat.f_iosize; 1375191783Srmacklem bp->b_blkno = bn * (biosize / DEV_BSIZE); 1376191783Srmacklem } 1377191783Srmacklem return (bp); 1378191783Srmacklem} 1379191783Srmacklem 1380191783Srmacklem/* 1381191783Srmacklem * Flush and invalidate all dirty buffers. If another process is already 1382191783Srmacklem * doing the flush, just wait for completion. 1383191783Srmacklem */ 1384191783Srmacklemint 1385191783Srmacklemncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1386191783Srmacklem{ 1387191783Srmacklem struct nfsnode *np = VTONFS(vp); 1388191783Srmacklem struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1389191783Srmacklem int error = 0, slpflag, slptimeo; 1390191783Srmacklem int old_lock = 0; 1391191783Srmacklem 1392191783Srmacklem ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf"); 1393191783Srmacklem 1394191783Srmacklem if ((nmp->nm_flag & NFSMNT_INT) == 0) 1395191783Srmacklem intrflg = 0; 1396191783Srmacklem if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)) 1397191783Srmacklem intrflg = 1; 1398191783Srmacklem if (intrflg) { 1399191783Srmacklem slpflag = PCATCH; 1400191783Srmacklem slptimeo = 2 * hz; 1401191783Srmacklem } else { 1402191783Srmacklem slpflag = 0; 1403191783Srmacklem slptimeo = 0; 1404191783Srmacklem } 1405191783Srmacklem 1406191783Srmacklem old_lock = ncl_upgrade_vnlock(vp); 1407193955Srmacklem if (vp->v_iflag & VI_DOOMED) { 1408193955Srmacklem /* 1409193955Srmacklem * Since vgonel() uses the generic vinvalbuf() to flush 1410193955Srmacklem * dirty buffers and it does not call this function, it 1411193955Srmacklem * is safe to just return OK when VI_DOOMED is set. 1412193955Srmacklem */ 1413193955Srmacklem ncl_downgrade_vnlock(vp, old_lock); 1414193955Srmacklem return (0); 1415193955Srmacklem } 1416193955Srmacklem 1417191783Srmacklem /* 1418191783Srmacklem * Now, flush as required. 1419191783Srmacklem */ 1420191783Srmacklem if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1421191783Srmacklem VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1422191783Srmacklem vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1423191783Srmacklem VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1424191783Srmacklem /* 1425191783Srmacklem * If the page clean was interrupted, fail the invalidation. 1426191783Srmacklem * Not doing so, we run the risk of losing dirty pages in the 1427191783Srmacklem * vinvalbuf() call below. 1428191783Srmacklem */ 1429191783Srmacklem if (intrflg && (error = newnfs_sigintr(nmp, td))) 1430191783Srmacklem goto out; 1431191783Srmacklem } 1432191783Srmacklem 1433191783Srmacklem error = vinvalbuf(vp, flags, slpflag, 0); 1434191783Srmacklem while (error) { 1435191783Srmacklem if (intrflg && (error = newnfs_sigintr(nmp, td))) 1436191783Srmacklem goto out; 1437191783Srmacklem error = vinvalbuf(vp, flags, 0, slptimeo); 1438191783Srmacklem } 1439191783Srmacklem mtx_lock(&np->n_mtx); 1440191783Srmacklem if (np->n_directio_asyncwr == 0) 1441191783Srmacklem np->n_flag &= ~NMODIFIED; 1442191783Srmacklem mtx_unlock(&np->n_mtx); 1443191783Srmacklemout: 1444191783Srmacklem ncl_downgrade_vnlock(vp, old_lock); 1445191783Srmacklem return error; 1446191783Srmacklem} 1447191783Srmacklem 1448191783Srmacklem/* 1449191783Srmacklem * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1450191783Srmacklem * This is mainly to avoid queueing async I/O requests when the nfsiods 1451191783Srmacklem * are all hung on a dead server. 1452191783Srmacklem * 1453191783Srmacklem * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1454191783Srmacklem * is eventually dequeued by the async daemon, ncl_doio() *will*. 1455191783Srmacklem */ 1456191783Srmacklemint 1457191783Srmacklemncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1458191783Srmacklem{ 1459191783Srmacklem int iod; 1460191783Srmacklem int gotiod; 1461191783Srmacklem int slpflag = 0; 1462191783Srmacklem int slptimeo = 0; 1463191783Srmacklem int error, error2; 1464191783Srmacklem 1465191783Srmacklem /* 1466191783Srmacklem * Unless iothreadcnt is set > 0, don't bother with async I/O 1467191783Srmacklem * threads. For LAN environments, they don't buy any significant 1468191783Srmacklem * performance improvement that you can't get with large block 1469191783Srmacklem * sizes. 1470191783Srmacklem */ 1471191783Srmacklem if (nmp->nm_readahead == 0) 1472191783Srmacklem return (EPERM); 1473191783Srmacklem 1474191783Srmacklem /* 1475191783Srmacklem * Commits are usually short and sweet so lets save some cpu and 1476191783Srmacklem * leave the async daemons for more important rpc's (such as reads 1477191783Srmacklem * and writes). 1478191783Srmacklem */ 1479191783Srmacklem mtx_lock(&ncl_iod_mutex); 1480191783Srmacklem if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1481191783Srmacklem (nmp->nm_bufqiods > ncl_numasync / 2)) { 1482191783Srmacklem mtx_unlock(&ncl_iod_mutex); 1483191783Srmacklem return(EIO); 1484191783Srmacklem } 1485191783Srmacklemagain: 1486191783Srmacklem if (nmp->nm_flag & NFSMNT_INT) 1487191783Srmacklem slpflag = PCATCH; 1488191783Srmacklem gotiod = FALSE; 1489191783Srmacklem 1490191783Srmacklem /* 1491191783Srmacklem * Find a free iod to process this request. 1492191783Srmacklem */ 1493191783Srmacklem for (iod = 0; iod < ncl_numasync; iod++) 1494191783Srmacklem if (ncl_iodwant[iod]) { 1495191783Srmacklem gotiod = TRUE; 1496191783Srmacklem break; 1497191783Srmacklem } 1498191783Srmacklem 1499191783Srmacklem /* 1500191783Srmacklem * Try to create one if none are free. 1501191783Srmacklem */ 1502191783Srmacklem if (!gotiod) { 1503191783Srmacklem iod = ncl_nfsiodnew(); 1504191783Srmacklem if (iod != -1) 1505191783Srmacklem gotiod = TRUE; 1506191783Srmacklem } 1507191783Srmacklem 1508191783Srmacklem if (gotiod) { 1509191783Srmacklem /* 1510191783Srmacklem * Found one, so wake it up and tell it which 1511191783Srmacklem * mount to process. 1512191783Srmacklem */ 1513191783Srmacklem NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n", 1514191783Srmacklem iod, nmp)); 1515191783Srmacklem ncl_iodwant[iod] = NULL; 1516191783Srmacklem ncl_iodmount[iod] = nmp; 1517191783Srmacklem nmp->nm_bufqiods++; 1518191783Srmacklem wakeup(&ncl_iodwant[iod]); 1519191783Srmacklem } 1520191783Srmacklem 1521191783Srmacklem /* 1522191783Srmacklem * If none are free, we may already have an iod working on this mount 1523191783Srmacklem * point. If so, it will process our request. 1524191783Srmacklem */ 1525191783Srmacklem if (!gotiod) { 1526191783Srmacklem if (nmp->nm_bufqiods > 0) { 1527191783Srmacklem NFS_DPF(ASYNCIO, 1528191783Srmacklem ("ncl_asyncio: %d iods are already processing mount %p\n", 1529191783Srmacklem nmp->nm_bufqiods, nmp)); 1530191783Srmacklem gotiod = TRUE; 1531191783Srmacklem } 1532191783Srmacklem } 1533191783Srmacklem 1534191783Srmacklem /* 1535191783Srmacklem * If we have an iod which can process the request, then queue 1536191783Srmacklem * the buffer. 1537191783Srmacklem */ 1538191783Srmacklem if (gotiod) { 1539191783Srmacklem /* 1540191783Srmacklem * Ensure that the queue never grows too large. We still want 1541191783Srmacklem * to asynchronize so we block rather then return EIO. 1542191783Srmacklem */ 1543191783Srmacklem while (nmp->nm_bufqlen >= 2*ncl_numasync) { 1544191783Srmacklem NFS_DPF(ASYNCIO, 1545191783Srmacklem ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp)); 1546191783Srmacklem nmp->nm_bufqwant = TRUE; 1547191783Srmacklem error = ncl_msleep(td, &nmp->nm_bufq, &ncl_iod_mutex, 1548191783Srmacklem slpflag | PRIBIO, 1549191783Srmacklem "nfsaio", slptimeo); 1550191783Srmacklem if (error) { 1551191783Srmacklem error2 = newnfs_sigintr(nmp, td); 1552191783Srmacklem if (error2) { 1553191783Srmacklem mtx_unlock(&ncl_iod_mutex); 1554191783Srmacklem return (error2); 1555191783Srmacklem } 1556191783Srmacklem if (slpflag == PCATCH) { 1557191783Srmacklem slpflag = 0; 1558191783Srmacklem slptimeo = 2 * hz; 1559191783Srmacklem } 1560191783Srmacklem } 1561191783Srmacklem /* 1562191783Srmacklem * We might have lost our iod while sleeping, 1563191783Srmacklem * so check and loop if nescessary. 1564191783Srmacklem */ 1565191783Srmacklem if (nmp->nm_bufqiods == 0) { 1566191783Srmacklem NFS_DPF(ASYNCIO, 1567191783Srmacklem ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1568191783Srmacklem goto again; 1569191783Srmacklem } 1570191783Srmacklem } 1571191783Srmacklem 1572191783Srmacklem /* We might have lost our nfsiod */ 1573191783Srmacklem if (nmp->nm_bufqiods == 0) { 1574191783Srmacklem NFS_DPF(ASYNCIO, 1575191783Srmacklem ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1576191783Srmacklem goto again; 1577191783Srmacklem } 1578191783Srmacklem 1579191783Srmacklem if (bp->b_iocmd == BIO_READ) { 1580191783Srmacklem if (bp->b_rcred == NOCRED && cred != NOCRED) 1581191783Srmacklem bp->b_rcred = crhold(cred); 1582191783Srmacklem } else { 1583191783Srmacklem if (bp->b_wcred == NOCRED && cred != NOCRED) 1584191783Srmacklem bp->b_wcred = crhold(cred); 1585191783Srmacklem } 1586191783Srmacklem 1587191783Srmacklem if (bp->b_flags & B_REMFREE) 1588191783Srmacklem bremfreef(bp); 1589191783Srmacklem BUF_KERNPROC(bp); 1590191783Srmacklem TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1591191783Srmacklem nmp->nm_bufqlen++; 1592191783Srmacklem if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1593191783Srmacklem mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1594191783Srmacklem VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1595191783Srmacklem VTONFS(bp->b_vp)->n_directio_asyncwr++; 1596191783Srmacklem mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1597191783Srmacklem } 1598191783Srmacklem mtx_unlock(&ncl_iod_mutex); 1599191783Srmacklem return (0); 1600191783Srmacklem } 1601191783Srmacklem 1602191783Srmacklem mtx_unlock(&ncl_iod_mutex); 1603191783Srmacklem 1604191783Srmacklem /* 1605191783Srmacklem * All the iods are busy on other mounts, so return EIO to 1606191783Srmacklem * force the caller to process the i/o synchronously. 1607191783Srmacklem */ 1608191783Srmacklem NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n")); 1609191783Srmacklem return (EIO); 1610191783Srmacklem} 1611191783Srmacklem 1612191783Srmacklemvoid 1613191783Srmacklemncl_doio_directwrite(struct buf *bp) 1614191783Srmacklem{ 1615191783Srmacklem int iomode, must_commit; 1616191783Srmacklem struct uio *uiop = (struct uio *)bp->b_caller1; 1617191783Srmacklem char *iov_base = uiop->uio_iov->iov_base; 1618191783Srmacklem 1619191783Srmacklem iomode = NFSWRITE_FILESYNC; 1620191783Srmacklem uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1621191783Srmacklem ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit); 1622191783Srmacklem KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write")); 1623191783Srmacklem free(iov_base, M_NFSDIRECTIO); 1624191783Srmacklem free(uiop->uio_iov, M_NFSDIRECTIO); 1625191783Srmacklem free(uiop, M_NFSDIRECTIO); 1626191783Srmacklem if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1627191783Srmacklem struct nfsnode *np = VTONFS(bp->b_vp); 1628191783Srmacklem mtx_lock(&np->n_mtx); 1629191783Srmacklem np->n_directio_asyncwr--; 1630191783Srmacklem if (np->n_directio_asyncwr == 0) { 1631191783Srmacklem np->n_flag &= ~NMODIFIED; 1632191783Srmacklem if ((np->n_flag & NFSYNCWAIT)) { 1633191783Srmacklem np->n_flag &= ~NFSYNCWAIT; 1634191783Srmacklem wakeup((caddr_t)&np->n_directio_asyncwr); 1635191783Srmacklem } 1636191783Srmacklem } 1637191783Srmacklem mtx_unlock(&np->n_mtx); 1638191783Srmacklem } 1639191783Srmacklem bp->b_vp = NULL; 1640191783Srmacklem relpbuf(bp, &ncl_pbuf_freecnt); 1641191783Srmacklem} 1642191783Srmacklem 1643191783Srmacklem/* 1644191783Srmacklem * Do an I/O operation to/from a cache block. This may be called 1645191783Srmacklem * synchronously or from an nfsiod. 1646191783Srmacklem */ 1647191783Srmacklemint 1648191783Srmacklemncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td) 1649191783Srmacklem{ 1650191783Srmacklem struct uio *uiop; 1651191783Srmacklem struct nfsnode *np; 1652191783Srmacklem struct nfsmount *nmp; 1653191783Srmacklem int error = 0, iomode, must_commit = 0; 1654191783Srmacklem struct uio uio; 1655191783Srmacklem struct iovec io; 1656191783Srmacklem struct proc *p = td ? td->td_proc : NULL; 1657191783Srmacklem uint8_t iocmd; 1658191783Srmacklem 1659191783Srmacklem np = VTONFS(vp); 1660191783Srmacklem nmp = VFSTONFS(vp->v_mount); 1661191783Srmacklem uiop = &uio; 1662191783Srmacklem uiop->uio_iov = &io; 1663191783Srmacklem uiop->uio_iovcnt = 1; 1664191783Srmacklem uiop->uio_segflg = UIO_SYSSPACE; 1665191783Srmacklem uiop->uio_td = td; 1666191783Srmacklem 1667191783Srmacklem /* 1668191783Srmacklem * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1669191783Srmacklem * do this here so we do not have to do it in all the code that 1670191783Srmacklem * calls us. 1671191783Srmacklem */ 1672191783Srmacklem bp->b_flags &= ~B_INVAL; 1673191783Srmacklem bp->b_ioflags &= ~BIO_ERROR; 1674191783Srmacklem 1675191783Srmacklem KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp)); 1676191783Srmacklem iocmd = bp->b_iocmd; 1677191783Srmacklem if (iocmd == BIO_READ) { 1678191783Srmacklem io.iov_len = uiop->uio_resid = bp->b_bcount; 1679191783Srmacklem io.iov_base = bp->b_data; 1680191783Srmacklem uiop->uio_rw = UIO_READ; 1681191783Srmacklem 1682191783Srmacklem switch (vp->v_type) { 1683191783Srmacklem case VREG: 1684191783Srmacklem uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1685191783Srmacklem NFSINCRGLOBAL(newnfsstats.read_bios); 1686191783Srmacklem error = ncl_readrpc(vp, uiop, cr); 1687191783Srmacklem 1688191783Srmacklem if (!error) { 1689191783Srmacklem if (uiop->uio_resid) { 1690191783Srmacklem /* 1691191783Srmacklem * If we had a short read with no error, we must have 1692191783Srmacklem * hit a file hole. We should zero-fill the remainder. 1693191783Srmacklem * This can also occur if the server hits the file EOF. 1694191783Srmacklem * 1695191783Srmacklem * Holes used to be able to occur due to pending 1696191783Srmacklem * writes, but that is not possible any longer. 1697191783Srmacklem */ 1698191783Srmacklem int nread = bp->b_bcount - uiop->uio_resid; 1699191783Srmacklem int left = uiop->uio_resid; 1700191783Srmacklem 1701191783Srmacklem if (left > 0) 1702191783Srmacklem bzero((char *)bp->b_data + nread, left); 1703191783Srmacklem uiop->uio_resid = 0; 1704191783Srmacklem } 1705191783Srmacklem } 1706191783Srmacklem /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */ 1707191783Srmacklem if (p && (vp->v_vflag & VV_TEXT)) { 1708191783Srmacklem mtx_lock(&np->n_mtx); 1709191783Srmacklem if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) { 1710191783Srmacklem mtx_unlock(&np->n_mtx); 1711191783Srmacklem PROC_LOCK(p); 1712191783Srmacklem killproc(p, "text file modification"); 1713191783Srmacklem PROC_UNLOCK(p); 1714191783Srmacklem } else 1715191783Srmacklem mtx_unlock(&np->n_mtx); 1716191783Srmacklem } 1717191783Srmacklem break; 1718191783Srmacklem case VLNK: 1719191783Srmacklem uiop->uio_offset = (off_t)0; 1720191783Srmacklem NFSINCRGLOBAL(newnfsstats.readlink_bios); 1721191783Srmacklem error = ncl_readlinkrpc(vp, uiop, cr); 1722191783Srmacklem break; 1723191783Srmacklem case VDIR: 1724191783Srmacklem NFSINCRGLOBAL(newnfsstats.readdir_bios); 1725191783Srmacklem uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1726191783Srmacklem if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1727191783Srmacklem error = ncl_readdirplusrpc(vp, uiop, cr, td); 1728191783Srmacklem if (error == NFSERR_NOTSUPP) 1729191783Srmacklem nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1730191783Srmacklem } 1731191783Srmacklem if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1732191783Srmacklem error = ncl_readdirrpc(vp, uiop, cr, td); 1733191783Srmacklem /* 1734191783Srmacklem * end-of-directory sets B_INVAL but does not generate an 1735191783Srmacklem * error. 1736191783Srmacklem */ 1737191783Srmacklem if (error == 0 && uiop->uio_resid == bp->b_bcount) 1738191783Srmacklem bp->b_flags |= B_INVAL; 1739191783Srmacklem break; 1740191783Srmacklem default: 1741191783Srmacklem ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type); 1742191783Srmacklem break; 1743191783Srmacklem }; 1744191783Srmacklem if (error) { 1745191783Srmacklem bp->b_ioflags |= BIO_ERROR; 1746191783Srmacklem bp->b_error = error; 1747191783Srmacklem } 1748191783Srmacklem } else { 1749191783Srmacklem /* 1750191783Srmacklem * If we only need to commit, try to commit 1751191783Srmacklem */ 1752191783Srmacklem if (bp->b_flags & B_NEEDCOMMIT) { 1753191783Srmacklem int retv; 1754191783Srmacklem off_t off; 1755191783Srmacklem 1756191783Srmacklem off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1757191783Srmacklem retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1758191783Srmacklem bp->b_wcred, td); 1759191783Srmacklem if (retv == 0) { 1760191783Srmacklem bp->b_dirtyoff = bp->b_dirtyend = 0; 1761191783Srmacklem bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1762191783Srmacklem bp->b_resid = 0; 1763191783Srmacklem bufdone(bp); 1764191783Srmacklem return (0); 1765191783Srmacklem } 1766191783Srmacklem if (retv == NFSERR_STALEWRITEVERF) { 1767191783Srmacklem ncl_clearcommit(vp->v_mount); 1768191783Srmacklem } 1769191783Srmacklem } 1770191783Srmacklem 1771191783Srmacklem /* 1772191783Srmacklem * Setup for actual write 1773191783Srmacklem */ 1774191783Srmacklem mtx_lock(&np->n_mtx); 1775191783Srmacklem if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1776191783Srmacklem bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1777191783Srmacklem mtx_unlock(&np->n_mtx); 1778191783Srmacklem 1779191783Srmacklem if (bp->b_dirtyend > bp->b_dirtyoff) { 1780191783Srmacklem io.iov_len = uiop->uio_resid = bp->b_dirtyend 1781191783Srmacklem - bp->b_dirtyoff; 1782191783Srmacklem uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1783191783Srmacklem + bp->b_dirtyoff; 1784191783Srmacklem io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1785191783Srmacklem uiop->uio_rw = UIO_WRITE; 1786191783Srmacklem NFSINCRGLOBAL(newnfsstats.write_bios); 1787191783Srmacklem 1788191783Srmacklem if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1789191783Srmacklem iomode = NFSWRITE_UNSTABLE; 1790191783Srmacklem else 1791191783Srmacklem iomode = NFSWRITE_FILESYNC; 1792191783Srmacklem 1793191783Srmacklem error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit); 1794191783Srmacklem 1795191783Srmacklem /* 1796191783Srmacklem * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1797191783Srmacklem * to cluster the buffers needing commit. This will allow 1798191783Srmacklem * the system to submit a single commit rpc for the whole 1799191783Srmacklem * cluster. We can do this even if the buffer is not 100% 1800191783Srmacklem * dirty (relative to the NFS blocksize), so we optimize the 1801191783Srmacklem * append-to-file-case. 1802191783Srmacklem * 1803191783Srmacklem * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1804191783Srmacklem * cleared because write clustering only works for commit 1805191783Srmacklem * rpc's, not for the data portion of the write). 1806191783Srmacklem */ 1807191783Srmacklem 1808191783Srmacklem if (!error && iomode == NFSWRITE_UNSTABLE) { 1809191783Srmacklem bp->b_flags |= B_NEEDCOMMIT; 1810191783Srmacklem if (bp->b_dirtyoff == 0 1811191783Srmacklem && bp->b_dirtyend == bp->b_bcount) 1812191783Srmacklem bp->b_flags |= B_CLUSTEROK; 1813191783Srmacklem } else { 1814191783Srmacklem bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1815191783Srmacklem } 1816191783Srmacklem 1817191783Srmacklem /* 1818191783Srmacklem * For an interrupted write, the buffer is still valid 1819191783Srmacklem * and the write hasn't been pushed to the server yet, 1820191783Srmacklem * so we can't set BIO_ERROR and report the interruption 1821191783Srmacklem * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1822191783Srmacklem * is not relevant, so the rpc attempt is essentially 1823191783Srmacklem * a noop. For the case of a V3 write rpc not being 1824191783Srmacklem * committed to stable storage, the block is still 1825191783Srmacklem * dirty and requires either a commit rpc or another 1826191783Srmacklem * write rpc with iomode == NFSV3WRITE_FILESYNC before 1827191783Srmacklem * the block is reused. This is indicated by setting 1828191783Srmacklem * the B_DELWRI and B_NEEDCOMMIT flags. 1829191783Srmacklem * 1830191783Srmacklem * If the buffer is marked B_PAGING, it does not reside on 1831191783Srmacklem * the vp's paging queues so we cannot call bdirty(). The 1832191783Srmacklem * bp in this case is not an NFS cache block so we should 1833191783Srmacklem * be safe. XXX 1834191783Srmacklem * 1835191783Srmacklem * The logic below breaks up errors into recoverable and 1836191783Srmacklem * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1837191783Srmacklem * and keep the buffer around for potential write retries. 1838191783Srmacklem * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1839191783Srmacklem * and save the error in the nfsnode. This is less than ideal 1840191783Srmacklem * but necessary. Keeping such buffers around could potentially 1841191783Srmacklem * cause buffer exhaustion eventually (they can never be written 1842191783Srmacklem * out, so will get constantly be re-dirtied). It also causes 1843191783Srmacklem * all sorts of vfs panics. For non-recoverable write errors, 1844191783Srmacklem * also invalidate the attrcache, so we'll be forced to go over 1845191783Srmacklem * the wire for this object, returning an error to user on next 1846191783Srmacklem * call (most of the time). 1847191783Srmacklem */ 1848191783Srmacklem if (error == EINTR || error == EIO || error == ETIMEDOUT 1849191783Srmacklem || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1850191783Srmacklem int s; 1851191783Srmacklem 1852191783Srmacklem s = splbio(); 1853191783Srmacklem bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1854191783Srmacklem if ((bp->b_flags & B_PAGING) == 0) { 1855191783Srmacklem bdirty(bp); 1856191783Srmacklem bp->b_flags &= ~B_DONE; 1857191783Srmacklem } 1858191783Srmacklem if (error && (bp->b_flags & B_ASYNC) == 0) 1859191783Srmacklem bp->b_flags |= B_EINTR; 1860191783Srmacklem splx(s); 1861191783Srmacklem } else { 1862191783Srmacklem if (error) { 1863191783Srmacklem bp->b_ioflags |= BIO_ERROR; 1864191783Srmacklem bp->b_flags |= B_INVAL; 1865191783Srmacklem bp->b_error = np->n_error = error; 1866191783Srmacklem mtx_lock(&np->n_mtx); 1867191783Srmacklem np->n_flag |= NWRITEERR; 1868191783Srmacklem np->n_attrstamp = 0; 1869191783Srmacklem mtx_unlock(&np->n_mtx); 1870191783Srmacklem } 1871191783Srmacklem bp->b_dirtyoff = bp->b_dirtyend = 0; 1872191783Srmacklem } 1873191783Srmacklem } else { 1874191783Srmacklem bp->b_resid = 0; 1875191783Srmacklem bufdone(bp); 1876191783Srmacklem return (0); 1877191783Srmacklem } 1878191783Srmacklem } 1879191783Srmacklem bp->b_resid = uiop->uio_resid; 1880191783Srmacklem if (must_commit) 1881191783Srmacklem ncl_clearcommit(vp->v_mount); 1882191783Srmacklem bufdone(bp); 1883191783Srmacklem return (error); 1884191783Srmacklem} 1885191783Srmacklem 1886191783Srmacklem/* 1887191783Srmacklem * Used to aid in handling ftruncate() operations on the NFS client side. 1888191783Srmacklem * Truncation creates a number of special problems for NFS. We have to 1889191783Srmacklem * throw away VM pages and buffer cache buffers that are beyond EOF, and 1890191783Srmacklem * we have to properly handle VM pages or (potentially dirty) buffers 1891191783Srmacklem * that straddle the truncation point. 1892191783Srmacklem */ 1893191783Srmacklem 1894191783Srmacklemint 1895191783Srmacklemncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1896191783Srmacklem{ 1897191783Srmacklem struct nfsnode *np = VTONFS(vp); 1898191783Srmacklem u_quad_t tsize; 1899191783Srmacklem int biosize = vp->v_mount->mnt_stat.f_iosize; 1900191783Srmacklem int error = 0; 1901191783Srmacklem 1902191783Srmacklem mtx_lock(&np->n_mtx); 1903191783Srmacklem tsize = np->n_size; 1904191783Srmacklem np->n_size = nsize; 1905191783Srmacklem mtx_unlock(&np->n_mtx); 1906191783Srmacklem 1907191783Srmacklem if (nsize < tsize) { 1908191783Srmacklem struct buf *bp; 1909191783Srmacklem daddr_t lbn; 1910191783Srmacklem int bufsize; 1911191783Srmacklem 1912191783Srmacklem /* 1913191783Srmacklem * vtruncbuf() doesn't get the buffer overlapping the 1914191783Srmacklem * truncation point. We may have a B_DELWRI and/or B_CACHE 1915191783Srmacklem * buffer that now needs to be truncated. 1916191783Srmacklem */ 1917191783Srmacklem error = vtruncbuf(vp, cred, td, nsize, biosize); 1918191783Srmacklem lbn = nsize / biosize; 1919191783Srmacklem bufsize = nsize & (biosize - 1); 1920191783Srmacklem bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1921191783Srmacklem if (!bp) 1922191783Srmacklem return EINTR; 1923191783Srmacklem if (bp->b_dirtyoff > bp->b_bcount) 1924191783Srmacklem bp->b_dirtyoff = bp->b_bcount; 1925191783Srmacklem if (bp->b_dirtyend > bp->b_bcount) 1926191783Srmacklem bp->b_dirtyend = bp->b_bcount; 1927191783Srmacklem bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1928191783Srmacklem brelse(bp); 1929191783Srmacklem } else { 1930191783Srmacklem vnode_pager_setsize(vp, nsize); 1931191783Srmacklem } 1932191783Srmacklem return(error); 1933191783Srmacklem} 1934191783Srmacklem 1935