| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| In the Linux kernel, the following vulnerability has been resolved:
ALSA: pcm: Serialize PCM mmap with buffer reallocation to fix page UAF
snd_pcm_hw_params() and snd_pcm_hw_free() guard buffer reallocation
with an mmap_count check performed under the PCM stream lock, but the
lock is released long before the buffer is actually freed:
snd_pcm_sync_stop(), constraint refinement and do_free_pages() all
happen in between. snd_pcm_mmap_data(), on the other hand, takes no
lock at all: it validates against the old buffer's state and
dma_bytes, remaps its pages into the VMA, and only then increments
mmap_count.
A concurrent mmap() can therefore slip in between the check and the
free. remap_pfn_range() installs writable PTEs for the old buffer's
pages without taking page references, and the subsequent
do_free_pages() returns those pages to the page allocator while the
VMA still maps them. This leaves a stale, writable mapping of freed
pages: a page-level use-after-free that can be leveraged for local
privilege escalation.
Make snd_pcm_mmap_data() participate in the buffer-access scheme
introduced for hw_params/hw_free: acquire runtime->buffer_accessing
before validating and remapping, and release it afterwards. Buffer
reallocation already fails with -EBUSY while accessors are active,
and the mmap side now fails with -EBUSY while a reallocation is in
progress, so the validate/remap sequence and the check/free sequence
can no longer interleave.
A reproducer that turns this race into a stale writable mapping of
the freed DMA buffer pages is available on request. |
| In the Linux kernel, the following vulnerability has been resolved:
ksmbd: rate limit unmapped SID errors
A client can include many structurally valid but unmapped SIDs in a DACL.
Logging every mapping failure lets one request generate hundreds of kernel
error messages.
Rate limit the message to prevent an authenticated client from flooding
the kernel log. |
| In the Linux kernel, the following vulnerability has been resolved:
Bluetooth: btintel: validate version TLV value lengths
btintel_parse_version_tlv() verifies that a complete TLV is present in
the response, but it does not ensure that the value is long enough for
the specific TLV type. A short value can therefore cause an
out-of-bounds read through get_unaligned_le16(), get_unaligned_le32(),
or memcpy().
Reject values shorter than the minimum required by each known TLV type.
Also reject responses that do not contain the Command Complete Status
field. |
| In the Linux kernel, the following vulnerability has been resolved:
Bluetooth: hci_core: Fix race condition during device registration
In hci_register_dev(), the power_on work item is queued to
hdev->req_workqueue before initializing hdev->adv_monitors_idr and
registering the MSFT extension via msft_register(). For devices marked with
quirks such as HCI_QUIRK_RAW_DEVICE, the HCI_UNCONFIGURED flag is set on
the device. When the power_on work item runs concurrently on another CPU,
hci_power_on() detects that the device is unconfigured and immediately
invokes hci_dev_do_close(), which calls msft_do_close().
Concurrently, msft_register() allocates the msft structure and exposes it
to hdev->msft_data prior to calling mutex_init(&msft->filter_lock). If
msft_do_close() executes while hdev->msft_data is already assigned but the
mutex has not yet been initialized, mutex_lock(&msft->filter_lock) operates
on an uninitialized mutex, triggering a DEBUG_LOCKS warning:
DEBUG_LOCKS_WARN_ON(lock->magic != lock)
WARNING: kernel/locking/mutex.c:625 at __mutex_lock_common
kernel/locking/mutex.c:625 [inline]
WARNING: kernel/locking/mutex.c:625 at __mutex_lock+0x12d8/0x1550
kernel/locking/mutex.c:821
...
Call Trace:
<TASK>
msft_do_close+0x308/0x7b0 net/bluetooth/msft.c:693
hci_dev_close_sync+0x86b/0x10a0 net/bluetooth/hci_sync.c:5522
hci_dev_do_close net/bluetooth/hci_core.c:499 [inline]
hci_power_on+0x32c/0x750 net/bluetooth/hci_core.c:937
process_one_work kernel/workqueue.c:3322 [inline]
process_scheduled_works+0xa8e/0x14e0 kernel/workqueue.c:3405
worker_thread+0x92d/0xe10 kernel/workqueue.c:3486
kthread+0x388/0x470 kernel/kthread.c:436
ret_from_fork+0x514/0xb70 arch/x86/kernel/process.c:158
ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:245
</TASK>
Fix this by moving the queue_work() call in hci_register_dev() to after
idr_init(&hdev->adv_monitors_idr) and msft_register(hdev) so that device
structures and extensions are fully initialized before asynchronous tasks
can access them. Additionally, assign hdev->msft_data in msft_register()
only after mutex_init(&msft->filter_lock) has completed. |
| In the Linux kernel, the following vulnerability has been resolved:
Bluetooth: L2CAP: fix chan mode for LE_CONN_REQ + EXT_FLOWCTL pchan
l2cap_new_connection() sets default value of channel mode to match the
parent channel. l2cap_le_connect_req() left this at the default, and
created L2CAP_MODE_EXT_FLOWCTL channels if listening pchan has that
mode. This causes FLAG_DEFER_SETUP channels to reply to
L2CAP_LE_CONN_REQ with L2CAP_ECRED_CONN_RSP, which is incorrect.
It can also result to stack OOB write (of l2cap_alloc_cid determined
values) in l2cap_ecred_rsp_defer(), as l2cap_le_connect_req() does not
limit maximum number of deferred channels or check for duplicate ident.
Fix by setting chan->mode correctly in l2cap_le_connect_req().
Also check channel mode in l2cap_ecred_rsp_defer(), and do WARN_ON_ONCE
instead of OOB write to make it less brittle. |
| In the Linux kernel, the following vulnerability has been resolved:
net/sched: cls_u32: fix duplicate handle when node ID pool is exhausted
gen_new_kid() falls back to returning max (htid | 0xFFF) when both
idr_alloc_u32() ranges are full, instead of reporting an error.
u32_change() trusts that value and inserts a new knode with a handle
that is already live in the hash table, breaking handle uniqueness
within the table's node ID space.
The handle was never reserved in ht->handle_idr, so every later error
path that does idr_remove(&ht->handle_idr, handle) removes the
reservation of a different, live knode, which is then reused — one
failed add compounds into further duplicates.
The 4095 limit is per (table, bucket) — ht->handle_idr is per hash
table and the range is derived from htid (bucketid), so a table with
divisor 256 can legitimately hold 256*4095 knodes.
The sibling helper gen_new_htid() has the same silent in-band failure:
it returns 0 when the tp_c handle pool (1..0x7FF) is full, and
u32_init() publishes the root hash table with handle 0 without
checking. Two root tables with handle 0 alias in u32_lookup_ht(),
allowing cross-tcf_proto knode add/lookup/delete. Add the same
exhaustion check that the divisor path already has.
Return an error so u32_change() fails with ENOSPC/ENOMEM when the
node ID space is exhausted, and so u32_init() fails with -ENOMEM
when the hash table ID space is exhausted. The extack message
distinguishes pool exhaustion (-ENOSPC) from a transient allocation
failure (-ENOMEM).
Conditions to recreate the bug:
- CONFIG_NET_SCHED=y, CONFIG_CLS_U32=y (or =m with module loaded)
- Create a clsact qdisc on a device, then add 4095 u32 filters with
auto-generated handles to fill the node ID space for the root hash
table (single bucket). The 4096th auto-handle filter add triggers
the duplicate handle (fh 800::fff reused). Reachable at Level 2
(unshare -Urn, namespace-local CAP_NET_ADMIN).
- For gen_new_htid: create 2047 u32 proto entries on the same block
to fill the tp_c handle pool, then create one more. The root table
gets handle 0 and aliases with other handle-0 root tables. |
| In the Linux kernel, the following vulnerability has been resolved:
ipv6: mcast: fix RCU list diversion in ip6_mc_del1_src()
When removing a source filter whose count reaches zero, ip6_mc_del1_src()
unlinks psf from pmc->mca_sources. If the filter was previously active,
the code moved psf directly into pmc->mca_tomb by updating psf->sf_next.
Because pmc->mca_sources is traversed locklessly under RCU (e.g. by
ipv6_chk_mcast_addr()), mutating psf->sf_next before a grace period
elapses diverts concurrent readers to the tombstone list. Consequently,
readers miss remaining active sources in pmc->mca_sources and improperly
examine deleted tombstone entries.
Fix this by allocating a new tombstone node for pmc->mca_tomb (as done
in sf_setstate()) and retiring the original psf via kfree_rcu(). |
| In the Linux kernel, the following vulnerability has been resolved:
tipc: fix NULL deref in tipc_named_node_up() on empty publication list
User-space applications can bind a large number of service addresses to
one or more sockets. Each binding of a local-scope service address inserts
one entry (publication) into the TIPC name table. If the number of these
publications exceeds TIPC_MAX_PUBL (65535), protocol service types
(such as node state and link state) are no longer inserted into the name
table. This causes two issues:
1. User-space applications subscribing to node or link up/down events
stop receiving notifications.
2. A NULL pointer dereference can occur:
BUG: kernel NULL pointer dereference, address: 00000000000000d0
...
CPU: 0 UID: 0 PID: 0 Comm: swapper/0 Not tainted 7.2.0-rc4-default+ #5 PREEMPT(full)
...
RIP: 0010:tipc_named_node_up (./include/linux/skbuff.h:2251 net/tipc/name_distr.c:195 net/tipc/name_distr.c:221)
...
Call Trace:
<IRQ>
tipc_node_write_unlock (net/tipc/node.c:428)
tipc_rcv (net/tipc/node.c:934 net/tipc/node.c:2189)
tipc_udp_recv (net/tipc/udp_media.c:389)
Thread 1 (tipc_net_finalize) | Thread 2 (named_distribute)
-----------------------------|-----------------------------
| ...
| list_for_each_entry(publ, pls, binding_node) {
| ...
| __skb_queue_tail(list, skb);
| ...
| }
| ...
| hdr = buf_msg(skb_peek_tail(list));
... |
tipc_nametbl_publish(); |
If 'tipc_nametbl_publish()' (Thread 1) fails because the number of
local publications reaches TIPC_MAX_PUBL, list (Thread 2) will be empty. As a
result, NULL is passed to 'buf_msg()', leading to a NULL pointer dereference.
Fix these issues by allowing protocol service types (node state, link state,
and topology server) to be inserted into the name table unconditionally.
This ensures that users subscribing to these types always receive
notifications. In addition, the maximum number of local user publications is
reduced to (TIPC_MAX_PUBL - 1). This ensures that the maximum bulk size
calculated in tipc_link_set_queue_limits() remains valid. |
| In the Linux kernel, the following vulnerability has been resolved:
ipv6: sr: restore network header before routing and forwarding
ipv6_srh_rcv() runs with skb->data at the Segment Routing Header (SRH)
while skb_network_header() points at the IPv6 header.
When segments_left > 0, ipv6_srh_rcv() previously restored the skb->data
position by pushing sizeof(struct ipv6hdr), assuming the SRH immediately
followed the fixed IPv6 header. If another extension header (such as a
Hop-by-Hop options header) precedes the SRH, skb_network_offset()
remained negative.
This led to two problems:
1. During ip6_route_input(), fib6_rules_early_flow_dissect() invokes
__skb_flow_dissect() which passes the negative skb_network_offset()
to flow dissection, breaking BPF and C flow dissector logic.
2. If forwarded via ip6_forward() or redirected via act_mirred, downstream
handlers (like sch_fragment() or neighbour output) pass the negative
offset as an unsigned length, triggering OOB memcpy or buffer overflows.
Fix this by pushing -skb_network_offset(skb) before routing, ensuring
skb_network_offset(skb) is 0 for route lookup / flow dissection as well as
downstream forwarding. On the loopback path, pull skb_transport_offset(skb)
to restore skb->data to the SRH before looping back. |
| In the Linux kernel, the following vulnerability has been resolved:
bonding: alb: fix uninitialized transport header access in alb_determine_nd()
alb_determine_nd() uses icmp6_hdr(skb) to inspect ICMPv6 headers.
However, in xmit paths (e.g. packets sent via AF_PACKET / raw sockets
or forwarded packets), skb->transport_header is not guaranteed to be
initialized. While pskb_network_may_pull() ensures the packet data is
linear starting from the network header, it does not set or adjust the
transport header offset.
Dereferencing icmp6_hdr(skb) can therefore access out-of-bounds memory.
Fetch the icmp6hdr directly after ipv6hdr following pskb_network_may_pull(),
and reload ipv6hdr in case pskb_may_pull() reallocated skb->head.
Also remove the unused bond argument from alb_determine_nd(). |
| In the Linux kernel, the following vulnerability has been resolved:
scsi: mpt3sas: Avoid out-of-bounds cpumask_of_node() call in _base_assign_reply_queues()
dev_to_node() can return NUMA_NO_NODE (-1) on systems without NUMA
topology information for the PCI device, such as single-socket boards
that don't expose device-to-node affinity. Passing -1 directly into
cpumask_of_node() indexes node_to_cpumask_map[-1], an out-of-bounds
array read caught by UBSAN:
UBSAN: array-index-out-of-bounds in arch/x86/include/asm/topology.h:72:28
index -1 is out of range for type 'cpumask *[1024]'
Fall back to cpu_online_mask when no NUMA node is available, rather than
assuming dev_to_node() always returns a valid node index. |
| In the Linux kernel, the following vulnerability has been resolved:
ALSA: ump: do not touch legacy_rmidi before it exists
snd_ump_parse_endpoint() sets ump->parsed on every exit, including
error, before the caller attaches the legacy rawmidi device.
ump_handle_ep_name_msg() then treats parsed as "legacy_rmidi is live"
and calls ump_legacy_set_rawmidi_name(), which snprintf()s into
ump->legacy_rmidi->name. If a UMP packet arrives in that window
(IRQ path from snd_ump_receive), legacy_rmidi is still NULL
(KASAN null-ptr-deref in snprintf).
Guard the legacy helpers. parsed only means endpoint info was
parsed, not that legacy_rmidi exists. |
| In the Linux kernel, the following vulnerability has been resolved:
netfilter: nf_conntrack_sip: fix OOB read in sip_skip_whitespace()
sip_skip_whitespace() returns dptr unchanged when its own loop
exhausts the buffer (dptr == limit), instead of NULL like its sibling
sip_follow_continuation() returns on its own "no more data" path.
ct_sip_get_header() only checks for NULL after calling it:
dptr = sip_skip_whitespace(dptr, limit);
if (dptr == NULL)
break;
if (*dptr != ':' || ++dptr >= limit)
break;
so a recognized header name followed only by spaces/tabs running to
the exact end of the SIP payload, with no colon, makes the very next
statement read one byte past the buffer.
Make both "no more data" outcomes return NULL, matching the
convention sip_follow_continuation() already uses and that both
existing callers already check for. |
| In the Linux kernel, the following vulnerability has been resolved:
tracing/probes: Fix use-after-free on field name/type of events with multiple probes
The fields of a probe-based dynamic event (kprobe, uprobe, eprobe and
fprobe events) are created in traceprobe_define_arg_fields() by handing
the probe_arg name/type strings to trace_define_field(), which only
stores the pointers without copying. Those strings are owned by the
trace_probe and are freed when that probe is removed.
An event can have several probes attached. The field list is defined
only once, by the first probe that registers the event, but it is kept
alive by any surviving sibling probe. Deleting just that first probe by
symbol -
# primary A: fields are defined from A's args
echo 'p:kprobes/ev vfs_read a1=$arg1' > kprobe_events
# append B: shares A's event call
echo 'p:kprobes/ev vfs_write a1=$arg1' >> kprobe_events
# delete only A (matched by symbol), B survives
echo '-:kprobes/ev vfs_read' >> kprobe_events
frees A's args (trace_probe_cleanup() -> traceprobe_free_probe_arg()),
but trace_probe_unlink() keeps the trace_probe_event because the probe
list is not empty. The event call stays registered via B while its
fields now reference freed memory. Any field lookup then reads it, e.g.
echo 'a1 == 1' > events/kprobes/ev/filter
BUG: KASAN: slab-use-after-free in strcmp+0xa7/0xb0
Call Trace:
strcmp
trace_find_event_field
parse_pred
process_preds
create_filter
apply_event_filter
event_filter_write
field->name references parg->name (kstrdup'd, freed with the probe) and,
for array arguments, field->type references parg->fmt (kmalloc'd, freed
with the probe) - the scalar type otherwise points at the static
fmttype rodata, which is safe.
Have traceprobe_define_arg_fields() duplicate the name and type strings
and anchor the copies on the trace_probe_event, which embeds the event
call and outlives every individual probe; trace_probe_event_free()
releases them.
The reproducer above triggers reliably; the field lookup and the delete
both run under event_mutex, so this is a dangling reference after
removal rather than a race.
The issue was found by the autokbug dynamic kernel fuzzer at Tencent
Yunding Lab. |
| In the Linux kernel, the following vulnerability has been resolved:
bpf: reject BPF_PSEUDO_FUNC reference to the main program
fixups.c:jit_subprogs() rewrites BPF_PSEUDO_FUNC loads to contain real
function addresses. This function is invoked from bpf_jit_subprogs()
only when env->subprog_cnt > 1. Meaning that for any program like
below:
int main(void *ctx) {
void *ptr = main;
...
bpf_timer_set_callback(..., ptr);
...
}
The 'ptr' won't be ever converted to contain an address.
In combination with e.g. bpf_timer_set_callback() this would lead to a
function call at a bogus address.
Instead of complicating the implementation, just assume that no useful
program needs main to be a sync or async callback and reject
BPF_PSEUDO_FUNC loads for the main subprogram. |
| In the Linux kernel, the following vulnerability has been resolved:
bonding: do not clear curr_active_slave prematurely when releasing all slaves
When releasing all slaves during bond destruction (all == true),
__bond_release_one() unconditionally clears bond->curr_active_slave to
NULL in every iteration.
If a backup slave is released before the active slave,
bond_alb_deinit_slave() triggers rlb_teach_disabled_mac_on_primary(),
which increments the active slave dev promiscuity counter and sets
bond_info->primary_is_promisc = 1.
Because bond->curr_active_slave was prematurely cleared to NULL when
releasing the backup slave, the subsequent iteration releasing the active
slave evaluates oldcurrent as NULL, so bond_change_active_slave(bond, NULL)
is skipped. Consequently, bond_alb_handle_active_change() is never called
to decrement the promiscuity counter, permanently leaking promiscuous
mode on the physical device after bond teardown.
When oldcurrent == slave, bond_change_active_slave(bond, NULL) already sets
bond->curr_active_slave to NULL. We only need to avoid selecting a new
active slave when all == true. Replace the if (all) branch with
if (!all && oldcurrent == slave). |
| In the Linux kernel, the following vulnerability has been resolved:
net/rds: use wq_has_sleeper() in release_in_xmit()
release_in_xmit() clears RDS_IN_XMIT with clear_bit_unlock() and then
checks waitqueue_active() to decide whether anyone needs waking.
clear_bit_unlock() is only a release operation: it orders the
critical section before the bit clear, but does not order the
subsequent plain load of the wait queue head after it. The waiter
side does the mirror image - it adds itself to the wait queue and
then tests the bit. That is the classic store-buffering pattern: the
releasing CPU can read the wait queue as empty while the waiting CPU
still reads the bit as set, so the sleeper is never woken.
The waiters are rds_conn_shutdown() and rds_tcp_reset_callbacks(),
both in uninterruptible wait_event() with no timeout. A lost wake-up
strands the shutdown worker on its single-threaded workqueue until
some other sender releases the bit again - and on a connection that
is being torn down precisely because it failed, there may never be
another sender.
The barrier used to be there: release_in_xmit() did clear_bit()
followed by smp_mb__after_atomic() until commit 1422f28826d2 ("rds:
introduce acquire/release ordering in acquire/release_in_xmit()")
folded both into clear_bit_unlock(), which strengthened the lock
hand-off but silently dropped the full barrier the wake-up check
depends on. The refill counterpart, release_refill() in
net/rds/ib_recv.c, still carries its smp_mb__after_atomic() for
exactly this reason.
Use wq_has_sleeper(), which is waitqueue_active() preceded by the
required full barrier. |
| In the Linux kernel, the following vulnerability has been resolved:
net/rds: clear cp_flags bits individually in rds_conn_path_reset()
rds_conn_path_reset() wipes the whole flag word with a plain
cp->cp_flags = 0 store. Every other accessor of that word uses
atomic bitops, and some of them can run concurrently with the reset:
RDS_LL_SEND_FULL is set from rds_send_xmit() and cleared from the
transport completion paths, neither of which holds anything that
excludes the shutdown worker. A plain store racing an atomic
read-modify-write on the same word is a data race, and whichever
side loses has its update silently discarded.
Clear the two bits the reset is actually responsible for instead.
RDS_IN_XMIT and RDS_RECV_REFILL need no store at all here: they
belong to the caller, rds_conn_shutdown(), which waits for both to be
clear before calling the transport shutdown and this reset.
This also gives every bit in cp_flags a single well-defined writer
discipline, which the following patches rely on when they turn
RDS_IN_XMIT and RDS_RECV_REFILL into bit locks held across the
teardown: a blanket store mid-teardown would destroy lock ownership
that an atomic clear preserves.
Oracle UEK carries the same conversion ("net/rds: Preserve essential
connection state flags"), motivated by its asynchronous shutdown
state machine, whose progress and destroy flags must survive the
reset. UEK's variant also clears RDS_IN_XMIT and RDS_RECV_REFILL
because there the reset runs as the final step of a teardown that
owns both bits, making those clears its unlock. Upstream that
release belongs in rds_conn_shutdown(): once a later patch in this
series turns the two bits into locks held across the teardown, ending
ownership needs release semantics and a wake-up that a plain clear
inside the reset would not provide.
Based on Oracle UEK commit "net/rds: Preserve essential connection
state flags" by Gerd Rausch. |
| In the Linux kernel, the following vulnerability has been resolved:
net/rds: acquire RDS_IN_XMIT in rds_tcp_reset_callbacks()
rds_tcp_reset_callbacks() quiesces the transmit path by setting the
path state to RDS_CONN_RESETTING and then waiting for RDS_IN_XMIT to
be sampled clear before swapping the underlying socket and calling
rds_send_path_reset().
Sampling the bit clear is not the same as owning it: rds_send_xmit()
can re-acquire RDS_IN_XMIT right after the wait_event() returns. Its
state recheck after taking the lock is a store-buffering pattern (the
resetter writes the state and reads the bit, the sender writes the
bit and reads the state) and acquire_in_xmit() is only an acquire
operation, so on weakly ordered architectures both sides can miss
each other's write and the transmit path then runs concurrently with
rds_send_path_reset() rewriting cp_xmit_* state - which is exactly
what the comment above rds_send_path_reset() tells its callers to
prevent.
Take the lock instead, hold it across the socket swap and
rds_send_path_reset(), and release it with a wake-up at the end. The
lock-ordering constraint documented above the wait still holds: the
lock is acquired before lock_sock(), so a sender inside tcp_sendmsg()
can never be waited on while we hold the socket lock.
Two details of the old code go away with the same change:
- t_sock is now read only after the lock is acquired. The old code
cached it before waiting; the teardown in rds_conn_shutdown()
releases that socket and clears t_sock, so a pointer cached before
the wait can be stale by the time the accept path resumes. Reading
it under RDS_IN_XMIT is what makes the exclusion complete once the
teardown owns the same lock, which the next patch arranges; until
then the teardown still only samples the bit, and the two paths
remain as exposed to each other as they are today.
- The old !osock early path called rds_send_path_reset() with no
serialization at all. It now runs under the lock like the normal
path. The conditional RDS_CONN_RESETTING transition of the
previous patch happens before the socket check either way: a path
found without a socket is either still connecting (its reconnect
worker blocked on t_conn_path_lock) and legitimately goes
RESETTING -> UP on the new socket, or it has been torn down
meanwhile and is dropped.
The in-function comment describing the old wait-based quiesce is
rewritten to describe the lock-based one, and the stale block comment
above the function (which still described a return value and an
incomplete list of t_sock writers) is refreshed to name all four
writers - the connect, accept, teardown and swap paths - and what
serializes each of them. |
| In the Linux kernel, the following vulnerability has been resolved:
net/rds: acquire the fastpath locks in rds_conn_shutdown()
rds_conn_shutdown() quiesces the transmit and receive-refill paths by
waiting for RDS_IN_XMIT and RDS_RECV_REFILL to be sampled clear, and
then runs the transport shutdown and rds_conn_path_reset(). Sampling
the bits clear is not the same as owning them: the moment after the
wait_event() returns, rds_send_xmit() can re-acquire RDS_IN_XMIT (or
rds_ib_recv_refill() can re-acquire RDS_RECV_REFILL) and run
concurrently with the teardown.
The sender does recheck the connection state after taking the lock,
but that recheck is a classic store-buffering pattern: teardown writes
the state and reads the bit while the sender writes the bit and reads
the state. acquire_in_xmit() is only an acquire operation, so on
weakly ordered architectures both sides can miss each other's write,
and the transmit path then runs while the transport zeroes its rings
(e.g. rds_ib_ring_init()) and rds_send_path_reset() rewrites the
transmit state under it.
Oracle UEK fixed the same class of crashes - a 14-year tail of
BUG_ON()s in rds_ib_sub_signaled(), unexpected op-codes and NULL
dereferences in rds_ib_send_cqe_handler() during failover testing -
by making the teardown path *acquire* the fastpath bit locks instead
of testing them ("rds: Make sure transmit path and connection
tear-down does not run concurrently"). Ownership of a single word is
decided by RMW atomicity, so no cross-variable ordering is needed.
Do the same here: take both locks before calling the transport
shutdown, hold them across rds_conn_path_reset(), and release them
explicitly with a wake-up afterwards. Both are released with
clear_bit_unlock(), so that the ring re-initialization done by the
transport shutdown and the transmit state rewritten by
rds_send_path_reset() are ordered before either bit is seen clear by
the next acquire_in_xmit() or acquire_refill().
The fastpath users of these bits - rds_send_xmit() and
rds_ib_recv_refill() - are trylock style and back off while teardown
owns the locks, so no new lock dependency is introduced for them.
rds_tcp_reset_callbacks() is different: since the previous patch it
acquires RDS_IN_XMIT as well, and it blocks doing so, so its wait now
spans the teardown instead of at most one send batch. That waiter
runs from rds_tcp_accept_one() on the single-threaded krdsd workqueue
and holds rds_tcp_accept_lock and t_conn_path_lock while it waits, so
a duelling SYN accepted while its path is being torn down parks
accept processing for the duration of the teardown - for TCP bounded
by the (up to 5 s) drain loop in rds_tcp_conn_path_shutdown(). An IB
path's drain in rds_ib_conn_path_shutdown() has no round cap, but no
blocking waiter either: rds_tcp_reset_callbacks() is the only blocking
acquirer of these bits and waits only on its own TCP path, and the
fastpaths are trylock-and-back-off on both transports, so a long IB
drain lengthens only that path's own quiesce. The
window is narrow: the accept-side state check has to pass before the
teardown moves the path to RDS_CONN_DISCONNECTING.
Because krdsd is a single global workqueue, everything else queued
there - accept processing for other connections and network
namespaces, and the flush_workqueue(rds_wq) in rds_tcp_listen_stop()
during namespace teardown - waits behind the parked accept worker for
that time. It cannot deadlock, although the waits do point at each
other: the teardown blocks until the bit's holder releases it, and
the holder may be that krdsd accept worker. The holder finishes
without needing anything the teardown owns: the sync cancels
rds_tcp_reset_callbacks() issues target cp_send_w and cp_recv_w on
the path's ordered cp_wq, whose only execution slot is occupied by
the blocked cp_down_w itself, so they are pending at most and cancel
without flushing - a reliance on cp_wq being ordered that is now
noted next to those cancels (on
---truncated--- |
