Index: projects/fuse2/sys/fs/fuse/fuse_io.c =================================================================== --- projects/fuse2/sys/fs/fuse/fuse_io.c (revision 349377) +++ projects/fuse2/sys/fs/fuse/fuse_io.c (revision 349378) @@ -1,1075 +1,1096 @@ /*- * 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" +/* + * Set in a struct buf to indicate that the write came from the buffer cache + * and the originating cred and pid are no longer known. + */ +#define B_FUSEFS_WRITE_CACHE B_FS_FLAG1 + 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*"); -SDT_PROBE_DEFINE5(fusefs, , io, io_dispatch_filehandles_closed, "struct vnode*", - "struct uio*", "int", "bool", "struct ucred*"); +SDT_PROBE_DEFINE4(fusefs, , io, io_dispatch_filehandles_closed, "struct vnode*", + "struct uio*", "int", "struct ucred*"); int -fuse_io_dispatch(struct vnode *vp, struct uio *uio, int ioflag, bool pages, +fuse_io_dispatch(struct vnode *vp, struct uio *uio, int ioflag, 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) { - SDT_PROBE5(fusefs, , io, io_dispatch_filehandles_closed, - vp, uio, ioflag, pages, cred); + SDT_PROBE4(fusefs, , io, io_dispatch_filehandles_closed, + vp, uio, ioflag, cred); 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); + KASSERT((ioflag & (IO_VMIO | IO_DIRECT)) != + (IO_VMIO | IO_DIRECT), + ("IO_DIRECT used for a cache flush?")); + /* Invalidate the write cache when writing directly */ + v_inval_buf_range(vp, start, end, iosize); err = fuse_write_directbackend(vp, uio, cred, fufh, - filesize, ioflag, pages); + filesize, ioflag, false); } 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 - (intptr_t)bp->b_fsprivate1) - n = MIN((unsigned)(bcount - (intptr_t)bp->b_fsprivate1 - on), + if (on < bcount - bp->b_resid) + n = MIN((unsigned)(bcount - bp->b_resid - 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 ((intptr_t)bp->b_fsprivate1 > 0) { + if (bp->b_resid > 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_resid > 0) { + /* + * Short read indicates EOF. Update file size + * from the server and try again. + */ + SDT_PROBE2(fusefs, , io, trace, 1, + "Short read during a RMW"); + brelse(bp); + err = fuse_vnode_size(vp, &filesize, cred, + curthread); + if (err) + break; + else + goto again; + } } 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); + bp->b_flags |= B_FUSEFS_WRITE_CACHE; if (ioflag & IO_SYNC) { SDT_PROBE2(fusefs, , io, write_biobackend_issue, 2, bp); + if (!(ioflag & IO_VMIO)) + bp->b_flags &= ~B_FUSEFS_WRITE_CACHE; 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); - bp->b_fsprivate1 = (void*)(intptr_t)0; 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; + left = uiop->uio_resid; bzero((char *)bp->b_data + nread, left); - if (fuse_data_cache_mode != FUSE_CACHE_WB || - (fvdat->flag & FN_SIZECHANGE) == 0) { + if ((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. + * indicates EOF caused by a server-side + * truncation. Clear the attr cache so we'll + * pick up the new file size and timestamps. + * + * We must still bzero the remaining buffer so + * uninitialized data doesn't get exposed by a + * future truncate that extends the file. * * To prevent lock order problems, we must - * truncate the file upstack + * truncate the file upstack, not here. */ SDT_PROBE2(fusefs, , io, trace, 1, "Short read of a clean file"); - uiop->uio_resid = 0; + fuse_vnode_clear_attr_cache(vp); } else { /* * If dirty writes _are_ cached beyond EOF, * that indicates a newly created hole that the - * server doesn't know about. + * server doesn't know about. Those don't pose + * any problem. * 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; + bool pages = bp->b_flags & B_FUSEFS_WRITE_CACHE; error = fuse_write_directbackend(vp, uiop, cred, fufh, - filesize, 0, false); + filesize, 0, pages); 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/sys/fs/fuse/fuse_io.h =================================================================== --- projects/fuse2/sys/fs/fuse/fuse_io.h (revision 349377) +++ projects/fuse2/sys/fs/fuse/fuse_io.h (revision 349378) @@ -1,69 +1,69 @@ /*- * 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. * * $FreeBSD$ */ #ifndef _FUSE_IO_H_ #define _FUSE_IO_H_ -int fuse_io_dispatch(struct vnode *vp, struct uio *uio, int ioflag, bool pages, +int fuse_io_dispatch(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred, pid_t pid); int fuse_io_strategy(struct vnode *vp, struct buf *bp); int fuse_io_flushbuf(struct vnode *vp, int waitfor, struct thread *td); int fuse_io_invalbuf(struct vnode *vp, struct thread *td); #endif /* _FUSE_IO_H_ */ Index: projects/fuse2/sys/fs/fuse/fuse_vnops.c =================================================================== --- projects/fuse2/sys/fs/fuse/fuse_vnops.c (revision 349377) +++ projects/fuse2/sys/fs/fuse/fuse_vnops.c (revision 349378) @@ -1,2670 +1,2499 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 2007-2009 Google Inc. and Amit Singh * 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 #include #include #include #include "fuse.h" #include "fuse_file.h" #include "fuse_internal.h" #include "fuse_ipc.h" #include "fuse_node.h" #include "fuse_io.h" #include /* Maximum number of hardlinks to a single FUSE file */ #define FUSE_LINK_MAX UINT32_MAX SDT_PROVIDER_DECLARE(fusefs); /* * Fuse trace probe: * arg0: verbosity. Higher numbers give more verbose messages * arg1: Textual message */ SDT_PROBE_DEFINE2(fusefs, , vnops, trace, "int", "char*"); /* vnode ops */ static vop_access_t fuse_vnop_access; static vop_advlock_t fuse_vnop_advlock; static vop_bmap_t fuse_vnop_bmap; static vop_close_t fuse_fifo_close; static vop_close_t fuse_vnop_close; static vop_create_t fuse_vnop_create; static vop_deleteextattr_t fuse_vnop_deleteextattr; static vop_fdatasync_t fuse_vnop_fdatasync; static vop_fsync_t fuse_vnop_fsync; static vop_getattr_t fuse_vnop_getattr; static vop_getextattr_t fuse_vnop_getextattr; static vop_inactive_t fuse_vnop_inactive; static vop_link_t fuse_vnop_link; static vop_listextattr_t fuse_vnop_listextattr; static vop_lookup_t fuse_vnop_lookup; static vop_mkdir_t fuse_vnop_mkdir; static vop_mknod_t fuse_vnop_mknod; static vop_open_t fuse_vnop_open; static vop_pathconf_t fuse_vnop_pathconf; static vop_read_t fuse_vnop_read; static vop_readdir_t fuse_vnop_readdir; static vop_readlink_t fuse_vnop_readlink; static vop_reclaim_t fuse_vnop_reclaim; static vop_remove_t fuse_vnop_remove; static vop_rename_t fuse_vnop_rename; static vop_rmdir_t fuse_vnop_rmdir; static vop_setattr_t fuse_vnop_setattr; static vop_setextattr_t fuse_vnop_setextattr; static vop_strategy_t fuse_vnop_strategy; static vop_symlink_t fuse_vnop_symlink; static vop_write_t fuse_vnop_write; static vop_getpages_t fuse_vnop_getpages; -static vop_putpages_t fuse_vnop_putpages; static vop_print_t fuse_vnop_print; static vop_vptofh_t fuse_vnop_vptofh; struct vop_vector fuse_fifoops = { .vop_default = &fifo_specops, .vop_access = fuse_vnop_access, .vop_close = fuse_fifo_close, .vop_fsync = fuse_vnop_fsync, .vop_getattr = fuse_vnop_getattr, .vop_inactive = fuse_vnop_inactive, .vop_pathconf = fuse_vnop_pathconf, .vop_print = fuse_vnop_print, .vop_read = VOP_PANIC, .vop_reclaim = fuse_vnop_reclaim, .vop_setattr = fuse_vnop_setattr, .vop_write = VOP_PANIC, .vop_vptofh = fuse_vnop_vptofh, }; struct vop_vector fuse_vnops = { .vop_allocate = VOP_EINVAL, .vop_default = &default_vnodeops, .vop_access = fuse_vnop_access, .vop_advlock = fuse_vnop_advlock, .vop_bmap = fuse_vnop_bmap, .vop_close = fuse_vnop_close, .vop_create = fuse_vnop_create, .vop_deleteextattr = fuse_vnop_deleteextattr, .vop_fsync = fuse_vnop_fsync, .vop_fdatasync = fuse_vnop_fdatasync, .vop_getattr = fuse_vnop_getattr, .vop_getextattr = fuse_vnop_getextattr, .vop_inactive = fuse_vnop_inactive, /* * TODO: implement vop_ioctl after upgrading to protocol 7.16. * FUSE_IOCTL was added in 7.11, but 32-bit compat is broken until * 7.16. */ .vop_link = fuse_vnop_link, .vop_listextattr = fuse_vnop_listextattr, .vop_lookup = fuse_vnop_lookup, .vop_mkdir = fuse_vnop_mkdir, .vop_mknod = fuse_vnop_mknod, .vop_open = fuse_vnop_open, .vop_pathconf = fuse_vnop_pathconf, /* * TODO: implement vop_poll after upgrading to protocol 7.21. * FUSE_POLL was added in protocol 7.11, but it's kind of broken until * 7.21, which adds the ability for the client to choose which poll * events it wants, and for a client to deregister a file handle */ .vop_read = fuse_vnop_read, .vop_readdir = fuse_vnop_readdir, .vop_readlink = fuse_vnop_readlink, .vop_reclaim = fuse_vnop_reclaim, .vop_remove = fuse_vnop_remove, .vop_rename = fuse_vnop_rename, .vop_rmdir = fuse_vnop_rmdir, .vop_setattr = fuse_vnop_setattr, .vop_setextattr = fuse_vnop_setextattr, .vop_strategy = fuse_vnop_strategy, .vop_symlink = fuse_vnop_symlink, .vop_write = fuse_vnop_write, .vop_getpages = fuse_vnop_getpages, - .vop_putpages = fuse_vnop_putpages, .vop_print = fuse_vnop_print, .vop_vptofh = fuse_vnop_vptofh, }; u_long fuse_lookup_cache_hits = 0; SYSCTL_ULONG(_vfs_fusefs, OID_AUTO, lookup_cache_hits, CTLFLAG_RD, &fuse_lookup_cache_hits, 0, "number of positive cache hits in lookup"); u_long fuse_lookup_cache_misses = 0; SYSCTL_ULONG(_vfs_fusefs, OID_AUTO, lookup_cache_misses, CTLFLAG_RD, &fuse_lookup_cache_misses, 0, "number of cache misses in lookup"); /* * XXX: This feature is highly experimental and can bring to instabilities, * needs revisiting before to be enabled by default. */ static int fuse_reclaim_revoked = 0; SYSCTL_INT(_vfs_fusefs, OID_AUTO, reclaim_revoked, CTLFLAG_RW, &fuse_reclaim_revoked, 0, ""); uma_zone_t fuse_pbuf_zone; #define fuse_vm_page_lock(m) vm_page_lock((m)); #define fuse_vm_page_unlock(m) vm_page_unlock((m)); #define fuse_vm_page_lock_queues() ((void)0) #define fuse_vm_page_unlock_queues() ((void)0) /* Check permission for extattr operations, much like extattr_check_cred */ static int fuse_extattr_check_cred(struct vnode *vp, int ns, struct ucred *cred, struct thread *td, accmode_t accmode) { struct mount *mp = vnode_mount(vp); struct fuse_data *data = fuse_get_mpdata(mp); /* * Kernel-invoked always succeeds. */ if (cred == NOCRED) return (0); /* * Do not allow privileged processes in jail to directly manipulate * system attributes. */ switch (ns) { case EXTATTR_NAMESPACE_SYSTEM: if (data->dataflags & FSESS_DEFAULT_PERMISSIONS) { return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); } /* FALLTHROUGH */ case EXTATTR_NAMESPACE_USER: return (fuse_internal_access(vp, accmode, td, cred)); default: return (EPERM); } } /* Get a filehandle for a directory */ static int fuse_filehandle_get_dir(struct vnode *vp, struct fuse_filehandle **fufhp, struct ucred *cred, pid_t pid) { if (fuse_filehandle_get(vp, FREAD, fufhp, cred, pid) == 0) return 0; return fuse_filehandle_get(vp, FEXEC, fufhp, cred, pid); } /* Send FUSE_FLUSH for this vnode */ static int fuse_flush(struct vnode *vp, struct ucred *cred, pid_t pid, int fflag) { struct fuse_flush_in *ffi; struct fuse_filehandle *fufh; struct fuse_dispatcher fdi; struct thread *td = curthread; struct mount *mp = vnode_mount(vp); int err; if (!fsess_isimpl(vnode_mount(vp), FUSE_FLUSH)) return 0; err = fuse_filehandle_getrw(vp, fflag, &fufh, cred, pid); if (err) return err; fdisp_init(&fdi, sizeof(*ffi)); fdisp_make_vp(&fdi, FUSE_FLUSH, vp, td, cred); ffi = fdi.indata; ffi->fh = fufh->fh_id; /* * If the file has a POSIX lock then we're supposed to set lock_owner. * If not, then lock_owner is undefined. So we may as well always set * it. */ ffi->lock_owner = td->td_proc->p_pid; err = fdisp_wait_answ(&fdi); if (err == ENOSYS) { fsess_set_notimpl(mp, FUSE_FLUSH); err = 0; } fdisp_destroy(&fdi); return err; } /* Close wrapper for fifos. */ static int fuse_fifo_close(struct vop_close_args *ap) { return (fifo_specops.vop_close(ap)); } /* struct vnop_access_args { struct vnode *a_vp; #if VOP_ACCESS_TAKES_ACCMODE_T accmode_t a_accmode; #else int a_mode; #endif struct ucred *a_cred; struct thread *a_td; }; */ static int fuse_vnop_access(struct vop_access_args *ap) { struct vnode *vp = ap->a_vp; int accmode = ap->a_accmode; struct ucred *cred = ap->a_cred; struct fuse_data *data = fuse_get_mpdata(vnode_mount(vp)); int err; if (fuse_isdeadfs(vp)) { if (vnode_isvroot(vp)) { return 0; } return ENXIO; } if (!(data->dataflags & FSESS_INITED)) { if (vnode_isvroot(vp)) { if (priv_check_cred(cred, PRIV_VFS_ADMIN) || (fuse_match_cred(data->daemoncred, cred) == 0)) { return 0; } } return EBADF; } if (vnode_islnk(vp)) { return 0; } err = fuse_internal_access(vp, accmode, ap->a_td, ap->a_cred); return err; } /* * struct vop_advlock_args { * struct vop_generic_args a_gen; * struct vnode *a_vp; * void *a_id; * int a_op; * struct flock *a_fl; * int a_flags; * } */ static int fuse_vnop_advlock(struct vop_advlock_args *ap) { struct vnode *vp = ap->a_vp; struct flock *fl = ap->a_fl; struct thread *td = curthread; struct ucred *cred = td->td_ucred; pid_t pid = td->td_proc->p_pid; struct fuse_filehandle *fufh; struct fuse_dispatcher fdi; struct fuse_lk_in *fli; struct fuse_lk_out *flo; enum fuse_opcode op; int dataflags, err; int flags = ap->a_flags; dataflags = fuse_get_mpdata(vnode_mount(vp))->dataflags; if (fuse_isdeadfs(vp)) { return ENXIO; } if (!(dataflags & FSESS_POSIX_LOCKS)) return vop_stdadvlock(ap); /* FUSE doesn't properly support flock until protocol 7.17 */ if (flags & F_FLOCK) return vop_stdadvlock(ap); err = fuse_filehandle_get_anyflags(vp, &fufh, cred, pid); if (err) return err; fdisp_init(&fdi, sizeof(*fli)); switch(ap->a_op) { case F_GETLK: op = FUSE_GETLK; break; case F_SETLK: op = FUSE_SETLK; break; case F_SETLKW: op = FUSE_SETLKW; break; default: return EINVAL; } fdisp_make_vp(&fdi, op, vp, td, cred); fli = fdi.indata; fli->fh = fufh->fh_id; fli->owner = fl->l_pid; fli->lk.start = fl->l_start; if (fl->l_len != 0) fli->lk.end = fl->l_start + fl->l_len - 1; else fli->lk.end = INT64_MAX; fli->lk.type = fl->l_type; fli->lk.pid = fl->l_pid; err = fdisp_wait_answ(&fdi); fdisp_destroy(&fdi); if (err == 0 && op == FUSE_GETLK) { flo = fdi.answ; fl->l_type = flo->lk.type; fl->l_pid = flo->lk.pid; if (flo->lk.type != F_UNLCK) { fl->l_start = flo->lk.start; if (flo->lk.end == INT64_MAX) fl->l_len = 0; else fl->l_len = flo->lk.end - flo->lk.start + 1; fl->l_start = flo->lk.start; } } return err; } /* { struct vnode *a_vp; daddr_t a_bn; struct bufobj **a_bop; daddr_t *a_bnp; int *a_runp; int *a_runb; } */ static int fuse_vnop_bmap(struct vop_bmap_args *ap) { struct vnode *vp = ap->a_vp; struct bufobj **bo = ap->a_bop; struct thread *td = curthread; struct mount *mp; struct fuse_dispatcher fdi; struct fuse_bmap_in *fbi; struct fuse_bmap_out *fbo; struct fuse_data *data; uint64_t biosize; off_t filesize; daddr_t lbn = ap->a_bn; daddr_t *pbn = ap->a_bnp; int *runp = ap->a_runp; int *runb = ap->a_runb; int error = 0; int maxrun; if (fuse_isdeadfs(vp)) { return ENXIO; } mp = vnode_mount(vp); data = fuse_get_mpdata(mp); biosize = fuse_iosize(vp); maxrun = MIN(vp->v_mount->mnt_iosize_max / biosize - 1, data->max_readahead_blocks); if (bo != NULL) *bo = &vp->v_bufobj; /* * The FUSE_BMAP operation does not include the runp and runb * variables, so we must guess. Report nonzero contiguous runs so * cluster_read will combine adjacent reads. It's worthwhile to reduce * upcalls even if we don't know the true physical layout of the file. * * FUSE file systems may opt out of read clustering in two ways: * * mounting with -onoclusterr * * Setting max_readahead <= maxbcachebuf during FUSE_INIT */ if (runb != NULL) *runb = MIN(lbn, maxrun); if (runp != NULL) { error = fuse_vnode_size(vp, &filesize, td->td_ucred, td); if (error == 0) *runp = MIN(MAX(0, filesize / biosize - lbn - 1), maxrun); else *runp = 0; } if (fsess_isimpl(mp, FUSE_BMAP)) { fdisp_init(&fdi, sizeof(*fbi)); fdisp_make_vp(&fdi, FUSE_BMAP, vp, td, td->td_ucred); fbi = fdi.indata; fbi->block = lbn; fbi->blocksize = biosize; error = fdisp_wait_answ(&fdi); if (error == ENOSYS) { fsess_set_notimpl(mp, FUSE_BMAP); error = 0; } else { fbo = fdi.answ; if (error == 0 && pbn != NULL) *pbn = fbo->block; return error; } } /* If the daemon doesn't support BMAP, make up a sensible default */ if (pbn != NULL) *pbn = lbn * btodb(biosize); return (error); } /* struct vop_close_args { struct vnode *a_vp; int a_fflag; struct ucred *a_cred; struct thread *a_td; }; */ static int fuse_vnop_close(struct vop_close_args *ap) { struct vnode *vp = ap->a_vp; struct ucred *cred = ap->a_cred; int fflag = ap->a_fflag; struct thread *td = ap->a_td; pid_t pid = td->td_proc->p_pid; int err = 0; if (fuse_isdeadfs(vp)) return 0; if (vnode_isdir(vp)) return 0; if (fflag & IO_NDELAY) return 0; err = fuse_flush(vp, cred, pid, fflag); /* TODO: close the file handle, if we're sure it's no longer used */ if ((VTOFUD(vp)->flag & FN_SIZECHANGE) != 0) { fuse_vnode_savesize(vp, cred, td->td_proc->p_pid); } return err; } static void fdisp_make_mknod_for_fallback( struct fuse_dispatcher *fdip, struct componentname *cnp, struct vnode *dvp, uint64_t parentnid, struct thread *td, struct ucred *cred, mode_t mode, enum fuse_opcode *op) { struct fuse_mknod_in *fmni; fdisp_init(fdip, sizeof(*fmni) + cnp->cn_namelen + 1); *op = FUSE_MKNOD; fdisp_make(fdip, *op, vnode_mount(dvp), parentnid, td, cred); fmni = fdip->indata; fmni->mode = mode; fmni->rdev = 0; memcpy((char *)fdip->indata + sizeof(*fmni), cnp->cn_nameptr, cnp->cn_namelen); ((char *)fdip->indata)[sizeof(*fmni) + cnp->cn_namelen] = '\0'; } /* struct vnop_create_args { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; struct vattr *a_vap; }; */ static int fuse_vnop_create(struct vop_create_args *ap) { struct vnode *dvp = ap->a_dvp; struct vnode **vpp = ap->a_vpp; struct componentname *cnp = ap->a_cnp; struct vattr *vap = ap->a_vap; struct thread *td = cnp->cn_thread; struct ucred *cred = cnp->cn_cred; struct fuse_data *data; struct fuse_create_in *fci; struct fuse_entry_out *feo; struct fuse_open_out *foo; struct fuse_dispatcher fdi, fdi2; struct fuse_dispatcher *fdip = &fdi; struct fuse_dispatcher *fdip2 = NULL; int err; struct mount *mp = vnode_mount(dvp); data = fuse_get_mpdata(mp); uint64_t parentnid = VTOFUD(dvp)->nid; mode_t mode = MAKEIMODE(vap->va_type, vap->va_mode); enum fuse_opcode op; int flags; if (fuse_isdeadfs(dvp)) return ENXIO; /* FUSE expects sockets to be created with FUSE_MKNOD */ if (vap->va_type == VSOCK) return fuse_internal_mknod(dvp, vpp, cnp, vap); /* * VOP_CREATE doesn't tell us the open(2) flags, so we guess. Only a * writable mode makes sense, and we might as well include readability * too. */ flags = O_RDWR; bzero(&fdi, sizeof(fdi)); if (vap->va_type != VREG) return (EINVAL); if (!fsess_isimpl(mp, FUSE_CREATE) || vap->va_type == VSOCK) { /* Fallback to FUSE_MKNOD/FUSE_OPEN */ fdisp_make_mknod_for_fallback(fdip, cnp, dvp, parentnid, td, cred, mode, &op); } else { /* Use FUSE_CREATE */ size_t insize; op = FUSE_CREATE; fdisp_init(fdip, sizeof(*fci) + cnp->cn_namelen + 1); fdisp_make(fdip, op, vnode_mount(dvp), parentnid, td, cred); fci = fdip->indata; fci->mode = mode; fci->flags = O_CREAT | flags; if (fuse_libabi_geq(data, 7, 12)) { insize = sizeof(*fci); fci->umask = td->td_proc->p_fd->fd_cmask; } else { insize = sizeof(struct fuse_open_in); } memcpy((char *)fdip->indata + insize, cnp->cn_nameptr, cnp->cn_namelen); ((char *)fdip->indata)[insize + cnp->cn_namelen] = '\0'; } err = fdisp_wait_answ(fdip); if (err) { if (err == ENOSYS && op == FUSE_CREATE) { fsess_set_notimpl(mp, FUSE_CREATE); fdisp_make_mknod_for_fallback(fdip, cnp, dvp, parentnid, td, cred, mode, &op); err = fdisp_wait_answ(fdip); } if (err) goto out; } feo = fdip->answ; if ((err = fuse_internal_checkentry(feo, vap->va_type))) { goto out; } if (op == FUSE_CREATE) { foo = (struct fuse_open_out*)(feo + 1); } else { /* Issue a separate FUSE_OPEN */ struct fuse_open_in *foi; fdip2 = &fdi2; fdisp_init(fdip2, sizeof(*foi)); fdisp_make(fdip2, FUSE_OPEN, vnode_mount(dvp), feo->nodeid, td, cred); foi = fdip2->indata; foi->flags = flags; err = fdisp_wait_answ(fdip2); if (err) goto out; foo = fdip2->answ; } err = fuse_vnode_get(mp, feo, feo->nodeid, dvp, vpp, cnp, vap->va_type); if (err) { struct fuse_release_in *fri; uint64_t nodeid = feo->nodeid; uint64_t fh_id = foo->fh; fdisp_init(fdip, sizeof(*fri)); fdisp_make(fdip, FUSE_RELEASE, mp, nodeid, td, cred); fri = fdip->indata; fri->fh = fh_id; fri->flags = flags; fuse_insert_callback(fdip->tick, fuse_internal_forget_callback); fuse_insert_message(fdip->tick, false); goto out; } ASSERT_VOP_ELOCKED(*vpp, "fuse_vnop_create"); fuse_internal_cache_attrs(*vpp, &feo->attr, feo->attr_valid, feo->attr_valid_nsec, NULL); fuse_filehandle_init(*vpp, FUFH_RDWR, NULL, td, cred, foo); fuse_vnode_open(*vpp, foo->open_flags, td); /* * Purge the parent's attribute cache because the daemon should've * updated its mtime and ctime */ fuse_vnode_clear_attr_cache(dvp); cache_purge_negative(dvp); out: if (fdip2) fdisp_destroy(fdip2); fdisp_destroy(fdip); return err; } /* struct vnop_fdatasync_args { struct vop_generic_args a_gen; struct vnode * a_vp; struct thread * a_td; }; */ static int fuse_vnop_fdatasync(struct vop_fdatasync_args *ap) { struct vnode *vp = ap->a_vp; struct thread *td = ap->a_td; int waitfor = MNT_WAIT; int err = 0; if (fuse_isdeadfs(vp)) { return 0; } if ((err = vop_stdfdatasync_buf(ap))) return err; return fuse_internal_fsync(vp, td, waitfor, true); } /* struct vnop_fsync_args { struct vop_generic_args a_gen; struct vnode * a_vp; int a_waitfor; struct thread * a_td; }; */ static int fuse_vnop_fsync(struct vop_fsync_args *ap) { struct vnode *vp = ap->a_vp; struct thread *td = ap->a_td; int waitfor = ap->a_waitfor; int err = 0; if (fuse_isdeadfs(vp)) { return 0; } if ((err = vop_stdfsync(ap))) return err; return fuse_internal_fsync(vp, td, waitfor, false); } /* struct vnop_getattr_args { struct vnode *a_vp; struct vattr *a_vap; struct ucred *a_cred; struct thread *a_td; }; */ static int fuse_vnop_getattr(struct vop_getattr_args *ap) { struct vnode *vp = ap->a_vp; struct vattr *vap = ap->a_vap; struct ucred *cred = ap->a_cred; struct thread *td = curthread; int err = 0; int dataflags; dataflags = fuse_get_mpdata(vnode_mount(vp))->dataflags; /* Note that we are not bailing out on a dead file system just yet. */ if (!(dataflags & FSESS_INITED)) { if (!vnode_isvroot(vp)) { fdata_set_dead(fuse_get_mpdata(vnode_mount(vp))); err = ENOTCONN; return err; } else { goto fake; } } err = fuse_internal_getattr(vp, vap, cred, td); if (err == ENOTCONN && vnode_isvroot(vp)) { /* see comment in fuse_vfsop_statfs() */ goto fake; } else { return err; } fake: bzero(vap, sizeof(*vap)); vap->va_type = vnode_vtype(vp); return 0; } /* struct vnop_inactive_args { struct vnode *a_vp; struct thread *a_td; }; */ static int fuse_vnop_inactive(struct vop_inactive_args *ap) { struct vnode *vp = ap->a_vp; struct thread *td = ap->a_td; struct fuse_vnode_data *fvdat = VTOFUD(vp); struct fuse_filehandle *fufh, *fufh_tmp; int need_flush = 1; LIST_FOREACH_SAFE(fufh, &fvdat->handles, next, fufh_tmp) { if (need_flush && vp->v_type == VREG) { if ((VTOFUD(vp)->flag & FN_SIZECHANGE) != 0) { fuse_vnode_savesize(vp, NULL, 0); } if ((fvdat->flag & FN_REVOKED) != 0) fuse_io_invalbuf(vp, td); else fuse_io_flushbuf(vp, MNT_WAIT, td); need_flush = 0; } fuse_filehandle_close(vp, fufh, td, NULL); } if ((fvdat->flag & FN_REVOKED) != 0 && fuse_reclaim_revoked) { vrecycle(vp); } return 0; } /* struct vnop_link_args { struct vnode *a_tdvp; struct vnode *a_vp; struct componentname *a_cnp; }; */ static int fuse_vnop_link(struct vop_link_args *ap) { struct vnode *vp = ap->a_vp; struct vnode *tdvp = ap->a_tdvp; struct componentname *cnp = ap->a_cnp; struct vattr *vap = VTOVA(vp); struct fuse_dispatcher fdi; struct fuse_entry_out *feo; struct fuse_link_in fli; int err; if (fuse_isdeadfs(vp)) { return ENXIO; } if (vnode_mount(tdvp) != vnode_mount(vp)) { return EXDEV; } /* * This is a seatbelt check to protect naive userspace filesystems from * themselves and the limitations of the FUSE IPC protocol. If a * filesystem does not allow attribute caching, assume it is capable of * validating that nlink does not overflow. */ if (vap != NULL && vap->va_nlink >= FUSE_LINK_MAX) return EMLINK; fli.oldnodeid = VTOI(vp); fdisp_init(&fdi, 0); fuse_internal_newentry_makerequest(vnode_mount(tdvp), VTOI(tdvp), cnp, FUSE_LINK, &fli, sizeof(fli), &fdi); if ((err = fdisp_wait_answ(&fdi))) { goto out; } feo = fdi.answ; err = fuse_internal_checkentry(feo, vnode_vtype(vp)); if (!err) { /* * Purge the parent's attribute cache because the daemon * should've updated its mtime and ctime */ fuse_vnode_clear_attr_cache(tdvp); fuse_internal_cache_attrs(vp, &feo->attr, feo->attr_valid, feo->attr_valid_nsec, NULL); } out: fdisp_destroy(&fdi); return err; } struct fuse_lookup_alloc_arg { struct fuse_entry_out *feo; struct componentname *cnp; uint64_t nid; enum vtype vtyp; }; /* Callback for vn_get_ino */ static int fuse_lookup_alloc(struct mount *mp, void *arg, int lkflags, struct vnode **vpp) { struct fuse_lookup_alloc_arg *flaa = arg; return fuse_vnode_get(mp, flaa->feo, flaa->nid, NULL, vpp, flaa->cnp, flaa->vtyp); } SDT_PROBE_DEFINE3(fusefs, , vnops, cache_lookup, "int", "struct timespec*", "struct timespec*"); /* struct vnop_lookup_args { struct vnodeop_desc *a_desc; struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; }; */ int fuse_vnop_lookup(struct vop_lookup_args *ap) { struct vnode *dvp = ap->a_dvp; struct vnode **vpp = ap->a_vpp; struct componentname *cnp = ap->a_cnp; struct thread *td = cnp->cn_thread; struct ucred *cred = cnp->cn_cred; int nameiop = cnp->cn_nameiop; int flags = cnp->cn_flags; int wantparent = flags & (LOCKPARENT | WANTPARENT); int islastcn = flags & ISLASTCN; struct mount *mp = vnode_mount(dvp); int err = 0; int lookup_err = 0; struct vnode *vp = NULL; struct fuse_dispatcher fdi; bool did_lookup = false; struct fuse_entry_out *feo = NULL; enum vtype vtyp; /* vnode type of target */ off_t filesize; /* filesize of target */ uint64_t nid; if (fuse_isdeadfs(dvp)) { *vpp = NULL; return ENXIO; } if (!vnode_isdir(dvp)) return ENOTDIR; if (islastcn && vfs_isrdonly(mp) && (nameiop != LOOKUP)) return EROFS; if ((err = fuse_internal_access(dvp, VEXEC, td, cred))) return err; if (flags & ISDOTDOT) { KASSERT(VTOFUD(dvp)->flag & FN_PARENT_NID, ("Looking up .. is TODO")); nid = VTOFUD(dvp)->parent_nid; if (nid == 0) return ENOENT; /* .. is obviously a directory */ vtyp = VDIR; filesize = 0; } else if (cnp->cn_namelen == 1 && *(cnp->cn_nameptr) == '.') { nid = VTOI(dvp); /* . is obviously a directory */ vtyp = VDIR; filesize = 0; } else { struct timespec now, timeout; err = cache_lookup(dvp, vpp, cnp, &timeout, NULL); getnanouptime(&now); SDT_PROBE3(fusefs, , vnops, cache_lookup, err, &timeout, &now); switch (err) { case -1: /* positive match */ if (timespeccmp(&timeout, &now, >)) { atomic_add_acq_long(&fuse_lookup_cache_hits, 1); } else { /* Cache timeout */ atomic_add_acq_long(&fuse_lookup_cache_misses, 1); bintime_clear( &VTOFUD(*vpp)->entry_cache_timeout); cache_purge(*vpp); if (dvp != *vpp) vput(*vpp); else vrele(*vpp); *vpp = NULL; break; } return 0; case 0: /* no match in cache */ atomic_add_acq_long(&fuse_lookup_cache_misses, 1); break; case ENOENT: /* negative match */ getnanouptime(&now); if (timespeccmp(&timeout, &now, <=)) { /* Cache timeout */ cache_purge_negative(dvp); break; } /* fall through */ default: return err; } nid = VTOI(dvp); fdisp_init(&fdi, cnp->cn_namelen + 1); fdisp_make(&fdi, FUSE_LOOKUP, mp, nid, td, cred); memcpy(fdi.indata, cnp->cn_nameptr, cnp->cn_namelen); ((char *)fdi.indata)[cnp->cn_namelen] = '\0'; lookup_err = fdisp_wait_answ(&fdi); did_lookup = true; if (!lookup_err) { /* lookup call succeeded */ feo = (struct fuse_entry_out *)fdi.answ; nid = feo->nodeid; if (nid == 0) { /* zero nodeid means ENOENT and cache it */ struct timespec timeout; fdi.answ_stat = ENOENT; lookup_err = ENOENT; if (cnp->cn_flags & MAKEENTRY) { fuse_validity_2_timespec(feo, &timeout); cache_enter_time(dvp, *vpp, cnp, &timeout, NULL); } } else if (nid == FUSE_ROOT_ID) { lookup_err = EINVAL; } vtyp = IFTOVT(feo->attr.mode); filesize = feo->attr.size; } if (lookup_err && (!fdi.answ_stat || lookup_err != ENOENT)) { fdisp_destroy(&fdi); return lookup_err; } } /* lookup_err, if non-zero, must be ENOENT at this point */ if (lookup_err) { /* Entry not found */ if ((nameiop == CREATE || nameiop == RENAME) && islastcn) { err = fuse_internal_access(dvp, VWRITE, td, cred); if (!err) { /* * Set the SAVENAME flag to hold onto the * pathname for use later in VOP_CREATE or * VOP_RENAME. */ cnp->cn_flags |= SAVENAME; err = EJUSTRETURN; } } else { err = ENOENT; } } else { /* Entry was found */ if (flags & ISDOTDOT) { struct fuse_lookup_alloc_arg flaa; flaa.nid = nid; flaa.feo = feo; flaa.cnp = cnp; flaa.vtyp = vtyp; err = vn_vget_ino_gen(dvp, fuse_lookup_alloc, &flaa, 0, &vp); *vpp = vp; } else if (nid == VTOI(dvp)) { vref(dvp); *vpp = dvp; } else { struct fuse_vnode_data *fvdat; err = fuse_vnode_get(vnode_mount(dvp), feo, nid, dvp, &vp, cnp, vtyp); if (err) goto out; *vpp = vp; /* * In the case where we are looking up a FUSE node * represented by an existing cached vnode, and the * true size reported by FUSE_LOOKUP doesn't match * the vnode's cached size, then any cached writes * beyond the file's current size are lost. * * We can get here: * * following attribute cache expiration, or * * due a bug in the daemon, or */ fvdat = VTOFUD(vp); if (vnode_isreg(vp) && filesize != fvdat->cached_attrs.va_size && fvdat->flag & FN_SIZECHANGE) { /* * The FN_SIZECHANGE flag reflects a dirty * append. If userspace lets us know our cache * is invalid, that write was lost. (Dirty * writes that do not cause append are also * lost, but we don't detect them here.) * * XXX: Maybe disable WB caching on this mount. */ printf("%s: WB cache incoherent on %s!\n", __func__, vnode_mount(vp)->mnt_stat.f_mntonname); fvdat->flag &= ~FN_SIZECHANGE; } MPASS(feo != NULL); fuse_internal_cache_attrs(*vpp, &feo->attr, feo->attr_valid, feo->attr_valid_nsec, NULL); fuse_validity_2_bintime(feo->entry_valid, feo->entry_valid_nsec, &fvdat->entry_cache_timeout); if ((nameiop == DELETE || nameiop == RENAME) && islastcn) { struct vattr dvattr; err = fuse_internal_access(dvp, VWRITE, td, cred); if (err != 0) goto out; /* * if the parent's sticky bit is set, check * whether we're allowed to remove the file. * Need to figure out the vnode locking to make * this work. */ fuse_internal_getattr(dvp, &dvattr, cred, td); if ((dvattr.va_mode & S_ISTXT) && fuse_internal_access(dvp, VADMIN, td, cred) && fuse_internal_access(*vpp, VADMIN, td, cred)) { err = EPERM; goto out; } } if (islastcn && ( (nameiop == DELETE) || (nameiop == RENAME && wantparent))) { cnp->cn_flags |= SAVENAME; } } } out: if (err) { if (vp != NULL && dvp != vp) vput(vp); else if (vp != NULL) vrele(vp); *vpp = NULL; } if (did_lookup) fdisp_destroy(&fdi); return err; } /* struct vnop_mkdir_args { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; struct vattr *a_vap; }; */ static int fuse_vnop_mkdir(struct vop_mkdir_args *ap) { struct vnode *dvp = ap->a_dvp; struct vnode **vpp = ap->a_vpp; struct componentname *cnp = ap->a_cnp; struct vattr *vap = ap->a_vap; struct fuse_mkdir_in fmdi; if (fuse_isdeadfs(dvp)) { return ENXIO; } fmdi.mode = MAKEIMODE(vap->va_type, vap->va_mode); fmdi.umask = curthread->td_proc->p_fd->fd_cmask; return (fuse_internal_newentry(dvp, vpp, cnp, FUSE_MKDIR, &fmdi, sizeof(fmdi), VDIR)); } /* struct vnop_mknod_args { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; struct vattr *a_vap; }; */ static int fuse_vnop_mknod(struct vop_mknod_args *ap) { struct vnode *dvp = ap->a_dvp; struct vnode **vpp = ap->a_vpp; struct componentname *cnp = ap->a_cnp; struct vattr *vap = ap->a_vap; if (fuse_isdeadfs(dvp)) return ENXIO; return fuse_internal_mknod(dvp, vpp, cnp, vap); } /* struct vop_open_args { struct vnode *a_vp; int a_mode; struct ucred *a_cred; struct thread *a_td; int a_fdidx; / struct file *a_fp; }; */ static int fuse_vnop_open(struct vop_open_args *ap) { struct vnode *vp = ap->a_vp; int a_mode = ap->a_mode; struct thread *td = ap->a_td; struct ucred *cred = ap->a_cred; pid_t pid = td->td_proc->p_pid; struct fuse_vnode_data *fvdat; if (fuse_isdeadfs(vp)) return ENXIO; if (vp->v_type == VCHR || vp->v_type == VBLK || vp->v_type == VFIFO) return (EOPNOTSUPP); if ((a_mode & (FREAD | FWRITE | FEXEC)) == 0) return EINVAL; fvdat = VTOFUD(vp); if (fuse_filehandle_validrw(vp, a_mode, cred, pid)) { fuse_vnode_open(vp, 0, td); return 0; } return fuse_filehandle_open(vp, a_mode, NULL, td, cred); } static int fuse_vnop_pathconf(struct vop_pathconf_args *ap) { switch (ap->a_name) { case _PC_FILESIZEBITS: *ap->a_retval = 64; return (0); case _PC_NAME_MAX: *ap->a_retval = NAME_MAX; return (0); case _PC_LINK_MAX: *ap->a_retval = MIN(LONG_MAX, FUSE_LINK_MAX); return (0); case _PC_SYMLINK_MAX: *ap->a_retval = MAXPATHLEN; return (0); case _PC_NO_TRUNC: *ap->a_retval = 1; return (0); default: return (vop_stdpathconf(ap)); } } /* struct vnop_read_args { struct vnode *a_vp; struct uio *a_uio; int a_ioflag; struct ucred *a_cred; }; */ static int fuse_vnop_read(struct vop_read_args *ap) { struct vnode *vp = ap->a_vp; struct uio *uio = ap->a_uio; int ioflag = ap->a_ioflag; struct ucred *cred = ap->a_cred; pid_t pid = curthread->td_proc->p_pid; if (fuse_isdeadfs(vp)) { return ENXIO; } if (VTOFUD(vp)->flag & FN_DIRECTIO) { ioflag |= IO_DIRECT; } - return fuse_io_dispatch(vp, uio, ioflag, false, cred, pid); + return fuse_io_dispatch(vp, uio, ioflag, cred, pid); } /* struct vnop_readdir_args { struct vnode *a_vp; struct uio *a_uio; struct ucred *a_cred; int *a_eofflag; int *a_ncookies; u_long **a_cookies; }; */ static int fuse_vnop_readdir(struct vop_readdir_args *ap) { struct vnode *vp = ap->a_vp; struct uio *uio = ap->a_uio; struct ucred *cred = ap->a_cred; struct fuse_filehandle *fufh = NULL; struct fuse_iov cookediov; int err = 0; u_long *cookies; off_t startoff; ssize_t tresid; int ncookies; bool closefufh = false; pid_t pid = curthread->td_proc->p_pid; if (ap->a_eofflag) *ap->a_eofflag = 0; if (fuse_isdeadfs(vp)) { return ENXIO; } if ( /* XXXIP ((uio_iovcnt(uio) > 1)) || */ (uio_resid(uio) < sizeof(struct dirent))) { return EINVAL; } tresid = uio->uio_resid; startoff = uio->uio_offset; err = fuse_filehandle_get_dir(vp, &fufh, cred, pid); if (err == EBADF && vnode_mount(vp)->mnt_flag & MNT_EXPORTED) { /* * nfsd will do VOP_READDIR without first doing VOP_OPEN. We * must implicitly open the directory here */ err = fuse_filehandle_open(vp, FREAD, &fufh, curthread, cred); if (err == 0) { /* * When a directory is opened, it must be read from * the beginning. Hopefully, the "startoff" still * exists as an offset cookie for the directory. * If not, it will read the entire directory without * returning any entries and just return eof. */ uio->uio_offset = 0; } closefufh = true; } if (err) return (err); if (ap->a_ncookies != NULL) { ncookies = uio->uio_resid / (offsetof(struct dirent, d_name) + 4) + 1; cookies = malloc(ncookies * sizeof(*cookies), M_TEMP, M_WAITOK); *ap->a_ncookies = ncookies; *ap->a_cookies = cookies; } else { ncookies = 0; cookies = NULL; } #define DIRCOOKEDSIZE FUSE_DIRENT_ALIGN(FUSE_NAME_OFFSET + MAXNAMLEN + 1) fiov_init(&cookediov, DIRCOOKEDSIZE); err = fuse_internal_readdir(vp, uio, startoff, fufh, &cookediov, &ncookies, cookies); fiov_teardown(&cookediov); if (closefufh) fuse_filehandle_close(vp, fufh, curthread, cred); if (ap->a_ncookies != NULL) { if (err == 0) { *ap->a_ncookies -= ncookies; } else { free(*ap->a_cookies, M_TEMP); *ap->a_ncookies = 0; *ap->a_cookies = NULL; } } if (err == 0 && tresid == uio->uio_resid) *ap->a_eofflag = 1; return err; } /* struct vnop_readlink_args { struct vnode *a_vp; struct uio *a_uio; struct ucred *a_cred; }; */ static int fuse_vnop_readlink(struct vop_readlink_args *ap) { struct vnode *vp = ap->a_vp; struct uio *uio = ap->a_uio; struct ucred *cred = ap->a_cred; struct fuse_dispatcher fdi; int err; if (fuse_isdeadfs(vp)) { return ENXIO; } if (!vnode_islnk(vp)) { return EINVAL; } fdisp_init(&fdi, 0); err = fdisp_simple_putget_vp(&fdi, FUSE_READLINK, vp, curthread, cred); if (err) { goto out; } if (((char *)fdi.answ)[0] == '/' && fuse_get_mpdata(vnode_mount(vp))->dataflags & FSESS_PUSH_SYMLINKS_IN) { char *mpth = vnode_mount(vp)->mnt_stat.f_mntonname; err = uiomove(mpth, strlen(mpth), uio); } if (!err) { err = uiomove(fdi.answ, fdi.iosize, uio); } out: fdisp_destroy(&fdi); return err; } /* struct vnop_reclaim_args { struct vnode *a_vp; struct thread *a_td; }; */ static int fuse_vnop_reclaim(struct vop_reclaim_args *ap) { struct vnode *vp = ap->a_vp; struct thread *td = ap->a_td; struct fuse_vnode_data *fvdat = VTOFUD(vp); struct fuse_filehandle *fufh, *fufh_tmp; if (!fvdat) { panic("FUSE: no vnode data during recycling"); } LIST_FOREACH_SAFE(fufh, &fvdat->handles, next, fufh_tmp) { printf("FUSE: vnode being reclaimed with open fufh " "(type=%#x)", fufh->fufh_type); fuse_filehandle_close(vp, fufh, td, NULL); } if ((!fuse_isdeadfs(vp)) && (fvdat->nlookup)) { fuse_internal_forget_send(vnode_mount(vp), td, NULL, VTOI(vp), fvdat->nlookup); } fuse_vnode_setparent(vp, NULL); cache_purge(vp); vfs_hash_remove(vp); vnode_destroy_vobject(vp); fuse_vnode_destroy(vp); return 0; } /* struct vnop_remove_args { struct vnode *a_dvp; struct vnode *a_vp; struct componentname *a_cnp; }; */ static int fuse_vnop_remove(struct vop_remove_args *ap) { struct vnode *dvp = ap->a_dvp; struct vnode *vp = ap->a_vp; struct componentname *cnp = ap->a_cnp; int err; if (fuse_isdeadfs(vp)) { return ENXIO; } if (vnode_isdir(vp)) { return EPERM; } cache_purge(vp); err = fuse_internal_remove(dvp, vp, cnp, FUSE_UNLINK); if (err == 0) { fuse_internal_vnode_disappear(vp); /* * Purge the parent's attribute cache because the daemon * should've updated its mtime and ctime */ fuse_vnode_clear_attr_cache(dvp); } return err; } /* struct vnop_rename_args { struct vnode *a_fdvp; struct vnode *a_fvp; struct componentname *a_fcnp; struct vnode *a_tdvp; struct vnode *a_tvp; struct componentname *a_tcnp; }; */ static int fuse_vnop_rename(struct vop_rename_args *ap) { struct vnode *fdvp = ap->a_fdvp; struct vnode *fvp = ap->a_fvp; struct componentname *fcnp = ap->a_fcnp; struct vnode *tdvp = ap->a_tdvp; struct vnode *tvp = ap->a_tvp; struct componentname *tcnp = ap->a_tcnp; struct fuse_data *data; bool newparent = fdvp != tdvp; bool isdir = fvp->v_type == VDIR; int err = 0; if (fuse_isdeadfs(fdvp)) { return ENXIO; } if (fvp->v_mount != tdvp->v_mount || (tvp && fvp->v_mount != tvp->v_mount)) { SDT_PROBE2(fusefs, , vnops, trace, 1, "cross-device rename"); err = EXDEV; goto out; } cache_purge(fvp); /* * FUSE library is expected to check if target directory is not * under the source directory in the file system tree. * Linux performs this check at VFS level. */ /* * If source is a directory, and it will get a new parent, user must * have write permission to it, so ".." can be modified. */ data = fuse_get_mpdata(vnode_mount(tdvp)); if (data->dataflags & FSESS_DEFAULT_PERMISSIONS && isdir && newparent) { err = fuse_internal_access(fvp, VWRITE, tcnp->cn_thread, tcnp->cn_cred); if (err) goto out; } sx_xlock(&data->rename_lock); err = fuse_internal_rename(fdvp, fcnp, tdvp, tcnp); if (err == 0) { if (tdvp != fdvp) fuse_vnode_setparent(fvp, tdvp); if (tvp != NULL) fuse_vnode_setparent(tvp, NULL); } sx_unlock(&data->rename_lock); if (tvp != NULL && tvp != fvp) { cache_purge(tvp); } if (vnode_isdir(fvp)) { if ((tvp != NULL) && vnode_isdir(tvp)) { cache_purge(tdvp); } cache_purge(fdvp); } out: if (tdvp == tvp) { vrele(tdvp); } else { vput(tdvp); } if (tvp != NULL) { vput(tvp); } vrele(fdvp); vrele(fvp); return err; } /* struct vnop_rmdir_args { struct vnode *a_dvp; struct vnode *a_vp; struct componentname *a_cnp; } *ap; */ static int fuse_vnop_rmdir(struct vop_rmdir_args *ap) { struct vnode *dvp = ap->a_dvp; struct vnode *vp = ap->a_vp; int err; if (fuse_isdeadfs(vp)) { return ENXIO; } if (VTOFUD(vp) == VTOFUD(dvp)) { return EINVAL; } err = fuse_internal_remove(dvp, vp, ap->a_cnp, FUSE_RMDIR); if (err == 0) { fuse_internal_vnode_disappear(vp); /* * Purge the parent's attribute cache because the daemon * should've updated its mtime and ctime */ fuse_vnode_clear_attr_cache(dvp); } return err; } /* struct vnop_setattr_args { struct vnode *a_vp; struct vattr *a_vap; struct ucred *a_cred; struct thread *a_td; }; */ static int fuse_vnop_setattr(struct vop_setattr_args *ap) { struct vnode *vp = ap->a_vp; struct vattr *vap = ap->a_vap; struct ucred *cred = ap->a_cred; struct thread *td = curthread; struct mount *mp; struct fuse_data *data; struct vattr old_va; int dataflags; int err = 0, err2; accmode_t accmode = 0; bool checkperm; bool drop_suid = false; gid_t cr_gid; mp = vnode_mount(vp); data = fuse_get_mpdata(mp); dataflags = data->dataflags; checkperm = dataflags & FSESS_DEFAULT_PERMISSIONS; if (cred->cr_ngroups > 0) cr_gid = cred->cr_groups[0]; else cr_gid = 0; if (fuse_isdeadfs(vp)) { return ENXIO; } if (vap->va_uid != (uid_t)VNOVAL) { if (checkperm) { /* Only root may change a file's owner */ err = priv_check_cred(cred, PRIV_VFS_CHOWN); if (err) { /* As a special case, allow the null chown */ err2 = fuse_internal_getattr(vp, &old_va, cred, td); if (err2) return (err2); if (vap->va_uid != old_va.va_uid) return err; else accmode |= VADMIN; drop_suid = true; } else accmode |= VADMIN; } else accmode |= VADMIN; } if (vap->va_gid != (gid_t)VNOVAL) { if (checkperm && priv_check_cred(cred, PRIV_VFS_CHOWN)) drop_suid = true; if (checkperm && !groupmember(vap->va_gid, cred)) { /* * Non-root users may only chgrp to one of their own * groups */ err = priv_check_cred(cred, PRIV_VFS_CHOWN); if (err) { /* As a special case, allow the null chgrp */ err2 = fuse_internal_getattr(vp, &old_va, cred, td); if (err2) return (err2); if (vap->va_gid != old_va.va_gid) return err; accmode |= VADMIN; } else accmode |= VADMIN; } else accmode |= VADMIN; } if (vap->va_size != VNOVAL) { switch (vp->v_type) { case VDIR: return (EISDIR); case VLNK: case VREG: if (vfs_isrdonly(mp)) return (EROFS); break; default: /* * According to POSIX, the result is unspecified * for file types other than regular files, * directories and shared memory objects. We * don't support shared memory objects in the file * system, and have dubious support for truncating * symlinks. Just ignore the request in other cases. */ return (0); } /* Don't set accmode. Permission to trunc is checked upstack */ } if (vap->va_atime.tv_sec != VNOVAL || vap->va_mtime.tv_sec != VNOVAL) { if (vap->va_vaflags & VA_UTIMES_NULL) accmode |= VWRITE; else accmode |= VADMIN; } if (drop_suid) { if (vap->va_mode != (mode_t)VNOVAL) vap->va_mode &= ~(S_ISUID | S_ISGID); else { err = fuse_internal_getattr(vp, &old_va, cred, td); if (err) return (err); vap->va_mode = old_va.va_mode & ~(S_ISUID | S_ISGID); } } if (vap->va_mode != (mode_t)VNOVAL) { /* Only root may set the sticky bit on non-directories */ if (checkperm && vp->v_type != VDIR && (vap->va_mode & S_ISTXT) && priv_check_cred(cred, PRIV_VFS_STICKYFILE)) return EFTYPE; if (checkperm && (vap->va_mode & S_ISGID)) { err = fuse_internal_getattr(vp, &old_va, cred, td); if (err) return (err); if (!groupmember(old_va.va_gid, cred)) { err = priv_check_cred(cred, PRIV_VFS_SETGID); if (err) return (err); } } accmode |= VADMIN; } if (vfs_isrdonly(mp)) return EROFS; err = fuse_internal_access(vp, accmode, td, cred); if (err) return err; else return fuse_internal_setattr(vp, vap, td, cred); } /* struct vnop_strategy_args { struct vnode *a_vp; struct buf *a_bp; }; */ static int fuse_vnop_strategy(struct vop_strategy_args *ap) { struct vnode *vp = ap->a_vp; struct buf *bp = ap->a_bp; if (!vp || fuse_isdeadfs(vp)) { bp->b_ioflags |= BIO_ERROR; bp->b_error = ENXIO; bufdone(bp); return 0; } /* * VOP_STRATEGY always returns zero and signals error via bp->b_ioflags. * fuse_io_strategy sets bp's error fields */ (void)fuse_io_strategy(vp, bp); return 0; } /* struct vnop_symlink_args { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; struct vattr *a_vap; char *a_target; }; */ static int fuse_vnop_symlink(struct vop_symlink_args *ap) { struct vnode *dvp = ap->a_dvp; struct vnode **vpp = ap->a_vpp; struct componentname *cnp = ap->a_cnp; const char *target = ap->a_target; struct fuse_dispatcher fdi; int err; size_t len; if (fuse_isdeadfs(dvp)) { return ENXIO; } /* * Unlike the other creator type calls, here we have to create a message * where the name of the new entry comes first, and the data describing * the entry comes second. * Hence we can't rely on our handy fuse_internal_newentry() routine, * but put together the message manually and just call the core part. */ len = strlen(target) + 1; fdisp_init(&fdi, len + cnp->cn_namelen + 1); fdisp_make_vp(&fdi, FUSE_SYMLINK, dvp, curthread, NULL); memcpy(fdi.indata, cnp->cn_nameptr, cnp->cn_namelen); ((char *)fdi.indata)[cnp->cn_namelen] = '\0'; memcpy((char *)fdi.indata + cnp->cn_namelen + 1, target, len); err = fuse_internal_newentry_core(dvp, vpp, cnp, VLNK, &fdi); fdisp_destroy(&fdi); return err; } /* struct vnop_write_args { struct vnode *a_vp; struct uio *a_uio; int a_ioflag; struct ucred *a_cred; }; */ static int fuse_vnop_write(struct vop_write_args *ap) { struct vnode *vp = ap->a_vp; struct uio *uio = ap->a_uio; int ioflag = ap->a_ioflag; struct ucred *cred = ap->a_cred; pid_t pid = curthread->td_proc->p_pid; if (fuse_isdeadfs(vp)) { return ENXIO; } if (VTOFUD(vp)->flag & FN_DIRECTIO) { ioflag |= IO_DIRECT; } - return fuse_io_dispatch(vp, uio, ioflag, false, cred, pid); + return fuse_io_dispatch(vp, uio, ioflag, cred, pid); } -SDT_PROBE_DEFINE1(fusefs, , vnops, vnop_getpages_error, "int"); -/* - struct vnop_getpages_args { - struct vnode *a_vp; - vm_page_t *a_m; - int a_count; - int a_reqpage; - }; -*/ -static int -fuse_vnop_getpages(struct vop_getpages_args *ap) +static daddr_t +fuse_gbp_getblkno(struct vnode *vp, vm_ooffset_t off) { - int i, error, nextoff, size, toff, count, npages; - struct uio uio; - struct iovec iov; - vm_offset_t kva; - struct buf *bp; - struct vnode *vp; - struct thread *td; - struct ucred *cred; - vm_page_t *pages; - pid_t pid = curthread->td_proc->p_pid; + const int biosize = fuse_iosize(vp); - vp = ap->a_vp; - KASSERT(vp->v_object, ("objectless vp passed to getpages")); - td = curthread; /* XXX */ - cred = curthread->td_ucred; /* XXX */ - pages = ap->a_m; - npages = ap->a_count; + return (off / biosize); +} - if (!fsess_opt_mmap(vnode_mount(vp))) { - SDT_PROBE2(fusefs, , vnops, trace, 1, - "called on non-cacheable vnode??\n"); - return (VM_PAGER_ERROR); - } +static int +fuse_gbp_getblksz(struct vnode *vp, daddr_t lbn) +{ + off_t filesize; + int blksz, err; + const int biosize = fuse_iosize(vp); - /* - * If the last page is partially valid, just return it and allow - * the pager to zero-out the blanks. Partially valid pages can - * only occur at the file EOF. - * - * XXXGL: is that true for FUSE, which is a local filesystem, - * but still somewhat disconnected from the kernel? - */ - VM_OBJECT_WLOCK(vp->v_object); - if (pages[npages - 1]->valid != 0 && --npages == 0) - goto out; - VM_OBJECT_WUNLOCK(vp->v_object); + err = fuse_vnode_size(vp, &filesize, NULL, NULL); + KASSERT(err == 0, ("vfs_bio_getpages can't handle errors here")); + if (err) + return biosize; - /* - * We use only the kva address for the buffer, but this is extremely - * convenient and fast. - */ - bp = uma_zalloc(fuse_pbuf_zone, M_WAITOK); - - kva = (vm_offset_t)bp->b_data; - pmap_qenter(kva, pages, npages); - VM_CNT_INC(v_vnodein); - VM_CNT_ADD(v_vnodepgsin, npages); - - count = npages << PAGE_SHIFT; - iov.iov_base = (caddr_t)kva; - iov.iov_len = count; - uio.uio_iov = &iov; - uio.uio_iovcnt = 1; - uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); - uio.uio_resid = count; - uio.uio_segflg = UIO_SYSSPACE; - uio.uio_rw = UIO_READ; - uio.uio_td = td; - - error = fuse_io_dispatch(vp, &uio, IO_DIRECT, true, cred, pid); - pmap_qremove(kva, npages); - - uma_zfree(fuse_pbuf_zone, bp); - - if (error && (uio.uio_resid == count)) { - SDT_PROBE1(fusefs, , vnops, vnop_getpages_error, error); - return VM_PAGER_ERROR; + if ((off_t)lbn * biosize >= filesize) { + blksz = 0; + } else if ((off_t)(lbn + 1) * biosize > filesize) { + blksz = filesize - (off_t)lbn *biosize; + } else { + blksz = biosize; } - /* - * Calculate the number of bytes read and validate only that number - * of bytes. Note that due to pending writes, size may be 0. This - * does not mean that the remaining data is invalid! - */ - - size = count - uio.uio_resid; - VM_OBJECT_WLOCK(vp->v_object); - fuse_vm_page_lock_queues(); - for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { - vm_page_t m; - - nextoff = toff + PAGE_SIZE; - m = pages[i]; - - if (nextoff <= size) { - /* - * Read operation filled an entire page - */ - m->valid = VM_PAGE_BITS_ALL; - KASSERT(m->dirty == 0, - ("fuse_getpages: page %p is dirty", m)); - } else if (size > toff) { - /* - * Read operation filled a partial page. - */ - m->valid = 0; - vm_page_set_valid_range(m, 0, size - toff); - KASSERT(m->dirty == 0, - ("fuse_getpages: page %p is dirty", m)); - } else { - /* - * Read operation was short. If no error occurred - * we may have hit a zero-fill section. We simply - * leave valid set to 0. - */ - ; - } - } - fuse_vm_page_unlock_queues(); -out: - VM_OBJECT_WUNLOCK(vp->v_object); - if (ap->a_rbehind) - *ap->a_rbehind = 0; - if (ap->a_rahead) - *ap->a_rahead = 0; - return (VM_PAGER_OK); + return (blksz); } /* - struct vnop_putpages_args { + struct vnop_getpages_args { struct vnode *a_vp; vm_page_t *a_m; int a_count; - int a_sync; - int *a_rtvals; - vm_ooffset_t a_offset; + int a_reqpage; }; */ static int -fuse_vnop_putpages(struct vop_putpages_args *ap) +fuse_vnop_getpages(struct vop_getpages_args *ap) { - struct uio uio; - struct iovec iov; - vm_offset_t kva; - struct buf *bp; - int i, error, npages, count; - off_t offset; - int *rtvals; - struct vnode *vp; - struct thread *td; - struct ucred *cred; - vm_page_t *pages; - vm_ooffset_t fsize; - pid_t pid = curthread->td_proc->p_pid; + struct vnode *vp = ap->a_vp; - vp = ap->a_vp; - KASSERT(vp->v_object, ("objectless vp passed to putpages")); - fsize = vp->v_object->un_pager.vnp.vnp_size; - td = curthread; /* XXX */ - cred = curthread->td_ucred; /* XXX */ - pages = ap->a_m; - count = ap->a_count; - rtvals = ap->a_rtvals; - npages = btoc(count); - offset = IDX_TO_OFF(pages[0]->pindex); - if (!fsess_opt_mmap(vnode_mount(vp))) { SDT_PROBE2(fusefs, , vnops, trace, 1, "called on non-cacheable vnode??\n"); + return (VM_PAGER_ERROR); } - for (i = 0; i < npages; i++) - rtvals[i] = VM_PAGER_AGAIN; - /* - * When putting pages, do not extend file past EOF. - */ - - if (offset + count > fsize) { - count = fsize - offset; - if (count < 0) - count = 0; - } - /* - * We use only the kva address for the buffer, but this is extremely - * convenient and fast. - */ - bp = uma_zalloc(fuse_pbuf_zone, M_WAITOK); - - kva = (vm_offset_t)bp->b_data; - pmap_qenter(kva, pages, npages); - VM_CNT_INC(v_vnodeout); - VM_CNT_ADD(v_vnodepgsout, count); - - iov.iov_base = (caddr_t)kva; - iov.iov_len = count; - uio.uio_iov = &iov; - uio.uio_iovcnt = 1; - uio.uio_offset = offset; - uio.uio_resid = count; - uio.uio_segflg = UIO_SYSSPACE; - uio.uio_rw = UIO_WRITE; - uio.uio_td = td; - - error = fuse_io_dispatch(vp, &uio, IO_DIRECT, true, cred, pid); - - pmap_qremove(kva, npages); - uma_zfree(fuse_pbuf_zone, bp); - - if (!error) { - int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; - - for (i = 0; i < nwritten; i++) { - rtvals[i] = VM_PAGER_OK; - VM_OBJECT_WLOCK(pages[i]->object); - vm_page_undirty(pages[i]); - VM_OBJECT_WUNLOCK(pages[i]->object); - } - } - return rtvals[0]; + return (vfs_bio_getpages(vp, ap->a_m, ap->a_count, ap->a_rbehind, + ap->a_rahead, fuse_gbp_getblkno, fuse_gbp_getblksz)); } static const char extattr_namespace_separator = '.'; /* struct vop_getextattr_args { struct vop_generic_args a_gen; struct vnode *a_vp; int a_attrnamespace; const char *a_name; struct uio *a_uio; size_t *a_size; struct ucred *a_cred; struct thread *a_td; }; */ static int fuse_vnop_getextattr(struct vop_getextattr_args *ap) { struct vnode *vp = ap->a_vp; struct uio *uio = ap->a_uio; struct fuse_dispatcher fdi; struct fuse_getxattr_in *get_xattr_in; struct fuse_getxattr_out *get_xattr_out; struct mount *mp = vnode_mount(vp); struct thread *td = ap->a_td; struct ucred *cred = ap->a_cred; char *prefix; char *attr_str; size_t len; int err; if (fuse_isdeadfs(vp)) return (ENXIO); if (!fsess_isimpl(mp, FUSE_GETXATTR)) return EOPNOTSUPP; err = fuse_extattr_check_cred(vp, ap->a_attrnamespace, cred, td, VREAD); if (err) return err; /* Default to looking for user attributes. */ if (ap->a_attrnamespace == EXTATTR_NAMESPACE_SYSTEM) prefix = EXTATTR_NAMESPACE_SYSTEM_STRING; else prefix = EXTATTR_NAMESPACE_USER_STRING; len = strlen(prefix) + sizeof(extattr_namespace_separator) + strlen(ap->a_name) + 1; fdisp_init(&fdi, len + sizeof(*get_xattr_in)); fdisp_make_vp(&fdi, FUSE_GETXATTR, vp, td, cred); get_xattr_in = fdi.indata; /* * Check to see whether we're querying the available size or * issuing the actual request. If we pass in 0, we get back struct * fuse_getxattr_out. If we pass in a non-zero size, we get back * that much data, without the struct fuse_getxattr_out header. */ if (uio == NULL) get_xattr_in->size = 0; else get_xattr_in->size = uio->uio_resid; attr_str = (char *)fdi.indata + sizeof(*get_xattr_in); snprintf(attr_str, len, "%s%c%s", prefix, extattr_namespace_separator, ap->a_name); err = fdisp_wait_answ(&fdi); if (err != 0) { if (err == ENOSYS) { fsess_set_notimpl(mp, FUSE_GETXATTR); err = EOPNOTSUPP; } goto out; } get_xattr_out = fdi.answ; if (ap->a_size != NULL) *ap->a_size = get_xattr_out->size; if (uio != NULL) err = uiomove(fdi.answ, fdi.iosize, uio); out: fdisp_destroy(&fdi); return (err); } /* struct vop_setextattr_args { struct vop_generic_args a_gen; struct vnode *a_vp; int a_attrnamespace; const char *a_name; struct uio *a_uio; struct ucred *a_cred; struct thread *a_td; }; */ static int fuse_vnop_setextattr(struct vop_setextattr_args *ap) { struct vnode *vp = ap->a_vp; struct uio *uio = ap->a_uio; struct fuse_dispatcher fdi; struct fuse_setxattr_in *set_xattr_in; struct mount *mp = vnode_mount(vp); struct thread *td = ap->a_td; struct ucred *cred = ap->a_cred; char *prefix; size_t len; char *attr_str; int err; if (fuse_isdeadfs(vp)) return (ENXIO); if (!fsess_isimpl(mp, FUSE_SETXATTR)) return EOPNOTSUPP; if (vfs_isrdonly(mp)) return EROFS; /* Deleting xattrs must use VOP_DELETEEXTATTR instead */ if (ap->a_uio == NULL) { /* * If we got here as fallback from VOP_DELETEEXTATTR, then * return EOPNOTSUPP. */ if (!fsess_isimpl(mp, FUSE_REMOVEXATTR)) return (EOPNOTSUPP); else return (EINVAL); } err = fuse_extattr_check_cred(vp, ap->a_attrnamespace, cred, td, VWRITE); if (err) return err; /* Default to looking for user attributes. */ if (ap->a_attrnamespace == EXTATTR_NAMESPACE_SYSTEM) prefix = EXTATTR_NAMESPACE_SYSTEM_STRING; else prefix = EXTATTR_NAMESPACE_USER_STRING; len = strlen(prefix) + sizeof(extattr_namespace_separator) + strlen(ap->a_name) + 1; fdisp_init(&fdi, len + sizeof(*set_xattr_in) + uio->uio_resid); fdisp_make_vp(&fdi, FUSE_SETXATTR, vp, td, cred); set_xattr_in = fdi.indata; set_xattr_in->size = uio->uio_resid; attr_str = (char *)fdi.indata + sizeof(*set_xattr_in); snprintf(attr_str, len, "%s%c%s", prefix, extattr_namespace_separator, ap->a_name); err = uiomove((char *)fdi.indata + sizeof(*set_xattr_in) + len, uio->uio_resid, uio); if (err != 0) { goto out; } err = fdisp_wait_answ(&fdi); if (err == ENOSYS) { fsess_set_notimpl(mp, FUSE_SETXATTR); err = EOPNOTSUPP; } if (err == ERESTART) { /* Can't restart after calling uiomove */ err = EINTR; } out: fdisp_destroy(&fdi); return (err); } /* * The Linux / FUSE extended attribute list is simply a collection of * NUL-terminated strings. The FreeBSD extended attribute list is a single * byte length followed by a non-NUL terminated string. So, this allows * conversion of the Linux / FUSE format to the FreeBSD format in place. * Linux attribute names are reported with the namespace as a prefix (e.g. * "user.attribute_name"), but in FreeBSD they are reported without the * namespace prefix (e.g. "attribute_name"). So, we're going from: * * user.attr_name1\0user.attr_name2\0 * * to: * * attr_name1attr_name2 * * Where "" is a single byte number of characters in the attribute name. * * Args: * prefix - exattr namespace prefix string * list, list_len - input list with namespace prefixes * bsd_list, bsd_list_len - output list compatible with bsd vfs */ static int fuse_xattrlist_convert(char *prefix, const char *list, int list_len, char *bsd_list, int *bsd_list_len) { int len, pos, dist_to_next, prefix_len; pos = 0; *bsd_list_len = 0; prefix_len = strlen(prefix); while (pos < list_len && list[pos] != '\0') { dist_to_next = strlen(&list[pos]) + 1; if (bcmp(&list[pos], prefix, prefix_len) == 0 && list[pos + prefix_len] == extattr_namespace_separator) { len = dist_to_next - (prefix_len + sizeof(extattr_namespace_separator)) - 1; if (len >= EXTATTR_MAXNAMELEN) return (ENAMETOOLONG); bsd_list[*bsd_list_len] = len; memcpy(&bsd_list[*bsd_list_len + 1], &list[pos + prefix_len + sizeof(extattr_namespace_separator)], len); *bsd_list_len += len + 1; } pos += dist_to_next; } return (0); } /* struct vop_listextattr_args { struct vop_generic_args a_gen; struct vnode *a_vp; int a_attrnamespace; struct uio *a_uio; size_t *a_size; struct ucred *a_cred; struct thread *a_td; }; */ static int fuse_vnop_listextattr(struct vop_listextattr_args *ap) { struct vnode *vp = ap->a_vp; struct uio *uio = ap->a_uio; struct fuse_dispatcher fdi; struct fuse_listxattr_in *list_xattr_in; struct fuse_listxattr_out *list_xattr_out; struct mount *mp = vnode_mount(vp); struct thread *td = ap->a_td; struct ucred *cred = ap->a_cred; size_t len; char *prefix; char *attr_str; char *bsd_list = NULL; char *linux_list; int bsd_list_len; int linux_list_len; int err; if (fuse_isdeadfs(vp)) return (ENXIO); if (!fsess_isimpl(mp, FUSE_LISTXATTR)) return EOPNOTSUPP; err = fuse_extattr_check_cred(vp, ap->a_attrnamespace, cred, td, VREAD); if (err) return err; /* * Add space for a NUL and the period separator if enabled. * Default to looking for user attributes. */ if (ap->a_attrnamespace == EXTATTR_NAMESPACE_SYSTEM) prefix = EXTATTR_NAMESPACE_SYSTEM_STRING; else prefix = EXTATTR_NAMESPACE_USER_STRING; len = strlen(prefix) + sizeof(extattr_namespace_separator) + 1; fdisp_init(&fdi, sizeof(*list_xattr_in) + len); fdisp_make_vp(&fdi, FUSE_LISTXATTR, vp, td, cred); /* * Retrieve Linux / FUSE compatible list size. */ list_xattr_in = fdi.indata; list_xattr_in->size = 0; attr_str = (char *)fdi.indata + sizeof(*list_xattr_in); snprintf(attr_str, len, "%s%c", prefix, extattr_namespace_separator); err = fdisp_wait_answ(&fdi); if (err != 0) { if (err == ENOSYS) { fsess_set_notimpl(mp, FUSE_LISTXATTR); err = EOPNOTSUPP; } goto out; } list_xattr_out = fdi.answ; linux_list_len = list_xattr_out->size; if (linux_list_len == 0) { if (ap->a_size != NULL) *ap->a_size = linux_list_len; goto out; } /* * Retrieve Linux / FUSE compatible list values. */ fdisp_refresh_vp(&fdi, FUSE_LISTXATTR, vp, td, cred); list_xattr_in = fdi.indata; list_xattr_in->size = linux_list_len + sizeof(*list_xattr_out); attr_str = (char *)fdi.indata + sizeof(*list_xattr_in); snprintf(attr_str, len, "%s%c", prefix, extattr_namespace_separator); err = fdisp_wait_answ(&fdi); if (err != 0) goto out; linux_list = fdi.answ; linux_list_len = fdi.iosize; /* * Retrieve the BSD compatible list values. * The Linux / FUSE attribute list format isn't the same * as FreeBSD's format. So we need to transform it into * FreeBSD's format before giving it to the user. */ bsd_list = malloc(linux_list_len, M_TEMP, M_WAITOK); err = fuse_xattrlist_convert(prefix, linux_list, linux_list_len, bsd_list, &bsd_list_len); if (err != 0) goto out; if (ap->a_size != NULL) *ap->a_size = bsd_list_len; if (uio != NULL) err = uiomove(bsd_list, bsd_list_len, uio); out: free(bsd_list, M_TEMP); fdisp_destroy(&fdi); return (err); } /* struct vop_deleteextattr_args { struct vop_generic_args a_gen; struct vnode *a_vp; int a_attrnamespace; const char *a_name; struct ucred *a_cred; struct thread *a_td; }; */ static int fuse_vnop_deleteextattr(struct vop_deleteextattr_args *ap) { struct vnode *vp = ap->a_vp; struct fuse_dispatcher fdi; struct mount *mp = vnode_mount(vp); struct thread *td = ap->a_td; struct ucred *cred = ap->a_cred; char *prefix; size_t len; char *attr_str; int err; if (fuse_isdeadfs(vp)) return (ENXIO); if (!fsess_isimpl(mp, FUSE_REMOVEXATTR)) return EOPNOTSUPP; if (vfs_isrdonly(mp)) return EROFS; err = fuse_extattr_check_cred(vp, ap->a_attrnamespace, cred, td, VWRITE); if (err) return err; /* Default to looking for user attributes. */ if (ap->a_attrnamespace == EXTATTR_NAMESPACE_SYSTEM) prefix = EXTATTR_NAMESPACE_SYSTEM_STRING; else prefix = EXTATTR_NAMESPACE_USER_STRING; len = strlen(prefix) + sizeof(extattr_namespace_separator) + strlen(ap->a_name) + 1; fdisp_init(&fdi, len); fdisp_make_vp(&fdi, FUSE_REMOVEXATTR, vp, td, cred); attr_str = fdi.indata; snprintf(attr_str, len, "%s%c%s", prefix, extattr_namespace_separator, ap->a_name); err = fdisp_wait_answ(&fdi); if (err == ENOSYS) { fsess_set_notimpl(mp, FUSE_REMOVEXATTR); err = EOPNOTSUPP; } fdisp_destroy(&fdi); return (err); } /* struct vnop_print_args { struct vnode *a_vp; }; */ static int fuse_vnop_print(struct vop_print_args *ap) { struct fuse_vnode_data *fvdat = VTOFUD(ap->a_vp); printf("nodeid: %ju, parent nodeid: %ju, nlookup: %ju, flag: %#x\n", (uintmax_t)VTOILLU(ap->a_vp), (uintmax_t)fvdat->parent_nid, (uintmax_t)fvdat->nlookup, fvdat->flag); return 0; } /* * Get an NFS filehandle for a FUSE file. * * This will only work for FUSE file systems that guarantee the uniqueness of * nodeid:generation, which most don't. */ /* vop_vptofh { IN struct vnode *a_vp; IN struct fid *a_fhp; }; */ static int fuse_vnop_vptofh(struct vop_vptofh_args *ap) { struct vnode *vp = ap->a_vp; struct fuse_vnode_data *fvdat = VTOFUD(vp); struct fuse_fid *fhp = (struct fuse_fid *)(ap->a_fhp); _Static_assert(sizeof(struct fuse_fid) <= sizeof(struct fid), "FUSE fid type is too big"); struct mount *mp = vnode_mount(vp); struct fuse_data *data = fuse_get_mpdata(mp); struct vattr va; int err; if (!(data->dataflags & FSESS_EXPORT_SUPPORT)) return EOPNOTSUPP; err = fuse_internal_getattr(vp, &va, curthread->td_ucred, curthread); if (err) return err; /*ip = VTOI(ap->a_vp);*/ /*ufhp = (struct ufid *)ap->a_fhp;*/ fhp->len = sizeof(struct fuse_fid); fhp->nid = fvdat->nid; if (fvdat->generation <= UINT32_MAX) fhp->gen = fvdat->generation; else return EOVERFLOW; return (0); } Index: projects/fuse2/tests/sys/fs/fusefs/read.cc =================================================================== --- projects/fuse2/tests/sys/fs/fusefs/read.cc (revision 349377) +++ projects/fuse2/tests/sys/fs/fusefs/read.cc (revision 349378) @@ -1,864 +1,920 @@ /*- * 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 #include #include #include #include } #include "mockfs.hh" #include "utils.hh" using namespace testing; class Read: public FuseTest { public: void expect_lookup(const char *relpath, uint64_t ino, uint64_t size) { FuseTest::expect_lookup(relpath, ino, S_IFREG | 0644, size, 1); } }; class Read_7_8: public FuseTest { public: virtual void SetUp() { m_kernel_minor_version = 8; FuseTest::SetUp(); } void expect_lookup(const char *relpath, uint64_t ino, uint64_t size) { FuseTest::expect_lookup_7_8(relpath, ino, S_IFREG | 0644, size, 1); } }; class AioRead: public Read { public: virtual void SetUp() { const char *node = "vfs.aio.enable_unsafe"; int val = 0; size_t size = sizeof(val); FuseTest::SetUp(); ASSERT_EQ(0, sysctlbyname(node, &val, &size, NULL, 0)) << strerror(errno); if (!val) GTEST_SKIP() << "vfs.aio.enable_unsafe must be set for this test"; } }; class AsyncRead: public AioRead { virtual void SetUp() { m_init_flags = FUSE_ASYNC_READ; AioRead::SetUp(); } }; class ReadCacheable: public Read { public: virtual void SetUp() { const char *node = "vfs.fusefs.data_cache_mode"; int val = 0; size_t size = sizeof(val); FuseTest::SetUp(); ASSERT_EQ(0, sysctlbyname(node, &val, &size, NULL, 0)) << strerror(errno); if (val == 0) GTEST_SKIP() << "fusefs data caching must be enabled for this test"; } }; class ReadAhead: public ReadCacheable, public WithParamInterface> { virtual void SetUp() { m_maxreadahead = get<1>(GetParam()); m_noclusterr = get<0>(GetParam()); ReadCacheable::SetUp(); } }; /* AIO reads need to set the header's pid field correctly */ /* https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=236379 */ TEST_F(AioRead, aio_read) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); char buf[bufsize]; struct aiocb iocb, *piocb; expect_lookup(RELPATH, ino, bufsize); expect_open(ino, 0, 1); expect_read(ino, 0, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); iocb.aio_nbytes = bufsize; iocb.aio_fildes = fd; iocb.aio_buf = buf; iocb.aio_offset = 0; iocb.aio_sigevent.sigev_notify = SIGEV_NONE; ASSERT_EQ(0, aio_read(&iocb)) << strerror(errno); ASSERT_EQ(bufsize, aio_waitcomplete(&piocb, NULL)) << strerror(errno); ASSERT_EQ(0, memcmp(buf, CONTENTS, bufsize)); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * Without the FUSE_ASYNC_READ mount option, fuse(4) should ensure that there * is at most one outstanding read operation per file handle */ TEST_F(AioRead, async_read_disabled) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; uint64_t ino = 42; int fd; ssize_t bufsize = 50; char buf0[bufsize], buf1[bufsize]; off_t off0 = 0; off_t off1 = 65536; struct aiocb iocb0, iocb1; volatile sig_atomic_t read_count = 0; expect_lookup(RELPATH, ino, 131072); expect_open(ino, 0, 1); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_READ && in.header.nodeid == ino && in.body.read.fh == FH && in.body.read.offset == (uint64_t)off0); }, Eq(true)), _) ).WillRepeatedly(Invoke([&](auto in __unused, auto &out __unused) { read_count++; /* Filesystem is slow to respond */ })); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_READ && in.header.nodeid == ino && in.body.read.fh == FH && in.body.read.offset == (uint64_t)off1); }, Eq(true)), _) ).WillRepeatedly(Invoke([&](auto in __unused, auto &out __unused) { read_count++; /* Filesystem is slow to respond */ })); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); /* * Submit two AIO read requests, and respond to neither. If the * filesystem ever gets the second read request, then we failed to * limit outstanding reads. */ iocb0.aio_nbytes = bufsize; iocb0.aio_fildes = fd; iocb0.aio_buf = buf0; iocb0.aio_offset = off0; iocb0.aio_sigevent.sigev_notify = SIGEV_NONE; ASSERT_EQ(0, aio_read(&iocb0)) << strerror(errno); iocb1.aio_nbytes = bufsize; iocb1.aio_fildes = fd; iocb1.aio_buf = buf1; iocb1.aio_offset = off1; iocb1.aio_sigevent.sigev_notify = SIGEV_NONE; ASSERT_EQ(0, aio_read(&iocb1)) << strerror(errno); /* * Sleep for awhile to make sure the kernel has had a chance to issue * the second read, even though the first has not yet returned */ nap(); EXPECT_EQ(read_count, 1); m_mock->kill_daemon(); /* Wait for AIO activity to complete, but ignore errors */ (void)aio_waitcomplete(NULL, NULL); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * With the FUSE_ASYNC_READ mount option, fuse(4) may issue multiple * simultaneous read requests on the same file handle. */ TEST_F(AsyncRead, async_read) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; uint64_t ino = 42; int fd; ssize_t bufsize = 50; char buf0[bufsize], buf1[bufsize]; off_t off0 = 0; off_t off1 = 65536; struct aiocb iocb0, iocb1; sem_t sem; ASSERT_EQ(0, sem_init(&sem, 0, 0)) << strerror(errno); expect_lookup(RELPATH, ino, 131072); expect_open(ino, 0, 1); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_READ && in.header.nodeid == ino && in.body.read.fh == FH && in.body.read.offset == (uint64_t)off0); }, Eq(true)), _) ).WillOnce(Invoke([&](auto in __unused, auto &out __unused) { sem_post(&sem); /* Filesystem is slow to respond */ })); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_READ && in.header.nodeid == ino && in.body.read.fh == FH && in.body.read.offset == (uint64_t)off1); }, Eq(true)), _) ).WillOnce(Invoke([&](auto in __unused, auto &out __unused) { sem_post(&sem); /* Filesystem is slow to respond */ })); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); /* * Submit two AIO read requests, but respond to neither. Ensure that * we received both. */ iocb0.aio_nbytes = bufsize; iocb0.aio_fildes = fd; iocb0.aio_buf = buf0; iocb0.aio_offset = off0; iocb0.aio_sigevent.sigev_notify = SIGEV_NONE; ASSERT_EQ(0, aio_read(&iocb0)) << strerror(errno); iocb1.aio_nbytes = bufsize; iocb1.aio_fildes = fd; iocb1.aio_buf = buf1; iocb1.aio_offset = off1; iocb1.aio_sigevent.sigev_notify = SIGEV_NONE; ASSERT_EQ(0, aio_read(&iocb1)) << strerror(errno); /* Wait until both reads have reached the daemon */ ASSERT_EQ(0, sem_wait(&sem)) << strerror(errno); ASSERT_EQ(0, sem_wait(&sem)) << strerror(errno); m_mock->kill_daemon(); /* Wait for AIO activity to complete, but ignore errors */ (void)aio_waitcomplete(NULL, NULL); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* 0-length reads shouldn't cause any confusion */ TEST_F(Read, direct_io_read_nothing) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; uint64_t ino = 42; int fd; uint64_t offset = 100; char buf[80]; expect_lookup(RELPATH, ino, offset + 1000); expect_open(ino, FOPEN_DIRECT_IO, 1); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); ASSERT_EQ(0, pread(fd, buf, 0, offset)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * With direct_io, reads should not fill the cache. They should go straight to * the daemon */ TEST_F(Read, direct_io_pread) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; uint64_t offset = 100; ssize_t bufsize = strlen(CONTENTS); char buf[bufsize]; expect_lookup(RELPATH, ino, offset + bufsize); expect_open(ino, FOPEN_DIRECT_IO, 1); expect_read(ino, offset, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, pread(fd, buf, bufsize, offset)) << strerror(errno); ASSERT_EQ(0, memcmp(buf, CONTENTS, bufsize)); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * With direct_io, filesystems are allowed to return less data than is * requested. fuse(4) should return a short read to userland. */ TEST_F(Read, direct_io_short_read) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefghijklmnop"; uint64_t ino = 42; int fd; uint64_t offset = 100; ssize_t bufsize = strlen(CONTENTS); ssize_t halfbufsize = bufsize / 2; char buf[bufsize]; expect_lookup(RELPATH, ino, offset + bufsize); expect_open(ino, FOPEN_DIRECT_IO, 1); expect_read(ino, offset, bufsize, halfbufsize, CONTENTS); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); ASSERT_EQ(halfbufsize, pread(fd, buf, bufsize, offset)) << strerror(errno); ASSERT_EQ(0, memcmp(buf, CONTENTS, halfbufsize)); /* Deliberately leak fd. close(2) will be tested in release.cc */ } TEST_F(Read, eio) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); char buf[bufsize]; expect_lookup(RELPATH, ino, bufsize); expect_open(ino, 0, 1); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_READ); }, Eq(true)), _) ).WillOnce(Invoke(ReturnErrno(EIO))); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); ASSERT_EQ(-1, read(fd, buf, bufsize)) << strerror(errno); ASSERT_EQ(EIO, errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * If the server returns a short read when direct io is not in use, that - * indicates EOF and we should update the file size. + * indicates EOF, because of a server-side truncation. We should invalidate + * all cached attributes. We may update the file size, */ TEST_F(ReadCacheable, eof) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefghijklmnop"; uint64_t ino = 42; int fd; uint64_t offset = 100; ssize_t bufsize = strlen(CONTENTS); ssize_t partbufsize = 3 * bufsize / 4; + ssize_t r; char buf[bufsize]; struct stat sb; expect_lookup(RELPATH, ino, offset + bufsize); expect_open(ino, 0, 1); expect_read(ino, 0, offset + bufsize, offset + partbufsize, CONTENTS); + expect_getattr(ino, offset + partbufsize); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); - ASSERT_EQ(partbufsize, pread(fd, buf, bufsize, offset)) - << strerror(errno); + r = pread(fd, buf, bufsize, offset); + ASSERT_LE(0, r) << strerror(errno); + EXPECT_EQ(partbufsize, r) << strerror(errno); ASSERT_EQ(0, fstat(fd, &sb)); EXPECT_EQ((off_t)(offset + partbufsize), sb.st_size); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* Like ReadCacheable.eof, but causes an entire buffer to be invalidated */ TEST_F(ReadCacheable, eof_of_whole_buffer) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefghijklmnop"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); off_t old_filesize = m_maxbcachebuf * 2 + bufsize; char buf[bufsize]; struct stat sb; expect_lookup(RELPATH, ino, old_filesize); expect_open(ino, 0, 1); expect_read(ino, 2 * m_maxbcachebuf, bufsize, bufsize, CONTENTS); expect_read(ino, m_maxbcachebuf, m_maxbcachebuf, 0, CONTENTS); + expect_getattr(ino, m_maxbcachebuf); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); /* Cache the third block */ ASSERT_EQ(bufsize, pread(fd, buf, bufsize, m_maxbcachebuf * 2)) << strerror(errno); /* Try to read the 2nd block, but it's past EOF */ ASSERT_EQ(0, pread(fd, buf, bufsize, m_maxbcachebuf)) << strerror(errno); ASSERT_EQ(0, fstat(fd, &sb)); EXPECT_EQ((off_t)(m_maxbcachebuf), sb.st_size); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * With the keep_cache option, the kernel may keep its read cache across * multiple open(2)s. */ TEST_F(ReadCacheable, keep_cache) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd0, fd1; ssize_t bufsize = strlen(CONTENTS); char buf[bufsize]; FuseTest::expect_lookup(RELPATH, ino, S_IFREG | 0644, bufsize, 2); expect_open(ino, FOPEN_KEEP_CACHE, 2); expect_read(ino, 0, bufsize, bufsize, CONTENTS); fd0 = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd0) << strerror(errno); ASSERT_EQ(bufsize, read(fd0, buf, bufsize)) << strerror(errno); fd1 = open(FULLPATH, O_RDWR); ASSERT_LE(0, fd1) << strerror(errno); /* * This read should be serviced by cache, even though it's on the other * file descriptor */ ASSERT_EQ(bufsize, read(fd1, buf, bufsize)) << strerror(errno); /* Deliberately leak fd0 and fd1. */ } /* * Without the keep_cache option, the kernel should drop its read caches on * every open */ TEST_F(Read, keep_cache_disabled) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd0, fd1; ssize_t bufsize = strlen(CONTENTS); char buf[bufsize]; FuseTest::expect_lookup(RELPATH, ino, S_IFREG | 0644, bufsize, 2); expect_open(ino, 0, 2); expect_read(ino, 0, bufsize, bufsize, CONTENTS); fd0 = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd0) << strerror(errno); ASSERT_EQ(bufsize, read(fd0, buf, bufsize)) << strerror(errno); fd1 = open(FULLPATH, O_RDWR); ASSERT_LE(0, fd1) << strerror(errno); /* * This read should not be serviced by cache, even though it's on the * original file descriptor */ expect_read(ino, 0, bufsize, bufsize, CONTENTS); ASSERT_EQ(0, lseek(fd0, 0, SEEK_SET)) << strerror(errno); ASSERT_EQ(bufsize, read(fd0, buf, bufsize)) << strerror(errno); /* Deliberately leak fd0 and fd1. */ } TEST_F(ReadCacheable, mmap) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t len; size_t bufsize = strlen(CONTENTS); void *p; len = getpagesize(); expect_lookup(RELPATH, ino, bufsize); expect_open(ino, 0, 1); /* mmap may legitimately try to read more data than is available */ EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_READ && in.header.nodeid == ino && in.body.read.fh == Read::FH && in.body.read.offset == 0 && in.body.read.size >= bufsize); }, Eq(true)), _) ).WillOnce(Invoke(ReturnImmediate([=](auto in __unused, auto& out) { out.header.len = sizeof(struct fuse_out_header) + bufsize; memmove(out.body.bytes, CONTENTS, bufsize); }))); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); p = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, 0); ASSERT_NE(MAP_FAILED, p) << strerror(errno); ASSERT_EQ(0, memcmp(p, CONTENTS, bufsize)); + + ASSERT_EQ(0, munmap(p, len)) << strerror(errno); + /* Deliberately leak fd. close(2) will be tested in release.cc */ +} + +/* + * A read via mmap comes up short, indicating that the file was truncated + * server-side. + */ +TEST_F(ReadCacheable, mmap_eof) +{ + const char FULLPATH[] = "mountpoint/some_file.txt"; + const char RELPATH[] = "some_file.txt"; + const char *CONTENTS = "abcdefgh"; + uint64_t ino = 42; + int fd; + ssize_t len; + size_t bufsize = strlen(CONTENTS); + struct stat sb; + void *p; + + len = getpagesize(); + + expect_lookup(RELPATH, ino, 100000); + expect_open(ino, 0, 1); + /* mmap may legitimately try to read more data than is available */ + EXPECT_CALL(*m_mock, process( + ResultOf([=](auto in) { + return (in.header.opcode == FUSE_READ && + in.header.nodeid == ino && + in.body.read.fh == Read::FH && + in.body.read.offset == 0 && + in.body.read.size >= bufsize); + }, Eq(true)), + _) + ).WillOnce(Invoke(ReturnImmediate([=](auto in __unused, auto& out) { + out.header.len = sizeof(struct fuse_out_header) + bufsize; + memmove(out.body.bytes, CONTENTS, bufsize); + }))); + expect_getattr(ino, bufsize); + + fd = open(FULLPATH, O_RDONLY); + ASSERT_LE(0, fd) << strerror(errno); + + p = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, 0); + ASSERT_NE(MAP_FAILED, p) << strerror(errno); + + /* The file size should be automatically truncated */ + ASSERT_EQ(0, memcmp(p, CONTENTS, bufsize)); + ASSERT_EQ(0, fstat(fd, &sb)) << strerror(errno); + EXPECT_EQ((off_t)bufsize, sb.st_size); ASSERT_EQ(0, munmap(p, len)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * Just as when FOPEN_DIRECT_IO is used, reads with O_DIRECT should bypass * cache and to straight to the daemon */ TEST_F(Read, o_direct) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); char buf[bufsize]; expect_lookup(RELPATH, ino, bufsize); expect_open(ino, 0, 1); expect_read(ino, 0, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); // Fill the cache ASSERT_EQ(bufsize, read(fd, buf, bufsize)) << strerror(errno); ASSERT_EQ(0, memcmp(buf, CONTENTS, bufsize)); // Reads with o_direct should bypass the cache expect_read(ino, 0, bufsize, bufsize, CONTENTS); ASSERT_EQ(0, fcntl(fd, F_SETFL, O_DIRECT)) << strerror(errno); ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno); ASSERT_EQ(bufsize, read(fd, buf, bufsize)) << strerror(errno); ASSERT_EQ(0, memcmp(buf, CONTENTS, bufsize)); /* Deliberately leak fd. close(2) will be tested in release.cc */ } TEST_F(Read, pread) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; /* * Set offset to a maxbcachebuf boundary so we'll be sure what offset * to read from. Without this, the read might start at a lower offset. */ uint64_t offset = m_maxbcachebuf; ssize_t bufsize = strlen(CONTENTS); char buf[bufsize]; expect_lookup(RELPATH, ino, offset + bufsize); expect_open(ino, 0, 1); expect_read(ino, offset, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, pread(fd, buf, bufsize, offset)) << strerror(errno); ASSERT_EQ(0, memcmp(buf, CONTENTS, bufsize)); /* Deliberately leak fd. close(2) will be tested in release.cc */ } TEST_F(Read, read) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); char buf[bufsize]; expect_lookup(RELPATH, ino, bufsize); expect_open(ino, 0, 1); expect_read(ino, 0, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, read(fd, buf, bufsize)) << strerror(errno); ASSERT_EQ(0, memcmp(buf, CONTENTS, bufsize)); /* Deliberately leak fd. close(2) will be tested in release.cc */ } TEST_F(Read_7_8, read) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); char buf[bufsize]; expect_lookup(RELPATH, ino, bufsize); expect_open(ino, 0, 1); expect_read(ino, 0, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, read(fd, buf, bufsize)) << strerror(errno); ASSERT_EQ(0, memcmp(buf, CONTENTS, bufsize)); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * If cacheing is enabled, the kernel should try to read an entire cache block * at a time. */ TEST_F(ReadCacheable, cache_block) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS0 = "abcdefghijklmnop"; uint64_t ino = 42; int fd; ssize_t bufsize = 8; ssize_t filesize = m_maxbcachebuf * 2; char *contents; char buf[bufsize]; const char *contents1 = CONTENTS0 + bufsize; contents = (char*)calloc(1, filesize); ASSERT_NE(NULL, contents); memmove(contents, CONTENTS0, strlen(CONTENTS0)); expect_lookup(RELPATH, ino, filesize); expect_open(ino, 0, 1); expect_read(ino, 0, m_maxbcachebuf, m_maxbcachebuf, contents); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, read(fd, buf, bufsize)) << strerror(errno); ASSERT_EQ(0, memcmp(buf, CONTENTS0, bufsize)); /* A subsequent read should be serviced by cache */ ASSERT_EQ(bufsize, read(fd, buf, bufsize)) << strerror(errno); ASSERT_EQ(0, memcmp(buf, contents1, bufsize)); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* Reading with sendfile should work (though it obviously won't be 0-copy) */ TEST_F(ReadCacheable, sendfile) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; size_t bufsize = strlen(CONTENTS); char buf[bufsize]; int sp[2]; off_t sbytes; expect_lookup(RELPATH, ino, bufsize); expect_open(ino, 0, 1); /* Like mmap, sendfile may request more data than is available */ EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_READ && in.header.nodeid == ino && in.body.read.fh == Read::FH && in.body.read.offset == 0 && in.body.read.size >= bufsize); }, Eq(true)), _) ).WillOnce(Invoke(ReturnImmediate([=](auto in __unused, auto& out) { out.header.len = sizeof(struct fuse_out_header) + bufsize; memmove(out.body.bytes, CONTENTS, bufsize); }))); ASSERT_EQ(0, socketpair(PF_LOCAL, SOCK_STREAM, 0, sp)) << strerror(errno); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); ASSERT_EQ(0, sendfile(fd, sp[1], 0, bufsize, NULL, &sbytes, 0)) << strerror(errno); ASSERT_EQ(static_cast(bufsize), read(sp[0], buf, bufsize)) << strerror(errno); ASSERT_EQ(0, memcmp(buf, CONTENTS, bufsize)); close(sp[1]); close(sp[0]); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* sendfile should fail gracefully if fuse declines the read */ /* https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=236466 */ TEST_F(ReadCacheable, DISABLED_sendfile_eio) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); int sp[2]; off_t sbytes; expect_lookup(RELPATH, ino, bufsize); expect_open(ino, 0, 1); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_READ); }, Eq(true)), _) ).WillOnce(Invoke(ReturnErrno(EIO))); ASSERT_EQ(0, socketpair(PF_LOCAL, SOCK_STREAM, 0, sp)) << strerror(errno); fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); ASSERT_NE(0, sendfile(fd, sp[1], 0, bufsize, NULL, &sbytes, 0)); close(sp[1]); close(sp[0]); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * Sequential reads should use readahead. And if allowed, large reads should * be clustered. */ TEST_P(ReadAhead, readahead) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; uint64_t ino = 42; int fd, maxcontig, clustersize; ssize_t bufsize = 4 * m_maxbcachebuf; ssize_t filesize = bufsize; uint64_t len; char *rbuf, *contents; off_t offs; contents = (char*)malloc(filesize); ASSERT_NE(NULL, contents); memset(contents, 'X', filesize); rbuf = (char*)calloc(1, bufsize); expect_lookup(RELPATH, ino, filesize); expect_open(ino, 0, 1); maxcontig = m_noclusterr ? m_maxbcachebuf : m_maxbcachebuf + (int)get<1>(GetParam()); clustersize = MIN(maxcontig, MAXPHYS); for (offs = 0; offs < bufsize; offs += clustersize) { len = std::min((size_t)clustersize, (size_t)(filesize - offs)); expect_read(ino, offs, len, len, contents + offs); } fd = open(FULLPATH, O_RDONLY); ASSERT_LE(0, fd) << strerror(errno); /* Set the internal readahead counter to a "large" value */ ASSERT_EQ(0, fcntl(fd, F_READAHEAD, 1'000'000'000)) << strerror(errno); ASSERT_EQ(bufsize, read(fd, rbuf, bufsize)) << strerror(errno); ASSERT_EQ(0, memcmp(rbuf, contents, bufsize)); /* Deliberately leak fd. close(2) will be tested in release.cc */ } INSTANTIATE_TEST_CASE_P(RA, ReadAhead, Values(tuple(false, 0u), tuple(false, 0x10000), tuple(false, 0x20000), tuple(false, 0x30000), tuple(true, 0u), tuple(true, 0x10000), tuple(true, 0x20000))); Index: projects/fuse2/tests/sys/fs/fusefs/write.cc =================================================================== --- projects/fuse2/tests/sys/fs/fusefs/write.cc (revision 349377) +++ projects/fuse2/tests/sys/fs/fusefs/write.cc (revision 349378) @@ -1,1034 +1,1067 @@ /*- * 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 #include #include #include #include #include #include } #include "mockfs.hh" #include "utils.hh" using namespace testing; class Write: public FuseTest { public: static sig_atomic_t s_sigxfsz; void SetUp() { s_sigxfsz = 0; FuseTest::SetUp(); } void TearDown() { struct sigaction sa; bzero(&sa, sizeof(sa)); sa.sa_handler = SIG_DFL; sigaction(SIGXFSZ, &sa, NULL); FuseTest::TearDown(); } void expect_lookup(const char *relpath, uint64_t ino, uint64_t size) { FuseTest::expect_lookup(relpath, ino, S_IFREG | 0644, size, 1); } void expect_release(uint64_t ino, ProcessMockerT r) { EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_RELEASE && in.header.nodeid == ino); }, Eq(true)), _) ).WillRepeatedly(Invoke(r)); } void expect_write(uint64_t ino, uint64_t offset, uint64_t isize, uint64_t osize, const void *contents) { FuseTest::expect_write(ino, offset, isize, osize, 0, 0, contents); } /* Expect a write that may or may not come, depending on the cache mode */ void maybe_expect_write(uint64_t ino, uint64_t offset, uint64_t size, const void *contents) { EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { const char *buf = (const char*)in.body.bytes + sizeof(struct fuse_write_in); return (in.header.opcode == FUSE_WRITE && in.header.nodeid == ino && in.body.write.offset == offset && in.body.write.size == size && 0 == bcmp(buf, contents, size)); }, Eq(true)), _) ).Times(AtMost(1)) .WillRepeatedly(Invoke( ReturnImmediate([=](auto in __unused, auto& out) { SET_OUT_HEADER_LEN(out, write); out.body.write.size = size; }) )); } }; class WriteCacheable: public Write { public: virtual void SetUp() { const char *node = "vfs.fusefs.data_cache_mode"; int val = 0; size_t size = sizeof(val); FuseTest::SetUp(); ASSERT_EQ(0, sysctlbyname(node, &val, &size, NULL, 0)) << strerror(errno); if (val == 0) GTEST_SKIP() << "fusefs data caching must be enabled for this test"; } }; sig_atomic_t Write::s_sigxfsz = 0; class Write_7_8: public FuseTest { public: virtual void SetUp() { m_kernel_minor_version = 8; FuseTest::SetUp(); } void expect_lookup(const char *relpath, uint64_t ino, uint64_t size) { FuseTest::expect_lookup_7_8(relpath, ino, S_IFREG | 0644, size, 1); } }; class AioWrite: public Write { virtual void SetUp() { const char *node = "vfs.aio.enable_unsafe"; int val = 0; size_t size = sizeof(val); FuseTest::SetUp(); ASSERT_EQ(0, sysctlbyname(node, &val, &size, NULL, 0)) << strerror(errno); if (!val) GTEST_SKIP() << "vfs.aio.enable_unsafe must be set for this test"; } }; /* Tests for the write-through cache mode */ class WriteThrough: public Write { public: virtual void SetUp() { const char *cache_mode_node = "vfs.fusefs.data_cache_mode"; int val = 0; size_t size = sizeof(val); FuseTest::SetUp(); if (IsSkipped()) return; ASSERT_EQ(0, sysctlbyname(cache_mode_node, &val, &size, NULL, 0)) << strerror(errno); if (val != 1) GTEST_SKIP() << "vfs.fusefs.data_cache_mode must be set to 1 " "(writethrough) for this test"; } void expect_write(uint64_t ino, uint64_t offset, uint64_t isize, uint64_t osize, const void *contents) { FuseTest::expect_write(ino, offset, isize, osize, 0, FUSE_WRITE_CACHE, contents); } }; /* Tests for the writeback cache mode */ class WriteBack: public Write { public: virtual void SetUp() { const char *node = "vfs.fusefs.data_cache_mode"; int val = 0; size_t size = sizeof(val); FuseTest::SetUp(); if (IsSkipped()) return; ASSERT_EQ(0, sysctlbyname(node, &val, &size, NULL, 0)) << strerror(errno); if (val != 2) GTEST_SKIP() << "vfs.fusefs.data_cache_mode must be set to 2 " "(writeback) for this test"; } void expect_write(uint64_t ino, uint64_t offset, uint64_t isize, uint64_t osize, const void *contents) { FuseTest::expect_write(ino, offset, isize, osize, FUSE_WRITE_CACHE, 0, contents); } }; class WriteBackAsync: public WriteBack { public: virtual void SetUp() { m_async = true; WriteBack::SetUp(); } }; /* Tests for clustered writes with WriteBack cacheing */ class WriteCluster: public WriteBack { public: virtual void SetUp() { if (MAXPHYS < 2 * DFLTPHYS) GTEST_SKIP() << "MAXPHYS must be at least twice DFLTPHYS" << "for this test"; m_async = true; m_maxwrite = MAXPHYS; WriteBack::SetUp(); } }; void sigxfsz_handler(int __unused sig) { Write::s_sigxfsz = 1; } /* AIO writes need to set the header's pid field correctly */ /* https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=236379 */ TEST_F(AioWrite, DISABLED_aio_write) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; uint64_t offset = 4096; int fd; ssize_t bufsize = strlen(CONTENTS); struct aiocb iocb, *piocb; expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); expect_write(ino, offset, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_WRONLY); EXPECT_LE(0, fd) << strerror(errno); iocb.aio_nbytes = bufsize; iocb.aio_fildes = fd; iocb.aio_buf = (void *)CONTENTS; iocb.aio_offset = offset; iocb.aio_sigevent.sigev_notify = SIGEV_NONE; ASSERT_EQ(0, aio_write(&iocb)) << strerror(errno); ASSERT_EQ(bufsize, aio_waitcomplete(&piocb, NULL)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * When a file is opened with O_APPEND, we should forward that flag to * FUSE_OPEN (tested by Open.o_append) but still attempt to calculate the * offset internally. That way we'll work both with filesystems that * understand O_APPEND (and ignore the offset) and filesystems that don't (and * simply use the offset). * * Note that verifying the O_APPEND flag in FUSE_OPEN is done in the * Open.o_append test. */ TEST_F(Write, append) { const ssize_t BUFSIZE = 9; const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char CONTENTS[BUFSIZE] = "abcdefgh"; uint64_t ino = 42; /* * Set offset to a maxbcachebuf boundary so we don't need to RMW when * using writeback caching */ uint64_t initial_offset = m_maxbcachebuf; int fd; expect_lookup(RELPATH, ino, initial_offset); expect_open(ino, 0, 1); expect_write(ino, initial_offset, BUFSIZE, BUFSIZE, CONTENTS); /* Must open O_RDWR or fuse(4) implicitly sets direct_io */ fd = open(FULLPATH, O_RDWR | O_APPEND); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(BUFSIZE, write(fd, CONTENTS, BUFSIZE)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* If a file is cached, then appending to the end should not cause a read */ TEST_F(Write, append_to_cached) { const ssize_t BUFSIZE = 9; const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; char *oldcontents, *oldbuf; const char CONTENTS[BUFSIZE] = "abcdefgh"; uint64_t ino = 42; /* * Set offset in between maxbcachebuf boundary to test buffer handling */ uint64_t oldsize = m_maxbcachebuf / 2; int fd; oldcontents = (char*)calloc(1, oldsize); ASSERT_NE(NULL, oldcontents) << strerror(errno); oldbuf = (char*)malloc(oldsize); ASSERT_NE(NULL, oldbuf) << strerror(errno); expect_lookup(RELPATH, ino, oldsize); expect_open(ino, 0, 1); expect_read(ino, 0, oldsize, oldsize, oldcontents); maybe_expect_write(ino, oldsize, BUFSIZE, CONTENTS); /* Must open O_RDWR or fuse(4) implicitly sets direct_io */ fd = open(FULLPATH, O_RDWR | O_APPEND); EXPECT_LE(0, fd) << strerror(errno); /* Read the old data into the cache */ ASSERT_EQ((ssize_t)oldsize, read(fd, oldbuf, oldsize)) << strerror(errno); /* Write the new data. There should be no more read operations */ ASSERT_EQ(BUFSIZE, write(fd, CONTENTS, BUFSIZE)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } TEST_F(Write, append_direct_io) { const ssize_t BUFSIZE = 9; const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char CONTENTS[BUFSIZE] = "abcdefgh"; uint64_t ino = 42; uint64_t initial_offset = 4096; int fd; expect_lookup(RELPATH, ino, initial_offset); expect_open(ino, FOPEN_DIRECT_IO, 1); expect_write(ino, initial_offset, BUFSIZE, BUFSIZE, CONTENTS); fd = open(FULLPATH, O_WRONLY | O_APPEND); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(BUFSIZE, write(fd, CONTENTS, BUFSIZE)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* A direct write should evict any overlapping cached data */ TEST_F(Write, direct_io_evicts_cache) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char CONTENTS0[] = "abcdefgh"; const char CONTENTS1[] = "ijklmnop"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS0) + 1; char readbuf[bufsize]; expect_lookup(RELPATH, ino, bufsize); expect_open(ino, 0, 1); expect_read(ino, 0, bufsize, bufsize, CONTENTS0); expect_write(ino, 0, bufsize, bufsize, CONTENTS1); fd = open(FULLPATH, O_RDWR); EXPECT_LE(0, fd) << strerror(errno); // Prime cache ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno); // Write directly, evicting cache ASSERT_EQ(0, fcntl(fd, F_SETFL, O_DIRECT)) << strerror(errno); ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno); ASSERT_EQ(bufsize, write(fd, CONTENTS1, bufsize)) << strerror(errno); // Read again. Cache should be bypassed expect_read(ino, 0, bufsize, bufsize, CONTENTS1); ASSERT_EQ(0, fcntl(fd, F_SETFL, 0)) << strerror(errno); ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno); ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno); ASSERT_STREQ(readbuf, CONTENTS1); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * If the server doesn't return FOPEN_DIRECT_IO during FUSE_OPEN, then it's not * allowed to return a short write for that file handle. However, if it does * then we should still do our darndest to handle it by resending the unwritten * portion. */ TEST_F(Write, indirect_io_short_write) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefghijklmnop"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); ssize_t bufsize0 = 11; ssize_t bufsize1 = strlen(CONTENTS) - bufsize0; const char *contents1 = CONTENTS + bufsize0; expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); expect_write(ino, 0, bufsize, bufsize0, CONTENTS); expect_write(ino, bufsize0, bufsize1, bufsize1, contents1); fd = open(FULLPATH, O_WRONLY); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * When the direct_io option is used, filesystems are allowed to write less * data than requested. We should return the short write to userland. */ TEST_F(Write, direct_io_short_write) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefghijklmnop"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); ssize_t halfbufsize = bufsize / 2; expect_lookup(RELPATH, ino, 0); expect_open(ino, FOPEN_DIRECT_IO, 1); expect_write(ino, 0, bufsize, halfbufsize, CONTENTS); fd = open(FULLPATH, O_WRONLY); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(halfbufsize, write(fd, CONTENTS, bufsize)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * An insidious edge case: the filesystem returns a short write, and the * difference between what we requested and what it actually wrote crosses an * iov element boundary */ TEST_F(Write, direct_io_short_write_iov) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS0 = "abcdefgh"; const char *CONTENTS1 = "ijklmnop"; const char *EXPECTED0 = "abcdefghijklmnop"; uint64_t ino = 42; int fd; ssize_t size0 = strlen(CONTENTS0) - 1; ssize_t size1 = strlen(CONTENTS1) + 1; ssize_t totalsize = size0 + size1; struct iovec iov[2]; expect_lookup(RELPATH, ino, 0); expect_open(ino, FOPEN_DIRECT_IO, 1); expect_write(ino, 0, totalsize, size0, EXPECTED0); fd = open(FULLPATH, O_WRONLY); EXPECT_LE(0, fd) << strerror(errno); iov[0].iov_base = (void*)CONTENTS0; iov[0].iov_len = strlen(CONTENTS0); iov[1].iov_base = (void*)CONTENTS1; iov[1].iov_len = strlen(CONTENTS1); ASSERT_EQ(size0, writev(fd, iov, 2)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* fusefs should respect RLIMIT_FSIZE */ TEST_F(Write, rlimit_fsize) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; struct rlimit rl; ssize_t bufsize = strlen(CONTENTS); off_t offset = 1'000'000'000; uint64_t ino = 42; int fd; expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); rl.rlim_cur = offset; rl.rlim_max = 10 * offset; ASSERT_EQ(0, setrlimit(RLIMIT_FSIZE, &rl)) << strerror(errno); ASSERT_NE(SIG_ERR, signal(SIGXFSZ, sigxfsz_handler)) << strerror(errno); fd = open(FULLPATH, O_WRONLY); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(-1, pwrite(fd, CONTENTS, bufsize, offset)); EXPECT_EQ(EFBIG, errno); EXPECT_EQ(1, s_sigxfsz); /* Deliberately leak fd. close(2) will be tested in release.cc */ } +/* + * A short read indicates EOF. Test that nothing bad happens if we get EOF + * during the R of a RMW operation. + */ +TEST_F(WriteCacheable, eof_during_rmw) +{ + const char FULLPATH[] = "mountpoint/some_file.txt"; + const char RELPATH[] = "some_file.txt"; + const char *CONTENTS = "abcdefgh"; + const char *INITIAL = "XXXXXXXXXX"; + uint64_t ino = 42; + uint64_t offset = 1; + ssize_t bufsize = strlen(CONTENTS); + off_t orig_fsize = 10; + off_t truncated_fsize = 5; + off_t final_fsize = bufsize; + int fd; + + FuseTest::expect_lookup(RELPATH, ino, S_IFREG | 0644, orig_fsize, 1); + expect_open(ino, 0, 1); + expect_read(ino, 0, orig_fsize, truncated_fsize, INITIAL, O_RDWR); + expect_getattr(ino, truncated_fsize); + expect_read(ino, 0, final_fsize, final_fsize, INITIAL, O_RDWR); + maybe_expect_write(ino, offset, bufsize, CONTENTS); + + fd = open(FULLPATH, O_RDWR); + EXPECT_LE(0, fd) << strerror(errno); + + ASSERT_EQ(bufsize, pwrite(fd, CONTENTS, bufsize, offset)) + << strerror(errno); + /* Deliberately leak fd. close(2) will be tested in release.cc */ +} + /* * If the kernel cannot be sure which uid, gid, or pid was responsible for a * write, then it must set the FUSE_WRITE_CACHE bit */ /* https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=236378 */ TEST_F(WriteCacheable, mmap) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); void *p; uint64_t offset = 10; size_t len; void *zeros, *expected; len = getpagesize(); zeros = calloc(1, len); ASSERT_NE(NULL, zeros); expected = calloc(1, len); ASSERT_NE(NULL, expected); memmove((uint8_t*)expected + offset, CONTENTS, bufsize); expect_lookup(RELPATH, ino, len); expect_open(ino, 0, 1); expect_read(ino, 0, len, len, zeros); /* * Writes from the pager may or may not be associated with the correct * pid, so they must set FUSE_WRITE_CACHE. */ FuseTest::expect_write(ino, 0, len, len, FUSE_WRITE_CACHE, 0, expected); expect_flush(ino, 1, ReturnErrno(0)); expect_release(ino, ReturnErrno(0)); fd = open(FULLPATH, O_RDWR); EXPECT_LE(0, fd) << strerror(errno); p = mmap(NULL, len, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0); ASSERT_NE(MAP_FAILED, p) << strerror(errno); memmove((uint8_t*)p + offset, CONTENTS, bufsize); ASSERT_EQ(0, munmap(p, len)) << strerror(errno); close(fd); // Write mmap'd data on close free(expected); free(zeros); } TEST_F(WriteThrough, pwrite) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; uint64_t offset = 65536; int fd; ssize_t bufsize = strlen(CONTENTS); expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); expect_write(ino, offset, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_WRONLY); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, pwrite(fd, CONTENTS, bufsize, offset)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } TEST_F(Write, write) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); expect_write(ino, 0, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_WRONLY); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* fuse(4) should not issue writes of greater size than the daemon requests */ TEST_F(Write, write_large) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; int *contents; uint64_t ino = 42; int fd; ssize_t halfbufsize, bufsize; halfbufsize = m_mock->m_maxwrite; bufsize = halfbufsize * 2; contents = (int*)malloc(bufsize); ASSERT_NE(NULL, contents); for (int i = 0; i < (int)bufsize / (int)sizeof(i); i++) { contents[i] = i; } expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); expect_write(ino, 0, halfbufsize, halfbufsize, contents); maybe_expect_write(ino, halfbufsize, halfbufsize, &contents[halfbufsize / sizeof(int)]); fd = open(FULLPATH, O_WRONLY); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, write(fd, contents, bufsize)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ free(contents); } TEST_F(Write, write_nothing) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = ""; uint64_t ino = 42; int fd; ssize_t bufsize = 0; expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); fd = open(FULLPATH, O_WRONLY); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } TEST_F(Write_7_8, write) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); expect_write_7_8(ino, 0, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_WRONLY); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* In writeback mode, dirty data should be written on close */ TEST_F(WriteBackAsync, close) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); expect_write(ino, 0, bufsize, bufsize, CONTENTS); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_SETATTR); }, Eq(true)), _) ).WillRepeatedly(Invoke(ReturnImmediate([=](auto i __unused, auto& out) { SET_OUT_HEADER_LEN(out, attr); out.body.attr.attr.ino = ino; // Must match nodeid }))); expect_flush(ino, 1, ReturnErrno(0)); expect_release(ino, ReturnErrno(0)); fd = open(FULLPATH, O_RDWR); ASSERT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno); close(fd); } /* In writeback mode, adjacent writes will be clustered together */ TEST_F(WriteCluster, clustering) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; uint64_t ino = 42; int i, fd; void *wbuf, *wbuf2x; ssize_t bufsize = 65536; off_t filesize = 327680; wbuf = malloc(bufsize); ASSERT_NE(NULL, wbuf) << strerror(errno); memset(wbuf, 'X', bufsize); wbuf2x = malloc(2 * bufsize); ASSERT_NE(NULL, wbuf2x) << strerror(errno); memset(wbuf2x, 'X', 2 * bufsize); expect_lookup(RELPATH, ino, filesize); expect_open(ino, 0, 1); /* * Writes of bufsize-bytes each should be clustered into greater sizes. * The amount of clustering is adaptive, so the first write actually * issued will be 2x bufsize and subsequent writes may be larger */ expect_write(ino, 0, 2 * bufsize, 2 * bufsize, wbuf2x); expect_write(ino, 2 * bufsize, 2 * bufsize, 2 * bufsize, wbuf2x); expect_flush(ino, 1, ReturnErrno(0)); expect_release(ino, ReturnErrno(0)); fd = open(FULLPATH, O_RDWR); ASSERT_LE(0, fd) << strerror(errno); for (i = 0; i < 4; i++) { ASSERT_EQ(bufsize, write(fd, wbuf, bufsize)) << strerror(errno); } close(fd); } /* * When clustering writes, an I/O error to any of the cluster's children should * not panic the system on unmount */ /* * Disabled because it panics. * https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=238565 */ TEST_F(WriteCluster, DISABLED_cluster_write_err) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; uint64_t ino = 42; int i, fd; void *wbuf; ssize_t bufsize = 65536; off_t filesize = 262144; wbuf = malloc(bufsize); ASSERT_NE(NULL, wbuf) << strerror(errno); memset(wbuf, 'X', bufsize); expect_lookup(RELPATH, ino, filesize); expect_open(ino, 0, 1); EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_WRITE); }, Eq(true)), _) ).WillRepeatedly(Invoke(ReturnErrno(EIO))); expect_flush(ino, 1, ReturnErrno(0)); expect_release(ino, ReturnErrno(0)); fd = open(FULLPATH, O_RDWR); ASSERT_LE(0, fd) << strerror(errno); for (i = 0; i < 3; i++) { ASSERT_EQ(bufsize, write(fd, wbuf, bufsize)) << strerror(errno); } close(fd); } /* * In writeback mode, writes to an O_WRONLY file could trigger reads from the * server. The FUSE protocol explicitly allows that. */ TEST_F(WriteBack, rmw) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; const char *INITIAL = "XXXXXXXXXX"; uint64_t ino = 42; uint64_t offset = 1; off_t fsize = 10; int fd; ssize_t bufsize = strlen(CONTENTS); FuseTest::expect_lookup(RELPATH, ino, S_IFREG | 0644, fsize, 1); expect_open(ino, 0, 1); expect_read(ino, 0, fsize, fsize, INITIAL, O_WRONLY); maybe_expect_write(ino, offset, bufsize, CONTENTS); fd = open(FULLPATH, O_WRONLY); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, pwrite(fd, CONTENTS, bufsize, offset)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * Without direct_io, writes should be committed to cache */ TEST_F(WriteBack, cache) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); char readbuf[bufsize]; expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); expect_write(ino, 0, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_RDWR); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno); /* * A subsequent read should be serviced by cache, without querying the * filesystem daemon */ ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno); ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * With O_DIRECT, writes should be not committed to cache. Admittedly this is * an odd test, because it would be unusual to use O_DIRECT for writes but not * reads. */ TEST_F(WriteBack, o_direct) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); char readbuf[bufsize]; expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); FuseTest::expect_write(ino, 0, bufsize, bufsize, 0, FUSE_WRITE_CACHE, CONTENTS); expect_read(ino, 0, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_RDWR | O_DIRECT); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno); /* A subsequent read must query the daemon because cache is empty */ ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno); ASSERT_EQ(0, fcntl(fd, F_SETFL, 0)) << strerror(errno); ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * When mounted with -o async, the writeback cache mode should delay writes */ TEST_F(WriteBackAsync, delay) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); /* Write should be cached, but FUSE_WRITE shouldn't be sent */ EXPECT_CALL(*m_mock, process( ResultOf([=](auto in) { return (in.header.opcode == FUSE_WRITE); }, Eq(true)), _) ).Times(0); fd = open(FULLPATH, O_RDWR); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno); /* Don't close the file because that would flush the cache */ } /* * In WriteBack mode, writes may be cached beyond what the server thinks is the * EOF. In this case, a short read at EOF should _not_ cause fusefs to update * the file's size. */ TEST_F(WriteBackAsync, eof) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS0 = "abcdefgh"; const char *CONTENTS1 = "ijklmnop"; uint64_t ino = 42; int fd; off_t offset = m_maxbcachebuf; ssize_t wbufsize = strlen(CONTENTS1); off_t old_filesize = (off_t)strlen(CONTENTS0); ssize_t rbufsize = 2 * old_filesize; char readbuf[rbufsize]; size_t holesize = rbufsize - old_filesize; char hole[holesize]; struct stat sb; ssize_t r; expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); expect_read(ino, 0, m_maxbcachebuf, old_filesize, CONTENTS0); fd = open(FULLPATH, O_RDWR); EXPECT_LE(0, fd) << strerror(errno); /* Write and cache data beyond EOF */ ASSERT_EQ(wbufsize, pwrite(fd, CONTENTS1, wbufsize, offset)) << strerror(errno); /* Read from the old EOF */ r = pread(fd, readbuf, rbufsize, 0); ASSERT_LE(0, r) << strerror(errno); EXPECT_EQ(rbufsize, r) << "read should've synthesized a hole"; EXPECT_EQ(0, memcmp(CONTENTS0, readbuf, old_filesize)); bzero(hole, holesize); EXPECT_EQ(0, memcmp(hole, readbuf + old_filesize, holesize)); /* The file's size should still be what was established by pwrite */ ASSERT_EQ(0, fstat(fd, &sb)) << strerror(errno); EXPECT_EQ(offset + wbufsize, sb.st_size); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* * Without direct_io, writes should be committed to cache */ TEST_F(WriteThrough, writethrough) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); char readbuf[bufsize]; expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); expect_write(ino, 0, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_RDWR); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno); /* * A subsequent read should be serviced by cache, without querying the * filesystem daemon */ ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno); ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno); /* Deliberately leak fd. close(2) will be tested in release.cc */ } /* With writethrough caching, writes update the cached file size */ TEST_F(WriteThrough, update_file_size) { const char FULLPATH[] = "mountpoint/some_file.txt"; const char RELPATH[] = "some_file.txt"; const char *CONTENTS = "abcdefgh"; struct stat sb; uint64_t ino = 42; int fd; ssize_t bufsize = strlen(CONTENTS); expect_lookup(RELPATH, ino, 0); expect_open(ino, 0, 1); expect_write(ino, 0, bufsize, bufsize, CONTENTS); fd = open(FULLPATH, O_RDWR); EXPECT_LE(0, fd) << strerror(errno); ASSERT_EQ(bufsize, write(fd, CONTENTS, bufsize)) << strerror(errno); /* Get cached attributes */ ASSERT_EQ(0, fstat(fd, &sb)) << strerror(errno); ASSERT_EQ(bufsize, sb.st_size); /* Deliberately leak fd. close(2) will be tested in release.cc */ }