Index: projects/fuse2/sys/fs/fuse/fuse_io.c =================================================================== --- projects/fuse2/sys/fs/fuse/fuse_io.c (revision 349331) +++ projects/fuse2/sys/fs/fuse/fuse_io.c (revision 349332) @@ -1,1061 +1,1070 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 2007-2009 Google Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are * met: * * * Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * * Redistributions in binary form must reproduce the above * copyright notice, this list of conditions and the following disclaimer * in the documentation and/or other materials provided with the * distribution. * * Neither the name of Google Inc. nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * Copyright (C) 2005 Csaba Henk. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "fuse.h" #include "fuse_file.h" #include "fuse_node.h" #include "fuse_internal.h" #include "fuse_ipc.h" #include "fuse_io.h" SDT_PROVIDER_DECLARE(fusefs); /* * Fuse trace probe: * arg0: verbosity. Higher numbers give more verbose messages * arg1: Textual message */ SDT_PROBE_DEFINE2(fusefs, , io, trace, "int", "char*"); static void fuse_io_clear_suid_on_write(struct vnode *vp, struct ucred *cred, struct thread *td); static int fuse_read_directbackend(struct vnode *vp, struct uio *uio, struct ucred *cred, struct fuse_filehandle *fufh); static int fuse_read_biobackend(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred, struct fuse_filehandle *fufh, pid_t pid); static int fuse_write_directbackend(struct vnode *vp, struct uio *uio, struct ucred *cred, struct fuse_filehandle *fufh, off_t filesize, int ioflag, bool pages); static int fuse_write_biobackend(struct vnode *vp, struct uio *uio, struct ucred *cred, struct fuse_filehandle *fufh, int ioflag, pid_t pid); /* * FreeBSD clears the SUID and SGID bits on any write by a non-root user. */ static void fuse_io_clear_suid_on_write(struct vnode *vp, struct ucred *cred, struct thread *td) { struct fuse_data *data; struct mount *mp; struct vattr va; int dataflags; mp = vnode_mount(vp); data = fuse_get_mpdata(mp); dataflags = data->dataflags; if (dataflags & FSESS_DEFAULT_PERMISSIONS) { if (priv_check_cred(cred, PRIV_VFS_RETAINSUGID)) { fuse_internal_getattr(vp, &va, cred, td); if (va.va_mode & (S_ISUID | S_ISGID)) { mode_t mode = va.va_mode & ~(S_ISUID | S_ISGID); /* Clear all vattr fields except mode */ vattr_null(&va); va.va_mode = mode; /* * Ignore fuse_internal_setattr's return value, * because at this point the write operation has * already succeeded and we don't want to return * failing status for that. */ (void)fuse_internal_setattr(vp, &va, td, NULL); } } } } SDT_PROBE_DEFINE5(fusefs, , io, io_dispatch, "struct vnode*", "struct uio*", "int", "struct ucred*", "struct fuse_filehandle*"); int fuse_io_dispatch(struct vnode *vp, struct uio *uio, int ioflag, bool pages, struct ucred *cred, pid_t pid) { struct fuse_filehandle *fufh; int err, directio; int fflag; bool closefufh = false; MPASS(vp->v_type == VREG || vp->v_type == VDIR); fflag = (uio->uio_rw == UIO_READ) ? FREAD : FWRITE; err = fuse_filehandle_getrw(vp, fflag, &fufh, cred, pid); if (err == EBADF && vnode_mount(vp)->mnt_flag & MNT_EXPORTED) { /* * nfsd will do I/O without first doing VOP_OPEN. We * must implicitly open the file here */ err = fuse_filehandle_open(vp, fflag, &fufh, curthread, cred); closefufh = true; } else if (err) { printf("FUSE: io dispatch: filehandles are closed\n"); return err; } if (err) goto out; SDT_PROBE5(fusefs, , io, io_dispatch, vp, uio, ioflag, cred, fufh); /* * Ideally, when the daemon asks for direct io at open time, the * standard file flag should be set according to this, so that would * just change the default mode, which later on could be changed via * fcntl(2). * But this doesn't work, the O_DIRECT flag gets cleared at some point * (don't know where). So to make any use of the Fuse direct_io option, * we hardwire it into the file's private data (similarly to Linux, * btw.). */ directio = (ioflag & IO_DIRECT) || !fsess_opt_datacache(vnode_mount(vp)); switch (uio->uio_rw) { case UIO_READ: if (directio) { SDT_PROBE2(fusefs, , io, trace, 1, "direct read of vnode"); err = fuse_read_directbackend(vp, uio, cred, fufh); } else { SDT_PROBE2(fusefs, , io, trace, 1, "buffered read of vnode"); err = fuse_read_biobackend(vp, uio, ioflag, cred, fufh, pid); } break; case UIO_WRITE: if (directio) { const int iosize = fuse_iosize(vp); off_t start, end, filesize; SDT_PROBE2(fusefs, , io, trace, 1, "direct write of vnode"); err = fuse_vnode_size(vp, &filesize, cred, curthread); if (err) goto out; start = uio->uio_offset; end = start + uio->uio_resid; /* * Invalidate the write cache unless we're coming from * VOP_PUTPAGES, in which case we're writing _from_ the * write cache */ if (!pages ) v_inval_buf_range(vp, start, end, iosize); err = fuse_write_directbackend(vp, uio, cred, fufh, filesize, ioflag, pages); } else { SDT_PROBE2(fusefs, , io, trace, 1, "buffered write of vnode"); if (fuse_data_cache_mode == FUSE_CACHE_WT) ioflag |= IO_SYNC; err = fuse_write_biobackend(vp, uio, cred, fufh, ioflag, pid); } fuse_io_clear_suid_on_write(vp, cred, uio->uio_td); break; default: panic("uninterpreted mode passed to fuse_io_dispatch"); } out: if (closefufh) fuse_filehandle_close(vp, fufh, curthread, cred); return (err); } SDT_PROBE_DEFINE4(fusefs, , io, read_bio_backend_start, "int", "int", "int", "int"); SDT_PROBE_DEFINE2(fusefs, , io, read_bio_backend_feed, "int", "struct buf*"); SDT_PROBE_DEFINE4(fusefs, , io, read_bio_backend_end, "int", "ssize_t", "int", "struct buf*"); static int fuse_read_biobackend(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred, struct fuse_filehandle *fufh, pid_t pid) { struct buf *bp; struct mount *mp; struct fuse_data *data; daddr_t lbn, nextlbn; int bcount, nextsize; int err, n = 0, on = 0, seqcount; off_t filesize; const int biosize = fuse_iosize(vp); mp = vnode_mount(vp); data = fuse_get_mpdata(mp); if (uio->uio_offset < 0) return (EINVAL); seqcount = ioflag >> IO_SEQSHIFT; err = fuse_vnode_size(vp, &filesize, cred, curthread); if (err) return err; for (err = 0, bp = NULL; uio->uio_resid > 0; bp = NULL) { if (fuse_isdeadfs(vp)) { err = ENXIO; break; } if (filesize - uio->uio_offset <= 0) break; lbn = uio->uio_offset / biosize; on = uio->uio_offset & (biosize - 1); if ((off_t)lbn * biosize >= filesize) { bcount = 0; } else if ((off_t)(lbn + 1) * biosize > filesize) { bcount = filesize - (off_t)lbn *biosize; } else { bcount = biosize; } nextlbn = lbn + 1; nextsize = MIN(biosize, filesize - nextlbn * biosize); SDT_PROBE4(fusefs, , io, read_bio_backend_start, biosize, (int)lbn, on, bcount); if (bcount < biosize) { /* If near EOF, don't do readahead */ err = bread(vp, lbn, bcount, NOCRED, &bp); } else if ((vp->v_mount->mnt_flag & MNT_NOCLUSTERR) == 0) { /* Try clustered read */ long totread = uio->uio_resid + on; seqcount = MIN(seqcount, data->max_readahead_blocks + 1); err = cluster_read(vp, filesize, lbn, bcount, NOCRED, totread, seqcount, 0, &bp); } else if (seqcount > 1 && data->max_readahead_blocks >= 1) { /* Try non-clustered readahead */ err = breadn(vp, lbn, bcount, &nextlbn, &nextsize, 1, NOCRED, &bp); } else { /* Just read what was requested */ err = bread(vp, lbn, bcount, NOCRED, &bp); } if (err) { brelse(bp); bp = NULL; break; } /* * on is the offset into the current bp. Figure out how many * bytes we can copy out of the bp. Note that bcount is * NOT DEV_BSIZE aligned. * * Then figure out how many bytes we can copy into the uio. */ n = 0; - if (on < bcount - bp->b_resid) - n = MIN((unsigned)(bcount - bp->b_resid - on), + if (on < bcount - (intptr_t)bp->b_fsprivate1) + n = MIN((unsigned)(bcount - (intptr_t)bp->b_fsprivate1 - on), uio->uio_resid); if (n > 0) { SDT_PROBE2(fusefs, , io, read_bio_backend_feed, n, bp); err = uiomove(bp->b_data + on, n, uio); } vfs_bio_brelse(bp, ioflag); SDT_PROBE4(fusefs, , io, read_bio_backend_end, err, uio->uio_resid, n, bp); - if (bp->b_resid > 0) { + if ((intptr_t)bp->b_fsprivate1 > 0) { /* Short read indicates EOF */ (void)fuse_vnode_setsize(vp, uio->uio_offset); + bp->b_fsprivate1 = (void*)0; break; } } return (err); } SDT_PROBE_DEFINE1(fusefs, , io, read_directbackend_start, "struct fuse_read_in*"); SDT_PROBE_DEFINE3(fusefs, , io, read_directbackend_complete, "struct fuse_dispatcher*", "struct fuse_read_in*", "struct uio*"); static int fuse_read_directbackend(struct vnode *vp, struct uio *uio, struct ucred *cred, struct fuse_filehandle *fufh) { struct fuse_data *data; struct fuse_dispatcher fdi; struct fuse_read_in *fri; int err = 0; data = fuse_get_mpdata(vp->v_mount); if (uio->uio_resid == 0) return (0); fdisp_init(&fdi, 0); /* * XXX In "normal" case we use an intermediate kernel buffer for * transmitting data from daemon's context to ours. Eventually, we should * get rid of this. Anyway, if the target uio lives in sysspace (we are * called from pageops), and the input data doesn't need kernel-side * processing (we are not called from readdir) we can already invoke * an optimized, "peer-to-peer" I/O routine. */ while (uio->uio_resid > 0) { fdi.iosize = sizeof(*fri); fdisp_make_vp(&fdi, FUSE_READ, vp, uio->uio_td, cred); fri = fdi.indata; fri->fh = fufh->fh_id; fri->offset = uio->uio_offset; fri->size = MIN(uio->uio_resid, fuse_get_mpdata(vp->v_mount)->max_read); if (fuse_libabi_geq(data, 7, 9)) { /* See comment regarding FUSE_WRITE_LOCKOWNER */ fri->read_flags = 0; fri->flags = fufh_type_2_fflags(fufh->fufh_type); } SDT_PROBE1(fusefs, , io, read_directbackend_start, fri); if ((err = fdisp_wait_answ(&fdi))) goto out; SDT_PROBE3(fusefs, , io, read_directbackend_complete, &fdi, fri, uio); if ((err = uiomove(fdi.answ, MIN(fri->size, fdi.iosize), uio))) break; if (fdi.iosize < fri->size) { /* * Short read. Should only happen at EOF or with * direct io. */ break; } } out: fdisp_destroy(&fdi); return (err); } static int fuse_write_directbackend(struct vnode *vp, struct uio *uio, struct ucred *cred, struct fuse_filehandle *fufh, off_t filesize, int ioflag, bool pages) { struct fuse_vnode_data *fvdat = VTOFUD(vp); struct fuse_data *data; struct fuse_write_in *fwi; struct fuse_write_out *fwo; struct fuse_dispatcher fdi; size_t chunksize; void *fwi_data; off_t as_written_offset; int diff; int err = 0; bool direct_io = fufh->fuse_open_flags & FOPEN_DIRECT_IO; uint32_t write_flags; data = fuse_get_mpdata(vp->v_mount); /* * Don't set FUSE_WRITE_LOCKOWNER in write_flags. It can't be set * accurately when using POSIX AIO, libfuse doesn't use it, and I'm not * aware of any file systems that do. It was an attempt to add * Linux-style mandatory locking to the FUSE protocol, but mandatory * locking is deprecated even on Linux. See Linux commit * f33321141b273d60cbb3a8f56a5489baad82ba5e . */ /* * Set FUSE_WRITE_CACHE whenever we don't know the uid, gid, and/or pid * that originated a write. For example when writing from the * writeback cache. I don't know of a single file system that cares, * but the protocol says we're supposed to do this. */ write_flags = !pages && ( (ioflag & IO_DIRECT) || !fsess_opt_datacache(vnode_mount(vp)) || fuse_data_cache_mode != FUSE_CACHE_WB) ? 0 : FUSE_WRITE_CACHE; if (uio->uio_resid == 0) return (0); if (ioflag & IO_APPEND) uio_setoffset(uio, filesize); if (vn_rlimit_fsize(vp, uio, uio->uio_td)) return (EFBIG); fdisp_init(&fdi, 0); while (uio->uio_resid > 0) { chunksize = MIN(uio->uio_resid, data->max_write); fdi.iosize = sizeof(*fwi) + chunksize; fdisp_make_vp(&fdi, FUSE_WRITE, vp, uio->uio_td, cred); fwi = fdi.indata; fwi->fh = fufh->fh_id; fwi->offset = uio->uio_offset; fwi->size = chunksize; fwi->write_flags = write_flags; if (fuse_libabi_geq(data, 7, 9)) { fwi->flags = fufh_type_2_fflags(fufh->fufh_type); fwi_data = (char *)fdi.indata + sizeof(*fwi); } else { fwi_data = (char *)fdi.indata + FUSE_COMPAT_WRITE_IN_SIZE; } if ((err = uiomove(fwi_data, chunksize, uio))) break; retry: err = fdisp_wait_answ(&fdi); if (err == ERESTART || err == EINTR || err == EWOULDBLOCK) { /* * Rewind the uio so dofilewrite will know it's * incomplete */ uio->uio_resid += fwi->size; uio->uio_offset -= fwi->size; /* * Change ERESTART into EINTR because we can't rewind * uio->uio_iov. Basically, once uiomove(9) has been * called, it's impossible to restart a syscall. */ if (err == ERESTART) err = EINTR; break; } else if (err) { break; } fwo = ((struct fuse_write_out *)fdi.answ); /* Adjust the uio in the case of short writes */ diff = fwi->size - fwo->size; as_written_offset = uio->uio_offset - diff; if (as_written_offset - diff > filesize && fuse_data_cache_mode != FUSE_CACHE_UC) fuse_vnode_setsize(vp, as_written_offset); if (as_written_offset - diff >= filesize) fvdat->flag &= ~FN_SIZECHANGE; if (diff < 0) { printf("WARNING: misbehaving FUSE filesystem " "wrote more data than we provided it\n"); err = EINVAL; break; } else if (diff > 0) { /* Short write */ if (!direct_io) { printf("WARNING: misbehaving FUSE filesystem: " "short writes are only allowed with " "direct_io\n"); } if (ioflag & IO_DIRECT) { /* Return early */ uio->uio_resid += diff; uio->uio_offset -= diff; break; } else { /* Resend the unwritten portion of data */ fdi.iosize = sizeof(*fwi) + diff; /* Refresh fdi without clearing data buffer */ fdisp_refresh_vp(&fdi, FUSE_WRITE, vp, uio->uio_td, cred); fwi = fdi.indata; MPASS2(fwi == fdi.indata, "FUSE dispatcher " "reallocated despite no increase in " "size?"); void *src = (char*)fwi_data + fwo->size; memmove(fwi_data, src, diff); fwi->fh = fufh->fh_id; fwi->offset = as_written_offset; fwi->size = diff; fwi->write_flags = write_flags; goto retry; } } } fdisp_destroy(&fdi); return (err); } SDT_PROBE_DEFINE6(fusefs, , io, write_biobackend_start, "int64_t", "int", "int", "struct uio*", "int", "bool"); SDT_PROBE_DEFINE2(fusefs, , io, write_biobackend_append_race, "long", "int"); SDT_PROBE_DEFINE2(fusefs, , io, write_biobackend_issue, "int", "struct buf*"); static int fuse_write_biobackend(struct vnode *vp, struct uio *uio, struct ucred *cred, struct fuse_filehandle *fufh, int ioflag, pid_t pid) { struct fuse_vnode_data *fvdat = VTOFUD(vp); struct buf *bp; daddr_t lbn; off_t filesize; int bcount; int n, on, seqcount, err = 0; bool last_page; const int biosize = fuse_iosize(vp); seqcount = ioflag >> IO_SEQSHIFT; KASSERT(uio->uio_rw == UIO_WRITE, ("fuse_write_biobackend mode")); if (vp->v_type != VREG) return (EIO); if (uio->uio_offset < 0) return (EINVAL); if (uio->uio_resid == 0) return (0); err = fuse_vnode_size(vp, &filesize, cred, curthread); if (err) return err; if (ioflag & IO_APPEND) uio_setoffset(uio, filesize); if (vn_rlimit_fsize(vp, uio, uio->uio_td)) return (EFBIG); do { bool direct_append, extending; if (fuse_isdeadfs(vp)) { err = ENXIO; break; } lbn = uio->uio_offset / biosize; on = uio->uio_offset & (biosize - 1); n = MIN((unsigned)(biosize - on), uio->uio_resid); again: /* Get or create a buffer for the write */ direct_append = uio->uio_offset == filesize && n; if (uio->uio_offset + n < filesize) { extending = false; if ((off_t)(lbn + 1) * biosize < filesize) { /* Not the file's last block */ bcount = biosize; } else { /* The file's last block */ bcount = filesize - (off_t)lbn * biosize; } } else { extending = true; bcount = on + n; } if (howmany(((off_t)lbn * biosize + on + n - 1), PAGE_SIZE) >= howmany(filesize, PAGE_SIZE)) last_page = true; else last_page = false; if (direct_append) { /* * Take care to preserve the buffer's B_CACHE state so * as not to cause an unnecessary read. */ bp = getblk(vp, lbn, on, PCATCH, 0, 0); if (bp != NULL) { uint32_t save = bp->b_flags & B_CACHE; allocbuf(bp, bcount); bp->b_flags |= save; } } else { bp = getblk(vp, lbn, bcount, PCATCH, 0, 0); } if (!bp) { err = EINTR; break; } if (extending) { /* * Extend file _after_ locking buffer so we won't race * with other readers */ err = fuse_vnode_setsize(vp, uio->uio_offset + n); filesize = uio->uio_offset + n; fvdat->flag |= FN_SIZECHANGE; if (err) { brelse(bp); break; } } SDT_PROBE6(fusefs, , io, write_biobackend_start, lbn, on, n, uio, bcount, direct_append); /* * Issue a READ if B_CACHE is not set. In special-append * mode, B_CACHE is based on the buffer prior to the write * op and is typically set, avoiding the read. If a read * is required in special append mode, the server will * probably send us a short-read since we extended the file * on our end, resulting in b_resid == 0 and, thusly, * B_CACHE getting set. * * We can also avoid issuing the read if the write covers * the entire buffer. We have to make sure the buffer state * is reasonable in this case since we will not be initiating * I/O. See the comments in kern/vfs_bio.c's getblk() for * more information. * * B_CACHE may also be set due to the buffer being cached * normally. */ if (on == 0 && n == bcount) { bp->b_flags |= B_CACHE; bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); fuse_io_strategy(vp, bp); if ((err = bp->b_error)) { brelse(bp); break; } } if (bp->b_wcred == NOCRED) bp->b_wcred = crhold(cred); /* * If dirtyend exceeds file size, chop it down. This should * not normally occur but there is an append race where it * might occur XXX, so we log it. * * If the chopping creates a reverse-indexed or degenerate * situation with dirtyoff/end, we 0 both of them. */ if (bp->b_dirtyend > bcount) { SDT_PROBE2(fusefs, , io, write_biobackend_append_race, (long)bp->b_blkno * biosize, bp->b_dirtyend - bcount); bp->b_dirtyend = bcount; } if (bp->b_dirtyoff >= bp->b_dirtyend) bp->b_dirtyoff = bp->b_dirtyend = 0; /* * If the new write will leave a contiguous dirty * area, just update the b_dirtyoff and b_dirtyend, * otherwise force a write rpc of the old dirty area. * * While it is possible to merge discontiguous writes due to * our having a B_CACHE buffer ( and thus valid read data * for the hole), we don't because it could lead to * significant cache coherency problems with multiple clients, * especially if locking is implemented later on. * * as an optimization we could theoretically maintain * a linked list of discontinuous areas, but we would still * have to commit them separately so there isn't much * advantage to it except perhaps a bit of asynchronization. */ if (bp->b_dirtyend > 0 && (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { /* * Yes, we mean it. Write out everything to "storage" * immediately, without hesitation. (Apart from other * reasons: the only way to know if a write is valid * if its actually written out.) */ SDT_PROBE2(fusefs, , io, write_biobackend_issue, 0, bp); bwrite(bp); if (bp->b_error == EINTR) { err = EINTR; break; } goto again; } err = uiomove((char *)bp->b_data + on, n, uio); if (err) { bp->b_ioflags |= BIO_ERROR; bp->b_error = err; brelse(bp); break; /* TODO: vfs_bio_clrbuf like ffs_write does? */ } /* * Only update dirtyoff/dirtyend if not a degenerate * condition. */ if (n) { if (bp->b_dirtyend > 0) { bp->b_dirtyoff = MIN(on, bp->b_dirtyoff); bp->b_dirtyend = MAX((on + n), bp->b_dirtyend); } else { bp->b_dirtyoff = on; bp->b_dirtyend = on + n; } vfs_bio_set_valid(bp, on, n); } vfs_bio_set_flags(bp, ioflag); if (ioflag & IO_SYNC) { SDT_PROBE2(fusefs, , io, write_biobackend_issue, 2, bp); err = bwrite(bp); } else if (vm_page_count_severe() || buf_dirty_count_severe() || (ioflag & IO_ASYNC)) { bp->b_flags |= B_CLUSTEROK; SDT_PROBE2(fusefs, , io, write_biobackend_issue, 3, bp); bawrite(bp); } else if (on == 0 && n == bcount) { if ((vp->v_mount->mnt_flag & MNT_NOCLUSTERW) == 0) { bp->b_flags |= B_CLUSTEROK; SDT_PROBE2(fusefs, , io, write_biobackend_issue, 4, bp); cluster_write(vp, bp, filesize, seqcount, 0); } else { SDT_PROBE2(fusefs, , io, write_biobackend_issue, 5, bp); bawrite(bp); } } else if (ioflag & IO_DIRECT) { bp->b_flags |= B_CLUSTEROK; SDT_PROBE2(fusefs, , io, write_biobackend_issue, 6, bp); bawrite(bp); } else { bp->b_flags &= ~B_CLUSTEROK; SDT_PROBE2(fusefs, , io, write_biobackend_issue, 7, bp); bdwrite(bp); } if (err) break; } while (uio->uio_resid > 0 && n > 0); return (err); } int fuse_io_strategy(struct vnode *vp, struct buf *bp) { struct fuse_vnode_data *fvdat = VTOFUD(vp); struct fuse_filehandle *fufh; struct ucred *cred; struct uio *uiop; struct uio uio; struct iovec io; off_t filesize; int error = 0; int fflag; /* We don't know the true pid when we're dealing with the cache */ pid_t pid = 0; const int biosize = fuse_iosize(vp); MPASS(vp->v_type == VREG || vp->v_type == VDIR); MPASS(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE); fflag = bp->b_iocmd == BIO_READ ? FREAD : FWRITE; cred = bp->b_iocmd == BIO_READ ? bp->b_rcred : bp->b_wcred; error = fuse_filehandle_getrw(vp, fflag, &fufh, cred, pid); if (bp->b_iocmd == BIO_READ && error == EBADF) { /* * This may be a read-modify-write operation on a cached file * opened O_WRONLY. The FUSE protocol allows this. */ error = fuse_filehandle_get(vp, FWRITE, &fufh, cred, pid); } if (error) { printf("FUSE: strategy: filehandles are closed\n"); bp->b_ioflags |= BIO_ERROR; bp->b_error = error; bufdone(bp); return (error); } uiop = &uio; uiop->uio_iov = &io; uiop->uio_iovcnt = 1; uiop->uio_segflg = UIO_SYSSPACE; uiop->uio_td = curthread; /* * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We * do this here so we do not have to do it in all the code that * calls us. */ bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; KASSERT(!(bp->b_flags & B_DONE), ("fuse_io_strategy: bp %p already marked done", bp)); if (bp->b_iocmd == BIO_READ) { + ssize_t left; + io.iov_len = uiop->uio_resid = bp->b_bcount; io.iov_base = bp->b_data; uiop->uio_rw = UIO_READ; uiop->uio_offset = ((off_t)bp->b_lblkno) * biosize; error = fuse_read_directbackend(vp, uiop, cred, fufh); + left = uiop->uio_resid; + /* + * Store the amount we failed to read in the buffer's private + * field, so callers can truncate the file if necessary' + */ + bp->b_fsprivate1 = (void*)(intptr_t)left; if (!error && uiop->uio_resid) { int nread = bp->b_bcount - uiop->uio_resid; - int left = uiop->uio_resid; bzero((char *)bp->b_data + nread, left); if (fuse_data_cache_mode != FUSE_CACHE_WB || (fvdat->flag & FN_SIZECHANGE) == 0) { /* * A short read with no error, when not using * direct io, and when no writes are cached, * indicates EOF. Update the file size * accordingly. We must still bzero the * remaining buffer so uninitialized data * doesn't get exposed by a future truncate * that extends the file. * - * XXX To prevent lock order problems, we must + * To prevent lock order problems, we must * truncate the file upstack */ SDT_PROBE2(fusefs, , io, trace, 1, "Short read of a clean file"); + uiop->uio_resid = 0; } else { /* * If dirty writes _are_ cached beyond EOF, * that indicates a newly created hole that the * server doesn't know about. * XXX: we don't currently track whether dirty * writes are cached beyond EOF, before EOF, or * both. */ SDT_PROBE2(fusefs, , io, trace, 1, "Short read of a dirty file"); uiop->uio_resid = 0; } } if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_error = error; } } else { /* * Setup for actual write */ error = fuse_vnode_size(vp, &filesize, cred, curthread); if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_error = error; bufdone(bp); return (error); } if ((off_t)bp->b_lblkno * biosize + bp->b_dirtyend > filesize) bp->b_dirtyend = filesize - (off_t)bp->b_lblkno * biosize; if (bp->b_dirtyend > bp->b_dirtyoff) { io.iov_len = uiop->uio_resid = bp->b_dirtyend - bp->b_dirtyoff; uiop->uio_offset = (off_t)bp->b_lblkno * biosize + bp->b_dirtyoff; io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; uiop->uio_rw = UIO_WRITE; error = fuse_write_directbackend(vp, uiop, cred, fufh, filesize, 0, false); if (error == EINTR || error == ETIMEDOUT) { bp->b_flags &= ~(B_INVAL | B_NOCACHE); if ((bp->b_flags & B_PAGING) == 0) { bdirty(bp); bp->b_flags &= ~B_DONE; } if ((error == EINTR || error == ETIMEDOUT) && (bp->b_flags & B_ASYNC) == 0) bp->b_flags |= B_EINTR; } else { if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_flags |= B_INVAL; bp->b_error = error; } bp->b_dirtyoff = bp->b_dirtyend = 0; } } else { bp->b_resid = 0; bufdone(bp); return (0); } } bp->b_resid = uiop->uio_resid; bufdone(bp); return (error); } int fuse_io_flushbuf(struct vnode *vp, int waitfor, struct thread *td) { return (vn_fsync_buf(vp, waitfor)); } /* * Flush and invalidate all dirty buffers. If another process is already * doing the flush, just wait for completion. */ int fuse_io_invalbuf(struct vnode *vp, struct thread *td) { struct fuse_vnode_data *fvdat = VTOFUD(vp); int error = 0; if (vp->v_iflag & VI_DOOMED) return 0; ASSERT_VOP_ELOCKED(vp, "fuse_io_invalbuf"); while (fvdat->flag & FN_FLUSHINPROG) { struct proc *p = td->td_proc; if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) return EIO; fvdat->flag |= FN_FLUSHWANT; tsleep(&fvdat->flag, PRIBIO + 2, "fusevinv", 2 * hz); error = 0; if (p != NULL) { PROC_LOCK(p); if (SIGNOTEMPTY(p->p_siglist) || SIGNOTEMPTY(td->td_siglist)) error = EINTR; PROC_UNLOCK(p); } if (error == EINTR) return EINTR; } fvdat->flag |= FN_FLUSHINPROG; if (vp->v_bufobj.bo_object != NULL) { VM_OBJECT_WLOCK(vp->v_bufobj.bo_object); vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object); } error = vinvalbuf(vp, V_SAVE, PCATCH, 0); while (error) { if (error == ERESTART || error == EINTR) { fvdat->flag &= ~FN_FLUSHINPROG; if (fvdat->flag & FN_FLUSHWANT) { fvdat->flag &= ~FN_FLUSHWANT; wakeup(&fvdat->flag); } return EINTR; } error = vinvalbuf(vp, V_SAVE, PCATCH, 0); } fvdat->flag &= ~FN_FLUSHINPROG; if (fvdat->flag & FN_FLUSHWANT) { fvdat->flag &= ~FN_FLUSHWANT; wakeup(&fvdat->flag); } return (error); } Index: projects/fuse2/tests/sys/fs/fusefs/io.cc =================================================================== --- projects/fuse2/tests/sys/fs/fusefs/io.cc (revision 349331) +++ projects/fuse2/tests/sys/fs/fusefs/io.cc (revision 349332) @@ -1,374 +1,466 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2019 The FreeBSD Foundation * * This software was developed by BFF Storage Systems, LLC under sponsorship * from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ extern "C" { +#include + #include #include #include } #include "mockfs.hh" #include "utils.hh" /* * For testing I/O like fsx does, but deterministically and without a real * underlying file system * * TODO: after fusefs gains the options to select cache mode for each mount * point, run each of these tests for all cache modes. */ using namespace testing; const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const uint64_t ino = 42; static void compare(const void *tbuf, const void *controlbuf, off_t baseofs, ssize_t size) { int i; for (i = 0; i < size; i++) { if (((const char*)tbuf)[i] != ((const char*)controlbuf)[i]) { off_t ofs = baseofs + i; FAIL() << "miscompare at offset " << std::hex << std::showbase << ofs << ". expected = " << std::setw(2) << (unsigned)((const uint8_t*)controlbuf)[i] << " got = " << (unsigned)((const uint8_t*)tbuf)[i]; } } } class Io: public FuseTest, public WithParamInterface> { public: int m_backing_fd, m_control_fd, m_test_fd; +off_t m_filesize; Io(): m_backing_fd(-1), m_control_fd(-1) {}; void SetUp() { + m_filesize = 0; m_backing_fd = open("backing_file", O_RDWR | O_CREAT | O_TRUNC, 0644); if (m_backing_fd < 0) FAIL() << strerror(errno); m_control_fd = open("control", O_RDWR | O_CREAT | O_TRUNC, 0644); if (m_control_fd < 0) FAIL() << strerror(errno); srandom(22'9'1982); // Seed with my birthday m_init_flags = get<0>(GetParam()); m_maxwrite = get<1>(GetParam()); m_async = get<2>(GetParam()); FuseTest::SetUp(); if (IsSkipped()) return; expect_lookup(RELPATH, ino, S_IFREG | 0644, 0, 1); expect_open(ino, 0, 1); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_WRITE && in.header.nodeid == ino); }, Eq(true)), _) ).WillRepeatedly(Invoke(ReturnImmediate([=](auto in, auto& out) { const char *buf = (const char*)in.body.bytes + sizeof(struct fuse_write_in); ssize_t isize = in.body.write.size; off_t iofs = in.body.write.offset; ASSERT_EQ(isize, pwrite(m_backing_fd, buf, isize, iofs)) << strerror(errno); SET_OUT_HEADER_LEN(out, write); out.body.write.size = isize; }))); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_READ && in.header.nodeid == ino); }, Eq(true)), _) ).WillRepeatedly(Invoke(ReturnImmediate([=](auto in, auto& out) { ssize_t isize = in.body.write.size; off_t iofs = in.body.write.offset; void *buf = out.body.bytes; + ssize_t osize; - ASSERT_LE(0, pread(m_backing_fd, buf, isize, iofs)) - << strerror(errno); - out.header.len = sizeof(struct fuse_out_header) + isize; + osize = pread(m_backing_fd, buf, isize, iofs); + ASSERT_LE(0, osize) << strerror(errno); + out.header.len = sizeof(struct fuse_out_header) + osize; }))); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { uint32_t valid = FATTR_SIZE | FATTR_FH; return (in.header.opcode == FUSE_SETATTR && in.header.nodeid == ino && in.body.setattr.valid == valid); }, Eq(true)), _) ).WillRepeatedly(Invoke(ReturnImmediate([=](auto in, auto& out) { ASSERT_EQ(0, ftruncate(m_backing_fd, in.body.setattr.size)) << strerror(errno); SET_OUT_HEADER_LEN(out, attr); out.body.attr.attr.ino = ino; out.body.attr.attr.mode = S_IFREG | 0755; out.body.attr.attr.size = in.body.setattr.size; out.body.attr.attr_valid = UINT64_MAX; }))); /* Any test that close()s will send FUSE_FLUSH and FUSE_RELEASE */ EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_FLUSH && in.header.nodeid == ino); }, Eq(true)), _) ).WillRepeatedly(Invoke(ReturnErrno(0))); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_RELEASE && in.header.nodeid == ino); }, Eq(true)), _) ).WillRepeatedly(Invoke(ReturnErrno(0))); m_test_fd = open(FULLPATH, O_RDWR ); EXPECT_LE(0, m_test_fd) << strerror(errno); } void TearDown() { if (m_backing_fd >= 0) close(m_backing_fd); if (m_control_fd >= 0) close(m_control_fd); FuseTest::TearDown(); /* Deliberately leak test_fd */ } void do_ftruncate(off_t offs) { ASSERT_EQ(0, ftruncate(m_test_fd, offs)) << strerror(errno); ASSERT_EQ(0, ftruncate(m_control_fd, offs)) << strerror(errno); + m_filesize = offs; } +void do_mapread(ssize_t size, off_t offs) +{ + void *control_buf, *p; + off_t pg_offset, page_mask; + size_t map_size; + + page_mask = getpagesize() - 1; + pg_offset = offs & page_mask; + map_size = pg_offset + size; + + p = mmap(NULL, map_size, PROT_READ, MAP_FILE | MAP_SHARED, m_test_fd, + offs - pg_offset); + ASSERT_NE(p, MAP_FAILED) << strerror(errno); + + control_buf = malloc(size); + ASSERT_NE(NULL, control_buf) << strerror(errno); + + ASSERT_EQ(size, pread(m_control_fd, control_buf, size, offs)) + << strerror(errno); + + compare((void*)((char*)p + pg_offset), control_buf, offs, size); + + ASSERT_EQ(0, munmap(p, map_size)) << strerror(errno); + free(control_buf); +} + void do_read(ssize_t size, off_t offs) { void *test_buf, *control_buf; + ssize_t r; test_buf = malloc(size); ASSERT_NE(NULL, test_buf) << strerror(errno); control_buf = malloc(size); ASSERT_NE(NULL, control_buf) << strerror(errno); - ASSERT_EQ(size, pread(m_test_fd, test_buf, size, offs)) - << strerror(errno); - ASSERT_EQ(size, pread(m_control_fd, control_buf, size, offs)) - << strerror(errno); + errno = 0; + r = pread(m_test_fd, test_buf, size, offs); + ASSERT_NE(-1, r) << strerror(errno); + ASSERT_EQ(size, r) << "unexpected short read"; + r = pread(m_control_fd, control_buf, size, offs); + ASSERT_NE(-1, r) << strerror(errno); + ASSERT_EQ(size, r) << "unexpected short read"; compare(test_buf, control_buf, offs, size); free(control_buf); free(test_buf); } +void do_mapwrite(ssize_t size, off_t offs) +{ + char *buf; + void *p; + off_t pg_offset, page_mask; + size_t map_size; + long i; + + page_mask = getpagesize() - 1; + pg_offset = offs & page_mask; + map_size = pg_offset + size; + + buf = (char*)malloc(size); + ASSERT_NE(NULL, buf) << strerror(errno); + for (i=0; i < size; i++) + buf[i] = random(); + + if (offs + size > m_filesize) { + /* + * Must manually extend. vm_mmap_vnode will not implicitly + * extend a vnode + */ + do_ftruncate(offs + size); + } + + p = mmap(NULL, map_size, PROT_READ | PROT_WRITE, + MAP_FILE | MAP_SHARED, m_test_fd, offs - pg_offset); + ASSERT_NE(p, MAP_FAILED) << strerror(errno); + + bcopy(buf, (char*)p + pg_offset, size); + ASSERT_EQ(size, pwrite(m_control_fd, buf, size, offs)) + << strerror(errno); + + free(buf); + ASSERT_EQ(0, munmap(p, map_size)) << strerror(errno); +} + void do_write(ssize_t size, off_t offs) { char *buf; long i; buf = (char*)malloc(size); ASSERT_NE(NULL, buf) << strerror(errno); for (i=0; i < size; i++) buf[i] = random(); ASSERT_EQ(size, pwrite(m_test_fd, buf, size, offs )) << strerror(errno); ASSERT_EQ(size, pwrite(m_control_fd, buf, size, offs)) << strerror(errno); + m_filesize = std::max(m_filesize, offs + size); + + free(buf); } }; /* * Extend a file with dirty data in the last page of the last block. * * fsx -WR -P /tmp -S8 -N3 fsx.bin */ TEST_P(Io, extend_from_dirty_page) { off_t wofs = 0x21a0; ssize_t wsize = 0xf0a8; off_t rofs = 0xb284; ssize_t rsize = 0x9b22; off_t truncsize = 0x28702; do_write(wsize, wofs); do_ftruncate(truncsize); do_read(rsize, rofs); } /* + * mapwrite into a newly extended part of a file. + * + * fsx -c 100 -i 100 -l 524288 -o 131072 -N5 -P /tmp -S19 fsx.bin + */ +TEST_P(Io, extend_by_mapwrite) +{ + do_mapwrite(0x849e, 0x29a3a); /* [0x29a3a, 0x31ed7] */ + do_mapwrite(0x3994, 0x3c7d8); /* [0x3c7d8, 0x4016b] */ + do_read(0xf556, 0x30c16); /* [0x30c16, 0x4016b] */ +} + +/* * When writing the last page of a file, it must be written synchronously. * Otherwise the cached page can become invalid by a subsequent extend * operation. * * fsx -WR -P /tmp -S642 -N3 fsx.bin */ TEST_P(Io, last_page) { - off_t wofs0 = 0x1134f; - ssize_t wsize0 = 0xcc77; - off_t wofs1 = 0x2096a; - ssize_t wsize1 = 0xdfa7; - off_t rofs = 0x1a3aa; - ssize_t rsize = 0xb5b7; + do_write(0xcc77, 0x1134f); /* [0x1134f, 0x1dfc5] */ + do_write(0xdfa7, 0x2096a); /* [0x2096a, 0x2e910] */ + do_read(0xb5b7, 0x1a3aa); /* [0x1a3aa, 0x25960] */ +} - do_write(wsize0, wofs0); - do_write(wsize1, wofs1); - do_read(rsize, rofs); +/* + * Read a hole using mmap + * + * fsx -c 100 -i 100 -l 524288 -o 131072 -N11 -P /tmp -S14 fsx.bin + */ +TEST_P(Io, mapread_hole) +{ + do_write(0x123b7, 0xf205); /* [0xf205, 0x215bb] */ + do_mapread(0xeeea, 0x2f4c); /* [0x2f4c, 0x11e35] */ } /* * Read a hole from a block that contains some cached data. * * fsx -WR -P /tmp -S55 fsx.bin */ TEST_P(Io, read_hole_from_cached_block) { off_t wofs = 0x160c5; ssize_t wsize = 0xa996; off_t rofs = 0x472e; ssize_t rsize = 0xd8d5; do_write(wsize, wofs); do_read(rsize, rofs); } /* * Truncating a file into a dirty buffer should not causing anything untoward * to happen when that buffer is eventually flushed. * * fsx -WR -P /tmp -S839 -d -N6 fsx.bin */ TEST_P(Io, truncate_into_dirty_buffer) { off_t wofs0 = 0x3bad7; ssize_t wsize0 = 0x4529; off_t wofs1 = 0xc30d; ssize_t wsize1 = 0x5f77; off_t truncsize0 = 0x10916; off_t rofs = 0xdf17; ssize_t rsize = 0x29ff; off_t truncsize1 = 0x152b4; do_write(wsize0, wofs0); do_write(wsize1, wofs1); do_ftruncate(truncsize0); do_read(rsize, rofs); do_ftruncate(truncsize1); close(m_test_fd); } /* * Truncating a file into a dirty buffer should not causing anything untoward * to happen when that buffer is eventually flushed, even when the buffer's * dirty_off is > 0. * * Based on this command with a few steps removed: * fsx -WR -P /tmp -S677 -d -N8 fsx.bin */ TEST_P(Io, truncate_into_dirty_buffer2) { off_t truncsize0 = 0x344f3; off_t wofs = 0x2790c; ssize_t wsize = 0xd86a; off_t truncsize1 = 0x2de38; off_t rofs2 = 0x1fd7a; ssize_t rsize2 = 0xc594; off_t truncsize2 = 0x31e71; /* Sets the file size to something larger than the next write */ do_ftruncate(truncsize0); /* * Creates a dirty buffer. The part in lbn 2 doesn't flush * synchronously. */ do_write(wsize, wofs); /* Truncates part of the dirty buffer created in step 2 */ do_ftruncate(truncsize1); /* XXX ?I don't know why this is necessary? */ do_read(rsize2, rofs2); /* Truncates the dirty buffer */ do_ftruncate(truncsize2); close(m_test_fd); } /* * Regression test for a bug introduced in r348931 * * Sequence of operations: * 1) The first write reads lbn so it can modify it * 2) The first write flushes lbn 3 immediately because it's the end of file * 3) The first write then flushes lbn 4 because it's the end of the file * 4) The second write modifies the cached versions of lbn 3 and 4 * 5) The third write's getblkx invalidates lbn 4's B_CACHE because it's * extending the buffer. Then it flushes lbn 4 because B_DELWRI was set but * B_CACHE was clear. * 6) fuse_write_biobackend erroneously called vfs_bio_clrbuf, putting the * buffer into a weird write-only state. All read operations would return * 0. Writes were apparently still processed, because the buffer's contents * were correct when examined in a core dump. * 7) The third write reads lbn 4 because cache is clear * 9) uiomove dutifully copies new data into the buffer * 10) The buffer's dirty is flushed to lbn 4 * 11) The read returns all zeros because of step 6. * * Based on: * fsx -WR -l 524388 -o 131072 -P /tmp -S6456 -q fsx.bin */ TEST_P(Io, resize_a_valid_buffer_while_extending) { do_write(0x14530, 0x36ee6); /* [0x36ee6, 0x4b415] */ do_write(0x1507c, 0x33256); /* [0x33256, 0x482d1] */ do_write(0x175c, 0x4c03d); /* [0x4c03d, 0x4d798] */ do_read(0xe277, 0x3599c); /* [0x3599c, 0x43c12] */ close(m_test_fd); } INSTANTIATE_TEST_CASE_P(Io, Io, Combine(Values(0, FUSE_ASYNC_READ), /* m_init_flags */ Values(0x1000, 0x10000, 0x20000), /* m_maxwrite */ Bool())); /* m_async */