Repository navigation
Open
Task
0 / 390 of 39 issues completed
Copy link
Labels
area: openvmmInitial title-keyword triage suggestionInitial title-keyword triage suggestion
Description
Objective. The microVM's virtio-fs shares serve every request that a guest can send without aborting the VMM or allocating what the guest declares, enforce the share's policy on every path to a host object, keep each node bound to the host object that it names across unlinks, renames, and hard links, and hold their state within the bounds that a capture and a restore accept. lxutil runs a caller-owned request with only the caller's identity and gives the thread back its own, and OpenVMM's fuse crate turns a file system's out-of-contract result into an error reply instead of an abort.
Children.
| Part | Issue | Scope | Blocked by |
|---|---|---|---|
| Transport | #1621 | Zero-length descriptors in a virtqueue chain end request reads and reply writes early, so valid requests fail. | — |
| Requests | #1596 | RequestReader::read_count allocates a guest-declared count, up to 4 GiB, before it checks the request's remaining bytes. |
— |
| Requests | #1607 | A FUSE_INIT whose header length is below 40 underflows the payload length, and a build with overflow checks panics. |
— |
| Requests | #1597 | A FUSE_SETATTR time with 10^9 or more nanoseconds panics in Duration::new at u64::MAX seconds, and otherwise carries into the seconds instead of failing with EINVAL. |
— |
| Replies | #1615 | A successful FUSE_IOCTL reply carries the request's input instead of the file system's output. |
— |
| Replies | #1616 | ReplySender::send_arg_data leaves the trailing data out of fuse_out_header.len. |
— |
| Replies | #1625 | ReplySender::send_error(i32::MIN) overflows its negation, and a build with overflow checks panics. |
— |
| Replies | #1628 | A reply longer than u32::MAX bytes panics in make_header. |
— |
| Replies | #1629 | A data reply longer than the guest requested panics in Session::send_max_size. |
— |
| Replies | #1631 | A write or copy_file_range count above u32::MAX panics. |
— |
| Session | #1630 | dispatch_init panics on init flags that the guest did not offer, and trusts capabilities that init can overwrite. |
— |
| Identity | #1604 | The inode dedup key omits the host device number, so objects with the same st_ino on different devices, such as btrfs subvolume roots, share one node. |
— |
| Identity | #1619 | Guest inode numbers omit the host device number, so such objects report the same (st_dev, st_ino) in the guest. |
— |
| Identity | #1626 | A dedup key that an object of another file type reuses is preflighted as an alias, so the node count passes MAX_INODES and a later capture fails. |
— |
| Identity | #1633 | A node's lookup count wraps from u64::MAX to 0, which a restored state can reach, so the next FUSE_FORGET evicts a node that the guest still holds. |
— |
| Aliases | #1599 | A file handle stays usable after FUSE_FORGET of its node, and with another node's ID, and its inode keeps frozen aliases that bypass the write policy. |
— |
| Aliases | #1608 | FUSE_LINK on an unlinked node links whatever file now occupies its stale path. |
— |
| Aliases | #1623 | Path-based operations on an unlinked or overwritten node act on the file that replaced it at its stale path. | — |
| Aliases | #1610 | Renaming a path onto itself drops every tracked alias at or below it, so a later capture fails. | — |
| Aliases | #1620 | Renaming a hard link onto another link of the same file drops the source's alias. | — |
| Policy | #1601 | Renaming a writable directory moves its denied descendants onto visible, writable paths. | — |
| Policy | #1602 | A non-exclusive FUSE_CREATE with O_TRUNC at a writable hard link truncates a file that another alias makes read-only. |
— |
| Policy | #1612 | FUSE_MKNOD and RENAME_WHITEOUT create host device nodes, 0:0 whiteouts included. |
— |
| Policy | #1603 | On Windows hosts, case-insensitive lookup and 8.3 short names bypass the share's denied, allowed, and writable paths. | — |
| Policy | #1609 | On Windows hosts, a handle opened with O_RDONLY | O_TRUNC accepts writes. |
— |
| Listing | #1611 | A microVM READDIRPLUS that fails partway leaks the lookup references of the entries that it had looked up. |
— |
| Listing | #1622 | Unlinking entries while the guest pages through a directory fails READDIRPLUS with ENOENT until the handle closes. |
— |
| Listing | #1605 | A directory that holds a btrfs subvolume root or snapshot placeholder, whose d_ino differs from its st_ino, fails READDIRPLUS with EIO. |
— |
| Listing | #1634 | A child entry that dedups onto the listed directory's own node rewrites the directory's path, and the listing continues inside the child. | — |
| Listing | #1613 | Directory listings are bounded per handle at run time but in aggregate only at save, so a guest can make a capture fail. | — |
| Aggregate | #1598 | READDIR on the synthetic root reserves the guest's buffer size, up to 4 GiB, infallibly. |
— |
| Aggregate | #1617 | The synthetic root answers a buffer that fits no entry with an empty reply, which means end of directory, instead of EINVAL. |
— |
| Aggregate | #1624 | FUSE_FSYNCDIR on the synthetic root's handle fails with EBADF. |
— |
| Aggregate | #1627 | HandleMap hands out u64::MAX, the synthetic root's handle, and a saved state that holds it fails to restore. |
— |
| Aggregate | #1632 | The synthetic root's st_nlink counts single-file children as subdirectories. |
— |
| Aggregate | #1614 | A single-file child whose host file name holds : or \ attaches, but every guest lookup of it fails with EINVAL. |
— |
| Aggregate | #1618 | A single-file child named by a bare relative path or a symlink passes validation but fails to attach. | — |
| Credentials | #1600 | lxutil::with_fs_identity keeps the thread's unmapped supplementary groups when the caller's GID is overflowgid. |
— |
| Credentials | #1606 | Where a user namespace permits setgroups, with_fs_identity restores the thread's unmapped groups as overflowgid: the VMM aborts if that GID is unmapped, and the thread keeps the wrong groups if it is mapped. |
— |
Order. No child blocks another.
Notes.
- The children were reported at OpenVMM
9d64daad, which NVX's pin,fe61f9cb, contains. The microVM mode, the aggregate, and the share policy are the microVM chain's (vm/devices/virtio/virtiofs/src/microvm/, absent from upstreammain). NVX's exports are aggregates, and on Linux hosts NVX passes--mount-owner callerwhen asked, under which openvmm: lxutil: with_fs_identity retains unmapped host supplementary groups when target GID equals overflowgid #1600 and openvmm: lxutil: with_fs_identity cannot safely restore unmapped supplementary groups when setgroups is permitted in a user namespace #1606 run. - No NVX microVM reaches openvmm: fuse: FUSE_IOCTL reply sends request input data instead of filesystem output #1615, openvmm: fuse: ReplySender::send_arg_data omits trailing data length from fuse_out_header.len #1616, openvmm: fuse: ReplySender::send_error panics on i32::MIN overflow in debug builds #1625, openvmm: fuse: reply exceeding u32 header length panics in make_header #1628, openvmm: fuse: data reply exceeding requested size panics in Session::send_max_size #1629, openvmm: fuse: dispatch_init asserts on unoffered Fuse::init flags and trusts mutable SessionInfo capabilities #1630, or openvmm: fuse: write and copy_file_range byte counts exceeding u32::MAX panic #1631:
VirtioFs,DormantMicrovmFs, andSectionFsnever return the results that trigger them, as the issues note. They harden thefusecrate's API for otherFuseimplementations, and stay here with the crate's reachable issues instead of in OpenVMM/microVM umbrella: OpenVMM issues outside the microVM and NVX implementation #1186. - openvmm: fuse: FUSE_IOCTL reply sends request input data instead of filesystem output #1615 and openvmm: fuse: ReplySender::send_arg_data omits trailing data length from fuse_out_header.len #1616 are both in the
FUSE_IOCTLreply, the only caller ofsend_arg_data; fix them together. openvmm: fuse: ReplySender::send_error panics on i32::MIN overflow in debug builds #1625, openvmm: fuse: reply exceeding u32 header length panics in make_header #1628, openvmm: fuse: data reply exceeding requested size panics in Session::send_max_size #1629, openvmm: fuse: dispatch_init asserts on unoffered Fuse::init flags and trusts mutable SessionInfo capabilities #1630, and openvmm: fuse: write and copy_file_range byte counts exceeding u32::MAX panic #1631 each turn an abort infuse::SessionorReplySenderinto an error reply, which one change can do for all of them. - openvmm: virtiofs: inode deduplication key omits host device number (st_dev) #1604 and openvmm: virtiofs: guest inode numbers omit host device ID and collide across btrfs subvolumes #1619 both add the host device number, openvmm: virtiofs: inode deduplication key omits host device number (st_dev) #1604 to the dedup key and openvmm: virtiofs: guest inode numbers omit host device ID and collide across btrfs subvolumes #1619 to the guest inode number; give them one mechanism. openvmm: virtiofs: microVM readdirplus fails with EIO on btrfs subvolumes and snapshot placeholders #1605 and openvmm: virtiofs: microVM readdirplus continues listing when child entry aliases open directory node #1634 also arise on btrfs subvolumes, and openvmm: virtiofs: microVM readdirplus continues listing when child entry aliases open directory node #1634 names openvmm: virtiofs: inode deduplication key omits host device number (st_dev) #1604's missing device number as one of its triggers.
- openvmm: virtiofs: FUSE_LINK on an unlinked node links whatever new file occupies its stale path #1608 and openvmm: virtiofs: path-based operations on an unlinked or overwritten node act on replacement files at its stale path #1623 have one cause: an unlinked node, whose aliases are empty, still resolves its stale path. openvmm: virtiofs: renaming a path onto itself drops all tracked aliases at or below that path #1610 and openvmm: virtiofs: renaming one hard link onto another hard link of the same file drops the source alias #1620 have another:
VirtioFs::renamedrops and moves aliases even when the host's rename does nothing. openvmm: virtiofs: open file handles remain usable after FUSE_FORGET and with mismatched node IDs #1599, openvmm: virtiofs: renaming a writable ancestor directory moves denied descendants into visible writable paths #1601, and openvmm: virtiofs: non-exclusive FUSE_CREATE on an existing file bypasses read-only alias checks and truncates protected hard links #1602 bypass the write policy through aliases. All of them changeInodeMap's alias tracking orVirtioFs::rename; land them one at a time. - openvmm: virtiofs: microVM readdirplus fails with EIO on btrfs subvolumes and snapshot placeholders #1605, openvmm: virtiofs: microVM readdirplus leaks lookup references on mid-listing errors #1611, openvmm: virtiofs: microVM readdirplus fails with ENOENT when directory entries are unlinked during pagination #1622, and openvmm: virtiofs: microVM readdirplus continues listing when child entry aliases open directory node #1634 all change
VirtioFsFile::read_dir_microvm; land them one at a time. Skipping an entry instead of failing (openvmm: virtiofs: microVM readdirplus fails with EIO on btrfs subvolumes and snapshot placeholders #1605, openvmm: virtiofs: microVM readdirplus fails with ENOENT when directory entries are unlinked during pagination #1622) removes most of the errors through which openvmm: virtiofs: microVM readdirplus leaks lookup references on mid-listing errors #1611 leaks, but not all of them. - openvmm: virtiofs: aggregate synthetic root readdir performs unbounded infallible Vec::with_capacity(size) allocation #1598 and openvmm: virtiofs: aggregate synthetic root returns empty EOF reply instead of EINVAL when buffer fits no entry #1617 both change
read_synthetic_root_dir_with, and openvmm: virtiofs: FUSE_FSYNCDIR on aggregate synthetic root handle fails with EBADF #1624 and openvmm: virtiofs: HandleMap::can_insert hands out u64::MAX, colliding with SYNTHETIC_ROOT_FH #1627 both concernSYNTHETIC_ROOT_FH. - A guest can make a capture fail through openvmm: virtiofs: renaming a path onto itself drops all tracked aliases at or below that path #1610, openvmm: virtiofs: microVM directory snapshot aggregate bounds are enforced only at save, allowing guests to break VM snapshot capture #1613, openvmm: virtiofs: renaming one hard link onto another hard link of the same file drops the source alias #1620, and openvmm: virtiofs: preflight_microvm_insert treats retyped dedup keys as alias additions, bypassing MAX_INODES #1626, and a restore fail through openvmm: virtiofs: HandleMap::can_insert hands out u64::MAX, colliding with SYNTHETIC_ROOT_FH #1627. openvmm: virtiofs: inode lookup count wraps to zero at u64::MAX #1633 needs a crafted saved state.
- openvmm: lxutil: with_fs_identity retains unmapped host supplementary groups when target GID equals overflowgid #1600 and openvmm: lxutil: with_fs_identity cannot safely restore unmapped supplementary groups when setgroups is permitted in a user namespace #1606 both change
SwitchedCredentialsinlxutil'sunix/credentials.rs, and both need a user namespace whosegid_mapleaves host groups unmapped; land them one at a time. openvmm: lxutil: with_fs_identity cannot safely restore unmapped supplementary groups when setgroups is permitted in a user namespace #1606 was found by inspection and has not been reproduced. openvmm: virtiofs: Windows case-insensitive lookup and NTFS 8.3 short names bypass microVM SubtreePolicy #1603 was found by inspection too, and it and openvmm: virtiofs: FUSE write succeeds on O_RDONLY | O_TRUNC handles on Windows hosts #1609 need a Windows host. - openvmm: virtiofs: open file handles remain usable after FUSE_FORGET and with mismatched node IDs #1599, openvmm: virtiofs: renaming a writable ancestor directory moves denied descendants into visible writable paths #1601, openvmm: virtiofs: non-exclusive FUSE_CREATE on an existing file bypasses read-only alias checks and truncates protected hard links #1602, openvmm: virtiofs: inode deduplication key omits host device number (st_dev) #1604, openvmm: fuse: FUSE_INIT with header len smaller than fuse_in_header underflows payload length #1607, openvmm: virtiofs: FUSE_LINK on an unlinked node links whatever new file occupies its stale path #1608, openvmm: virtiofs: FUSE write succeeds on O_RDONLY | O_TRUNC handles on Windows hosts #1609, openvmm: virtiofs: renaming a path onto itself drops all tracked aliases at or below that path #1610, openvmm: virtiofs: microVM readdirplus leaks lookup references on mid-listing errors #1611, openvmm: virtiofs: microVM FUSE mknod and rename allow creating host device nodes and 0:0 whiteouts #1612, openvmm: virtiofs: microVM aggregate file child with ':' or '\' in host filename attaches successfully but fails every guest lookup with EINVAL #1614, openvmm: virtiofs: preflight_microvm_insert treats retyped dedup keys as alias additions, bypassing MAX_INODES #1626, and openvmm: virtiofs: HandleMap::can_insert hands out u64::MAX, colliding with SYNTHETIC_ROOT_FH #1627 block their reporter's formal verification of the device.
Metadata
Metadata
Assignees
Labels
area: openvmmInitial title-keyword triage suggestionInitial title-keyword triage suggestion
Activity
12 remaining items