nfs_clbio.c revision 260170
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: stable/9/sys/fs/nfsclient/nfs_clbio.c 260170 2014-01-01 22:14:49Z rmacklem $"); 37191783Srmacklem 38223280Srmacklem#include "opt_kdtrace.h" 39223280Srmacklem 40191783Srmacklem#include <sys/param.h> 41191783Srmacklem#include <sys/systm.h> 42191783Srmacklem#include <sys/bio.h> 43191783Srmacklem#include <sys/buf.h> 44191783Srmacklem#include <sys/kernel.h> 45191783Srmacklem#include <sys/mount.h> 46191783Srmacklem#include <sys/vmmeter.h> 47191783Srmacklem#include <sys/vnode.h> 48191783Srmacklem 49191783Srmacklem#include <vm/vm.h> 50240238Skib#include <vm/vm_param.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> 61223280Srmacklem#include <fs/nfsclient/nfs_kdtrace.h> 62191783Srmacklem 63191783Srmacklemextern int newnfs_directio_allow_mmap; 64191783Srmacklemextern struct nfsstats newnfsstats; 65191783Srmacklemextern struct mtx ncl_iod_mutex; 66191783Srmacklemextern int ncl_numasync; 67220683Srmacklemextern enum nfsiod_state ncl_iodwant[NFS_MAXASYNCDAEMON]; 68220683Srmacklemextern struct nfsmount *ncl_iodmount[NFS_MAXASYNCDAEMON]; 69191783Srmacklemextern int newnfs_directio_enable; 70233730Skibextern int nfs_keep_dirty_on_error; 71191783Srmacklem 72191783Srmacklemint ncl_pbuf_freecnt = -1; /* start out unlimited */ 73191783Srmacklem 74191783Srmacklemstatic struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 75191783Srmacklem struct thread *td); 76249077Skibstatic int nfs_directio_write(struct vnode *vp, struct uio *uiop, 77191783Srmacklem struct ucred *cred, int ioflag); 78191783Srmacklem 79191783Srmacklem/* 80191783Srmacklem * Vnode op for VM getpages. 81191783Srmacklem */ 82191783Srmacklemint 83191783Srmacklemncl_getpages(struct vop_getpages_args *ap) 84191783Srmacklem{ 85191783Srmacklem int i, error, nextoff, size, toff, count, npages; 86191783Srmacklem struct uio uio; 87191783Srmacklem struct iovec iov; 88191783Srmacklem vm_offset_t kva; 89191783Srmacklem struct buf *bp; 90191783Srmacklem struct vnode *vp; 91191783Srmacklem struct thread *td; 92191783Srmacklem struct ucred *cred; 93191783Srmacklem struct nfsmount *nmp; 94191783Srmacklem vm_object_t object; 95191783Srmacklem vm_page_t *pages; 96191783Srmacklem struct nfsnode *np; 97191783Srmacklem 98191783Srmacklem vp = ap->a_vp; 99191783Srmacklem np = VTONFS(vp); 100191783Srmacklem td = curthread; /* XXX */ 101191783Srmacklem cred = curthread->td_ucred; /* XXX */ 102191783Srmacklem nmp = VFSTONFS(vp->v_mount); 103191783Srmacklem pages = ap->a_m; 104191783Srmacklem count = ap->a_count; 105191783Srmacklem 106191783Srmacklem if ((object = vp->v_object) == NULL) { 107191783Srmacklem ncl_printf("nfs_getpages: called with non-merged cache vnode??\n"); 108194425Salc return (VM_PAGER_ERROR); 109191783Srmacklem } 110191783Srmacklem 111191783Srmacklem if (newnfs_directio_enable && !newnfs_directio_allow_mmap) { 112191783Srmacklem mtx_lock(&np->n_mtx); 113191783Srmacklem if ((np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 114191783Srmacklem mtx_unlock(&np->n_mtx); 115191783Srmacklem ncl_printf("nfs_getpages: called on non-cacheable vnode??\n"); 116194425Salc return (VM_PAGER_ERROR); 117191783Srmacklem } else 118191783Srmacklem mtx_unlock(&np->n_mtx); 119191783Srmacklem } 120191783Srmacklem 121191783Srmacklem mtx_lock(&nmp->nm_mtx); 122191783Srmacklem if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 123249077Skib (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 124191783Srmacklem mtx_unlock(&nmp->nm_mtx); 125191783Srmacklem /* We'll never get here for v4, because we always have fsinfo */ 126191783Srmacklem (void)ncl_fsinfo(nmp, vp, cred, td); 127191783Srmacklem } else 128191783Srmacklem mtx_unlock(&nmp->nm_mtx); 129191783Srmacklem 130191783Srmacklem npages = btoc(count); 131191783Srmacklem 132191783Srmacklem /* 133191783Srmacklem * If the requested page is partially valid, just return it and 134191783Srmacklem * allow the pager to zero-out the blanks. Partially valid pages 135191783Srmacklem * can only occur at the file EOF. 136191783Srmacklem */ 137194425Salc VM_OBJECT_LOCK(object); 138194425Salc if (pages[ap->a_reqpage]->valid != 0) { 139194425Salc for (i = 0; i < npages; ++i) { 140207669Salc if (i != ap->a_reqpage) { 141207669Salc vm_page_lock(pages[i]); 142194425Salc vm_page_free(pages[i]); 143207669Salc vm_page_unlock(pages[i]); 144207669Salc } 145191783Srmacklem } 146191783Srmacklem VM_OBJECT_UNLOCK(object); 147194425Salc return (0); 148191783Srmacklem } 149194425Salc VM_OBJECT_UNLOCK(object); 150191783Srmacklem 151191783Srmacklem /* 152191783Srmacklem * We use only the kva address for the buffer, but this is extremely 153191783Srmacklem * convienient and fast. 154191783Srmacklem */ 155191783Srmacklem bp = getpbuf(&ncl_pbuf_freecnt); 156191783Srmacklem 157191783Srmacklem kva = (vm_offset_t) bp->b_data; 158191783Srmacklem pmap_qenter(kva, pages, npages); 159191783Srmacklem PCPU_INC(cnt.v_vnodein); 160191783Srmacklem PCPU_ADD(cnt.v_vnodepgsin, npages); 161191783Srmacklem 162191783Srmacklem iov.iov_base = (caddr_t) kva; 163191783Srmacklem iov.iov_len = count; 164191783Srmacklem uio.uio_iov = &iov; 165191783Srmacklem uio.uio_iovcnt = 1; 166191783Srmacklem uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 167191783Srmacklem uio.uio_resid = count; 168191783Srmacklem uio.uio_segflg = UIO_SYSSPACE; 169191783Srmacklem uio.uio_rw = UIO_READ; 170191783Srmacklem uio.uio_td = td; 171191783Srmacklem 172191783Srmacklem error = ncl_readrpc(vp, &uio, cred); 173191783Srmacklem pmap_qremove(kva, npages); 174191783Srmacklem 175191783Srmacklem relpbuf(bp, &ncl_pbuf_freecnt); 176191783Srmacklem 177191783Srmacklem if (error && (uio.uio_resid == count)) { 178191783Srmacklem ncl_printf("nfs_getpages: error %d\n", error); 179191783Srmacklem VM_OBJECT_LOCK(object); 180191783Srmacklem for (i = 0; i < npages; ++i) { 181207669Salc if (i != ap->a_reqpage) { 182207669Salc vm_page_lock(pages[i]); 183191783Srmacklem vm_page_free(pages[i]); 184207669Salc vm_page_unlock(pages[i]); 185207669Salc } 186191783Srmacklem } 187191783Srmacklem VM_OBJECT_UNLOCK(object); 188194425Salc return (VM_PAGER_ERROR); 189191783Srmacklem } 190191783Srmacklem 191191783Srmacklem /* 192191783Srmacklem * Calculate the number of bytes read and validate only that number 193191783Srmacklem * of bytes. Note that due to pending writes, size may be 0. This 194191783Srmacklem * does not mean that the remaining data is invalid! 195191783Srmacklem */ 196191783Srmacklem 197191783Srmacklem size = count - uio.uio_resid; 198191783Srmacklem VM_OBJECT_LOCK(object); 199191783Srmacklem for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 200191783Srmacklem vm_page_t m; 201191783Srmacklem nextoff = toff + PAGE_SIZE; 202191783Srmacklem m = pages[i]; 203191783Srmacklem 204191783Srmacklem if (nextoff <= size) { 205191783Srmacklem /* 206191783Srmacklem * Read operation filled an entire page 207191783Srmacklem */ 208191783Srmacklem m->valid = VM_PAGE_BITS_ALL; 209192065Srmacklem KASSERT(m->dirty == 0, 210192065Srmacklem ("nfs_getpages: page %p is dirty", m)); 211191783Srmacklem } else if (size > toff) { 212191783Srmacklem /* 213191783Srmacklem * Read operation filled a partial page. 214191783Srmacklem */ 215191783Srmacklem m->valid = 0; 216192231Srmacklem vm_page_set_valid(m, 0, size - toff); 217192986Salc KASSERT(m->dirty == 0, 218192231Srmacklem ("nfs_getpages: page %p is dirty", m)); 219191783Srmacklem } else { 220191783Srmacklem /* 221239554Skib * Read operation was short. If no error 222239554Skib * occured we may have hit a zero-fill 223239554Skib * section. We leave valid set to 0, and page 224239554Skib * is freed by vm_page_readahead_finish() if 225239554Skib * its index is not equal to requested, or 226239554Skib * page is zeroed and set valid by 227239554Skib * vm_pager_get_pages() for requested page. 228191783Srmacklem */ 229191783Srmacklem ; 230191783Srmacklem } 231239554Skib if (i != ap->a_reqpage) 232239554Skib vm_page_readahead_finish(m); 233191783Srmacklem } 234191783Srmacklem VM_OBJECT_UNLOCK(object); 235194425Salc return (0); 236191783Srmacklem} 237191783Srmacklem 238191783Srmacklem/* 239191783Srmacklem * Vnode op for VM putpages. 240191783Srmacklem */ 241191783Srmacklemint 242191783Srmacklemncl_putpages(struct vop_putpages_args *ap) 243191783Srmacklem{ 244191783Srmacklem struct uio uio; 245191783Srmacklem struct iovec iov; 246191783Srmacklem vm_offset_t kva; 247191783Srmacklem struct buf *bp; 248191783Srmacklem int iomode, must_commit, i, error, npages, count; 249191783Srmacklem off_t offset; 250191783Srmacklem int *rtvals; 251191783Srmacklem struct vnode *vp; 252191783Srmacklem struct thread *td; 253191783Srmacklem struct ucred *cred; 254191783Srmacklem struct nfsmount *nmp; 255191783Srmacklem struct nfsnode *np; 256191783Srmacklem vm_page_t *pages; 257191783Srmacklem 258191783Srmacklem vp = ap->a_vp; 259191783Srmacklem np = VTONFS(vp); 260191783Srmacklem td = curthread; /* XXX */ 261236096Srmacklem /* Set the cred to n_writecred for the write rpcs. */ 262236096Srmacklem if (np->n_writecred != NULL) 263236096Srmacklem cred = crhold(np->n_writecred); 264236096Srmacklem else 265236096Srmacklem cred = crhold(curthread->td_ucred); /* XXX */ 266191783Srmacklem nmp = VFSTONFS(vp->v_mount); 267191783Srmacklem pages = ap->a_m; 268191783Srmacklem count = ap->a_count; 269191783Srmacklem rtvals = ap->a_rtvals; 270191783Srmacklem npages = btoc(count); 271191783Srmacklem offset = IDX_TO_OFF(pages[0]->pindex); 272249077Skib 273191783Srmacklem mtx_lock(&nmp->nm_mtx); 274191783Srmacklem if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 275191783Srmacklem (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 276191783Srmacklem mtx_unlock(&nmp->nm_mtx); 277191783Srmacklem (void)ncl_fsinfo(nmp, vp, cred, td); 278191783Srmacklem } else 279191783Srmacklem mtx_unlock(&nmp->nm_mtx); 280191783Srmacklem 281191783Srmacklem mtx_lock(&np->n_mtx); 282249077Skib if (newnfs_directio_enable && !newnfs_directio_allow_mmap && 283191783Srmacklem (np->n_flag & NNONCACHE) && (vp->v_type == VREG)) { 284249077Skib mtx_unlock(&np->n_mtx); 285191783Srmacklem ncl_printf("ncl_putpages: called on noncache-able vnode??\n"); 286191783Srmacklem mtx_lock(&np->n_mtx); 287191783Srmacklem } 288191783Srmacklem 289191783Srmacklem for (i = 0; i < npages; i++) 290222586Skib rtvals[i] = VM_PAGER_ERROR; 291191783Srmacklem 292191783Srmacklem /* 293191783Srmacklem * When putting pages, do not extend file past EOF. 294191783Srmacklem */ 295191783Srmacklem if (offset + count > np->n_size) { 296191783Srmacklem count = np->n_size - offset; 297191783Srmacklem if (count < 0) 298191783Srmacklem count = 0; 299191783Srmacklem } 300191783Srmacklem mtx_unlock(&np->n_mtx); 301191783Srmacklem 302191783Srmacklem /* 303191783Srmacklem * We use only the kva address for the buffer, but this is extremely 304191783Srmacklem * convienient and fast. 305191783Srmacklem */ 306191783Srmacklem bp = getpbuf(&ncl_pbuf_freecnt); 307191783Srmacklem 308191783Srmacklem kva = (vm_offset_t) bp->b_data; 309191783Srmacklem pmap_qenter(kva, pages, npages); 310191783Srmacklem PCPU_INC(cnt.v_vnodeout); 311191783Srmacklem PCPU_ADD(cnt.v_vnodepgsout, count); 312191783Srmacklem 313191783Srmacklem iov.iov_base = (caddr_t) kva; 314191783Srmacklem iov.iov_len = count; 315191783Srmacklem uio.uio_iov = &iov; 316191783Srmacklem uio.uio_iovcnt = 1; 317191783Srmacklem uio.uio_offset = offset; 318191783Srmacklem uio.uio_resid = count; 319191783Srmacklem uio.uio_segflg = UIO_SYSSPACE; 320191783Srmacklem uio.uio_rw = UIO_WRITE; 321191783Srmacklem uio.uio_td = td; 322191783Srmacklem 323191783Srmacklem if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 324191783Srmacklem iomode = NFSWRITE_UNSTABLE; 325191783Srmacklem else 326191783Srmacklem iomode = NFSWRITE_FILESYNC; 327191783Srmacklem 328207082Srmacklem error = ncl_writerpc(vp, &uio, cred, &iomode, &must_commit, 0); 329236096Srmacklem crfree(cred); 330191783Srmacklem 331191783Srmacklem pmap_qremove(kva, npages); 332191783Srmacklem relpbuf(bp, &ncl_pbuf_freecnt); 333191783Srmacklem 334233730Skib if (error == 0 || !nfs_keep_dirty_on_error) { 335233730Skib vnode_pager_undirty_pages(pages, rtvals, count - uio.uio_resid); 336233730Skib if (must_commit) 337233730Skib ncl_clearcommit(vp->v_mount); 338233730Skib } 339191783Srmacklem return rtvals[0]; 340191783Srmacklem} 341191783Srmacklem 342191783Srmacklem/* 343191783Srmacklem * For nfs, cache consistency can only be maintained approximately. 344191783Srmacklem * Although RFC1094 does not specify the criteria, the following is 345191783Srmacklem * believed to be compatible with the reference port. 346191783Srmacklem * For nfs: 347191783Srmacklem * If the file's modify time on the server has changed since the 348191783Srmacklem * last read rpc or you have written to the file, 349191783Srmacklem * you may have lost data cache consistency with the 350191783Srmacklem * server, so flush all of the file's data out of the cache. 351191783Srmacklem * Then force a getattr rpc to ensure that you have up to date 352191783Srmacklem * attributes. 353191783Srmacklem * NB: This implies that cache data can be read when up to 354191783Srmacklem * NFS_ATTRTIMEO seconds out of date. If you find that you need current 355191783Srmacklem * attributes this could be forced by setting n_attrstamp to 0 before 356191783Srmacklem * the VOP_GETATTR() call. 357191783Srmacklem */ 358191783Srmacklemstatic inline int 359191783Srmacklemnfs_bioread_check_cons(struct vnode *vp, struct thread *td, struct ucred *cred) 360191783Srmacklem{ 361191783Srmacklem int error = 0; 362191783Srmacklem struct vattr vattr; 363191783Srmacklem struct nfsnode *np = VTONFS(vp); 364191783Srmacklem int old_lock; 365249077Skib 366191783Srmacklem /* 367191783Srmacklem * Grab the exclusive lock before checking whether the cache is 368191783Srmacklem * consistent. 369191783Srmacklem * XXX - We can make this cheaper later (by acquiring cheaper locks). 370191783Srmacklem * But for now, this suffices. 371191783Srmacklem */ 372191783Srmacklem old_lock = ncl_upgrade_vnlock(vp); 373193955Srmacklem if (vp->v_iflag & VI_DOOMED) { 374193955Srmacklem ncl_downgrade_vnlock(vp, old_lock); 375193955Srmacklem return (EBADF); 376193955Srmacklem } 377193955Srmacklem 378191783Srmacklem mtx_lock(&np->n_mtx); 379191783Srmacklem if (np->n_flag & NMODIFIED) { 380191783Srmacklem mtx_unlock(&np->n_mtx); 381191783Srmacklem if (vp->v_type != VREG) { 382191783Srmacklem if (vp->v_type != VDIR) 383191783Srmacklem panic("nfs: bioread, not dir"); 384191783Srmacklem ncl_invaldir(vp); 385191783Srmacklem error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 386191783Srmacklem if (error) 387191783Srmacklem goto out; 388191783Srmacklem } 389191783Srmacklem np->n_attrstamp = 0; 390223280Srmacklem KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 391191783Srmacklem error = VOP_GETATTR(vp, &vattr, cred); 392191783Srmacklem if (error) 393191783Srmacklem goto out; 394191783Srmacklem mtx_lock(&np->n_mtx); 395191783Srmacklem np->n_mtime = vattr.va_mtime; 396191783Srmacklem mtx_unlock(&np->n_mtx); 397191783Srmacklem } else { 398191783Srmacklem mtx_unlock(&np->n_mtx); 399191783Srmacklem error = VOP_GETATTR(vp, &vattr, cred); 400191783Srmacklem if (error) 401191783Srmacklem return (error); 402191783Srmacklem mtx_lock(&np->n_mtx); 403191783Srmacklem if ((np->n_flag & NSIZECHANGED) 404191783Srmacklem || (NFS_TIMESPEC_COMPARE(&np->n_mtime, &vattr.va_mtime))) { 405191783Srmacklem mtx_unlock(&np->n_mtx); 406191783Srmacklem if (vp->v_type == VDIR) 407191783Srmacklem ncl_invaldir(vp); 408191783Srmacklem error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 409191783Srmacklem if (error) 410191783Srmacklem goto out; 411191783Srmacklem mtx_lock(&np->n_mtx); 412191783Srmacklem np->n_mtime = vattr.va_mtime; 413191783Srmacklem np->n_flag &= ~NSIZECHANGED; 414191783Srmacklem } 415191783Srmacklem mtx_unlock(&np->n_mtx); 416191783Srmacklem } 417249077Skibout: 418191783Srmacklem ncl_downgrade_vnlock(vp, old_lock); 419191783Srmacklem return error; 420191783Srmacklem} 421191783Srmacklem 422191783Srmacklem/* 423191783Srmacklem * Vnode op for read using bio 424191783Srmacklem */ 425191783Srmacklemint 426191783Srmacklemncl_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 427191783Srmacklem{ 428191783Srmacklem struct nfsnode *np = VTONFS(vp); 429191783Srmacklem int biosize, i; 430191783Srmacklem struct buf *bp, *rabp; 431191783Srmacklem struct thread *td; 432191783Srmacklem struct nfsmount *nmp = VFSTONFS(vp->v_mount); 433191783Srmacklem daddr_t lbn, rabn; 434191783Srmacklem int bcount; 435191783Srmacklem int seqcount; 436191783Srmacklem int nra, error = 0, n = 0, on = 0; 437220877Srmacklem off_t tmp_off; 438191783Srmacklem 439209120Skib KASSERT(uio->uio_rw == UIO_READ, ("ncl_read mode")); 440191783Srmacklem if (uio->uio_resid == 0) 441191783Srmacklem return (0); 442191783Srmacklem if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 443191783Srmacklem return (EINVAL); 444191783Srmacklem td = uio->uio_td; 445191783Srmacklem 446191783Srmacklem mtx_lock(&nmp->nm_mtx); 447191783Srmacklem if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 448191783Srmacklem (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 449191783Srmacklem mtx_unlock(&nmp->nm_mtx); 450191783Srmacklem (void)ncl_fsinfo(nmp, vp, cred, td); 451191783Srmacklem mtx_lock(&nmp->nm_mtx); 452191783Srmacklem } 453191783Srmacklem if (nmp->nm_rsize == 0 || nmp->nm_readdirsize == 0) 454191783Srmacklem (void) newnfs_iosize(nmp); 455191783Srmacklem 456220877Srmacklem tmp_off = uio->uio_offset + uio->uio_resid; 457191783Srmacklem if (vp->v_type != VDIR && 458220877Srmacklem (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset)) { 459249077Skib mtx_unlock(&nmp->nm_mtx); 460191783Srmacklem return (EFBIG); 461220877Srmacklem } 462249077Skib mtx_unlock(&nmp->nm_mtx); 463191783Srmacklem 464191783Srmacklem if (newnfs_directio_enable && (ioflag & IO_DIRECT) && (vp->v_type == VREG)) 465191783Srmacklem /* No caching/ no readaheads. Just read data into the user buffer */ 466191783Srmacklem return ncl_readrpc(vp, uio, cred); 467191783Srmacklem 468231330Srmacklem biosize = vp->v_bufobj.bo_bsize; 469191783Srmacklem seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 470249077Skib 471191783Srmacklem error = nfs_bioread_check_cons(vp, td, cred); 472191783Srmacklem if (error) 473191783Srmacklem return error; 474191783Srmacklem 475191783Srmacklem do { 476191783Srmacklem u_quad_t nsize; 477249077Skib 478191783Srmacklem mtx_lock(&np->n_mtx); 479191783Srmacklem nsize = np->n_size; 480249077Skib mtx_unlock(&np->n_mtx); 481191783Srmacklem 482191783Srmacklem switch (vp->v_type) { 483191783Srmacklem case VREG: 484191783Srmacklem NFSINCRGLOBAL(newnfsstats.biocache_reads); 485191783Srmacklem lbn = uio->uio_offset / biosize; 486191783Srmacklem on = uio->uio_offset & (biosize - 1); 487191783Srmacklem 488191783Srmacklem /* 489191783Srmacklem * Start the read ahead(s), as required. 490191783Srmacklem */ 491191783Srmacklem if (nmp->nm_readahead > 0) { 492191783Srmacklem for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 493191783Srmacklem (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) { 494191783Srmacklem rabn = lbn + 1 + nra; 495191783Srmacklem if (incore(&vp->v_bufobj, rabn) == NULL) { 496191783Srmacklem rabp = nfs_getcacheblk(vp, rabn, biosize, td); 497191783Srmacklem if (!rabp) { 498191783Srmacklem error = newnfs_sigintr(nmp, td); 499212217Srmacklem return (error ? error : EINTR); 500191783Srmacklem } 501191783Srmacklem if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 502191783Srmacklem rabp->b_flags |= B_ASYNC; 503191783Srmacklem rabp->b_iocmd = BIO_READ; 504191783Srmacklem vfs_busy_pages(rabp, 0); 505191783Srmacklem if (ncl_asyncio(nmp, rabp, cred, td)) { 506191783Srmacklem rabp->b_flags |= B_INVAL; 507191783Srmacklem rabp->b_ioflags |= BIO_ERROR; 508191783Srmacklem vfs_unbusy_pages(rabp); 509191783Srmacklem brelse(rabp); 510191783Srmacklem break; 511191783Srmacklem } 512191783Srmacklem } else { 513191783Srmacklem brelse(rabp); 514191783Srmacklem } 515191783Srmacklem } 516191783Srmacklem } 517191783Srmacklem } 518191783Srmacklem 519191783Srmacklem /* Note that bcount is *not* DEV_BSIZE aligned. */ 520191783Srmacklem bcount = biosize; 521191783Srmacklem if ((off_t)lbn * biosize >= nsize) { 522191783Srmacklem bcount = 0; 523191783Srmacklem } else if ((off_t)(lbn + 1) * biosize > nsize) { 524191783Srmacklem bcount = nsize - (off_t)lbn * biosize; 525191783Srmacklem } 526191783Srmacklem bp = nfs_getcacheblk(vp, lbn, bcount, td); 527191783Srmacklem 528191783Srmacklem if (!bp) { 529191783Srmacklem error = newnfs_sigintr(nmp, td); 530191783Srmacklem return (error ? error : EINTR); 531191783Srmacklem } 532191783Srmacklem 533191783Srmacklem /* 534191783Srmacklem * If B_CACHE is not set, we must issue the read. If this 535191783Srmacklem * fails, we return an error. 536191783Srmacklem */ 537191783Srmacklem 538191783Srmacklem if ((bp->b_flags & B_CACHE) == 0) { 539191783Srmacklem bp->b_iocmd = BIO_READ; 540191783Srmacklem vfs_busy_pages(bp, 0); 541207082Srmacklem error = ncl_doio(vp, bp, cred, td, 0); 542191783Srmacklem if (error) { 543191783Srmacklem brelse(bp); 544191783Srmacklem return (error); 545191783Srmacklem } 546191783Srmacklem } 547191783Srmacklem 548191783Srmacklem /* 549191783Srmacklem * on is the offset into the current bp. Figure out how many 550191783Srmacklem * bytes we can copy out of the bp. Note that bcount is 551191783Srmacklem * NOT DEV_BSIZE aligned. 552191783Srmacklem * 553191783Srmacklem * Then figure out how many bytes we can copy into the uio. 554191783Srmacklem */ 555191783Srmacklem 556191783Srmacklem n = 0; 557191783Srmacklem if (on < bcount) 558233353Skib n = MIN((unsigned)(bcount - on), uio->uio_resid); 559191783Srmacklem break; 560191783Srmacklem case VLNK: 561191783Srmacklem NFSINCRGLOBAL(newnfsstats.biocache_readlinks); 562191783Srmacklem bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 563191783Srmacklem if (!bp) { 564191783Srmacklem error = newnfs_sigintr(nmp, td); 565191783Srmacklem return (error ? error : EINTR); 566191783Srmacklem } 567191783Srmacklem if ((bp->b_flags & B_CACHE) == 0) { 568191783Srmacklem bp->b_iocmd = BIO_READ; 569191783Srmacklem vfs_busy_pages(bp, 0); 570207082Srmacklem error = ncl_doio(vp, bp, cred, td, 0); 571191783Srmacklem if (error) { 572191783Srmacklem bp->b_ioflags |= BIO_ERROR; 573191783Srmacklem brelse(bp); 574191783Srmacklem return (error); 575191783Srmacklem } 576191783Srmacklem } 577233353Skib n = MIN(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 578191783Srmacklem on = 0; 579191783Srmacklem break; 580191783Srmacklem case VDIR: 581191783Srmacklem NFSINCRGLOBAL(newnfsstats.biocache_readdirs); 582191783Srmacklem if (np->n_direofoffset 583191783Srmacklem && uio->uio_offset >= np->n_direofoffset) { 584191783Srmacklem return (0); 585191783Srmacklem } 586191783Srmacklem lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 587191783Srmacklem on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 588191783Srmacklem bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 589191783Srmacklem if (!bp) { 590191783Srmacklem error = newnfs_sigintr(nmp, td); 591191783Srmacklem return (error ? error : EINTR); 592191783Srmacklem } 593191783Srmacklem if ((bp->b_flags & B_CACHE) == 0) { 594191783Srmacklem bp->b_iocmd = BIO_READ; 595191783Srmacklem vfs_busy_pages(bp, 0); 596207082Srmacklem error = ncl_doio(vp, bp, cred, td, 0); 597191783Srmacklem if (error) { 598191783Srmacklem brelse(bp); 599191783Srmacklem } 600191783Srmacklem while (error == NFSERR_BAD_COOKIE) { 601191783Srmacklem ncl_invaldir(vp); 602191783Srmacklem error = ncl_vinvalbuf(vp, 0, td, 1); 603191783Srmacklem /* 604191783Srmacklem * Yuck! The directory has been modified on the 605191783Srmacklem * server. The only way to get the block is by 606191783Srmacklem * reading from the beginning to get all the 607191783Srmacklem * offset cookies. 608191783Srmacklem * 609191783Srmacklem * Leave the last bp intact unless there is an error. 610191783Srmacklem * Loop back up to the while if the error is another 611191783Srmacklem * NFSERR_BAD_COOKIE (double yuch!). 612191783Srmacklem */ 613191783Srmacklem for (i = 0; i <= lbn && !error; i++) { 614191783Srmacklem if (np->n_direofoffset 615191783Srmacklem && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 616191783Srmacklem return (0); 617191783Srmacklem bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 618191783Srmacklem if (!bp) { 619191783Srmacklem error = newnfs_sigintr(nmp, td); 620191783Srmacklem return (error ? error : EINTR); 621191783Srmacklem } 622191783Srmacklem if ((bp->b_flags & B_CACHE) == 0) { 623191783Srmacklem bp->b_iocmd = BIO_READ; 624191783Srmacklem vfs_busy_pages(bp, 0); 625207082Srmacklem error = ncl_doio(vp, bp, cred, td, 0); 626191783Srmacklem /* 627191783Srmacklem * no error + B_INVAL == directory EOF, 628191783Srmacklem * use the block. 629191783Srmacklem */ 630191783Srmacklem if (error == 0 && (bp->b_flags & B_INVAL)) 631191783Srmacklem break; 632191783Srmacklem } 633191783Srmacklem /* 634191783Srmacklem * An error will throw away the block and the 635191783Srmacklem * for loop will break out. If no error and this 636191783Srmacklem * is not the block we want, we throw away the 637191783Srmacklem * block and go for the next one via the for loop. 638191783Srmacklem */ 639191783Srmacklem if (error || i < lbn) 640191783Srmacklem brelse(bp); 641191783Srmacklem } 642191783Srmacklem } 643191783Srmacklem /* 644191783Srmacklem * The above while is repeated if we hit another cookie 645191783Srmacklem * error. If we hit an error and it wasn't a cookie error, 646191783Srmacklem * we give up. 647191783Srmacklem */ 648191783Srmacklem if (error) 649191783Srmacklem return (error); 650191783Srmacklem } 651191783Srmacklem 652191783Srmacklem /* 653191783Srmacklem * If not eof and read aheads are enabled, start one. 654191783Srmacklem * (You need the current block first, so that you have the 655191783Srmacklem * directory offset cookie of the next block.) 656191783Srmacklem */ 657191783Srmacklem if (nmp->nm_readahead > 0 && 658191783Srmacklem (bp->b_flags & B_INVAL) == 0 && 659191783Srmacklem (np->n_direofoffset == 0 || 660191783Srmacklem (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 661191783Srmacklem incore(&vp->v_bufobj, lbn + 1) == NULL) { 662191783Srmacklem rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 663191783Srmacklem if (rabp) { 664191783Srmacklem if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 665191783Srmacklem rabp->b_flags |= B_ASYNC; 666191783Srmacklem rabp->b_iocmd = BIO_READ; 667191783Srmacklem vfs_busy_pages(rabp, 0); 668191783Srmacklem if (ncl_asyncio(nmp, rabp, cred, td)) { 669191783Srmacklem rabp->b_flags |= B_INVAL; 670191783Srmacklem rabp->b_ioflags |= BIO_ERROR; 671191783Srmacklem vfs_unbusy_pages(rabp); 672191783Srmacklem brelse(rabp); 673191783Srmacklem } 674191783Srmacklem } else { 675191783Srmacklem brelse(rabp); 676191783Srmacklem } 677191783Srmacklem } 678191783Srmacklem } 679191783Srmacklem /* 680191783Srmacklem * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 681191783Srmacklem * chopped for the EOF condition, we cannot tell how large 682191783Srmacklem * NFS directories are going to be until we hit EOF. So 683191783Srmacklem * an NFS directory buffer is *not* chopped to its EOF. Now, 684191783Srmacklem * it just so happens that b_resid will effectively chop it 685191783Srmacklem * to EOF. *BUT* this information is lost if the buffer goes 686191783Srmacklem * away and is reconstituted into a B_CACHE state ( due to 687191783Srmacklem * being VMIO ) later. So we keep track of the directory eof 688191783Srmacklem * in np->n_direofoffset and chop it off as an extra step 689191783Srmacklem * right here. 690191783Srmacklem */ 691191783Srmacklem n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 692191783Srmacklem if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 693191783Srmacklem n = np->n_direofoffset - uio->uio_offset; 694191783Srmacklem break; 695191783Srmacklem default: 696191783Srmacklem ncl_printf(" ncl_bioread: type %x unexpected\n", vp->v_type); 697191783Srmacklem bp = NULL; 698191783Srmacklem break; 699191783Srmacklem }; 700191783Srmacklem 701191783Srmacklem if (n > 0) { 702239852Skib error = vn_io_fault_uiomove(bp->b_data + on, (int)n, uio); 703191783Srmacklem } 704191783Srmacklem if (vp->v_type == VLNK) 705191783Srmacklem n = 0; 706191783Srmacklem if (bp != NULL) 707191783Srmacklem brelse(bp); 708191783Srmacklem } while (error == 0 && uio->uio_resid > 0 && n > 0); 709191783Srmacklem return (error); 710191783Srmacklem} 711191783Srmacklem 712191783Srmacklem/* 713249077Skib * The NFS write path cannot handle iovecs with len > 1. So we need to 714191783Srmacklem * break up iovecs accordingly (restricting them to wsize). 715249077Skib * For the SYNC case, we can do this with 1 copy (user buffer -> mbuf). 716249077Skib * For the ASYNC case, 2 copies are needed. The first a copy from the 717191783Srmacklem * user buffer to a staging buffer and then a second copy from the staging 718191783Srmacklem * buffer to mbufs. This can be optimized by copying from the user buffer 719249077Skib * directly into mbufs and passing the chain down, but that requires a 720191783Srmacklem * fair amount of re-working of the relevant codepaths (and can be done 721191783Srmacklem * later). 722191783Srmacklem */ 723191783Srmacklemstatic int 724191783Srmacklemnfs_directio_write(vp, uiop, cred, ioflag) 725191783Srmacklem struct vnode *vp; 726191783Srmacklem struct uio *uiop; 727191783Srmacklem struct ucred *cred; 728191783Srmacklem int ioflag; 729191783Srmacklem{ 730191783Srmacklem int error; 731191783Srmacklem struct nfsmount *nmp = VFSTONFS(vp->v_mount); 732191783Srmacklem struct thread *td = uiop->uio_td; 733191783Srmacklem int size; 734191783Srmacklem int wsize; 735249077Skib 736191783Srmacklem mtx_lock(&nmp->nm_mtx); 737191783Srmacklem wsize = nmp->nm_wsize; 738191783Srmacklem mtx_unlock(&nmp->nm_mtx); 739191783Srmacklem if (ioflag & IO_SYNC) { 740191783Srmacklem int iomode, must_commit; 741191783Srmacklem struct uio uio; 742191783Srmacklem struct iovec iov; 743191783Srmacklemdo_sync: 744191783Srmacklem while (uiop->uio_resid > 0) { 745233353Skib size = MIN(uiop->uio_resid, wsize); 746233353Skib size = MIN(uiop->uio_iov->iov_len, size); 747191783Srmacklem iov.iov_base = uiop->uio_iov->iov_base; 748191783Srmacklem iov.iov_len = size; 749191783Srmacklem uio.uio_iov = &iov; 750191783Srmacklem uio.uio_iovcnt = 1; 751191783Srmacklem uio.uio_offset = uiop->uio_offset; 752191783Srmacklem uio.uio_resid = size; 753191783Srmacklem uio.uio_segflg = UIO_USERSPACE; 754191783Srmacklem uio.uio_rw = UIO_WRITE; 755191783Srmacklem uio.uio_td = td; 756191783Srmacklem iomode = NFSWRITE_FILESYNC; 757191783Srmacklem error = ncl_writerpc(vp, &uio, cred, &iomode, 758207082Srmacklem &must_commit, 0); 759249077Skib KASSERT((must_commit == 0), 760191783Srmacklem ("ncl_directio_write: Did not commit write")); 761191783Srmacklem if (error) 762191783Srmacklem return (error); 763191783Srmacklem uiop->uio_offset += size; 764191783Srmacklem uiop->uio_resid -= size; 765191783Srmacklem if (uiop->uio_iov->iov_len <= size) { 766191783Srmacklem uiop->uio_iovcnt--; 767191783Srmacklem uiop->uio_iov++; 768191783Srmacklem } else { 769249077Skib uiop->uio_iov->iov_base = 770191783Srmacklem (char *)uiop->uio_iov->iov_base + size; 771191783Srmacklem uiop->uio_iov->iov_len -= size; 772191783Srmacklem } 773191783Srmacklem } 774191783Srmacklem } else { 775191783Srmacklem struct uio *t_uio; 776191783Srmacklem struct iovec *t_iov; 777191783Srmacklem struct buf *bp; 778249077Skib 779191783Srmacklem /* 780191783Srmacklem * Break up the write into blocksize chunks and hand these 781191783Srmacklem * over to nfsiod's for write back. 782249077Skib * Unfortunately, this incurs a copy of the data. Since 783249077Skib * the user could modify the buffer before the write is 784191783Srmacklem * initiated. 785249077Skib * 786191783Srmacklem * The obvious optimization here is that one of the 2 copies 787191783Srmacklem * in the async write path can be eliminated by copying the 788191783Srmacklem * data here directly into mbufs and passing the mbuf chain 789191783Srmacklem * down. But that will require a fair amount of re-working 790191783Srmacklem * of the code and can be done if there's enough interest 791191783Srmacklem * in NFS directio access. 792191783Srmacklem */ 793191783Srmacklem while (uiop->uio_resid > 0) { 794233353Skib size = MIN(uiop->uio_resid, wsize); 795233353Skib size = MIN(uiop->uio_iov->iov_len, size); 796191783Srmacklem bp = getpbuf(&ncl_pbuf_freecnt); 797191783Srmacklem t_uio = malloc(sizeof(struct uio), M_NFSDIRECTIO, M_WAITOK); 798191783Srmacklem t_iov = malloc(sizeof(struct iovec), M_NFSDIRECTIO, M_WAITOK); 799191783Srmacklem t_iov->iov_base = malloc(size, M_NFSDIRECTIO, M_WAITOK); 800191783Srmacklem t_iov->iov_len = size; 801191783Srmacklem t_uio->uio_iov = t_iov; 802191783Srmacklem t_uio->uio_iovcnt = 1; 803191783Srmacklem t_uio->uio_offset = uiop->uio_offset; 804191783Srmacklem t_uio->uio_resid = size; 805191783Srmacklem t_uio->uio_segflg = UIO_SYSSPACE; 806191783Srmacklem t_uio->uio_rw = UIO_WRITE; 807191783Srmacklem t_uio->uio_td = td; 808232682Srmacklem KASSERT(uiop->uio_segflg == UIO_USERSPACE || 809232682Srmacklem uiop->uio_segflg == UIO_SYSSPACE, 810232682Srmacklem ("nfs_directio_write: Bad uio_segflg")); 811232682Srmacklem if (uiop->uio_segflg == UIO_USERSPACE) { 812232682Srmacklem error = copyin(uiop->uio_iov->iov_base, 813232682Srmacklem t_iov->iov_base, size); 814232682Srmacklem if (error != 0) 815232682Srmacklem goto err_free; 816232682Srmacklem } else 817232682Srmacklem /* 818232682Srmacklem * UIO_SYSSPACE may never happen, but handle 819232682Srmacklem * it just in case it does. 820232682Srmacklem */ 821232682Srmacklem bcopy(uiop->uio_iov->iov_base, t_iov->iov_base, 822232682Srmacklem size); 823191783Srmacklem bp->b_flags |= B_DIRECT; 824191783Srmacklem bp->b_iocmd = BIO_WRITE; 825191783Srmacklem if (cred != NOCRED) { 826191783Srmacklem crhold(cred); 827191783Srmacklem bp->b_wcred = cred; 828249077Skib } else 829249077Skib bp->b_wcred = NOCRED; 830191783Srmacklem bp->b_caller1 = (void *)t_uio; 831191783Srmacklem bp->b_vp = vp; 832191783Srmacklem error = ncl_asyncio(nmp, bp, NOCRED, td); 833232682Srmacklemerr_free: 834191783Srmacklem if (error) { 835191783Srmacklem free(t_iov->iov_base, M_NFSDIRECTIO); 836191783Srmacklem free(t_iov, M_NFSDIRECTIO); 837191783Srmacklem free(t_uio, M_NFSDIRECTIO); 838191783Srmacklem bp->b_vp = NULL; 839191783Srmacklem relpbuf(bp, &ncl_pbuf_freecnt); 840191783Srmacklem if (error == EINTR) 841191783Srmacklem return (error); 842191783Srmacklem goto do_sync; 843191783Srmacklem } 844191783Srmacklem uiop->uio_offset += size; 845191783Srmacklem uiop->uio_resid -= size; 846191783Srmacklem if (uiop->uio_iov->iov_len <= size) { 847191783Srmacklem uiop->uio_iovcnt--; 848191783Srmacklem uiop->uio_iov++; 849191783Srmacklem } else { 850249077Skib uiop->uio_iov->iov_base = 851191783Srmacklem (char *)uiop->uio_iov->iov_base + size; 852191783Srmacklem uiop->uio_iov->iov_len -= size; 853191783Srmacklem } 854191783Srmacklem } 855191783Srmacklem } 856191783Srmacklem return (0); 857191783Srmacklem} 858191783Srmacklem 859191783Srmacklem/* 860191783Srmacklem * Vnode op for write using bio 861191783Srmacklem */ 862191783Srmacklemint 863191783Srmacklemncl_write(struct vop_write_args *ap) 864191783Srmacklem{ 865191783Srmacklem int biosize; 866191783Srmacklem struct uio *uio = ap->a_uio; 867191783Srmacklem struct thread *td = uio->uio_td; 868191783Srmacklem struct vnode *vp = ap->a_vp; 869191783Srmacklem struct nfsnode *np = VTONFS(vp); 870191783Srmacklem struct ucred *cred = ap->a_cred; 871191783Srmacklem int ioflag = ap->a_ioflag; 872191783Srmacklem struct buf *bp; 873191783Srmacklem struct vattr vattr; 874191783Srmacklem struct nfsmount *nmp = VFSTONFS(vp->v_mount); 875191783Srmacklem daddr_t lbn; 876260170Srmacklem int bcount, noncontig_write, obcount; 877239848Skib int bp_cached, n, on, error = 0, error1; 878239845Skib size_t orig_resid, local_resid; 879239845Skib off_t orig_size, tmp_off; 880191783Srmacklem 881209120Skib KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode")); 882209120Skib KASSERT(uio->uio_segflg != UIO_USERSPACE || uio->uio_td == curthread, 883209120Skib ("ncl_write proc")); 884191783Srmacklem if (vp->v_type != VREG) 885191783Srmacklem return (EIO); 886191783Srmacklem mtx_lock(&np->n_mtx); 887191783Srmacklem if (np->n_flag & NWRITEERR) { 888191783Srmacklem np->n_flag &= ~NWRITEERR; 889191783Srmacklem mtx_unlock(&np->n_mtx); 890191783Srmacklem return (np->n_error); 891191783Srmacklem } else 892191783Srmacklem mtx_unlock(&np->n_mtx); 893191783Srmacklem mtx_lock(&nmp->nm_mtx); 894191783Srmacklem if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 895191783Srmacklem (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 896191783Srmacklem mtx_unlock(&nmp->nm_mtx); 897191783Srmacklem (void)ncl_fsinfo(nmp, vp, cred, td); 898191783Srmacklem mtx_lock(&nmp->nm_mtx); 899191783Srmacklem } 900191783Srmacklem if (nmp->nm_wsize == 0) 901191783Srmacklem (void) newnfs_iosize(nmp); 902191783Srmacklem mtx_unlock(&nmp->nm_mtx); 903191783Srmacklem 904191783Srmacklem /* 905191783Srmacklem * Synchronously flush pending buffers if we are in synchronous 906191783Srmacklem * mode or if we are appending. 907191783Srmacklem */ 908191783Srmacklem if (ioflag & (IO_APPEND | IO_SYNC)) { 909191783Srmacklem mtx_lock(&np->n_mtx); 910191783Srmacklem if (np->n_flag & NMODIFIED) { 911191783Srmacklem mtx_unlock(&np->n_mtx); 912191783Srmacklem#ifdef notyet /* Needs matching nonblock semantics elsewhere, too. */ 913191783Srmacklem /* 914191783Srmacklem * Require non-blocking, synchronous writes to 915191783Srmacklem * dirty files to inform the program it needs 916191783Srmacklem * to fsync(2) explicitly. 917191783Srmacklem */ 918191783Srmacklem if (ioflag & IO_NDELAY) 919191783Srmacklem return (EAGAIN); 920191783Srmacklem#endif 921191783Srmacklemflush_and_restart: 922191783Srmacklem np->n_attrstamp = 0; 923223280Srmacklem KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 924191783Srmacklem error = ncl_vinvalbuf(vp, V_SAVE, td, 1); 925191783Srmacklem if (error) 926191783Srmacklem return (error); 927191783Srmacklem } else 928191783Srmacklem mtx_unlock(&np->n_mtx); 929191783Srmacklem } 930191783Srmacklem 931239845Skib orig_resid = uio->uio_resid; 932239845Skib mtx_lock(&np->n_mtx); 933239845Skib orig_size = np->n_size; 934239845Skib mtx_unlock(&np->n_mtx); 935239845Skib 936191783Srmacklem /* 937191783Srmacklem * If IO_APPEND then load uio_offset. We restart here if we cannot 938191783Srmacklem * get the append lock. 939191783Srmacklem */ 940191783Srmacklem if (ioflag & IO_APPEND) { 941191783Srmacklem np->n_attrstamp = 0; 942223280Srmacklem KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 943191783Srmacklem error = VOP_GETATTR(vp, &vattr, cred); 944191783Srmacklem if (error) 945191783Srmacklem return (error); 946191783Srmacklem mtx_lock(&np->n_mtx); 947191783Srmacklem uio->uio_offset = np->n_size; 948191783Srmacklem mtx_unlock(&np->n_mtx); 949191783Srmacklem } 950191783Srmacklem 951191783Srmacklem if (uio->uio_offset < 0) 952191783Srmacklem return (EINVAL); 953220877Srmacklem tmp_off = uio->uio_offset + uio->uio_resid; 954220928Srmacklem if (tmp_off > nmp->nm_maxfilesize || tmp_off < uio->uio_offset) 955191783Srmacklem return (EFBIG); 956191783Srmacklem if (uio->uio_resid == 0) 957191783Srmacklem return (0); 958191783Srmacklem 959191783Srmacklem if (newnfs_directio_enable && (ioflag & IO_DIRECT) && vp->v_type == VREG) 960191783Srmacklem return nfs_directio_write(vp, uio, cred, ioflag); 961191783Srmacklem 962191783Srmacklem /* 963191783Srmacklem * Maybe this should be above the vnode op call, but so long as 964191783Srmacklem * file servers have no limits, i don't think it matters 965191783Srmacklem */ 966207662Strasz if (vn_rlimit_fsize(vp, uio, td)) 967207662Strasz return (EFBIG); 968191783Srmacklem 969231330Srmacklem biosize = vp->v_bufobj.bo_bsize; 970191783Srmacklem /* 971191783Srmacklem * Find all of this file's B_NEEDCOMMIT buffers. If our writes 972191783Srmacklem * would exceed the local maximum per-file write commit size when 973191783Srmacklem * combined with those, we must decide whether to flush, 974191783Srmacklem * go synchronous, or return error. We don't bother checking 975191783Srmacklem * IO_UNIT -- we just make all writes atomic anyway, as there's 976191783Srmacklem * no point optimizing for something that really won't ever happen. 977191783Srmacklem */ 978191783Srmacklem if (!(ioflag & IO_SYNC)) { 979191783Srmacklem int nflag; 980191783Srmacklem 981191783Srmacklem mtx_lock(&np->n_mtx); 982191783Srmacklem nflag = np->n_flag; 983249077Skib mtx_unlock(&np->n_mtx); 984191783Srmacklem int needrestart = 0; 985191783Srmacklem if (nmp->nm_wcommitsize < uio->uio_resid) { 986191783Srmacklem /* 987191783Srmacklem * If this request could not possibly be completed 988191783Srmacklem * without exceeding the maximum outstanding write 989191783Srmacklem * commit size, see if we can convert it into a 990191783Srmacklem * synchronous write operation. 991191783Srmacklem */ 992191783Srmacklem if (ioflag & IO_NDELAY) 993191783Srmacklem return (EAGAIN); 994191783Srmacklem ioflag |= IO_SYNC; 995191783Srmacklem if (nflag & NMODIFIED) 996191783Srmacklem needrestart = 1; 997191783Srmacklem } else if (nflag & NMODIFIED) { 998191783Srmacklem int wouldcommit = 0; 999191783Srmacklem BO_LOCK(&vp->v_bufobj); 1000191783Srmacklem if (vp->v_bufobj.bo_dirty.bv_cnt != 0) { 1001191783Srmacklem TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, 1002191783Srmacklem b_bobufs) { 1003191783Srmacklem if (bp->b_flags & B_NEEDCOMMIT) 1004191783Srmacklem wouldcommit += bp->b_bcount; 1005191783Srmacklem } 1006191783Srmacklem } 1007191783Srmacklem BO_UNLOCK(&vp->v_bufobj); 1008191783Srmacklem /* 1009191783Srmacklem * Since we're not operating synchronously and 1010191783Srmacklem * bypassing the buffer cache, we are in a commit 1011191783Srmacklem * and holding all of these buffers whether 1012191783Srmacklem * transmitted or not. If not limited, this 1013191783Srmacklem * will lead to the buffer cache deadlocking, 1014191783Srmacklem * as no one else can flush our uncommitted buffers. 1015191783Srmacklem */ 1016191783Srmacklem wouldcommit += uio->uio_resid; 1017191783Srmacklem /* 1018191783Srmacklem * If we would initially exceed the maximum 1019191783Srmacklem * outstanding write commit size, flush and restart. 1020191783Srmacklem */ 1021191783Srmacklem if (wouldcommit > nmp->nm_wcommitsize) 1022191783Srmacklem needrestart = 1; 1023191783Srmacklem } 1024191783Srmacklem if (needrestart) 1025191783Srmacklem goto flush_and_restart; 1026191783Srmacklem } 1027191783Srmacklem 1028191783Srmacklem do { 1029191783Srmacklem NFSINCRGLOBAL(newnfsstats.biocache_writes); 1030191783Srmacklem lbn = uio->uio_offset / biosize; 1031191783Srmacklem on = uio->uio_offset & (biosize-1); 1032233353Skib n = MIN((unsigned)(biosize - on), uio->uio_resid); 1033191783Srmacklemagain: 1034191783Srmacklem /* 1035191783Srmacklem * Handle direct append and file extension cases, calculate 1036191783Srmacklem * unaligned buffer size. 1037191783Srmacklem */ 1038191783Srmacklem mtx_lock(&np->n_mtx); 1039260170Srmacklem if ((np->n_flag & NHASBEENLOCKED) == 0 && 1040260170Srmacklem (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0) 1041260170Srmacklem noncontig_write = 1; 1042260170Srmacklem else 1043260170Srmacklem noncontig_write = 0; 1044260170Srmacklem if ((uio->uio_offset == np->n_size || 1045260170Srmacklem (noncontig_write != 0 && 1046260170Srmacklem lbn == (np->n_size / biosize) && 1047260170Srmacklem uio->uio_offset + n > np->n_size)) && n) { 1048191783Srmacklem mtx_unlock(&np->n_mtx); 1049191783Srmacklem /* 1050191783Srmacklem * Get the buffer (in its pre-append state to maintain 1051191783Srmacklem * B_CACHE if it was previously set). Resize the 1052191783Srmacklem * nfsnode after we have locked the buffer to prevent 1053191783Srmacklem * readers from reading garbage. 1054191783Srmacklem */ 1055260170Srmacklem obcount = np->n_size - (lbn * biosize); 1056260170Srmacklem bp = nfs_getcacheblk(vp, lbn, obcount, td); 1057191783Srmacklem 1058191783Srmacklem if (bp != NULL) { 1059191783Srmacklem long save; 1060191783Srmacklem 1061191783Srmacklem mtx_lock(&np->n_mtx); 1062191783Srmacklem np->n_size = uio->uio_offset + n; 1063191783Srmacklem np->n_flag |= NMODIFIED; 1064191783Srmacklem vnode_pager_setsize(vp, np->n_size); 1065191783Srmacklem mtx_unlock(&np->n_mtx); 1066191783Srmacklem 1067191783Srmacklem save = bp->b_flags & B_CACHE; 1068260170Srmacklem bcount = on + n; 1069191783Srmacklem allocbuf(bp, bcount); 1070191783Srmacklem bp->b_flags |= save; 1071260170Srmacklem if (noncontig_write != 0 && on > obcount) 1072260170Srmacklem vfs_bio_bzero_buf(bp, obcount, on - 1073260170Srmacklem obcount); 1074191783Srmacklem } 1075191783Srmacklem } else { 1076191783Srmacklem /* 1077191783Srmacklem * Obtain the locked cache block first, and then 1078191783Srmacklem * adjust the file's size as appropriate. 1079191783Srmacklem */ 1080191783Srmacklem bcount = on + n; 1081191783Srmacklem if ((off_t)lbn * biosize + bcount < np->n_size) { 1082191783Srmacklem if ((off_t)(lbn + 1) * biosize < np->n_size) 1083191783Srmacklem bcount = biosize; 1084191783Srmacklem else 1085191783Srmacklem bcount = np->n_size - (off_t)lbn * biosize; 1086191783Srmacklem } 1087191783Srmacklem mtx_unlock(&np->n_mtx); 1088191783Srmacklem bp = nfs_getcacheblk(vp, lbn, bcount, td); 1089191783Srmacklem mtx_lock(&np->n_mtx); 1090191783Srmacklem if (uio->uio_offset + n > np->n_size) { 1091191783Srmacklem np->n_size = uio->uio_offset + n; 1092191783Srmacklem np->n_flag |= NMODIFIED; 1093191783Srmacklem vnode_pager_setsize(vp, np->n_size); 1094191783Srmacklem } 1095191783Srmacklem mtx_unlock(&np->n_mtx); 1096191783Srmacklem } 1097191783Srmacklem 1098191783Srmacklem if (!bp) { 1099191783Srmacklem error = newnfs_sigintr(nmp, td); 1100191783Srmacklem if (!error) 1101191783Srmacklem error = EINTR; 1102191783Srmacklem break; 1103191783Srmacklem } 1104191783Srmacklem 1105191783Srmacklem /* 1106191783Srmacklem * Issue a READ if B_CACHE is not set. In special-append 1107191783Srmacklem * mode, B_CACHE is based on the buffer prior to the write 1108191783Srmacklem * op and is typically set, avoiding the read. If a read 1109191783Srmacklem * is required in special append mode, the server will 1110191783Srmacklem * probably send us a short-read since we extended the file 1111191783Srmacklem * on our end, resulting in b_resid == 0 and, thusly, 1112191783Srmacklem * B_CACHE getting set. 1113191783Srmacklem * 1114191783Srmacklem * We can also avoid issuing the read if the write covers 1115191783Srmacklem * the entire buffer. We have to make sure the buffer state 1116191783Srmacklem * is reasonable in this case since we will not be initiating 1117191783Srmacklem * I/O. See the comments in kern/vfs_bio.c's getblk() for 1118191783Srmacklem * more information. 1119191783Srmacklem * 1120191783Srmacklem * B_CACHE may also be set due to the buffer being cached 1121191783Srmacklem * normally. 1122191783Srmacklem */ 1123191783Srmacklem 1124239845Skib bp_cached = 1; 1125191783Srmacklem if (on == 0 && n == bcount) { 1126239845Skib if ((bp->b_flags & B_CACHE) == 0) 1127239845Skib bp_cached = 0; 1128191783Srmacklem bp->b_flags |= B_CACHE; 1129191783Srmacklem bp->b_flags &= ~B_INVAL; 1130191783Srmacklem bp->b_ioflags &= ~BIO_ERROR; 1131191783Srmacklem } 1132191783Srmacklem 1133191783Srmacklem if ((bp->b_flags & B_CACHE) == 0) { 1134191783Srmacklem bp->b_iocmd = BIO_READ; 1135191783Srmacklem vfs_busy_pages(bp, 0); 1136207082Srmacklem error = ncl_doio(vp, bp, cred, td, 0); 1137191783Srmacklem if (error) { 1138191783Srmacklem brelse(bp); 1139191783Srmacklem break; 1140191783Srmacklem } 1141191783Srmacklem } 1142191783Srmacklem if (bp->b_wcred == NOCRED) 1143191783Srmacklem bp->b_wcred = crhold(cred); 1144191783Srmacklem mtx_lock(&np->n_mtx); 1145191783Srmacklem np->n_flag |= NMODIFIED; 1146191783Srmacklem mtx_unlock(&np->n_mtx); 1147191783Srmacklem 1148191783Srmacklem /* 1149191783Srmacklem * If dirtyend exceeds file size, chop it down. This should 1150191783Srmacklem * not normally occur but there is an append race where it 1151191783Srmacklem * might occur XXX, so we log it. 1152191783Srmacklem * 1153191783Srmacklem * If the chopping creates a reverse-indexed or degenerate 1154191783Srmacklem * situation with dirtyoff/end, we 0 both of them. 1155191783Srmacklem */ 1156191783Srmacklem 1157191783Srmacklem if (bp->b_dirtyend > bcount) { 1158191783Srmacklem ncl_printf("NFS append race @%lx:%d\n", 1159191783Srmacklem (long)bp->b_blkno * DEV_BSIZE, 1160191783Srmacklem bp->b_dirtyend - bcount); 1161191783Srmacklem bp->b_dirtyend = bcount; 1162191783Srmacklem } 1163191783Srmacklem 1164191783Srmacklem if (bp->b_dirtyoff >= bp->b_dirtyend) 1165191783Srmacklem bp->b_dirtyoff = bp->b_dirtyend = 0; 1166191783Srmacklem 1167191783Srmacklem /* 1168191783Srmacklem * If the new write will leave a contiguous dirty 1169191783Srmacklem * area, just update the b_dirtyoff and b_dirtyend, 1170191783Srmacklem * otherwise force a write rpc of the old dirty area. 1171191783Srmacklem * 1172260170Srmacklem * If there has been a file lock applied to this file 1173260170Srmacklem * or vfs.nfs.old_noncontig_writing is set, do the following: 1174191783Srmacklem * While it is possible to merge discontiguous writes due to 1175191783Srmacklem * our having a B_CACHE buffer ( and thus valid read data 1176191783Srmacklem * for the hole), we don't because it could lead to 1177191783Srmacklem * significant cache coherency problems with multiple clients, 1178191783Srmacklem * especially if locking is implemented later on. 1179191783Srmacklem * 1180260170Srmacklem * If vfs.nfs.old_noncontig_writing is not set and there has 1181260170Srmacklem * not been file locking done on this file: 1182260170Srmacklem * Relax coherency a bit for the sake of performance and 1183260170Srmacklem * expand the current dirty region to contain the new 1184260170Srmacklem * write even if it means we mark some non-dirty data as 1185260170Srmacklem * dirty. 1186191783Srmacklem */ 1187191783Srmacklem 1188260170Srmacklem if (noncontig_write == 0 && bp->b_dirtyend > 0 && 1189191783Srmacklem (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1190191783Srmacklem if (bwrite(bp) == EINTR) { 1191191783Srmacklem error = EINTR; 1192191783Srmacklem break; 1193191783Srmacklem } 1194191783Srmacklem goto again; 1195191783Srmacklem } 1196191783Srmacklem 1197239845Skib local_resid = uio->uio_resid; 1198239852Skib error = vn_io_fault_uiomove((char *)bp->b_data + on, n, uio); 1199191783Srmacklem 1200239845Skib if (error != 0 && !bp_cached) { 1201239845Skib /* 1202239845Skib * This block has no other content then what 1203239845Skib * possibly was written by the faulty uiomove. 1204239845Skib * Release it, forgetting the data pages, to 1205239845Skib * prevent the leak of uninitialized data to 1206239845Skib * usermode. 1207239845Skib */ 1208239845Skib bp->b_ioflags |= BIO_ERROR; 1209239845Skib brelse(bp); 1210239845Skib uio->uio_offset -= local_resid - uio->uio_resid; 1211239845Skib uio->uio_resid = local_resid; 1212239845Skib break; 1213239845Skib } 1214239845Skib 1215191783Srmacklem /* 1216191783Srmacklem * Since this block is being modified, it must be written 1217191783Srmacklem * again and not just committed. Since write clustering does 1218191783Srmacklem * not work for the stage 1 data write, only the stage 2 1219191783Srmacklem * commit rpc, we have to clear B_CLUSTEROK as well. 1220191783Srmacklem */ 1221191783Srmacklem bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1222191783Srmacklem 1223239845Skib /* 1224239845Skib * Get the partial update on the progress made from 1225239845Skib * uiomove, if an error occured. 1226239845Skib */ 1227239845Skib if (error != 0) 1228239845Skib n = local_resid - uio->uio_resid; 1229191783Srmacklem 1230191783Srmacklem /* 1231191783Srmacklem * Only update dirtyoff/dirtyend if not a degenerate 1232191783Srmacklem * condition. 1233191783Srmacklem */ 1234239845Skib if (n > 0) { 1235191783Srmacklem if (bp->b_dirtyend > 0) { 1236191783Srmacklem bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1237191783Srmacklem bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1238191783Srmacklem } else { 1239191783Srmacklem bp->b_dirtyoff = on; 1240191783Srmacklem bp->b_dirtyend = on + n; 1241191783Srmacklem } 1242193187Salc vfs_bio_set_valid(bp, on, n); 1243191783Srmacklem } 1244191783Srmacklem 1245191783Srmacklem /* 1246191783Srmacklem * If IO_SYNC do bwrite(). 1247191783Srmacklem * 1248191783Srmacklem * IO_INVAL appears to be unused. The idea appears to be 1249191783Srmacklem * to turn off caching in this case. Very odd. XXX 1250191783Srmacklem */ 1251191783Srmacklem if ((ioflag & IO_SYNC)) { 1252191783Srmacklem if (ioflag & IO_INVAL) 1253191783Srmacklem bp->b_flags |= B_NOCACHE; 1254239848Skib error1 = bwrite(bp); 1255239848Skib if (error1 != 0) { 1256239848Skib if (error == 0) 1257239848Skib error = error1; 1258191783Srmacklem break; 1259239848Skib } 1260191783Srmacklem } else if ((n + on) == biosize) { 1261191783Srmacklem bp->b_flags |= B_ASYNC; 1262191783Srmacklem (void) ncl_writebp(bp, 0, NULL); 1263191783Srmacklem } else { 1264191783Srmacklem bdwrite(bp); 1265191783Srmacklem } 1266239845Skib 1267239845Skib if (error != 0) 1268239845Skib break; 1269191783Srmacklem } while (uio->uio_resid > 0 && n > 0); 1270191783Srmacklem 1271239845Skib if (error != 0) { 1272239845Skib if (ioflag & IO_UNIT) { 1273239845Skib VATTR_NULL(&vattr); 1274239845Skib vattr.va_size = orig_size; 1275239845Skib /* IO_SYNC is handled implicitely */ 1276239845Skib (void)VOP_SETATTR(vp, &vattr, cred); 1277239845Skib uio->uio_offset -= orig_resid - uio->uio_resid; 1278239845Skib uio->uio_resid = orig_resid; 1279239845Skib } 1280239845Skib } 1281239845Skib 1282191783Srmacklem return (error); 1283191783Srmacklem} 1284191783Srmacklem 1285191783Srmacklem/* 1286191783Srmacklem * Get an nfs cache block. 1287191783Srmacklem * 1288191783Srmacklem * Allocate a new one if the block isn't currently in the cache 1289191783Srmacklem * and return the block marked busy. If the calling process is 1290191783Srmacklem * interrupted by a signal for an interruptible mount point, return 1291191783Srmacklem * NULL. 1292191783Srmacklem * 1293191783Srmacklem * The caller must carefully deal with the possible B_INVAL state of 1294191783Srmacklem * the buffer. ncl_doio() clears B_INVAL (and ncl_asyncio() clears it 1295191783Srmacklem * indirectly), so synchronous reads can be issued without worrying about 1296191783Srmacklem * the B_INVAL state. We have to be a little more careful when dealing 1297191783Srmacklem * with writes (see comments in nfs_write()) when extending a file past 1298191783Srmacklem * its EOF. 1299191783Srmacklem */ 1300191783Srmacklemstatic struct buf * 1301191783Srmacklemnfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1302191783Srmacklem{ 1303191783Srmacklem struct buf *bp; 1304191783Srmacklem struct mount *mp; 1305191783Srmacklem struct nfsmount *nmp; 1306191783Srmacklem 1307191783Srmacklem mp = vp->v_mount; 1308191783Srmacklem nmp = VFSTONFS(mp); 1309191783Srmacklem 1310191783Srmacklem if (nmp->nm_flag & NFSMNT_INT) { 1311249077Skib sigset_t oldset; 1312191783Srmacklem 1313249077Skib newnfs_set_sigmask(td, &oldset); 1314195821Srmacklem bp = getblk(vp, bn, size, NFS_PCATCH, 0, 0); 1315249077Skib newnfs_restore_sigmask(td, &oldset); 1316191783Srmacklem while (bp == NULL) { 1317191783Srmacklem if (newnfs_sigintr(nmp, td)) 1318191783Srmacklem return (NULL); 1319191783Srmacklem bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1320191783Srmacklem } 1321191783Srmacklem } else { 1322191783Srmacklem bp = getblk(vp, bn, size, 0, 0, 0); 1323191783Srmacklem } 1324191783Srmacklem 1325231330Srmacklem if (vp->v_type == VREG) 1326231330Srmacklem bp->b_blkno = bn * (vp->v_bufobj.bo_bsize / DEV_BSIZE); 1327191783Srmacklem return (bp); 1328191783Srmacklem} 1329191783Srmacklem 1330191783Srmacklem/* 1331191783Srmacklem * Flush and invalidate all dirty buffers. If another process is already 1332191783Srmacklem * doing the flush, just wait for completion. 1333191783Srmacklem */ 1334191783Srmacklemint 1335191783Srmacklemncl_vinvalbuf(struct vnode *vp, int flags, struct thread *td, int intrflg) 1336191783Srmacklem{ 1337191783Srmacklem struct nfsnode *np = VTONFS(vp); 1338191783Srmacklem struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1339191783Srmacklem int error = 0, slpflag, slptimeo; 1340249077Skib int old_lock = 0; 1341191783Srmacklem 1342191783Srmacklem ASSERT_VOP_LOCKED(vp, "ncl_vinvalbuf"); 1343191783Srmacklem 1344191783Srmacklem if ((nmp->nm_flag & NFSMNT_INT) == 0) 1345191783Srmacklem intrflg = 0; 1346191783Srmacklem if ((nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)) 1347191783Srmacklem intrflg = 1; 1348191783Srmacklem if (intrflg) { 1349195821Srmacklem slpflag = NFS_PCATCH; 1350191783Srmacklem slptimeo = 2 * hz; 1351191783Srmacklem } else { 1352191783Srmacklem slpflag = 0; 1353191783Srmacklem slptimeo = 0; 1354191783Srmacklem } 1355191783Srmacklem 1356191783Srmacklem old_lock = ncl_upgrade_vnlock(vp); 1357193955Srmacklem if (vp->v_iflag & VI_DOOMED) { 1358193955Srmacklem /* 1359193955Srmacklem * Since vgonel() uses the generic vinvalbuf() to flush 1360193955Srmacklem * dirty buffers and it does not call this function, it 1361193955Srmacklem * is safe to just return OK when VI_DOOMED is set. 1362193955Srmacklem */ 1363193955Srmacklem ncl_downgrade_vnlock(vp, old_lock); 1364193955Srmacklem return (0); 1365193955Srmacklem } 1366193955Srmacklem 1367191783Srmacklem /* 1368191783Srmacklem * Now, flush as required. 1369191783Srmacklem */ 1370191783Srmacklem if ((flags & V_SAVE) && (vp->v_bufobj.bo_object != NULL)) { 1371191783Srmacklem VM_OBJECT_LOCK(vp->v_bufobj.bo_object); 1372191783Srmacklem vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1373191783Srmacklem VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object); 1374191783Srmacklem /* 1375191783Srmacklem * If the page clean was interrupted, fail the invalidation. 1376249077Skib * Not doing so, we run the risk of losing dirty pages in the 1377191783Srmacklem * vinvalbuf() call below. 1378191783Srmacklem */ 1379191783Srmacklem if (intrflg && (error = newnfs_sigintr(nmp, td))) 1380191783Srmacklem goto out; 1381191783Srmacklem } 1382191783Srmacklem 1383191783Srmacklem error = vinvalbuf(vp, flags, slpflag, 0); 1384191783Srmacklem while (error) { 1385191783Srmacklem if (intrflg && (error = newnfs_sigintr(nmp, td))) 1386191783Srmacklem goto out; 1387191783Srmacklem error = vinvalbuf(vp, flags, 0, slptimeo); 1388191783Srmacklem } 1389191783Srmacklem mtx_lock(&np->n_mtx); 1390191783Srmacklem if (np->n_directio_asyncwr == 0) 1391191783Srmacklem np->n_flag &= ~NMODIFIED; 1392191783Srmacklem mtx_unlock(&np->n_mtx); 1393191783Srmacklemout: 1394191783Srmacklem ncl_downgrade_vnlock(vp, old_lock); 1395191783Srmacklem return error; 1396191783Srmacklem} 1397191783Srmacklem 1398191783Srmacklem/* 1399191783Srmacklem * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1400191783Srmacklem * This is mainly to avoid queueing async I/O requests when the nfsiods 1401191783Srmacklem * are all hung on a dead server. 1402191783Srmacklem * 1403191783Srmacklem * Note: ncl_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1404191783Srmacklem * is eventually dequeued by the async daemon, ncl_doio() *will*. 1405191783Srmacklem */ 1406191783Srmacklemint 1407191783Srmacklemncl_asyncio(struct nfsmount *nmp, struct buf *bp, struct ucred *cred, struct thread *td) 1408191783Srmacklem{ 1409191783Srmacklem int iod; 1410191783Srmacklem int gotiod; 1411191783Srmacklem int slpflag = 0; 1412191783Srmacklem int slptimeo = 0; 1413191783Srmacklem int error, error2; 1414191783Srmacklem 1415191783Srmacklem /* 1416191783Srmacklem * Commits are usually short and sweet so lets save some cpu and 1417191783Srmacklem * leave the async daemons for more important rpc's (such as reads 1418191783Srmacklem * and writes). 1419250257Srmacklem * 1420250257Srmacklem * Readdirplus RPCs do vget()s to acquire the vnodes for entries 1421250257Srmacklem * in the directory in order to update attributes. This can deadlock 1422250257Srmacklem * with another thread that is waiting for async I/O to be done by 1423250257Srmacklem * an nfsiod thread while holding a lock on one of these vnodes. 1424250257Srmacklem * To avoid this deadlock, don't allow the async nfsiod threads to 1425250257Srmacklem * perform Readdirplus RPCs. 1426191783Srmacklem */ 1427191783Srmacklem mtx_lock(&ncl_iod_mutex); 1428250257Srmacklem if ((bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1429250257Srmacklem (nmp->nm_bufqiods > ncl_numasync / 2)) || 1430250257Srmacklem (bp->b_vp->v_type == VDIR && (nmp->nm_flag & NFSMNT_RDIRPLUS))) { 1431191783Srmacklem mtx_unlock(&ncl_iod_mutex); 1432191783Srmacklem return(EIO); 1433191783Srmacklem } 1434191783Srmacklemagain: 1435191783Srmacklem if (nmp->nm_flag & NFSMNT_INT) 1436195821Srmacklem slpflag = NFS_PCATCH; 1437191783Srmacklem gotiod = FALSE; 1438191783Srmacklem 1439191783Srmacklem /* 1440191783Srmacklem * Find a free iod to process this request. 1441191783Srmacklem */ 1442191783Srmacklem for (iod = 0; iod < ncl_numasync; iod++) 1443203119Srmacklem if (ncl_iodwant[iod] == NFSIOD_AVAILABLE) { 1444191783Srmacklem gotiod = TRUE; 1445191783Srmacklem break; 1446191783Srmacklem } 1447191783Srmacklem 1448191783Srmacklem /* 1449191783Srmacklem * Try to create one if none are free. 1450191783Srmacklem */ 1451220683Srmacklem if (!gotiod) 1452220683Srmacklem ncl_nfsiodnew(); 1453220683Srmacklem else { 1454191783Srmacklem /* 1455191783Srmacklem * Found one, so wake it up and tell it which 1456191783Srmacklem * mount to process. 1457191783Srmacklem */ 1458191783Srmacklem NFS_DPF(ASYNCIO, ("ncl_asyncio: waking iod %d for mount %p\n", 1459191783Srmacklem iod, nmp)); 1460203119Srmacklem ncl_iodwant[iod] = NFSIOD_NOT_AVAILABLE; 1461191783Srmacklem ncl_iodmount[iod] = nmp; 1462191783Srmacklem nmp->nm_bufqiods++; 1463191783Srmacklem wakeup(&ncl_iodwant[iod]); 1464191783Srmacklem } 1465191783Srmacklem 1466191783Srmacklem /* 1467191783Srmacklem * If none are free, we may already have an iod working on this mount 1468191783Srmacklem * point. If so, it will process our request. 1469191783Srmacklem */ 1470191783Srmacklem if (!gotiod) { 1471191783Srmacklem if (nmp->nm_bufqiods > 0) { 1472191783Srmacklem NFS_DPF(ASYNCIO, 1473191783Srmacklem ("ncl_asyncio: %d iods are already processing mount %p\n", 1474191783Srmacklem nmp->nm_bufqiods, nmp)); 1475191783Srmacklem gotiod = TRUE; 1476191783Srmacklem } 1477191783Srmacklem } 1478191783Srmacklem 1479191783Srmacklem /* 1480191783Srmacklem * If we have an iod which can process the request, then queue 1481191783Srmacklem * the buffer. 1482191783Srmacklem */ 1483191783Srmacklem if (gotiod) { 1484191783Srmacklem /* 1485191783Srmacklem * Ensure that the queue never grows too large. We still want 1486191783Srmacklem * to asynchronize so we block rather then return EIO. 1487191783Srmacklem */ 1488191783Srmacklem while (nmp->nm_bufqlen >= 2*ncl_numasync) { 1489191783Srmacklem NFS_DPF(ASYNCIO, 1490191783Srmacklem ("ncl_asyncio: waiting for mount %p queue to drain\n", nmp)); 1491191783Srmacklem nmp->nm_bufqwant = TRUE; 1492249077Skib error = newnfs_msleep(td, &nmp->nm_bufq, 1493201029Srmacklem &ncl_iod_mutex, slpflag | PRIBIO, "nfsaio", 1494249077Skib slptimeo); 1495191783Srmacklem if (error) { 1496191783Srmacklem error2 = newnfs_sigintr(nmp, td); 1497191783Srmacklem if (error2) { 1498249077Skib mtx_unlock(&ncl_iod_mutex); 1499191783Srmacklem return (error2); 1500191783Srmacklem } 1501195821Srmacklem if (slpflag == NFS_PCATCH) { 1502191783Srmacklem slpflag = 0; 1503191783Srmacklem slptimeo = 2 * hz; 1504191783Srmacklem } 1505191783Srmacklem } 1506191783Srmacklem /* 1507191783Srmacklem * We might have lost our iod while sleeping, 1508191783Srmacklem * so check and loop if nescessary. 1509191783Srmacklem */ 1510220683Srmacklem goto again; 1511191783Srmacklem } 1512191783Srmacklem 1513191783Srmacklem /* We might have lost our nfsiod */ 1514191783Srmacklem if (nmp->nm_bufqiods == 0) { 1515191783Srmacklem NFS_DPF(ASYNCIO, 1516191783Srmacklem ("ncl_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1517191783Srmacklem goto again; 1518191783Srmacklem } 1519191783Srmacklem 1520191783Srmacklem if (bp->b_iocmd == BIO_READ) { 1521191783Srmacklem if (bp->b_rcred == NOCRED && cred != NOCRED) 1522191783Srmacklem bp->b_rcred = crhold(cred); 1523191783Srmacklem } else { 1524191783Srmacklem if (bp->b_wcred == NOCRED && cred != NOCRED) 1525191783Srmacklem bp->b_wcred = crhold(cred); 1526191783Srmacklem } 1527191783Srmacklem 1528191783Srmacklem if (bp->b_flags & B_REMFREE) 1529191783Srmacklem bremfreef(bp); 1530191783Srmacklem BUF_KERNPROC(bp); 1531191783Srmacklem TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1532191783Srmacklem nmp->nm_bufqlen++; 1533191783Srmacklem if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1534249077Skib mtx_lock(&(VTONFS(bp->b_vp))->n_mtx); 1535191783Srmacklem VTONFS(bp->b_vp)->n_flag |= NMODIFIED; 1536191783Srmacklem VTONFS(bp->b_vp)->n_directio_asyncwr++; 1537191783Srmacklem mtx_unlock(&(VTONFS(bp->b_vp))->n_mtx); 1538191783Srmacklem } 1539191783Srmacklem mtx_unlock(&ncl_iod_mutex); 1540191783Srmacklem return (0); 1541191783Srmacklem } 1542191783Srmacklem 1543191783Srmacklem mtx_unlock(&ncl_iod_mutex); 1544191783Srmacklem 1545191783Srmacklem /* 1546191783Srmacklem * All the iods are busy on other mounts, so return EIO to 1547191783Srmacklem * force the caller to process the i/o synchronously. 1548191783Srmacklem */ 1549191783Srmacklem NFS_DPF(ASYNCIO, ("ncl_asyncio: no iods available, i/o is synchronous\n")); 1550191783Srmacklem return (EIO); 1551191783Srmacklem} 1552191783Srmacklem 1553191783Srmacklemvoid 1554191783Srmacklemncl_doio_directwrite(struct buf *bp) 1555191783Srmacklem{ 1556191783Srmacklem int iomode, must_commit; 1557191783Srmacklem struct uio *uiop = (struct uio *)bp->b_caller1; 1558191783Srmacklem char *iov_base = uiop->uio_iov->iov_base; 1559249077Skib 1560191783Srmacklem iomode = NFSWRITE_FILESYNC; 1561191783Srmacklem uiop->uio_td = NULL; /* NULL since we're in nfsiod */ 1562207082Srmacklem ncl_writerpc(bp->b_vp, uiop, bp->b_wcred, &iomode, &must_commit, 0); 1563191783Srmacklem KASSERT((must_commit == 0), ("ncl_doio_directwrite: Did not commit write")); 1564191783Srmacklem free(iov_base, M_NFSDIRECTIO); 1565191783Srmacklem free(uiop->uio_iov, M_NFSDIRECTIO); 1566191783Srmacklem free(uiop, M_NFSDIRECTIO); 1567191783Srmacklem if ((bp->b_flags & B_DIRECT) && bp->b_iocmd == BIO_WRITE) { 1568191783Srmacklem struct nfsnode *np = VTONFS(bp->b_vp); 1569191783Srmacklem mtx_lock(&np->n_mtx); 1570191783Srmacklem np->n_directio_asyncwr--; 1571191783Srmacklem if (np->n_directio_asyncwr == 0) { 1572191783Srmacklem np->n_flag &= ~NMODIFIED; 1573191783Srmacklem if ((np->n_flag & NFSYNCWAIT)) { 1574191783Srmacklem np->n_flag &= ~NFSYNCWAIT; 1575191783Srmacklem wakeup((caddr_t)&np->n_directio_asyncwr); 1576191783Srmacklem } 1577191783Srmacklem } 1578191783Srmacklem mtx_unlock(&np->n_mtx); 1579191783Srmacklem } 1580191783Srmacklem bp->b_vp = NULL; 1581191783Srmacklem relpbuf(bp, &ncl_pbuf_freecnt); 1582191783Srmacklem} 1583191783Srmacklem 1584191783Srmacklem/* 1585191783Srmacklem * Do an I/O operation to/from a cache block. This may be called 1586191783Srmacklem * synchronously or from an nfsiod. 1587191783Srmacklem */ 1588191783Srmacklemint 1589207082Srmacklemncl_doio(struct vnode *vp, struct buf *bp, struct ucred *cr, struct thread *td, 1590207082Srmacklem int called_from_strategy) 1591191783Srmacklem{ 1592191783Srmacklem struct uio *uiop; 1593191783Srmacklem struct nfsnode *np; 1594191783Srmacklem struct nfsmount *nmp; 1595191783Srmacklem int error = 0, iomode, must_commit = 0; 1596191783Srmacklem struct uio uio; 1597191783Srmacklem struct iovec io; 1598191783Srmacklem struct proc *p = td ? td->td_proc : NULL; 1599191783Srmacklem uint8_t iocmd; 1600249077Skib 1601191783Srmacklem np = VTONFS(vp); 1602191783Srmacklem nmp = VFSTONFS(vp->v_mount); 1603191783Srmacklem uiop = &uio; 1604191783Srmacklem uiop->uio_iov = &io; 1605191783Srmacklem uiop->uio_iovcnt = 1; 1606191783Srmacklem uiop->uio_segflg = UIO_SYSSPACE; 1607191783Srmacklem uiop->uio_td = td; 1608191783Srmacklem 1609191783Srmacklem /* 1610191783Srmacklem * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1611191783Srmacklem * do this here so we do not have to do it in all the code that 1612191783Srmacklem * calls us. 1613191783Srmacklem */ 1614191783Srmacklem bp->b_flags &= ~B_INVAL; 1615191783Srmacklem bp->b_ioflags &= ~BIO_ERROR; 1616191783Srmacklem 1617191783Srmacklem KASSERT(!(bp->b_flags & B_DONE), ("ncl_doio: bp %p already marked done", bp)); 1618191783Srmacklem iocmd = bp->b_iocmd; 1619191783Srmacklem if (iocmd == BIO_READ) { 1620191783Srmacklem io.iov_len = uiop->uio_resid = bp->b_bcount; 1621191783Srmacklem io.iov_base = bp->b_data; 1622191783Srmacklem uiop->uio_rw = UIO_READ; 1623191783Srmacklem 1624191783Srmacklem switch (vp->v_type) { 1625191783Srmacklem case VREG: 1626191783Srmacklem uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1627191783Srmacklem NFSINCRGLOBAL(newnfsstats.read_bios); 1628191783Srmacklem error = ncl_readrpc(vp, uiop, cr); 1629191783Srmacklem 1630191783Srmacklem if (!error) { 1631191783Srmacklem if (uiop->uio_resid) { 1632191783Srmacklem /* 1633191783Srmacklem * If we had a short read with no error, we must have 1634191783Srmacklem * hit a file hole. We should zero-fill the remainder. 1635191783Srmacklem * This can also occur if the server hits the file EOF. 1636191783Srmacklem * 1637191783Srmacklem * Holes used to be able to occur due to pending 1638191783Srmacklem * writes, but that is not possible any longer. 1639191783Srmacklem */ 1640191783Srmacklem int nread = bp->b_bcount - uiop->uio_resid; 1641233353Skib ssize_t left = uiop->uio_resid; 1642191783Srmacklem 1643191783Srmacklem if (left > 0) 1644191783Srmacklem bzero((char *)bp->b_data + nread, left); 1645191783Srmacklem uiop->uio_resid = 0; 1646191783Srmacklem } 1647191783Srmacklem } 1648191783Srmacklem /* ASSERT_VOP_LOCKED(vp, "ncl_doio"); */ 1649191783Srmacklem if (p && (vp->v_vflag & VV_TEXT)) { 1650191783Srmacklem mtx_lock(&np->n_mtx); 1651191783Srmacklem if (NFS_TIMESPEC_COMPARE(&np->n_mtime, &np->n_vattr.na_mtime)) { 1652191783Srmacklem mtx_unlock(&np->n_mtx); 1653191783Srmacklem PROC_LOCK(p); 1654191783Srmacklem killproc(p, "text file modification"); 1655191783Srmacklem PROC_UNLOCK(p); 1656191783Srmacklem } else 1657191783Srmacklem mtx_unlock(&np->n_mtx); 1658191783Srmacklem } 1659191783Srmacklem break; 1660191783Srmacklem case VLNK: 1661191783Srmacklem uiop->uio_offset = (off_t)0; 1662191783Srmacklem NFSINCRGLOBAL(newnfsstats.readlink_bios); 1663191783Srmacklem error = ncl_readlinkrpc(vp, uiop, cr); 1664191783Srmacklem break; 1665191783Srmacklem case VDIR: 1666191783Srmacklem NFSINCRGLOBAL(newnfsstats.readdir_bios); 1667191783Srmacklem uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1668191783Srmacklem if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1669191783Srmacklem error = ncl_readdirplusrpc(vp, uiop, cr, td); 1670191783Srmacklem if (error == NFSERR_NOTSUPP) 1671191783Srmacklem nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1672191783Srmacklem } 1673191783Srmacklem if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1674191783Srmacklem error = ncl_readdirrpc(vp, uiop, cr, td); 1675191783Srmacklem /* 1676191783Srmacklem * end-of-directory sets B_INVAL but does not generate an 1677191783Srmacklem * error. 1678191783Srmacklem */ 1679191783Srmacklem if (error == 0 && uiop->uio_resid == bp->b_bcount) 1680191783Srmacklem bp->b_flags |= B_INVAL; 1681191783Srmacklem break; 1682191783Srmacklem default: 1683191783Srmacklem ncl_printf("ncl_doio: type %x unexpected\n", vp->v_type); 1684191783Srmacklem break; 1685191783Srmacklem }; 1686191783Srmacklem if (error) { 1687191783Srmacklem bp->b_ioflags |= BIO_ERROR; 1688191783Srmacklem bp->b_error = error; 1689191783Srmacklem } 1690191783Srmacklem } else { 1691191783Srmacklem /* 1692191783Srmacklem * If we only need to commit, try to commit 1693191783Srmacklem */ 1694191783Srmacklem if (bp->b_flags & B_NEEDCOMMIT) { 1695191783Srmacklem int retv; 1696191783Srmacklem off_t off; 1697191783Srmacklem 1698191783Srmacklem off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1699191783Srmacklem retv = ncl_commit(vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1700191783Srmacklem bp->b_wcred, td); 1701191783Srmacklem if (retv == 0) { 1702191783Srmacklem bp->b_dirtyoff = bp->b_dirtyend = 0; 1703191783Srmacklem bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1704191783Srmacklem bp->b_resid = 0; 1705191783Srmacklem bufdone(bp); 1706191783Srmacklem return (0); 1707191783Srmacklem } 1708191783Srmacklem if (retv == NFSERR_STALEWRITEVERF) { 1709191783Srmacklem ncl_clearcommit(vp->v_mount); 1710191783Srmacklem } 1711191783Srmacklem } 1712191783Srmacklem 1713191783Srmacklem /* 1714191783Srmacklem * Setup for actual write 1715191783Srmacklem */ 1716191783Srmacklem mtx_lock(&np->n_mtx); 1717191783Srmacklem if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1718191783Srmacklem bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1719191783Srmacklem mtx_unlock(&np->n_mtx); 1720191783Srmacklem 1721191783Srmacklem if (bp->b_dirtyend > bp->b_dirtyoff) { 1722191783Srmacklem io.iov_len = uiop->uio_resid = bp->b_dirtyend 1723191783Srmacklem - bp->b_dirtyoff; 1724191783Srmacklem uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1725191783Srmacklem + bp->b_dirtyoff; 1726191783Srmacklem io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1727191783Srmacklem uiop->uio_rw = UIO_WRITE; 1728191783Srmacklem NFSINCRGLOBAL(newnfsstats.write_bios); 1729191783Srmacklem 1730191783Srmacklem if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1731191783Srmacklem iomode = NFSWRITE_UNSTABLE; 1732191783Srmacklem else 1733191783Srmacklem iomode = NFSWRITE_FILESYNC; 1734191783Srmacklem 1735207082Srmacklem error = ncl_writerpc(vp, uiop, cr, &iomode, &must_commit, 1736207082Srmacklem called_from_strategy); 1737191783Srmacklem 1738191783Srmacklem /* 1739191783Srmacklem * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1740191783Srmacklem * to cluster the buffers needing commit. This will allow 1741191783Srmacklem * the system to submit a single commit rpc for the whole 1742191783Srmacklem * cluster. We can do this even if the buffer is not 100% 1743191783Srmacklem * dirty (relative to the NFS blocksize), so we optimize the 1744191783Srmacklem * append-to-file-case. 1745191783Srmacklem * 1746191783Srmacklem * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1747191783Srmacklem * cleared because write clustering only works for commit 1748191783Srmacklem * rpc's, not for the data portion of the write). 1749191783Srmacklem */ 1750191783Srmacklem 1751191783Srmacklem if (!error && iomode == NFSWRITE_UNSTABLE) { 1752191783Srmacklem bp->b_flags |= B_NEEDCOMMIT; 1753191783Srmacklem if (bp->b_dirtyoff == 0 1754191783Srmacklem && bp->b_dirtyend == bp->b_bcount) 1755191783Srmacklem bp->b_flags |= B_CLUSTEROK; 1756191783Srmacklem } else { 1757191783Srmacklem bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1758191783Srmacklem } 1759191783Srmacklem 1760191783Srmacklem /* 1761191783Srmacklem * For an interrupted write, the buffer is still valid 1762191783Srmacklem * and the write hasn't been pushed to the server yet, 1763191783Srmacklem * so we can't set BIO_ERROR and report the interruption 1764191783Srmacklem * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1765191783Srmacklem * is not relevant, so the rpc attempt is essentially 1766191783Srmacklem * a noop. For the case of a V3 write rpc not being 1767191783Srmacklem * committed to stable storage, the block is still 1768191783Srmacklem * dirty and requires either a commit rpc or another 1769191783Srmacklem * write rpc with iomode == NFSV3WRITE_FILESYNC before 1770191783Srmacklem * the block is reused. This is indicated by setting 1771191783Srmacklem * the B_DELWRI and B_NEEDCOMMIT flags. 1772191783Srmacklem * 1773207082Srmacklem * EIO is returned by ncl_writerpc() to indicate a recoverable 1774207082Srmacklem * write error and is handled as above, except that 1775207082Srmacklem * B_EINTR isn't set. One cause of this is a stale stateid 1776207082Srmacklem * error for the RPC that indicates recovery is required, 1777207082Srmacklem * when called with called_from_strategy != 0. 1778207082Srmacklem * 1779191783Srmacklem * If the buffer is marked B_PAGING, it does not reside on 1780191783Srmacklem * the vp's paging queues so we cannot call bdirty(). The 1781191783Srmacklem * bp in this case is not an NFS cache block so we should 1782191783Srmacklem * be safe. XXX 1783191783Srmacklem * 1784249077Skib * The logic below breaks up errors into recoverable and 1785191783Srmacklem * unrecoverable. For the former, we clear B_INVAL|B_NOCACHE 1786191783Srmacklem * and keep the buffer around for potential write retries. 1787191783Srmacklem * For the latter (eg ESTALE), we toss the buffer away (B_INVAL) 1788249077Skib * and save the error in the nfsnode. This is less than ideal 1789191783Srmacklem * but necessary. Keeping such buffers around could potentially 1790191783Srmacklem * cause buffer exhaustion eventually (they can never be written 1791191783Srmacklem * out, so will get constantly be re-dirtied). It also causes 1792249077Skib * all sorts of vfs panics. For non-recoverable write errors, 1793191783Srmacklem * also invalidate the attrcache, so we'll be forced to go over 1794191783Srmacklem * the wire for this object, returning an error to user on next 1795191783Srmacklem * call (most of the time). 1796191783Srmacklem */ 1797249077Skib if (error == EINTR || error == EIO || error == ETIMEDOUT 1798191783Srmacklem || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1799191783Srmacklem int s; 1800191783Srmacklem 1801191783Srmacklem s = splbio(); 1802191783Srmacklem bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1803191783Srmacklem if ((bp->b_flags & B_PAGING) == 0) { 1804191783Srmacklem bdirty(bp); 1805191783Srmacklem bp->b_flags &= ~B_DONE; 1806191783Srmacklem } 1807207082Srmacklem if ((error == EINTR || error == ETIMEDOUT) && 1808207082Srmacklem (bp->b_flags & B_ASYNC) == 0) 1809191783Srmacklem bp->b_flags |= B_EINTR; 1810191783Srmacklem splx(s); 1811249077Skib } else { 1812191783Srmacklem if (error) { 1813191783Srmacklem bp->b_ioflags |= BIO_ERROR; 1814191783Srmacklem bp->b_flags |= B_INVAL; 1815191783Srmacklem bp->b_error = np->n_error = error; 1816191783Srmacklem mtx_lock(&np->n_mtx); 1817191783Srmacklem np->n_flag |= NWRITEERR; 1818191783Srmacklem np->n_attrstamp = 0; 1819223280Srmacklem KDTRACE_NFS_ATTRCACHE_FLUSH_DONE(vp); 1820191783Srmacklem mtx_unlock(&np->n_mtx); 1821191783Srmacklem } 1822191783Srmacklem bp->b_dirtyoff = bp->b_dirtyend = 0; 1823191783Srmacklem } 1824191783Srmacklem } else { 1825191783Srmacklem bp->b_resid = 0; 1826191783Srmacklem bufdone(bp); 1827191783Srmacklem return (0); 1828191783Srmacklem } 1829191783Srmacklem } 1830191783Srmacklem bp->b_resid = uiop->uio_resid; 1831191783Srmacklem if (must_commit) 1832191783Srmacklem ncl_clearcommit(vp->v_mount); 1833191783Srmacklem bufdone(bp); 1834191783Srmacklem return (error); 1835191783Srmacklem} 1836191783Srmacklem 1837191783Srmacklem/* 1838191783Srmacklem * Used to aid in handling ftruncate() operations on the NFS client side. 1839191783Srmacklem * Truncation creates a number of special problems for NFS. We have to 1840191783Srmacklem * throw away VM pages and buffer cache buffers that are beyond EOF, and 1841191783Srmacklem * we have to properly handle VM pages or (potentially dirty) buffers 1842191783Srmacklem * that straddle the truncation point. 1843191783Srmacklem */ 1844191783Srmacklem 1845191783Srmacklemint 1846191783Srmacklemncl_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1847191783Srmacklem{ 1848191783Srmacklem struct nfsnode *np = VTONFS(vp); 1849191783Srmacklem u_quad_t tsize; 1850231330Srmacklem int biosize = vp->v_bufobj.bo_bsize; 1851191783Srmacklem int error = 0; 1852191783Srmacklem 1853191783Srmacklem mtx_lock(&np->n_mtx); 1854191783Srmacklem tsize = np->n_size; 1855191783Srmacklem np->n_size = nsize; 1856191783Srmacklem mtx_unlock(&np->n_mtx); 1857191783Srmacklem 1858191783Srmacklem if (nsize < tsize) { 1859191783Srmacklem struct buf *bp; 1860191783Srmacklem daddr_t lbn; 1861191783Srmacklem int bufsize; 1862191783Srmacklem 1863191783Srmacklem /* 1864249077Skib * vtruncbuf() doesn't get the buffer overlapping the 1865191783Srmacklem * truncation point. We may have a B_DELWRI and/or B_CACHE 1866191783Srmacklem * buffer that now needs to be truncated. 1867191783Srmacklem */ 1868191783Srmacklem error = vtruncbuf(vp, cred, td, nsize, biosize); 1869191783Srmacklem lbn = nsize / biosize; 1870191783Srmacklem bufsize = nsize & (biosize - 1); 1871191783Srmacklem bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1872249077Skib if (!bp) 1873249077Skib return EINTR; 1874191783Srmacklem if (bp->b_dirtyoff > bp->b_bcount) 1875191783Srmacklem bp->b_dirtyoff = bp->b_bcount; 1876191783Srmacklem if (bp->b_dirtyend > bp->b_bcount) 1877191783Srmacklem bp->b_dirtyend = bp->b_bcount; 1878191783Srmacklem bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1879191783Srmacklem brelse(bp); 1880191783Srmacklem } else { 1881191783Srmacklem vnode_pager_setsize(vp, nsize); 1882191783Srmacklem } 1883191783Srmacklem return(error); 1884191783Srmacklem} 1885191783Srmacklem 1886