diff --git a/sys/fs/fuse/fuse_io.c b/sys/fs/fuse/fuse_io.c index 06cfb2a933ba..e2a7c19d1cfa 100644 --- a/sys/fs/fuse/fuse_io.c +++ b/sys/fs/fuse/fuse_io.c @@ -1,1024 +1,1020 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 2007-2009 Google Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are * met: * * * Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * * Redistributions in binary form must reproduce the above * copyright notice, this list of conditions and the following disclaimer * in the documentation and/or other materials provided with the * distribution. * * Neither the name of Google Inc. nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * Copyright (C) 2005 Csaba Henk. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "fuse.h" #include "fuse_file.h" #include "fuse_node.h" #include "fuse_internal.h" #include "fuse_ipc.h" #include "fuse_io.h" SDT_PROVIDER_DECLARE(fusefs); /* * Fuse trace probe: * arg0: verbosity. Higher numbers give more verbose messages * arg1: Textual message */ SDT_PROBE_DEFINE2(fusefs, , io, trace, "int", "char*"); static void fuse_io_clear_suid_on_write(struct vnode *vp, struct ucred *cred, struct thread *td); static int fuse_read_directbackend(struct vnode *vp, struct uio *uio, struct ucred *cred, struct fuse_filehandle *fufh); static int fuse_read_biobackend(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred, struct fuse_filehandle *fufh, pid_t pid); static int fuse_write_directbackend(struct vnode *vp, struct uio *uio, struct ucred *cred, struct fuse_filehandle *fufh, off_t filesize, int ioflag, bool pages); static int fuse_write_biobackend(struct vnode *vp, struct uio *uio, struct ucred *cred, struct fuse_filehandle *fufh, int ioflag, pid_t pid); /* * FreeBSD clears the SUID and SGID bits on any write by a non-root user. */ static void fuse_io_clear_suid_on_write(struct vnode *vp, struct ucred *cred, struct thread *td) { struct fuse_data *data; struct mount *mp; struct vattr va; int dataflags; mp = vnode_mount(vp); data = fuse_get_mpdata(mp); dataflags = data->dataflags; if (dataflags & FSESS_DEFAULT_PERMISSIONS) { if (priv_check_cred(cred, PRIV_VFS_RETAINSUGID)) { fuse_internal_getattr(vp, &va, cred, td); if (va.va_mode & (S_ISUID | S_ISGID)) { mode_t mode = va.va_mode & ~(S_ISUID | S_ISGID); /* Clear all vattr fields except mode */ vattr_null(&va); va.va_mode = mode; /* * Ignore fuse_internal_setattr's return value, * because at this point the write operation has * already succeeded and we don't want to return * failing status for that. */ (void)fuse_internal_setattr(vp, &va, td, NULL); } } } } SDT_PROBE_DEFINE5(fusefs, , io, io_dispatch, "struct vnode*", "struct uio*", "int", "struct ucred*", "struct fuse_filehandle*"); int fuse_io_dispatch(struct vnode *vp, struct uio *uio, int ioflag, bool pages, struct ucred *cred, pid_t pid) { struct fuse_filehandle *fufh; int err, directio; int fflag; bool closefufh = false; MPASS(vp->v_type == VREG || vp->v_type == VDIR); fflag = (uio->uio_rw == UIO_READ) ? FREAD : FWRITE; err = fuse_filehandle_getrw(vp, fflag, &fufh, cred, pid); if (err == EBADF && vnode_mount(vp)->mnt_flag & MNT_EXPORTED) { /* * nfsd will do I/O without first doing VOP_OPEN. We * must implicitly open the file here */ err = fuse_filehandle_open(vp, fflag, &fufh, curthread, cred); closefufh = true; } else if (err) { printf("FUSE: io dispatch: filehandles are closed\n"); return err; } if (err) goto out; SDT_PROBE5(fusefs, , io, io_dispatch, vp, uio, ioflag, cred, fufh); /* * Ideally, when the daemon asks for direct io at open time, the * standard file flag should be set according to this, so that would * just change the default mode, which later on could be changed via * fcntl(2). * But this doesn't work, the O_DIRECT flag gets cleared at some point * (don't know where). So to make any use of the Fuse direct_io option, * we hardwire it into the file's private data (similarly to Linux, * btw.). */ directio = (ioflag & IO_DIRECT) || !fsess_opt_datacache(vnode_mount(vp)); switch (uio->uio_rw) { case UIO_READ: if (directio) { SDT_PROBE2(fusefs, , io, trace, 1, "direct read of vnode"); err = fuse_read_directbackend(vp, uio, cred, fufh); } else { SDT_PROBE2(fusefs, , io, trace, 1, "buffered read of vnode"); err = fuse_read_biobackend(vp, uio, ioflag, cred, fufh, pid); } break; case UIO_WRITE: - /* - * Kludge: simulate write-through caching via write-around - * caching. Same effect, as far as never caching dirty data, - * but slightly pessimal in that newly written data is not - * cached. - */ - if (directio || fuse_data_cache_mode == FUSE_CACHE_WT) { + if (directio) { const int iosize = fuse_iosize(vp); off_t start, end, filesize; SDT_PROBE2(fusefs, , io, trace, 1, "direct write of vnode"); err = fuse_vnode_size(vp, &filesize, cred, curthread); if (err) goto out; start = uio->uio_offset; end = start + uio->uio_resid; /* * Invalidate the write cache unless we're coming from * VOP_PUTPAGES, in which case we're writing _from_ the * write cache */ if (!pages ) v_inval_buf_range(vp, start, end, iosize); err = fuse_write_directbackend(vp, uio, cred, fufh, filesize, ioflag, pages); } else { SDT_PROBE2(fusefs, , io, trace, 1, "buffered write of vnode"); + if (fuse_data_cache_mode == FUSE_CACHE_WT) + ioflag |= IO_SYNC; err = fuse_write_biobackend(vp, uio, cred, fufh, ioflag, pid); } fuse_io_clear_suid_on_write(vp, cred, uio->uio_td); break; default: panic("uninterpreted mode passed to fuse_io_dispatch"); } out: if (closefufh) fuse_filehandle_close(vp, fufh, curthread, cred); return (err); } SDT_PROBE_DEFINE4(fusefs, , io, read_bio_backend_start, "int", "int", "int", "int"); SDT_PROBE_DEFINE2(fusefs, , io, read_bio_backend_feed, "int", "struct buf*"); SDT_PROBE_DEFINE4(fusefs, , io, read_bio_backend_end, "int", "ssize_t", "int", "struct buf*"); static int fuse_read_biobackend(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred, struct fuse_filehandle *fufh, pid_t pid) { struct buf *bp; daddr_t lbn; int bcount; int err, n = 0, on = 0; off_t filesize; const int biosize = fuse_iosize(vp); if (uio->uio_offset < 0) return (EINVAL); 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; } SDT_PROBE4(fusefs, , io, read_bio_backend_start, biosize, (int)lbn, on, bcount); /* TODO: readahead. See ext2_read for an example */ 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) n = MIN((unsigned)(bcount - 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); } 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) break; } out: fdisp_destroy(&fdi); return (err); } static int fuse_write_directbackend(struct vnode *vp, struct uio *uio, struct ucred *cred, struct fuse_filehandle *fufh, off_t filesize, int ioflag, bool pages) { struct fuse_vnode_data *fvdat = VTOFUD(vp); struct fuse_data *data; struct fuse_write_in *fwi; struct fuse_write_out *fwo; struct fuse_dispatcher fdi; size_t chunksize; void *fwi_data; off_t as_written_offset; int diff; int err = 0; bool direct_io = fufh->fuse_open_flags & FOPEN_DIRECT_IO; uint32_t write_flags; data = fuse_get_mpdata(vp->v_mount); /* * Don't set FUSE_WRITE_LOCKOWNER in write_flags. It can't be set * accurately when using POSIX AIO, libfuse doesn't use it, and I'm not * aware of any file systems that do. It was an attempt to add * Linux-style mandatory locking to the FUSE protocol, but mandatory * locking is deprecated even on Linux. See Linux commit * f33321141b273d60cbb3a8f56a5489baad82ba5e . */ /* * Set FUSE_WRITE_CACHE whenever we don't know the uid, gid, and/or pid * that originated a write. For example when writing from the * writeback cache. I don't know of a single file system that cares, * but the protocol says we're supposed to do this. */ write_flags = !pages && ( (ioflag & IO_DIRECT) || !fsess_opt_datacache(vnode_mount(vp)) || fuse_data_cache_mode != FUSE_CACHE_WB) ? 0 : FUSE_WRITE_CACHE; if (uio->uio_resid == 0) return (0); if (ioflag & IO_APPEND) uio_setoffset(uio, filesize); if (vn_rlimit_fsize(vp, uio, uio->uio_td)) return (EFBIG); fdisp_init(&fdi, 0); while (uio->uio_resid > 0) { chunksize = MIN(uio->uio_resid, data->max_write); fdi.iosize = sizeof(*fwi) + chunksize; fdisp_make_vp(&fdi, FUSE_WRITE, vp, uio->uio_td, cred); fwi = fdi.indata; fwi->fh = fufh->fh_id; fwi->offset = uio->uio_offset; fwi->size = chunksize; fwi->write_flags = write_flags; if (fuse_libabi_geq(data, 7, 9)) { fwi->flags = fufh_type_2_fflags(fufh->fufh_type); fwi_data = (char *)fdi.indata + sizeof(*fwi); } else { fwi_data = (char *)fdi.indata + FUSE_COMPAT_WRITE_IN_SIZE; } if ((err = uiomove(fwi_data, chunksize, uio))) break; retry: err = fdisp_wait_answ(&fdi); if (err == ERESTART || err == EINTR || err == EWOULDBLOCK) { /* * Rewind the uio so dofilewrite will know it's * incomplete */ uio->uio_resid += fwi->size; uio->uio_offset -= fwi->size; /* * Change ERESTART into EINTR because we can't rewind * uio->uio_iov. Basically, once uiomove(9) has been * called, it's impossible to restart a syscall. */ if (err == ERESTART) err = EINTR; break; } else if (err) { break; } fwo = ((struct fuse_write_out *)fdi.answ); /* Adjust the uio in the case of short writes */ diff = fwi->size - fwo->size; as_written_offset = uio->uio_offset - diff; if (as_written_offset - diff > filesize && fuse_data_cache_mode != FUSE_CACHE_UC) fuse_vnode_setsize(vp, as_written_offset); if (as_written_offset - diff >= filesize) fvdat->flag &= ~FN_SIZECHANGE; if (diff < 0) { printf("WARNING: misbehaving FUSE filesystem " "wrote more data than we provided it\n"); err = EINVAL; break; } else if (diff > 0) { /* Short write */ if (!direct_io) { printf("WARNING: misbehaving FUSE filesystem: " "short writes are only allowed with " "direct_io\n"); } if (ioflag & IO_DIRECT) { /* Return early */ uio->uio_resid += diff; uio->uio_offset -= diff; break; } else { /* Resend the unwritten portion of data */ fdi.iosize = sizeof(*fwi) + diff; /* Refresh fdi without clearing data buffer */ fdisp_refresh_vp(&fdi, FUSE_WRITE, vp, uio->uio_td, cred); fwi = fdi.indata; MPASS2(fwi == fdi.indata, "FUSE dispatcher " "reallocated despite no increase in " "size?"); void *src = (char*)fwi_data + fwo->size; memmove(fwi_data, src, diff); fwi->fh = fufh->fh_id; fwi->offset = as_written_offset; fwi->size = diff; fwi->write_flags = write_flags; goto retry; } } } fdisp_destroy(&fdi); return (err); } SDT_PROBE_DEFINE6(fusefs, , io, write_biobackend_start, "int64_t", "int", "int", "struct uio*", "int", "bool"); SDT_PROBE_DEFINE2(fusefs, , io, write_biobackend_append_race, "long", "int"); SDT_PROBE_DEFINE2(fusefs, , io, write_biobackend_issue, "int", "struct buf*"); static int fuse_write_biobackend(struct vnode *vp, struct uio *uio, struct ucred *cred, struct fuse_filehandle *fufh, int ioflag, pid_t pid) { struct fuse_vnode_data *fvdat = VTOFUD(vp); struct buf *bp; daddr_t lbn; off_t filesize; int bcount; int n, on, seqcount, err = 0; bool last_page; const int biosize = fuse_iosize(vp); seqcount = ioflag >> IO_SEQSHIFT; KASSERT(uio->uio_rw == UIO_WRITE, ("fuse_write_biobackend mode")); if (vp->v_type != VREG) return (EIO); if (uio->uio_offset < 0) return (EINVAL); if (uio->uio_resid == 0) return (0); err = fuse_vnode_size(vp, &filesize, cred, curthread); if (err) return err; if (ioflag & IO_APPEND) uio_setoffset(uio, filesize); if (vn_rlimit_fsize(vp, uio, uio->uio_td)) return (EFBIG); do { bool direct_append, extending; if (fuse_isdeadfs(vp)) { err = ENXIO; break; } lbn = uio->uio_offset / biosize; on = uio->uio_offset & (biosize - 1); n = MIN((unsigned)(biosize - on), uio->uio_resid); again: /* Get or create a buffer for the write */ direct_append = uio->uio_offset == filesize && n; if (uio->uio_offset + n < filesize) { extending = false; if ((off_t)(lbn + 1) * biosize < filesize) { /* Not the file's last block */ bcount = biosize; } else { /* The file's last block */ bcount = filesize - (off_t)lbn * biosize; } } else { extending = true; bcount = on + n; } if (howmany(((off_t)lbn * biosize + on + n - 1), PAGE_SIZE) >= howmany(filesize, PAGE_SIZE)) last_page = true; else last_page = false; if (direct_append) { /* * Take care to preserve the buffer's B_CACHE state so * as not to cause an unnecessary read. */ bp = getblk(vp, lbn, on, PCATCH, 0, 0); if (bp != NULL) { uint32_t save = bp->b_flags & B_CACHE; allocbuf(bp, bcount); bp->b_flags |= save; } } else { bp = getblk(vp, lbn, bcount, PCATCH, 0, 0); } if (!bp) { err = EINTR; break; } if (extending) { /* * Extend file _after_ locking buffer so we won't race * with other readers */ err = fuse_vnode_setsize(vp, uio->uio_offset + n); filesize = uio->uio_offset + n; fvdat->flag |= FN_SIZECHANGE; if (err) { brelse(bp); break; } } SDT_PROBE6(fusefs, , io, write_biobackend_start, lbn, on, n, uio, bcount, direct_append); /* * Issue a READ if B_CACHE is not set. In special-append * mode, B_CACHE is based on the buffer prior to the write * op and is typically set, avoiding the read. If a read * is required in special append mode, the server will * probably send us a short-read since we extended the file * on our end, resulting in b_resid == 0 and, thusly, * B_CACHE getting set. * * We can also avoid issuing the read if the write covers * the entire buffer. We have to make sure the buffer state * is reasonable in this case since we will not be initiating * I/O. See the comments in kern/vfs_bio.c's getblk() for * more information. * * B_CACHE may also be set due to the buffer being cached * normally. */ if (on == 0 && n == bcount) { bp->b_flags |= B_CACHE; bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; } if ((bp->b_flags & B_CACHE) == 0) { bp->b_iocmd = BIO_READ; vfs_busy_pages(bp, 0); fuse_io_strategy(vp, bp); if ((err = bp->b_error)) { brelse(bp); break; } } if (bp->b_wcred == NOCRED) bp->b_wcred = crhold(cred); /* * If dirtyend exceeds file size, chop it down. This should * not normally occur but there is an append race where it * might occur XXX, so we log it. * * If the chopping creates a reverse-indexed or degenerate * situation with dirtyoff/end, we 0 both of them. */ if (bp->b_dirtyend > bcount) { SDT_PROBE2(fusefs, , io, write_biobackend_append_race, (long)bp->b_blkno * biosize, bp->b_dirtyend - bcount); bp->b_dirtyend = bcount; } if (bp->b_dirtyoff >= bp->b_dirtyend) bp->b_dirtyoff = bp->b_dirtyend = 0; /* * If the new write will leave a contiguous dirty * area, just update the b_dirtyoff and b_dirtyend, * otherwise force a write rpc of the old dirty area. * * While it is possible to merge discontiguous writes due to * our having a B_CACHE buffer ( and thus valid read data * for the hole), we don't because it could lead to * significant cache coherency problems with multiple clients, * especially if locking is implemented later on. * * as an optimization we could theoretically maintain * a linked list of discontinuous areas, but we would still * have to commit them separately so there isn't much * advantage to it except perhaps a bit of asynchronization. */ if (bp->b_dirtyend > 0 && (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { /* * Yes, we mean it. Write out everything to "storage" * immediately, without hesitation. (Apart from other * reasons: the only way to know if a write is valid * if its actually written out.) */ SDT_PROBE2(fusefs, , io, write_biobackend_issue, 0, bp); bwrite(bp); if (bp->b_error == EINTR) { err = EINTR; break; } goto again; } err = uiomove((char *)bp->b_data + on, n, uio); if (err) { bp->b_ioflags |= BIO_ERROR; bp->b_error = err; brelse(bp); break; /* TODO: vfs_bio_clrbuf like ffs_write does? */ } /* * Only update dirtyoff/dirtyend if not a degenerate * condition. */ if (n) { if (bp->b_dirtyend > 0) { bp->b_dirtyoff = MIN(on, bp->b_dirtyoff); bp->b_dirtyend = MAX((on + n), bp->b_dirtyend); } else { bp->b_dirtyoff = on; bp->b_dirtyend = on + n; } vfs_bio_set_valid(bp, on, n); } vfs_bio_set_flags(bp, ioflag); if (last_page) { /* * When writing the last page of a file we must write * synchronously. If we didn't, then a subsequent * operation could extend the file, making the last * page of this buffer invalid because it would only be * partially cached. * * As an optimization, it would be allowable to only * write the last page synchronously. Or, it should be * possible to synchronously flush the last * already-written page whenever extending a file with * ftruncate or another write. */ SDT_PROBE2(fusefs, , io, write_biobackend_issue, 1, bp); err = bwrite(bp); } else if (ioflag & IO_SYNC) { SDT_PROBE2(fusefs, , io, write_biobackend_issue, 2, bp); err = bwrite(bp); } else if (vm_page_count_severe() || buf_dirty_count_severe() || (ioflag & IO_ASYNC)) { bp->b_flags |= B_CLUSTEROK; SDT_PROBE2(fusefs, , io, write_biobackend_issue, 3, bp); bawrite(bp); } else if (on == 0 && n == bcount) { if ((vp->v_mount->mnt_flag & MNT_NOCLUSTERW) == 0) { bp->b_flags |= B_CLUSTEROK; SDT_PROBE2(fusefs, , io, write_biobackend_issue, 4, bp); cluster_write(vp, bp, filesize, seqcount, 0); } else { SDT_PROBE2(fusefs, , io, write_biobackend_issue, 5, bp); bawrite(bp); } } else if (ioflag & IO_DIRECT) { bp->b_flags |= B_CLUSTEROK; SDT_PROBE2(fusefs, , io, write_biobackend_issue, 6, bp); bawrite(bp); } else { bp->b_flags &= ~B_CLUSTEROK; SDT_PROBE2(fusefs, , io, write_biobackend_issue, 7, bp); bdwrite(bp); } if (err) break; } while (uio->uio_resid > 0 && n > 0); return (err); } int fuse_io_strategy(struct vnode *vp, struct buf *bp) { struct fuse_filehandle *fufh; struct ucred *cred; struct uio *uiop; struct uio uio; struct iovec io; off_t filesize; int error = 0; int fflag; /* We don't know the true pid when we're dealing with the cache */ pid_t pid = 0; const int biosize = fuse_iosize(vp); MPASS(vp->v_type == VREG || vp->v_type == VDIR); MPASS(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE); fflag = bp->b_iocmd == BIO_READ ? FREAD : FWRITE; cred = bp->b_iocmd == BIO_READ ? bp->b_rcred : bp->b_wcred; error = fuse_filehandle_getrw(vp, fflag, &fufh, cred, pid); if (bp->b_iocmd == BIO_READ && error == EBADF) { /* * This may be a read-modify-write operation on a cached file * opened O_WRONLY. The FUSE protocol allows this. */ error = fuse_filehandle_get(vp, FWRITE, &fufh, cred, pid); } if (error) { printf("FUSE: strategy: filehandles are closed\n"); bp->b_ioflags |= BIO_ERROR; bp->b_error = error; bufdone(bp); return (error); } uiop = &uio; uiop->uio_iov = &io; uiop->uio_iovcnt = 1; uiop->uio_segflg = UIO_SYSSPACE; uiop->uio_td = curthread; /* * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We * do this here so we do not have to do it in all the code that * calls us. */ bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; KASSERT(!(bp->b_flags & B_DONE), ("fuse_io_strategy: bp %p already marked done", bp)); if (bp->b_iocmd == BIO_READ) { 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); if (!error && uiop->uio_resid) { /* * If we had a short read with no error, we must have * hit a file hole. We should zero-fill the remainder. * This can also occur if the server hits the file EOF. * * Holes used to be able to occur due to pending * writes, but that is not possible any longer. */ int nread = bp->b_bcount - uiop->uio_resid; int left = uiop->uio_resid; if (left > 0) bzero((char *)bp->b_data + nread, left); uiop->uio_resid = 0; } if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_error = error; } } else { /* * Setup for actual write */ error = fuse_vnode_size(vp, &filesize, cred, curthread); if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_error = error; bufdone(bp); return (error); } if ((off_t)bp->b_lblkno * biosize + bp->b_dirtyend > filesize) bp->b_dirtyend = filesize - (off_t)bp->b_lblkno * biosize; if (bp->b_dirtyend > bp->b_dirtyoff) { io.iov_len = uiop->uio_resid = bp->b_dirtyend - bp->b_dirtyoff; uiop->uio_offset = (off_t)bp->b_lblkno * biosize + bp->b_dirtyoff; io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; uiop->uio_rw = UIO_WRITE; error = fuse_write_directbackend(vp, uiop, cred, fufh, filesize, 0, false); if (error == EINTR || error == ETIMEDOUT) { bp->b_flags &= ~(B_INVAL | B_NOCACHE); if ((bp->b_flags & B_PAGING) == 0) { bdirty(bp); bp->b_flags &= ~B_DONE; } if ((error == EINTR || error == ETIMEDOUT) && (bp->b_flags & B_ASYNC) == 0) bp->b_flags |= B_EINTR; } else { if (error) { bp->b_ioflags |= BIO_ERROR; bp->b_flags |= B_INVAL; bp->b_error = error; } bp->b_dirtyoff = bp->b_dirtyend = 0; } } else { bp->b_resid = 0; bufdone(bp); return (0); } } bp->b_resid = uiop->uio_resid; bufdone(bp); return (error); } int fuse_io_flushbuf(struct vnode *vp, int waitfor, struct thread *td) { return (vn_fsync_buf(vp, waitfor)); } /* * Flush and invalidate all dirty buffers. If another process is already * doing the flush, just wait for completion. */ int fuse_io_invalbuf(struct vnode *vp, struct thread *td) { struct fuse_vnode_data *fvdat = VTOFUD(vp); int error = 0; if (vp->v_iflag & VI_DOOMED) return 0; ASSERT_VOP_ELOCKED(vp, "fuse_io_invalbuf"); while (fvdat->flag & FN_FLUSHINPROG) { struct proc *p = td->td_proc; if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) return EIO; fvdat->flag |= FN_FLUSHWANT; tsleep(&fvdat->flag, PRIBIO + 2, "fusevinv", 2 * hz); error = 0; if (p != NULL) { PROC_LOCK(p); if (SIGNOTEMPTY(p->p_siglist) || SIGNOTEMPTY(td->td_siglist)) error = EINTR; PROC_UNLOCK(p); } if (error == EINTR) return EINTR; } fvdat->flag |= FN_FLUSHINPROG; if (vp->v_bufobj.bo_object != NULL) { VM_OBJECT_WLOCK(vp->v_bufobj.bo_object); vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object); } error = vinvalbuf(vp, V_SAVE, PCATCH, 0); while (error) { if (error == ERESTART || error == EINTR) { fvdat->flag &= ~FN_FLUSHINPROG; if (fvdat->flag & FN_FLUSHWANT) { fvdat->flag &= ~FN_FLUSHWANT; wakeup(&fvdat->flag); } return EINTR; } error = vinvalbuf(vp, V_SAVE, PCATCH, 0); } fvdat->flag &= ~FN_FLUSHINPROG; if (fvdat->flag & FN_FLUSHWANT) { fvdat->flag &= ~FN_FLUSHWANT; wakeup(&fvdat->flag); } return (error); } diff --git a/tests/sys/fs/fusefs/write.cc b/tests/sys/fs/fusefs/write.cc index 407effab7db8..983c93d38eb0 100644 --- a/tests/sys/fs/fusefs/write.cc +++ b/tests/sys/fs/fusefs/write.cc @@ -1,957 +1,907 @@ /*- * 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); } }; 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); } }; /* 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); expect_write(ino, oldsize, 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); /* 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 */ } /* * 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 */ // TODO: check vfs.fusefs.mmap_enable TEST_F(Write, 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); } -/* In WriteThrough mode, a write should evict overlapping cached data */ -TEST_F(WriteThrough, evicts_read_cache) -{ - const char FULLPATH[] = "mountpoint/some_file.txt"; - const char RELPATH[] = "some_file.txt"; - ssize_t bufsize = 65536; - /* End the write in the middle of a page */ - ssize_t wrsize = bufsize - 1000; - char *contents0, *contents1, *readbuf, *expected; - uint64_t ino = 42; - int fd; - - contents0 = (char*)malloc(bufsize); - memset(contents0, 'X', bufsize); - contents0[bufsize - 1] = '\0'; // Null-terminate - contents1 = (char*)malloc(wrsize); - memset(contents1, 'Y', wrsize); - readbuf = (char*)calloc(bufsize, 1); - expected = (char*)malloc(bufsize); - memset(expected, 'Y', wrsize); - memset(expected + wrsize, 'X', bufsize - wrsize); - expected[bufsize - 1] = '\0'; // Null-terminate - - expect_lookup(RELPATH, ino, bufsize); - expect_open(ino, 0, 1); - expect_read(ino, 0, bufsize, bufsize, contents0); - expect_write(ino, 0, wrsize, wrsize, 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, lseek(fd, 0, SEEK_SET)) << strerror(errno); - ASSERT_EQ(wrsize, write(fd, contents1, wrsize)) << strerror(errno); - - // Read again. Cache should be bypassed - expect_read(ino, 0, bufsize, bufsize, expected); - ASSERT_EQ(0, lseek(fd, 0, SEEK_SET)) << strerror(errno); - ASSERT_EQ(bufsize, read(fd, readbuf, bufsize)) << strerror(errno); - ASSERT_STREQ(readbuf, expected); - - /* Deliberately leak fd. close(2) will be tested in release.cc */ -} - 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 = 4096; + 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); expect_write(ino, halfbufsize, 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(WriteBack, 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); 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 */ } /* * Without direct_io, writes should be committed to cache */ TEST_F(WriteBack, writeback) { 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 */ } /* * Without direct_io, writes should be committed to cache */ -/* - * Disabled because we don't yet implement write-through caching. No bugzilla - * entry, because that's a feature request, not a bug. - */ -TEST_F(WriteThrough, DISABLED_writethrough) +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 */ }