| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| In the Linux kernel, the following vulnerability has been resolved:
ocfs2: fix chunk number of the first chunk in a local quota file
Mounting a crafted OCFS2 image can corrupt kernel memory. When the
local quota file header of such an image claims zero chunks, the first
quota entry allocated during the mount gets a chunk number taken from a
kernel pointer, and releasing that entry clears a bit far outside the
chunk bitmap.
The write will corrupt the data stored at that address, and the
corruption may lead to a system crash or damage unrelated data.
Mounting requires CAP_SYS_ADMIN, so it takes an untrusted image, such
as removable media or a loop mount of a file from elsewhere.
Since the chunk number comes from a kernel pointer, the address of the
bit that gets cleared differs from boot to boot, and the issue has been
reported under several titles:
- BUG: unable to handle kernel paging request in ocfs2_local_release_dquot
- KASAN: use-after-free Write in ocfs2_local_release_dquot
- KASAN: slab-out-of-bounds Write in ocfs2_local_release_dquot
- KASAN: slab-use-after-free Write in ocfs2_local_release_dquot
- KFENCE: use-after-free write in ocfs2_local_release_dquot
Note that none of these is a use-after-free in the quota code: the
chunk and its buffer head are alive. The bit that gets cleared lies
far outside the bitmap. KASAN and KFENCE name each report based on the
object occupying that address, which explains why the same issue is
reported under so many different titles. In the report I sent, the
address fell in a free page, which KASAN labels use-after-free.
There are known reproducers. Both a syzkaller and a C reproducer are
available from the Google Drive link [1] in my report thread. The
issue was found by our modified syzkaller, not a theoretical thing
found by an LLM. I have not seen it outside fuzzing, so it is not
affecting me in the real world.
The local quota file in OCFS2 is divided into chunks, and each chunk
begins with a header block holding a bitmap of the quota entries that
chunk has handed out. Chunks are numbered from zero, and that number is
used to convert the file offset of an entry back into a bit position in
the bitmap. ocfs2_local_quota_add_chunk() appends a new chunk to the
in-memory list and numbers it one past the chunk that was last:
list_add_tail(&chunk->qc_chunk, &oinfo->dqi_chunk);
chunk->qc_num = list_entry(chunk->qc_chunk.prev,
struct ocfs2_quota_chunk,
qc_chunk)->qc_num + 1;
The predecessor is looked up after the new chunk is added to the list, so
if the list was empty, the prev pointer is the list head itself. The head
is the dqi_chunk member of struct ocfs2_mem_dqinfo and is not a chunk, so
reading qc_num through it lands 16 bytes past the start of the head, on
the dqi_gqinode pointer that follows it. The first chunk of the file is
then numbered with the lower half of a kernel pointer instead of 0.
The list is empty when the local quota file header claims the file has no
chunks. ocfs2_local_read_info() takes dqi_chunks from that header without
validating it, so an image with dqi_chunks == 0 takes this path when the
first quota entry is allocated.
ocfs2_create_local_dquot() turns the bad number into a file offset with
ol_dqblk_off(), which shifts a 32-bit block number left by the block size
bits, so the top bits of such a large block number are lost.
ocfs2_local_release_dquot() turns the offset back into a bit index with
ol_dqblk_chunk_off(), using the full chunk number, so the lost bits push
that index far outside the bitmap, and clearing it corrupts unrelated
memory.
Compute the chunk number before putting the chunk on the list, and use 0
when the list is empty. |
| In the Linux kernel, the following vulnerability has been resolved:
ocfs2: pin refcount trees before dropping the lookup lock
A refcount tree returned to ocfs2_lock_refcount_tree() must stay alive
until that caller releases its operation reference. However,
ocfs2_get_refcount_tree() returns the cached pointer after dropping
osb_lock, and the caller only takes its kref afterwards. The registry
reference does not protect this gap from concurrent removal.
This can happen when another node frees a refcount root and reuses its
block with a different generation while this mount retains the old
cached tree. Two local operations on different inodes sharing the new
root, such as data CoW and an external xattr update, can then look up
the same stale tree. Their inode locks are independent, and both the
REFCOUNT DLM lock and rf_sem are acquired after the first operation
reference.
The following ordering is possible, with no other operation references:
Data CoW Xattr update
-------- ------------
Look up the old tree under
osb_lock, then drop osb_lock
Look up the same tree and take a kref
Acquire the tree locks and read the
new on-disk generation
Erase the old tree under osb_lock
Unlock and drop the operation kref
Drop the registry kref, freeing tree
Take a kref through the
freed tree pointer
The final kref_get() accesses freed rf_getcnt storage, and subsequent
accesses in ocfs2_lock_refcount_tree() also use the freed allocation.
Take the operation reference in ocfs2_get_refcount_tree() before
releasing osb_lock, including when an insertion loses to another lookup.
Removal must acquire the same lock, so it cannot drop the registry
reference before the lookup has pinned the tree. Keep the existing
unlock and error path puts. The two callers that already hold the
refcount tree lock keep their borrowed lookup behavior and do not
acquire an extra reference.
KASAN report as below:
BUG: KASAN: slab-use-after-free in ocfs2_lock_refcount_tree+0x94a/0x9b0
Read of size 4 at addr ffff888117212820 by task python3/497
[Kernel and hardware identification omitted.]
Call Trace:
<TASK>
dump_stack_lvl+0x93/0xd0
print_report+0xce/0x630
? ocfs2_lock_refcount_tree+0x94a/0x9b0
? srso_alias_return_thunk+0x5/0xfbef5
? __virt_addr_valid+0x20e/0x420
? ocfs2_lock_refcount_tree+0x94a/0x9b0
kasan_report+0xe0/0x110
? ocfs2_lock_refcount_tree+0x94a/0x9b0
ocfs2_lock_refcount_tree+0x94a/0x9b0
? __pfx_ocfs2_lock_refcount_tree+0x10/0x10
? srso_alias_return_thunk+0x5/0xfbef5
? ocfs2_refcount_cow+0x3b6/0xd90
? srso_alias_return_thunk+0x5/0xfbef5
? __kmalloc_cache_noprof+0x28d/0x610
? ocfs2_refcount_cow+0x3b6/0xd90
ocfs2_refcount_cow+0x410/0xd90
? __pfx_ocfs2_refcount_cow+0x10/0x10
? __pfx___might_resched+0x10/0x10
? down_write+0x11a/0x1f0
? __pfx_down_write+0x10/0x10
? ocfs2_inode_unlock+0x1a9/0x260
? __pfx_ocfs2_inode_unlock+0x10/0x10
? ocfs2_file_write_iter+0x1413/0x1d30
ocfs2_file_write_iter+0x157b/0x1d30
? __pfx_ocfs2_file_write_iter+0x10/0x10
? lock_acquire+0x190/0x300
? __x64_sys_pwrite64+0x1b7/0x220
? __pfx___might_resched+0x10/0x10
? srso_alias_return_thunk+0x5/0xfbef5
? security_file_permission+0x26/0x80
? srso_alias_return_thunk+0x5/0xfbef5
? lock_is_held_type+0x8f/0x100
vfs_write+0x639/0x1010
? __pfx_ocfs2_file_write_iter+0x10/0x10
? lock_vma_under_rcu+0x154/0x4a0
? __pfx_vfs_write+0x10/0x10
? lock_release+0x121/0x290
? srso_alias_return_thunk+0x5/0xfbef5
__x64_sys_pwrite64+0x1b7/0x220
? __pfx___x64_sys_pwrite64+0x10/0x10
? irqentry_exit+0xbe/0x820
---truncated--- |
| In the Linux kernel, the following vulnerability has been resolved:
ovl: hold a reference on the parent in readdir
Listing a directory on an overlay uses a freed dentry and crashes the
kernel when a lookup moves the directory at the same time.
Code that can sleep may only use d_parent without a reference while
something stops a move of the dentry. ovl_iterate_real(),
ovl_need_adjust_d_ino() and ovl_cache_update() use the d_parent of the
listed directory and hold only the lock of that directory. When a
lookup finds the directory under another parent after a change to a
layer, d_splice_alias() moves it without that lock and drops its
reference on the old parent.
With a delay in ovl_iterate_real(), getdents64 on an overlay directory
raced a rename in the lower layer and a lookup that moved its dentry,
and after an rmdir of the old parent KASAN reported a use-after-free in
security_inode_getattr().
Fix this by using dget_parent() in the three functions.
[amir: fix comments in ovl_need_adjust_d_ino()] |
| In the Linux kernel, the following vulnerability has been resolved:
ovl: use the pinned upper dir in ovl_fix_origin()
Renames made directly on the upper filesystem during a lookup crash the
kernel.
A dentry that is passed to vfs_setxattr() must be pinned.
ovl_fix_origin() passes upper->d_parent to ovl_set_impure() without a
reference, and the lookup holds no lock of the upper directory. A
rename on the upper filesystem changes d_parent, and only the child
pins the new parent, so a second rename and an rmdir free it.
With delays in ovl_fix_origin() and ovl_set_impure(), renaming the upper
directory twice and removing the intermediate parent during the lookup
of a merge directory without an origin xattr gave a KASAN
use-after-free in vfs_setxattr().
Fix this by passing the upper dentry that the overlay parent holds.
[amir: add helper ovl_set_impure_dir()] |
| In the Linux kernel, the following vulnerability has been resolved:
rtnetlink: wait for RCU readers in rtnl_link_unregister()
rtnl_link_ops_get() finds ops on link_ops under rcu_read_lock() and only
then takes ops->srcu. rtnl_link_unregister() unlinks ops, then calls
synchronize_srcu() and cleanup_srcu_struct() without first waiting for
RCU readers. A reader that found ops before list_del_rcu() and has not yet
taken ops->srcu is not waited for.
Such a reader then locks a cleaned-up srcu_struct and keeps using ops
after the module that owns it is gone. rtnl_newlink() and
rtnl_dump_ifinfo() both look up ops without RTNL, the latter for any
user that dumps links filtered by kind.
With delays added between the lookup and srcu_read_lock() in
rtnl_link_ops_get(), and before rtnl_link_ops_put() in
rtnl_dump_ifinfo(), running "ip link show type dummy" alongside
"rmmod dummy" on a KASAN kernel gives:
BUG: KASAN: vmalloc-out-of-bounds in __srcu_read_unlock+0x57/0x70
Read of size 8 at addr ffffffffc02ff318 by task ip/147
Call Trace:
__srcu_read_unlock+0x57/0x70
rtnl_dump_ifinfo.cold+0x30/0x3b
rtnl_dumpit+0x7a/0x1b0
netlink_dump+0x486/0x11b0
__netlink_dump_start+0x661/0x910
rtnetlink_rcv_msg+0x611/0x9e0
netlink_rcv_skb+0x126/0x380
netlink_unicast+0x605/0x830
netlink_sendmsg+0x6d0/0xbb0
__sys_sendto+0x335/0x3a0
Call synchronize_rcu() before synchronize_srcu(), as
rtnl_af_unregister() already does. This should be enough because
rtnl_link_ops_get() is the only RCU walker of link_ops, and it takes
ops->srcu before rcu_read_unlock(), so every reader that found ops is
covered by synchronize_srcu() once the RCU grace period ends. The extra
grace period only costs time at module unload.
Readers and the writer were serialized by RTNL until link_ops got its
own mutex, which is why the missing RCU wait did not matter before. |
| In the Linux kernel, the following vulnerability has been resolved:
xfs: reject out-of-range attribute value lengths in xfs_attr_copy_value
xfs_attr_copy_value() takes the value length as a signed int and, for a
remote xattr, is handed args->rmtvaluelen. That field is filled from the
on-disk __be32 xfs_attr_leaf_name_remote.valuelen in
xfs_attr3_leaf_getvalue(), so a crafted length such as 0x80000000 is
stored into the signed rmtvaluelen as a negative number.
The "buffer too small" guard in xfs_attr_copy_value() is a signed
comparison:
if (args->valuelen < valuelen)
return -ERANGE;
A negative valuelen therefore compares as smaller than the caller's
buffer size, skips the -ERANGE path, and is then used as a copy length,
leading to an out-of-bounds copy of a full remote block into a small
getxattr(2) buffer on a mounted crafted image.
Reject a value length that is negative or larger than the maximum xattr
size before it is used, so a bogus on-disk length can no longer slip
through the value copier. |
| In the Linux kernel, the following vulnerability has been resolved:
xfs: reject remote xattr entries with an out-of-range value length
xfs_attr3_leaf_verify_entry() validates a remote attribute entry's name
but never bounds its on-disk value length (xfs_attr_leaf_name_remote.
valuelen, a __be32). A crafted leaf with valuelen = 0x80000000 passes
the verifier and the CRC.
That length is later assigned into the signed int args->rmtvaluelen
(xfs_attr3_leaf_getvalue), becoming negative, which slips past the signed
-ERANGE check in xfs_attr_copy_value(); a getxattr() with a small buffer
then memcpy()s a full remote block into the small kvalue buffer
(xfs_attr_rmtval_copyout) -- a heap out-of-bounds write with
attacker-controlled content, from an unprivileged getxattr(2) on a
mounted crafted image.
Reject remote entries whose value length exceeds XFS_XATTR_SIZE_MAX in
the leaf verifier, so the malicious block is rejected at read time. |
| In the Linux kernel, the following vulnerability has been resolved:
xfs: bound inode fork length against the fork size during log recovery
xlog_recover_inode_commit_pass2() copies an inode log item's data/attr
fork region into the inode buffer using the on-log region length without
bounding it against the fork capacity, e.g.:
len = item->ri_buf[2].iov_len;
memcpy(XFS_DFORK_DPTR(dip), src, len);
The only guard is an ASSERT, which is a no-op on production kernels
(CONFIG_XFS_DEBUG off), and xfs_dinode_verify() runs only after the copy.
A crafted image with a dirty log can therefore drive a heap out-of-bounds
write at mount time. The XFS_ILOG_DBROOT sibling already passes
XFS_DFORK_DSIZE as a bound; the DDATA/DEXT and ADATA/AEXT memcpy paths
did not.
Bound each logged fork region against the destination fork size before
copying it, and reject the log item with -EFSCORRUPTED when it does not
fit. Because the recovered inode is only verified after the fork data has
been copied in, the checks are done up front, before any memcpy into the
on-disk inode. |
| In the Linux kernel, the following vulnerability has been resolved:
smb: server: size directory name conversion buffer correctly
A 2048-byte ASCII directory name needs 4098 bytes when converted to
UTF-16, including the terminator. ksmbd_convert_dir_info_name() caps
its allocation at PATH_MAX and writes past the end of that buffer.
Allocate space for one UTF-16 code unit per input byte plus the
terminator. Use kmalloc_array() to check the size calculation. |
| In the Linux kernel, the following vulnerability has been resolved:
smb: client: don't resurrect an exiting tcon from cached dir workers
cfids_laundromat_worker() bumps tcon->tc_count for every expired open
cfid and queues cached_dir_offload_close() without checking whether the
tcon is already being torn down. Racing with the final cifs_put_tcon(),
the tcon comes back from zero and gets torn down twice:
T1 T2
cifs_put_tcon()
--tc_count == 0, TID_EXITING
tconInfoFree()
free_cached_dirs()
cancel_delayed_work_sync() cfids_laundromat_worker()
++tc_count == 1
queue_work(serverclose_wq)
kfree(tcon->origin_fullpath)
kfree(tcon)
cached_dir_offload_close()
close_cached_dir() [UAF]
cifs_put_tcon()
tconInfoFree() [double free]
cached_dir_lease_break() also bumps tc_count without tc_lock, racing with
the laundromat's locked update.
Take the tcon reference through a helper that refuses once the tcon is
TID_EXITING or its refcount has dropped to zero. Queued put_work and
close_work then always hold a tcon reference, so tconInfoFree() can't
run before they finish and free_cached_dirs() doesn't need to flush any
workqueue. If the laundromat can't get a reference, put the cfid back on
the dying list for free_cached_dirs() to free; its handle goes away with
the tree connection.
The lease break path finds the tcon on ses->tcon_list under
cifs_tcp_ses_lock, so it can't see TID_EXITING, but it now takes tc_lock. |
| In the Linux kernel, the following vulnerability has been resolved:
smb: client: fix buffer overflow in SMB2_tcon()
In SMB2_tcon(), unc_path is allocated with a fixed buffer size of
MAX_SHARENAME_LENGTH * 2 bytes (672 bytes). However, the user-controlled
UNC path (tree) length is passed to cifs_strtoUTF16(), which assumes the
destination buffer is large enough and writes strlen(tree) + 1 UTF-16 code
units.
Commit 62a1a439e0fd ("cifs: clean up handling of unc= option") removed the
length check on UNC mount options, allowing mount paths to exceed
MAX_SHARENAME_LENGTH while retaining the fixed-size buffer allocation in
SMB2_tcon(). Consequently, an oversized UNC string causes cifs_strtoUTF16()
to write past the end of the heap object, resulting in a slab-out-of-bounds
write. Furthermore, if a tree exceeds MAX_TREE_SIZE characters, strscpy()
truncates it in cifs_tcon::tree_name, causing subsequent reconnects or
failovers to fail.
Fix this by first verifying that the length of tree does not exceed
MAX_TREE_SIZE, returning -EINVAL if it does. Then, dynamically allocate
unc_path based on the length of tree so that it safely accommodates all
UTF-16 code units for paths within MAX_TREE_SIZE.
[pc: remove unused MAX_SHARENAME_LENGTH] |
| In the Linux kernel, the following vulnerability has been resolved:
smb: client: reclaim credits for decrypted reads with no matching mid
When a decrypted SMB2 READ response carries a MessageId that matches no
outstanding request, both the inline receive_encrypted_read() and the
offloaded smb2_decrypt_offload() paths drop the frame without accounting
the credits the server granted in it. The plaintext unknown-frame path in
cifs_demultiplex_thread() already reclaims them with
smb2_add_credits_from_hdr(), so encrypted reads leak credits that the
plaintext path would not.
This also happens for legitimate frames - e.g. a reply that arrives after
the client has timed out and dropped the mid - so over a long-lived
connection the client's credit count drifts down and can eventually starve
it of the ability to issue requests. The frame is AEAD-authenticated, so
the granted credits in the decrypted header come from the real server and
are safe to apply.
Reclaim the credits from the decrypted header before dropping the frame,
and export smb2_add_credits_from_hdr() so the SMB2 receive paths can use it. |
| In the Linux kernel, the following vulnerability has been resolved:
smb: client: validate NT transaction response layout and offsets
validate_ntransact() assumes that the response uses the fixed 18-word NT
transaction layout. A different WordCount makes get_bcc() locate the byte
count at a different offset from pSMBr->ByteCount, allowing the subsequent
bounds checks to use inconsistent response boundaries.
The function also turns the 32-bit ParameterOffset and DataOffset fields
into pointers before checking their bounds. On a 32-bit client, a large
offset can wrap the pointer below end_of_smb and pass the one-sided pointer
comparison. The caller then dereferences the parameter pointer or copies
from the data pointer.
Require the expected WordCount before reading the byte count. Check each
offset and count against the received SMB length as integers before forming
either pointer. Use subtraction after validating the offset so that the
range checks cannot overflow. |
| In the Linux kernel, the following vulnerability has been resolved:
smb: client: validate SMB1 EA list bounds before parsing
CIFSSMBQAllEAs() checks the server-provided EA list length by adding it
to the response pointer. On 32-bit systems, a positive length can wrap
the pointer and pass the comparison. The parser then walks outside the
received response while reading the next FEA header.
A response with a list length of 0x40000000 reproduces this as a native
i386 page fault and kernel panic from listxattr() against an SMB1
server.
Check that the EA list starts within the SMB byte area before reading
its header. Keep its length unsigned and compare it with the bytes left
in the response without pointer addition.
SMB1 is not negotiated by default and requires an explicit vers=1.0
mount. |
| In the Linux kernel, the following vulnerability has been resolved:
NFSD: copy SETXATTR data from all XDR buffer segments
nfsd4_vbuf_from_vector() manually copies the head followed by whole
pages starting at offset zero. This assumes that all remaining data
starts at the beginning of the page array and that no data is stored
in the tail.
However, nfsd4_decode_setxattr() passes an xdr_buf returned by
xdr_stream_subsegment(). Such a sub-buffer can have a nonzero
page_base and can span its page and tail segments. Ignoring that
layout can copy the wrong bytes into the value passed to
vfs_setxattr().
Use read_bytes_from_xdr_buf(), which observes the complete xdr_buf
layout. |
| In the Linux kernel, the following vulnerability has been resolved:
nfsd: don't modify a session slot when replaying its cached reply
nfsd4_sequence() claims a session slot by setting NFSD4_SLOT_INUSE
under nn->client_lock. A reply served from the slot's reply cache does
not claim it, and that distinction lived only in cstate->status, which
nfsd4_sequence() set to nfserr_replay_cache. Commit cc028a10a48c
("NFSD: Hoist status code encoding into XDR encoder functions") moved
nfsd4_proc_compound()'s cstate->status assignment below the out:
label, and the replay path's goto out was the one path that relied on
skipping it. The test in nfsd4_sequence_done() therefore no longer
identifies a replay, and every replay now stores its reply and clears
NFSD4_SLOT_INUSE as though it owned the slot.
NFSD4_SLOT_INUSE is the interlock: check_slot_seqid() rejects every
sequence id for a slot that is in use, before replay_matches_cache()
runs. A replay never sets it, so two replays can be in flight on the
same slot at once. One can read slot->sl_cred under nn->client_lock
while the other's completion frees and rebuilds it from the XDR
encoder without that lock. Two completions can also collide with each
other and release the same group_info twice. free_svc_cred() leaves
cr_uid, cr_gid and cr_flavor intact, so the reader passes every
earlier test in same_creds() and dereferences a NULL cr_group_info.
From a 6.12.91 production server:
BUG: kernel NULL pointer dereference, address: 0000000000000004
CPU: 63 UID: 0 PID: 39015 Comm: nfsd
RIP: 0010:same_creds+0x38/0xa0 [nfsd]
RDX: 0000000000000000
Call Trace:
nfsd4_sequence+0x6a8/0x910 [nfsd]
nfsd4_proc_compound+0x345/0x670 [nfsd]
nfsd_dispatch+0x100/0x220 [nfsd]
svc_process_common+0x311/0x700 [sunrpc]
svc_process+0x131/0x1c0 [sunrpc]
svc_recv+0x7ef/0x9c0 [sunrpc]
nfsd+0xa3/0x100 [nfsd]
Kernel panic - not syncing: Fatal exception
Since v6.14 a live session's slot table can shrink, and the unowned
store becomes a use-after-free write. nfsd4_sequence() defers the
shrink while a slot is in use, but it tests NFSD4_SLOT_INUSE, which a
replay does not set, so free_session_slots() can kfree() the slot the
replay still holds in cstate->slot.
Record the claim in cstate->slot_owned when nfsd4_sequence() accepts a
request and test that in nfsd4_sequence_done(), so only the request
that claimed the slot updates its cached reply and clears
NFSD4_SLOT_INUSE. svc_generic_init_request() zeroes the compound
response before each request, so the flag starts clear. |
| In the Linux kernel, the following vulnerability has been resolved:
SUNRPC: Do not credit control-record octets to the RPC stream
svc_tcp_sock_recv_cmsg() receives up to two octets into a local
buffer, and returns that count for any record type other than
TLS_RECORD_TYPE_ALERT. Nothing reached the caller's buffer, but
svc_tcp_read_marker() adds the count to sk_tcplen and
svc_tcp_read_msg()'s caller adds it to sk_datalen. The RPC stream
advances over octets it never received. The fragment marker is
assembled from stale sk_marker octets. The message body comes from
pages nothing wrote.
A conforming client reaches this. RFC 8446 Section 4.6.3 lets either
peer send KeyUpdate once it has sent its Finished, and svcsock has no
rekey path. kTLS leaves the partially consumed record on ctx->rx_list,
so the body drains two octets per svc_tcp_recvfrom() call. Each pair
is credited the same way.
Return -EAGAIN for a record that is not an alert. That is what
svc_tcp_sock_process_cmsg()'s default arm returned before the receive
moved into a local buffer. |
| In the Linux kernel, the following vulnerability has been resolved:
SUNRPC: fix oversized GSS proxy token copy
gss_read_proxy_verf() allocates one page per token chunk, but it copies the
entire linear head of an RPC request into in_token->pages[0]. A UDP datagram
can leave more than PAGE_SIZE bytes in the linear head, so a large
RPCSEC_GSS INIT token overwrites following pages.
Copy the linear head in page-sized chunks before switching to the request
page array. The linear head itself can span multiple pages. |
| In the Linux kernel, the following vulnerability has been resolved:
KVM: SVM: Trigger new ASID allocation in both VMCBs on pCPU switch
When the pCPU where the VMCB was most recently used is switched, reset
the ASID generation in both VMCBs, triggering new ASID allocation for
the immediate VMRUN as well as the next VMRUN on the other VMCB.
The ASID is shared between vmcb01 and vmcb02, and gets flushed on every
nested transition. However, since pCPU tracking is done per VMCB, it is
possible for one VMCB to allocate a new ASID when migrated to a new
pCPU, and then the other VMCB reuses that ASID on the old pCPU. This can
result in the same ASID being used by multiple vCPUs on the old pCPU.
Example scenario:
- vCPU runs on pCPU A, vmcb01 is active, asid=1.
- vCPU migrates to pCPU B, vmcb01 pCPU changes, new asid=2.
- Another vCPU runs on pCPU A and allocates asid=2 as well.
- vCPU migrates back to pCPU A, and then switches to vmcb02 before it
runs again with vmcb01.
- No pCPU switch is detected for vmcb02, so VMRUN is done with asid=2.
- Two vCPUs end up using asid=2 on the same pCPU.
Keep the VMCB dirtying to the active VMCB only. Clean bits are tracked
by a pCPU for each VMCB, so do not unnecessarily dirty a VMCB if its
pCPU does not change.
Additionally, initialize the tracked pCPU for vmcb02 to -1 on nested
enablement as hardening, so that a new ASID allocation is always
triggered when nested is disabled and re-enabled.
No performance regression was noticed when overcommitting L1 vCPUs in L0
(to force rescheduling), pinning L1 <-> L2 vCPUs, and running CPUID in a
tight loop bouncing between 2 vCPUs in L2.
An alternative (and perhaps more proper) fix would be tracking the ASID
per-VMCB instead (e.g. [1]). However, that's a more involved change, and
it would result in having different ASIDs for L1 and L2 without actually
properly maintaining them. It would probably work because all TLB
flushes target the current VMCB, and the other VMCB is always flushed on
nested transitions, but the code ends up in an arguably more fragile
state. Punt a proper clean fix to an incoming (and overdue) overhaul of
SVM's ASID usage [2].
[1] https://lore.kernel.org/lkml/20250205182402.2147495-2-yosry.ahmed@linux.dev
[2] https://lore.kernel.org/kvm/20260728003557.1136583-1-yosry@kernel.org |
| In the Linux kernel, the following vulnerability has been resolved:
KVM: nSVM: Reject KVM_SET_NESTED_STATE if L1 has EFER.LMA=1 && EFER.LME=0
Reject KVM_SET_NESTED_STATE if the incoming L1 host state has what is
effectively an impossible EFER combination of LMA=1 but LME=0, i.e. if the
state says long mode is active but not enabled. Unlike VMX, SVM doesn't
have an explicit consistent check for the illegal combination; presumably
hardware simply ignores EFER.LMA if EFER.LME=0.
Unfortunately, KVM doesn't ignore EFER.LMA in this case and consumes the
illegal state when constructing the shadow MMU for L2. E.g. if userspace
also clears CR4.PAE, then kvm_calc_cpu_role() will compute a role with 4 or
5 levels of paging, but shadow_mmu_init_context() will wire up the MMU to
use the paging32 template, which maxes out its levels at 2.
Note, the "real badness" is effectively the same as what happened with the
nVMX bug fixed by commit 112e66017bff ("KVM: nVMX: add missing consistency
checks for CR0 and CR4"). Unfortunately, the sanity check added by commit
72e2fb24a0b0 ("KVM: x86/mmu: Bug the VM if a vCPU ends up in long mode
without PAE enabled") doesn't work for this case, since L2 state is active
at the time of the page fault, but it's L1 that has the bad state. |