The internet is broken right now.
45 critical CVEs without a fix in the last 48 hours. Developers are crying quietly into their IDEs. 60 more already have a fix that someone still has to apply.
-
CVE-2026-99781
In the Linux kernel, the following vulnerability has been resolved: net/9p/usbg: set client to Disconnected on usb9pfs_disable This patch is setting the client status to Disconnected, when the client is still in use. Otherwise a disconnected usb cable would run any use of the mount to faults.
-
CVE-2026-99782
In the Linux kernel, the following vulnerability has been resolved: net/9p/usbg: also disable endpoints on p9_usbg_close The close function has to fully reverse the state change of 9p_create (mount) and the potential call of set_alt(1). This includes to ensure that the usage of the endpoints is not active any more.
-
CVE-2026-99786
In the Linux kernel, the following vulnerability has been resolved: xfrm: retry inexact policy lookup after node reinsertion An inexact policy lookup first records pointers to candidate hlist heads and then traverses the lists. A concurrent policy insertion can merge inexact tree nodes between those operations: lookup policy insertion ------ ---------------- find inexact candidates save obsolete hlist head write_seqcount_begin(&bin->count) merge inexact tree nodes hlist_del_rcu(&policy->bydst) reinsert policy->bydst in survivor write_seqcount_end(&bin->count) evaluate saved candidate list miss the moved policy The merge immediately reinserts the same hlist node into the surviving tree node. RCU keeps the policy alive, but it does not provide a consistent view while its list node is moved. A lookup that selected the obsolete list can observe it empty. A lookup already traversing a moved policy can instead follow the next pointer rewritten by the reinsertion. Either case can return an incorrect IPsec policy result. The per-bin sequence counter already brackets calls to xfrm_policy_inexact_insert_node(), including node merges. The read side, however, currently validates the counter only while searching an individual rb-tree. A successful search returns without validation, and the later candidate-list traversal is outside that read-side section. Snapshot the per-bin sequence before discovering candidate heads and validate it after evaluating all candidate lists. Retry the lookup if the inexact policy tree changes while it is being searched.
-
CVE-2026-99787
In the Linux kernel, the following vulnerability has been resolved: cpufreq: CPPC: Select the frequency-invariance callback per CPU cppc_cpufreq_cpu_fie_init() initializes deferred work only for CPUs whose feedback counters use PCC. However, if any CPU in a shared policy uses PCC, it installs the PCC tick callback for the entire policy. With FIE enabled, a non-PCC CPU then queues an uninitialized irq_work and can call a NULL function pointer when that work runs. Choose the frequency-invariance callback separately for each CPU using the recorded work-initialization state. Processor removal can unpublish the CPC descriptor before callback registration, so querying its counter transport again can incorrectly select direct reads for a PCC CPU. Register the callbacks after all counter initialization has completed, so a failed initial read for an online CPU cannot leave a partially registered policy. PCC CPUs use their initialized deferred work, while non-PCC CPUs read their own counters directly from the tick.
-
CVE-2026-99773
In the Linux kernel, the following vulnerability has been resolved: drm/drm_crtc: ensure dma_fence_ops remain valid during device unbind In [1], sashiko reported the following issue: === snip === Looking at how these fences are managed, drm_crtc_create_fence() creates a dma_fence without taking a reference to the drm_device or drm_crtc. Because the sync_file framework exposes this fence to userspace, the fence can outlive the CRTC. The dma_fence contract requires that data accessed by dma_fence_ops (like get_driver_name) must remain valid for an RCU grace period after the fence is signaled. However, drm_crtc_cleanup() and the subsequent freeing of the device do not wait for an RCU grace period via synchronize_rcu(). If userspace calls ioctl(SYNC_IOC_FILE_INFO) concurrently with a device hot-unplug: CPU1 (Userspace) sync_file_get_name() ops = rcu_dereference(fence->ops); if (!dma_fence_test_signaled_flag()) // Preempted or delayed here CPU2 (Driver Teardown) Signals the fence (setting fence->ops = NULL) Destroys and frees the CRTC without waiting for an RCU grace period CPU1 (Resumes) ops->get_driver_name(fence) -> drm_crtc_fence_get_driver_name() crtc = fence_to_crtc(fence); // Casts to the freed CRTC return crtc->dev->driver->name; // Use-after-free ... Does the CRTC or DRM device need to be kept alive for the RCU grace period, or should the fence hold a proper reference to prevent the use-after-free when get_driver_name() and get_timeline_name() access the freed CRTC structure? === snap === I believe this to be a correct observation and this patch implements the suggestion of waiting for an RCU grace period before proceeding with destruction of the drm_crtc, so that get_driver_name() and get_timeline_name() can still work.
-
CVE-2026-99774
In the Linux kernel, the following vulnerability has been resolved: xfrm: espintcp: reorder operations within espintcp_init_sk When enabled on a socket, espintcp sets the socket callbacks to its own before it has finished setting up its context and published it as icsk_ulp_data. Any one of those callbacks that gets called before will dereference a NULL icsk_ulp_data. Fix this by reording the operations, and add memory barriers to ensure callers will have a struct espintcp_ctx available. Add a few WRITE_ONCE/READ_ONCE too while we're there.
-
CVE-2026-99775
In the Linux kernel, the following vulnerability has been resolved: xfrm: interface: validate the IP header on xmit Packets injected into an xfrm interface (e.g. AF_PACKET on an xfrmi device) reach the xfrm output path without IP header validation, but xfrm4_transport_output() uses iph->ihl for __skb_pull() and memmove() and expects the header to be valid. vti_tunnel_xmit() already calls pskb_inet_may_pull() before handing packets to the xfrm output path. Do the same in xfrmi_xmit() and perform the header checks ip_rcv_core()/ip6_rcv_core() do on entry to the IP stack: reject wrong IP versions and malformed IPv4 header lengths.
-
CVE-2026-99777
In the Linux kernel, the following vulnerability has been resolved: net/9p/usbg: remove bogus context initialization in alloc_requests alloc_requests() initializes usb9pfs->in_req->context to point at the struct f_usb9pfs instance itself. usb9pfs_queue_tx() later overwrites this with the real struct p9_req_t pointer before every transmit, and usb9pfs_tx_complete() clears it back to NULL after each completion, so in_req->context only ever holds a valid p9_req_t once a request has actually been queued. usb9pfs_clear_tx(), however, can run at any time (mount close, gadget disable) independent of whether a request was ever queued. If it runs before the first usb9pfs_transmit(), it reads the leftover sentinel value, type-confuses the struct f_usb9pfs pointer as a struct p9_req_t, and both writes through it (req->t_err = -ECONNRESET) and hands it to p9_client_cb(), which manipulates req->wq and req->refcount at bogus offsets inside f_usb9pfs. This reliably corrupts memory or panics whenever the transport is torn down before any 9p request has been transmitted, e.g. mounting and immediately unmounting, or a cable disconnect racing the very first request. usb9pfs->out_req->context is set the same way but is never read by this transport (usb9pfs_rx_complete() identifies the instance via ep->driver_data instead), so it serves no purpose either. Both endpoints' ->driver_data are already set to usb9pfs in enable_endpoint(), which is what the completion handlers actually use to recover the f_usb9pfs instance. Just drop the leftover ->context assignments; usb_ep_alloc_request()/alloc_ep_req() already return zeroed requests, so in_req->context correctly starts out NULL.
-
CVE-2026-99779
In the Linux kernel, the following vulnerability has been resolved: net/9p/usbg: call disable_usb9pfs() from usb9pfs_disable() usb9pfs_disable() is the usb_function ->disable callback, invoked by the composite gadget core whenever the host resets or unconfigures the gadget. The gadget function API requires this callback to disable its endpoints; every other in-tree function driver does so from its ->disable hook. usb9pfs_disable() never called disable_usb9pfs(), so the IN/OUT endpoints and their usb_request objects were left active and allocated across a host-driven disable. When the host later reconfigures the device, enable_usb9pfs() calls alloc_requests() again and unconditionally overwrites usb9pfs->in_req/out_req, permanently leaking the previous allocations, while the endpoints themselves are left enabled underneath the function's own idea of being disabled. Call disable_usb9pfs() before reinit_completion(&usb9pfs->send), mirroring the same sequence already used in p9_usbg_close(). Placing it before the reinit_completion() also ensures usb_ep_disable() has synchronously flushed any requests still queued in hardware before the completion is reset, closing the same race that motivated resetting the completion here in the first place.
-
CVE-2026-99766
In the Linux kernel, the following vulnerability has been resolved: scsi: aic79xx: Fix NULL pointer dereference in device reset Since commit c67e63800446 ("scsi: aic79xx: Do not reference SCSI command when resetting device") the device reset SCB has no SCSI command, but ahd_linux_dev_reset() still sets its residual, and ahd_abort_scbs(), which completes it, sets its status. Seen on an AIC-7902 when a failing disk triggered a device reset, which killed the SCSI error handler: BUG: kernel NULL pointer dereference, address: 0000009c EIP: ahd_linux_dev_reset.cold+0xee/0x26e Add NULL command checks to the accessors in aic79xx_osm.h. Also skip autosense for an SCB without a command and complete it with ahd_done() right away: the REQUEST SENSE path can clear SCB_DEVICE_RESET, after which ahd_done() would dereference the missing command.
-
CVE-2026-99768
In the Linux kernel, the following vulnerability has been resolved: wifi: mac80211: reinstate IEEE80211_TX_CTL_AMPDU flag in ieee80211_tx_prep_agg() The commit 592234e941f1 ("wifi: mac80211: Fix iTXQ AMPDU fragmentation handling") changed the way handling of the IEEE80211_TX_CTL_AMPDU flag is done in mac80211, so the only place it is set is in ieee80211_tx_dequeue(). This means that any frames that are sent to the driver via drv_tx() will no longer have the flag set. In particular, this affects packets going through ieee80211_tx_pending(). Evan reported that this can result in the ath9k driver failing to account such packets in its block ack window, eventually overrunning the window and stalling TX. Fix this by reinstating the flag in ieee80211_tx_prep_agg() when inside a valid block ack window. This should be harmless for the TXQ path, and will reinstate the previous behaviour for drivers that use the flag as a signal to track the packet in the window.
-
CVE-2026-99769
In the Linux kernel, the following vulnerability has been resolved: drm/vc4: Drop the BO reservations outside of job_lock vc4_queue_submit() calls drm_exec_fini() with job_lock held and interrupts disabled. Besides unlocking the reservations, drm_exec_fini() frees the object array with kvfree(), which asks for preemptible task context or a non-NMI interrupt, and may reach might_sleep() through vfree(). Considering that an array that exceeds a page can take the vmalloc fallback, call drm_exec_fini() once job_lock is dropped. The reservations are then held across the job kick, which is harmless as nothing in the completion path takes a BO reservation. Note that vc4_attach_fences() has to stay inside the critical section, as it walks exec, which the job done worker is free to release as soon as job_lock is dropped.
-
CVE-2026-99770
In the Linux kernel, the following vulnerability has been resolved: drm/vc4: Publish the output fence outside of job_lock job_lock is taken by the V3D interrupt handler, so nesting an IRQ-unsafe lock under it creates an inversion. drm_syncobj_replace_fence() takes drm_syncobj::lock, which other paths acquire with interrupts enabled. For example, a thread interrupted while holding it blocks the handler on job_lock, and the handler can only make progress once the submit path releases job_lock, which it cannot do while waiting for drm_syncobj::lock. This is reported by lockdep: [ 188.139076] ======================================================== [ 188.139086] WARNING: possible irq lock inversion dependency detected [ 188.139096] 7.3.0-rc3-00067-gc4f3feeaf60f #67 Tainted: G C [ 188.139108] -------------------------------------------------------- [ 188.139116] swapper/0/0 just changed the state of lock: [ 188.139125] ffff0000054f8ff8 (&vc4->job_lock){-...}-{3:3}, at: vc4_irq+0xa8/0x28c [vc4] [ 188.139234] but this lock took another, HARDIRQ-unsafe lock in the past: [ 188.139242] (&syncobj->lock){+.+.}-{3:3} [ 188.139256] and interrupts could create inverse lock ordering between them. [ 188.139267] other info that might help us debug this: [ 188.139275] Possible interrupt unsafe locking scenario: [ 188.139284] CPU0 CPU1 [ 188.139291] ---- ---- [ 188.139299] lock(&syncobj->lock); [ 188.139312] local_irq_disable(); [ 188.139320] lock(&vc4->job_lock); [ 188.139335] lock(&syncobj->lock); [ 188.139350] <Interrupt> [ 188.139358] lock(&vc4->job_lock); [ 188.139372] *** DEADLOCK *** Hand the fence over after job_lock is dropped. The job can complete and drop exec->fence in that window, so take a reference first and publish through the local pointer, as exec may already be gone.
-
CVE-2026-99772
In the Linux kernel, the following vulnerability has been resolved: drm/drm_crtc: fix race with dma_fence_signal() in ::get_driver_name() Since commit 541c8f2468b9 ("dma-buf: detach fence ops on signal v3"), I'm seeing the BUG_ON() triggering in drm_crtc's fence_to_crtc() via drm_crtc_fence_get_driver_name() regularly: Call trace: panic+0x58/0x5c die+0x160/0x178 bug_brk_handler+0x70/0xa4 call_el1_break_hook+0x3c/0x1a0 do_el1_brk64+0x24/0x74 el1_brk64+0x34/0x54 el1h_64_sync_handler+0x80/0xfc el1h_64_sync+0x84/0x88 drm_crtc_fence_get_driver_name+0x60/0x68 (P) sync_file_get_name+0x184/0x45c sync_file_ioctl+0x404/0xf70 __arm64_sys_ioctl+0x124/0x1dc This looks to be caused by a code flow similar to the following: +++ snip +++ thread A thread B ioctl(SYNC_IOC_FILE_INFO) sync_file_ioctl() sync_file_get_name() dma_fence_signal_timestamp_locked() dma_fence_driver_name() ops = rcu_dereference(fence->ops) if (!dma_fence_test_signaled_flag()) ops->get_driver_name(fence) i.e. drm_crtc_fence_get_driver_name() test_and_set_bit(SIGNALED) RCU_INIT_POINTER(fence->ops, NULL) drm_crtc_fence_get_driver_name() BUG_ON(rcu_access_pointer(fence->ops) != &drm_crtc_fence_ops) +++ snap +++ In other words, a fence's ops pointer can now change to NULL on signal, which wasn't the case before the culprit commit and hence the BUG_ON() triggers. Simply drop the BUG_ON() as it's usage of fence ops is invalid now, and because using BUG_ON() and friends to take down the system is an unacceptable way to handle a failure. Note that the adjacent drm_crtc_fence_get_timeline_name() has the same problem and is fixed by this patch as well.
-
CVE-2026-99758
In the Linux kernel, the following vulnerability has been resolved: sctp: account for chunk padding in sctp_init_cause() sctp_init_cause() accepts a chunk when skb_tailroom() can hold the cause header and the payload. sctp_addto_chunk() then writes 0-3 bytes of padding before that header so the chunk stays 4-byte aligned. The check does not include the padding, and the following skb_put() calls skb_over_panic() from softirq. Count the padding sctp_addto_chunk() is about to write. The cause length on the wire stays the header plus the payload. Found by LLM inspection, triggered and tested the fix.
-
CVE-2026-99759
In the Linux kernel, the following vulnerability has been resolved: batman-adv: tt: don't drop local events queued during TVLV update batadv_tt_tvlv_container_update() read the number of local changes without holding changes_list_lock. batadv_tt_local_event() is not serialized by the commit_lock and can queue new events before the changes_list is flushed: CPU0 CPU1 batadv_tt_tvlv_container_update() num = ...local_changes; batadv_tt_local_event() list_add_tail(...); ...local_changes++; spin_lock_bh(&...changes_list_lock); /* only num events copied but all freed */ The additional events were freed without ever being sent. Other nodes will only notice this via a CRC mismatch and must request the full table. The number of changes and the changes_list must therefore be taken atomically. Events queued afterwards then stay in the tt.changes_list for the next TT round. The changes can no longer be kept for the next round when the TVLV buffer allocation fails. But they would be sent with a skipped TTVN anyway, so the other nodes have to request the full table in this case either way. The last changeset is only dropped in this case when the TTVN was incremented. Without new changes, it still describes the current TTVN. Previously, the unconsumed changes caused a retry of the TVLV container update in the next commit round. Since the changes are now consumed, the next round would only take the path without changes and wait for BATADV_TT_OGM_APPEND_MAX rounds before updating the container again. Until then, the outdated container with the previous TTVN would still be announced. The ogm_append_cnt is therefore reset on failure to retry the update in the next OGM interval. This doesn't make the CRC consistent with the announced changeset. The local table modifications are still not serialized with the commit and batadv_tt_local_transition_new() can still make a client part of the CRC for which the event is only part of the next TT round. Such an inconsistency is detected by the CRC check of the other nodes and solved via a full table request.
-
CVE-2026-99760
In the Linux kernel, the following vulnerability has been resolved: batman-adv: tt: ignore ROAM flag of received ADD changes A received TT change (OGM diff or full table response) with the ROAM flag but without DEL is handled by batadv_tt_global_add() like a ROAMING_ADV: * a newly created global entry gets BATADV_TT_CLIENT_ROAM and is therefore excluded from the global CRC. The sender has it in its local CRC and the receiver has to request the full table again * batadv_tt_global_purge_local() removes the local entry of the receiver. The receiver's own client is considered as roamed to the sender and frames for it are rerouted The ROAM flag only has a meaning together with DEL. Drop it on reception to keep it from populating the global table with a temporary ROAMING entry which is not part of the CRC - and therefore causing a CRC mismatch.
-
CVE-2026-99761
In the Linux kernel, the following vulnerability has been resolved: batman-adv: stop originator purge worker before freeing TT batadv_mesh_free() releases the translation tables via batadv_tt_free() before batadv_originator_free() stops orig_work. The purge worker can therefore still run while the TT is torn down. When an originator times out, batadv_purge_orig_ref() calls batadv_tt_global_del_orig(), which loads tt.global_hash and walks all buckets while taking their locks: CPU0 CPU1 batadv_tt_free() batadv_purge_orig() batadv_tt_global_del_orig() hash = ...tt.global_hash; batadv_tt_global_table_free() batadv_hash_destroy(hash); spin_lock_bh(&hash->list_locks[i]); Disable orig_work at the beginning of batadv_mesh_free(), directly after the mesh was marked as deactivating. The second disable_delayed_work_sync() in batadv_originator_free() is harmless because it is free'd anyway.
-
CVE-2026-99762
In the Linux kernel, the following vulnerability has been resolved: batman-adv: mcast: validate tracker padding before forwarding A multicast tracker with an even number of destinations needs two padding bytes after its destination list. batadv_mcast_forw_packet() currently only checks that the count field and the addresses fit in the tracker TVLV. If an unpadded even tracker is compacted to an odd number of destinations, batadv_mcast_forw_shrink_update_headers() includes those missing padding bytes in the amount to remove. The subsequent memmove can then write two bytes past the old packet tail. With no tailroom, this corrupts skb_shared_info. Include the required padding in the existing tracker length check. This rejects the malformed tracker before either received or locally generated packets can reach the compactor.
-
CVE-2026-99763
In the Linux kernel, the following vulnerability has been resolved: 9p: don't decrement a server-owned zero nlink A 9P2000.L server can report nlink=0 for a regular file that still exists. On a cache=loose or cache=fscache mount, lookup stores that value with set_nlink(). A later successful unlink calls v9fs_dec_count(), which still calls drop_nlink() for a non-directory. drop_nlink() warns, and the unsigned link count wraps to 4294967295. fstat() on a descriptor opened before the unlink shows 0 before and 4294967295 after. The warning on the unpatched kernel, trimmed: WARNING: fs/inode.c:408 at drop_nlink+0xac/0xd0, CPU#1: trigger/188 CPU: 1 UID: 0 PID: 188 Comm: trigger Tainted: G N 7.3.0-rc4-00537-ga3ff15db6820 #1 PREEMPT(lazy) RIP: 0010:drop_nlink+0xac/0xd0 Call Trace: <TASK> v9fs_remove+0x529/0x750 vfs_unlink+0x2f0/0xbd0 filename_unlinkat+0x340/0x6c0 __x64_sys_unlink+0x46/0x70 do_syscall_64+0xe8/0x4f0 entry_SYSCALL_64_after_hwframe+0x77/0x7f </TASK> ---[ end trace 0000000000000000 ]--- Directories whose cached nlink is at most 2 already skip this decrement, because the exported filesystem may not maintain nlink (commit ac89b2ef9b559). cache=none skips it as well (commit 574aa0b479947). cache=mmap does not take this path: it sets neither CACHE_META nor CACHE_LOOSE. A regular file on cache=loose or cache=fscache still does. Its count came from GETATTR and may already be zero, so skip the decrement when i_nlink is already 0. Verified on 7.3-rc4-00537-ga3ff15db6820 with only this change. A server that advertises nlink=0 and accepts the unlink no longer warns, and the open descriptor stays at nlink 0. An honest server on the same mount still goes from 1 to 0. A rejected unlink leaves the count unchanged. cache=none is unchanged. With panic_on_warn the unpatched kernel panics on the warning; the patched kernel does not.
-
CVE-2026-99764
In the Linux kernel, the following vulnerability has been resolved: 9p/xen: validate the response size from the backend p9_xen_response() takes the length of an incoming response from the ring -- the first field of the 9p header, which the Xen transport reuses as its framing header -- and uses it to advance the consumer index, but never checks that it is at least the header size or no larger than the number of bytes the backend has actually produced. A malicious or buggy backend can post a response whose size is smaller than the header (for example 0). The consumer index then never advances, so the response work re-reads the same ring contents instead of making progress. In testing this raced with the teardown that the malformed reply triggers and dereferenced a freed p9_client: 9pfs 9pfs-0: Wrong req tag=ffff BUG: kernel NULL pointer dereference, address: 0000000000000060 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: Oops: 0000 [#1] SMP NOPTI CPU: 0 UID: 0 PID: 11 Comm: kworker/0:1 Not tainted 7.2.0-rc5 #6 PREEMPT(lazy) Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.17.0-0-gb52ca86e094d-prebuilt.qemu.org 04/01/2014 Workqueue: events p9_xen_response RIP: 0010:idr_find+0x4/0x10 RSP: 0018:ffffbf5d80063de8 EFLAGS: 00010202 RAX: 0000000000000001 RBX: ffffa229c21195a0 RCX: ffffa229c2400000 RDX: ffffa229c11c2200 RSI: 000000000000ffff RDI: 0000000000000050 RBP: 000000000000ffff R08: 3fffffffffffdfff R09: ffffffffffffffff R10: 3fffffffffffdfff R11: ffffffff87a60e80 R12: 0000000000000050 R13: 0000000000000000 R14: 000000000000ffff R15: 0000000000100000 FS: 0000000000000000(0000) GS:ffffa22a7661c000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000000060 CR3: 000000003ac34002 CR4: 0000000000770ef0 PKRU: 55555554 Call Trace: <TASK> p9_tag_lookup+0x2b/0x90 p9_xen_response+0x17c/0x2e0 process_one_work+0x16a/0x3a0 worker_thread+0x172/0x2e0 kthread+0xdd/0x110 ret_from_fork+0x18b/0x240 ret_from_fork_asm+0x1a/0x30 </TASK> Modules linked in: CR2: 0000000000000060 ---[ end trace 0000000000000000 ]--- Reject a response whose size is outside [sizeof(header), queued] before using it, matching the validation the backend already applies to incoming requests.
-
CVE-2026-99751
In the Linux kernel, the following vulnerability has been resolved: bpf: Fix skb_pull() BUG in __bpf_redirect_neigh() The issue: This is a follow-up to the short link-layer header fixes from Eric. After vlan pop_eth, a privileged SCHED_ACT BPF program can call bpf_skb_change_head(skb, 1, 0) and then bpf_redirect_neigh(). __bpf_redirect_neigh() calls skb_pull(skb, sizeof(*ethh)) after checking only that mac_header < network_header. Nothing verifies that the MAC header is long enough, that it is linear, or that the IP header the redirect goes on to read is linear either. A SCHED_ACT BPF program can do: bpf_skb_change_head(skb, 1, 0); bpf_redirect_neigh(lo, NULL, 0, 0); after vlan pop_eth. The skb then has mac_len == 1 and only one byte of linear data; skb_pull(ETH_HLEN) trips the skb->len < skb->data_len BUG in __skb_pull(): kernel BUG at ./include/linux/skbuff.h:2839! RIP: skb_pull+0x40/0x50 Call Trace: skb_do_redirect -> packet_xmit -> packet_sendmsg Kernel panic - not syncing: Fatal exception in interrupt The v4 and v6 paths read ip_hdr() and ipv6_hdr() as soon as the MAC header is gone. On the AF_PACKET frame below that leaves headlen == 0 after the pull, so the network header sits at skb->tail and the route lookup reads ip4h->protocol, ip4h->saddr and ip4h->daddr past the linear area. The Fix: Use skb_network_offset(), which is the MAC header span, instead of skb->mac_len. mac_len is zero for locally generated tc egress and netkit packets, while the span still rejects the one-byte header produced by pop_eth + bpf_skb_change_head(skb, 1, 0). The single check also covers the mac_header >= network_header test it replaces. Use pskb_may_pull() to make the Ethernet header linear before it is read or pulled, and to make the IP header linear in both the v4 and the v6 path before ip_hdr()/ipv6_hdr() is read. The header pointers are taken after the pull, since pskb_may_pull() may reallocate skb->head. Conditions to recreate the bug: On x86-64 with 4 KiB pages, enable CONFIG_PACKET, CONFIG_BPF_SYSCALL, CONFIG_NET_SCH_INGRESS, CONFIG_NET_CLS_MATCHALL, CONFIG_NET_ACT_VLAN and CONFIG_NET_ACT_BPF. Run as root with BPF loading, network administration and raw-packet privileges. Bring lo up and attach a SCHED_ACT program that calls bpf_skb_change_head(skb, 1, 0), then returns bpf_redirect_neigh(skb->ifindex, NULL, 0, 0): ip link set lo up tc qdisc add dev lo clsact tc filter add dev lo egress matchall action vlan pop_eth pipe \ action bpf object-file neigh_chghead.o section action Send an untagged ETH_P_IP frame of ETH_HLEN + 8192 bytes through an AF_PACKET/SOCK_RAW socket bound to lo, without PACKET_VNET_HDR or PACKET_QDISC_BYPASS. Use a unicast destination MAC. The sender's skb has a 14-byte linear Ethernet prefix and a fragmented payload; pop_eth removes that prefix and change_head restores only one linear byte. The unfixed neighbor redirect then tries to pull 14 bytes and BUGs. Sending the same frame without pop_eth and change_head leaves the IP header in the frags, which is the route lookup reading past skb->tail.
-
CVE-2026-99756
In the Linux kernel, the following vulnerability has been resolved: btrfs: prevent deleting the wrong subvolume by ID When deleting a subvolume by ID, btrfs_ioctl_snap_destroy() first resolves the requested ID to a name, then looks up that name again under the parent directory lock. A concurrent rename can change which subvolume is associated with the name between these operations. Since the deletion path does not verify the ID of the resolved subvolume, it can delete a different subvolume while leaving the requested one intact. This was reproduced by repeatedly exchanging the names of two subvolumes using renameat2(RENAME_EXCHANGE) while deleting one of them by ID. The wrong subvolume was deleted in 9 out of 14 independent tests. Fix this by checking that the resolved subvolume ID matches the requested ID while holding the parent directory lock. Return -EAGAIN on a mismatch so that userspace can retry the operation instead of being told that the requested subvolume does not exist.
-
CVE-2026-99757
In the Linux kernel, the following vulnerability has been resolved: btrfs: sysfs: restore NOFS context on early return in btrfs_sysfs_add_block_group_type() btrfs_sysfs_add_block_group_type() sets up a NOFS context with memalloc_nofs_save() because kobject_add() performs GFP_KERNEL allocations deep in its call chain, and the function may be called while holding a transaction handle or a lock required for a transaction commit, where reclaim could deadlock. Every other return path of this function calls memalloc_nofs_restore() except the early return taken when another thread has already created the kobject for the same raid index (this can happen when restriping). Taking that path leaves PF_MEMALLOC_NOFS set on the calling task, so all subsequent memory allocations performed by that task remain restricted to the NOFS scope, which may lead to unexpected allocation failures or deadlocks later on. Restore the NOFS context before returning from that path.
-
CVE-2026-99743
In the Linux kernel, the following vulnerability has been resolved: ksmbd: grant a POSIX ACL group entry only to the group's members When no ACE of the DACL matches the caller, Everyone or Authenticated Users, smb_check_perm_dacl() falls back to the file's POSIX ACL. It takes the first ACL_USER or ACL_GROUP entry whose id equals the caller's uid. For ACL_GROUP, this compares a gid with the caller's uid. A numerical match grants the group's permissions even when the caller is not a member. Match a group entry only when the caller is also in the group.
-
CVE-2026-99744
In the Linux kernel, the following vulnerability has been resolved: net: atl1c: drop oversized RX frames instead of panicking atl1c sizes its receive buffers from the netdev MTU plus a fixed set of L2 overheads, which comes to 1522 bytes at the default MTU. A link partner that sends a longer frame - in practice a MikroTik CCR2004-1G-2XS-PCIe, whose RouterOS side defaults to l2mtu 1600 while the host stays at MTU 1500 - makes the length reported by the RX return descriptor exceed what the hardware wrote, and skb_put() then either includes bytes that were never received or trips skb_over_panic and takes the host down. Compare the descriptor length against buffer_info->length, the length the hardware was given, rather than against skb_tailroom(): the skb carries 64 bytes of slack for the DMA erratum, so tailroom is larger than the DMA buffer and a frame in between would pass the check and still be delivered with a tail the device never wrote. The same test rejects frames shorter than ETH_HLEN + ETH_FCS_LEN; before this change a reported length below ETH_FCS_LEN underflowed in skb_put(). The frame is dropped with kfree_skb_reason(SKB_DROP_REASON_PKT_TOO_BIG) so drop_monitor and the kfree_skb tracepoint see it, and counted through dev_core_stats_rx_dropped_inc(), which survives atl1c_get_stats() overwriting the hardware counters on every read. It also counts as work: atl1c_clean_rfd() has already recycled the RFD, and the refill at the end of the poll only runs when work_done is non-zero, so sustained oversized traffic would otherwise drain the ring and stop RX. Tested on a MikroTik CCR2004-1G-2XS-PCIe (4x AR8151) in a Proxmox VE host, with atl1c built from a DKMS tree carrying this change: the l2mtu-1600 traffic that panics the stock driver is dropped and counted with the patch applied, and received normally once the host MTU is raised to match. This complements commit 36c2009d90f2 ("net: atl1c: fix soft lockup on out-of-range tpd_cons read"), which fixed the TX side of the same "host dies while the card resets" problem.
-
CVE-2026-99745
In the Linux kernel, the following vulnerability has been resolved: x86/bugs: Select the BHI mitigation based on the attack vector properly The BHI mitigation selection is inconsistent with the documentation (attack_vector_controls.rst), which lists BHI under both the user->kernel and guest->host attack vectors. bhi_select_mitigation() only enables it when CPU_MITIGATE_GUEST_HOST is set; otherwise it selects BHI_MITIGATION_OFF, even if CPU_MITIGATE_USER_KERNEL is set. As a result, a kernel compiled with KVM disabled gets BHI automatically disabled, since CPU_MITIGATE_GUEST_HOST defaults to off without KVM. This is a bug, given that CPU_MITIGATE_USER_KERNEL still needs it: on affected Intel CPUs these kernels lose BHI_DIS_S or the BHB clearing loop by default. Booting with mitigations=auto,no_guest_host has the same effect. Select BHI_MITIGATION_ON whenever user->kernel is mitigated, and keep BHI_MITIGATION_VMEXIT_ONLY for the guest->host-only case. [ bp: Clarify commit title. ]
-
CVE-2026-99746
In the Linux kernel, the following vulnerability has been resolved: ASoC: SOF: Intel: hda-sdw-bpt: Stop active DMA before closing BPT streams When sdw_enable_stream() fails in intel_ace2x_bpt_send_async() after hda_sdw_bpt_send_async() has already started the TX and RX DMA streams, intel_ace2x_bpt_close_stream() invokes hda_sdw_bpt_close() directly without calling hda_sdw_bpt_wait(). As a result, hda_sdw_bpt_dma_deprepare() tears down the DMA buffers and BDL while the DMA hardware is still running. Disable any running TX or RX BPT DMA stream at the start of hda_sdw_bpt_close() before unmapping channels and depreparing DMA resources.
-
CVE-2026-99747
In the Linux kernel, the following vulnerability has been resolved: ksmbd: fix NULL pointer dereference in session_user_name() and session_account_type() session_user_name() and session_account_type() dereference session->user without checking for NULL. A session can be published and exposed via the procfs session interface before authentication completes, during which session->user is NULL: show_proc_session() -> session_user_name(sess) -> user_guest(session->user) -> NULL->flags -> oops Guard both functions by treating a NULL session->user as an explicit not-yet-authenticated state rather than dereferencing it.
-
CVE-2026-99748
In the Linux kernel, the following vulnerability has been resolved: ksmbd: revert fsids on durable reconnect and replay failure On a durable handle reconnect or a request replay, smb2_open() switches to the SMB user's fsids before calling ksmbd_vfs_getattr(). If that call fails, the code jumps to err_out2 and skips ksmbd_revert_fsids() at err_out1. The user's credentials then stay installed on the shared ksmbd worker. Later work on that worker, including other users' requests, runs with the first user's fsuid, fsgid and supplementary groups. ksmbd_free_work_struct() also warns, because work->saved_cred is still set. Jump to err_out1 instead. No path is held at this point, so skipping the path_put() at err_out is correct.
-
CVE-2026-99749
In the Linux kernel, the following vulnerability has been resolved: ksmbd: enforce the read-only share flag when creating a stream When a share is read-only, smb2_open() rejects opens that would create or truncate a file. Creating a named stream on an existing file with streams_xattr sets neither O_CREAT nor O_TRUNC, so it passes this check, and smb2_set_stream_name_xattr() then creates the stream. On a Time Machine share, opening a missing AFP_AfpInfo stream also writes one. A user who can write a file at the filesystem level can therefore add streams to it through a read-only share. Do not create a stream or synthesize AFP_AfpInfo when the share is not writable. Existing streams still open normally, and a missing stream opened with FILE_OPEN still returns STATUS_OBJECT_NAME_NOT_FOUND.
-
CVE-2026-99750
In the Linux kernel, the following vulnerability has been resolved: ksmbd: require DELETE access for delete-on-close opens smb2_open() accepts FILE_DELETE_ON_CLOSE on an existing file without DELETE in DesiredAccess. When the request asks only for FILE_READ_ATTRIBUTES or READ_CONTROL, it also skips inode_permission() and the parent directory MAY_WRITE | MAY_EXEC check that delete-on-close is meant to get. The file is then marked for deletion by a caller who could not delete it. The unlink happens later, in __ksmbd_inode_close(), when the last handle on the inode is closed, and it runs with the credentials of that handle. If the caller closes last, the unlink fails. If another user who can write the parent directory, such as the file's owner, still has the file open, the file is deleted when that user closes it. Until then new opens of the file fail with STATUS_DELETE_PENDING. So a user with only search permission on a directory can delete a file in it, or block opens of it, while the owner has it open. A MAXIMUM_ALLOWED open that also requests DELETE has the same problem. The requested access is replaced with the maximal access, which does not include DELETE when the caller cannot write the parent directory, and the later permission checks are skipped as already done. MS-FSA 2.1.5.1 requires an open with FILE_DELETE_ON_CLOSE and without DELETE in DesiredAccess to fail with STATUS_INVALID_PARAMETER, and Samba does the same. Add that check to the CreateOptions validation, using the generic-mapped DesiredAccess so that GENERIC_ALL still qualifies. This also refuses MAXIMUM_ALLOWED with FILE_DELETE_ON_CLOSE but no explicit DELETE, as MS-FSA and Samba do. When MAXIMUM_ALLOWED is requested with DELETE but the maximal access does not include DELETE, fail with STATUS_ACCESS_DENIED. With DELETE requested, delete-on-close opens go through the existing parent directory check.
-
CVE-2026-99736
In the Linux kernel, the following vulnerability has been resolved: mm/shmem: fix accounting of swapin-error markers A failed swap read replaces the swap entry with a swapin-error marker in the page-cache and subtracts its pages from "alloced" and "swapped"; however, it leaves their block charge in place. Removing the marker during truncate then subtracts from "swapped" again. This can underflow the counters or make swapped reach zero while other pages remain swapped out, causing swapoff to loop without finding them. When installing the marker, subtract only from "swapped" and let shmem_recalc_inode() reduce "alloced" and release the block charge. Make shmem_free_swap() report successful removal separately from the number of swap pages freed, so removing a swapin-error marker does not subtract from "swapped" again. It's a bit confusing, but replacing a swapped out page by a swapin-error marker in shmem_set_folio_swapin_error() behaves as follows: Initially: alloced=1, swapped=1, nrpages=0 Calling shmem_recalc_inode(inode, 0, -1) -> swapped = 0 -> freed = 1 - 0 - 0 = 1 -> alloced -= freed -> alloced=0, swapped=0, block charge released
-
CVE-2026-99737
In the Linux kernel, the following vulnerability has been resolved: mm/shmem: fix truncate loop after boundary swapin failure If swapping in a large entry at the new EOF fails, shmem_undo_range() leaves the entry in place. Its final scan skips the entry, then restarts from the same start index and finds it again. Truncation can loop forever while holding i_rwsem. Advance start past the skipped entry so that a restarted scan can finish, as we already do for folios that cannot be split. Note that this does not fix the missing zeroing when boundary swapin fails, it merely prevents the endless loop.
-
CVE-2026-99738
In the Linux kernel, the following vulnerability has been resolved: mm/shmem: preserve allocation errors when swapping in large entries shmem_swapin_folio() adjusts swap to the requested slot within a large swap entry before allocating a folio. If allocation or charging fails, the error path compares this adjusted value with the original entry still stored in the XArray in shmem_confirm_swap(). Unless the index is at the start of the entry, this turns -ENOMEM into -EEXIST even though nothing changed. The caller retries -EEXIST, so persistent allocation failures can leave the task spinning instead of returning the error. Restore swap to index_entry on allocation failure so the existing error path checks the original entry. This preserves the allocation error if the entry is unchanged and still retries if it changed concurrently.
-
CVE-2026-99740
In the Linux kernel, the following vulnerability has been resolved: mm/shmem: stop retrying folios that cannot be split A fault or collapse can insert a large folio crossing EOF after shmem_undo_range() has handled the boundary folios. If splitting that folio fails, truncate_inode_partial_folio() retains it when it is dirty. The final scan then repeatedly finds it because it restarts from an unchanged start. Truncation keeps the inode lock and ignores fatal signals while looping. A concurrent unlink() can then block other directory operations while holding the directory lock and waiting for the file's lock. Advance start past the retained folio, whose truncated part has already been zeroed. Do not move start backwards if the failed split has shrunk the folio. Fix generated by an LLM and sent by Chris Mason in private. [ david: split fixes, simplify descriptions, adjust comments ]
-
CVE-2026-99741
In the Linux kernel, the following vulnerability has been resolved: tmpfs: keep the quota soft limits of an unused id When the last reference to a dquot goes away, shmem_release_dquot() stores its limits in the rb tree of the mount, unless shmem_is_empty_dquot() says that there is nothing worth storing. That function looks at the usage and at the two hard limits only. So if the administrator sets nothing but soft limits for an id that owns no file yet, the entry is thrown away as soon as the dquot that Q_SETQUOTA used is released, and the id never gets to see a limit. To fix it, compare the soft limits as well: an entry can be empty only if both soft limits are zero as well, which is exactly what shmem_acquire_dquot() sets for a new entry. Fix generated by an LLM and sent by Josef Bacik in private. [ david: slightly adjusted description ]
-
CVE-2026-99742
In the Linux kernel, the following vulnerability has been resolved: mm/mempolicy: charge shared policy nodes Any local user can bypass memory.max with mbind() on a tmpfs file, and allocate unbounded kernel memory. mbind() sets a NUMA memory policy for a range of a mapping. If the mapping is of a shmem and friends, the kernel also stores the policy on the inode, in a tree with one entry per range of the file. Each mbind() on a single page of a mapping can add one shared policy node (struct sp_node) and one struct mempolicy to that tree. They outlive the mapping and stay around until the inode is destroyed. The tree is not bounded by the file size, so the file can be empty. Memory that user space can allocate must be charged to a memory cgroup. A slab cache does that for all of its objects if it is created with SLAB_ACCOUNT. numa_policy_init() creates the slab caches for both structs without SLAB_ACCOUNT, so they are not charged to any cgroup. Reproduced with mbind() on single pages of an empty tmpfs file, from a cgroup with memory.max=8M: the global OOM killer killed tasks owned by root, and the machine panicked. Fix this by creating both caches with SLAB_ACCOUNT. This change implies that mbind() can now fail when running into memcg limits. Further, the policy_cache also serves task and VMA policies, so they are similarly accounted now. Note that per-file limits are not really helpful, as the user could simply use many files; having limits per user is tricky as well. Fix generated by an LLM and submitted by Chris Mason in private. [ david: smaller description tweaks ]
-
CVE-2026-99729
In the Linux kernel, the following vulnerability has been resolved: mm/rmap: handle device-exclusive entries in folio_referenced_one() page_vma_mapped_walk() can return non-present device-exclusive entries, but folio_referenced_one() passes them to PTE young-bit accessors that require present PTEs. We must neither call clear_flush_young_ptes_notify() nor folio_pte_batch() with non-present entries. For example, clearing the young bit on non-present PTEs can corrupt the entry: on arm64 PTE_AF overlaps the swap type, so clearing it turns a device-exclusive entry into an ordinary swap entry. folio_pte_batch() will VM_WARN_ON_FOLIO. Apply the same approach as the previous fixes to e.g., damon_ptep_mkold(): treat these entries as old from the CPU's perspective and obtain their young state through the MMU notifier. Skip PTE batching and MGLRU look-around for non-present entries, and don't modify non-present PTEs.
-
CVE-2026-99730
In the Linux kernel, the following vulnerability has been resolved: mm/filemap: finish fault cleanup before dropping the fault lock On a retried fault, a fatal signal interrupting the folio lock wait makes lock_folio_maybe_drop_mmap() drop the fault lock without pinning the file. However, filemap_fault() may still need to release mapping->invalidate_lock. A concurrent munmap() can release the last file reference and free the mapping before that unlock. Defer dropping the fault lock to filemap_fault()'s retry cleanup, after releasing invalidate_lock. If the fault lock was already dropped for I/O, release the existing file reference instead. Keep the fault lock held for FAULT_FLAG_RETRY_NOWAIT as before.
-
CVE-2026-99731
In the Linux kernel, the following vulnerability has been resolved: mm/filemap: don't lose a folio lock wakeup to a signal An exclusive waiter in folio_wait_bit_common() can leave the wait loop because of a fatal signal, then be woken by folio_unlock() before finish_wait(). If the waker did not hand over the lock, the waiter returns -EINTR without trying to acquire it. folio_wake_bit() stops after waking the first exclusive waiter, so this loses the wakeup and can leave other waiters asleep indefinitely on an unlocked folio. After finish_wait(), let a waiter woken without a lock handoff try once more to acquire the lock. If successful, return success so the caller's eventual unlock wakes the next waiter. Otherwise, the current lock owner will do so. Fix generated by an LLM and sent by Josef Bacik in private. [ david: clarify and simplify comment + description ]
-
CVE-2026-99732
In the Linux kernel, the following vulnerability has been resolved: mm/filemap: read mapping->a_ops->free_folio under i_lock memory_failure() -> ... -> filemap_remove_folio() can read mapping->a_ops after the inode has been freed. Once the folio is deleted from the pagecache it no longer pins the mapping, so after i_lock is dropped e.g., a concurrent final iput() can evict and free the inode before filemap_free_folio() looks up mapping->a_ops->free_folio. Most callers hold a reference on the inode, but not all. Documentation/filesystems/vfs.rst explicitly states that ->free_folio() must not assume the original address_space mapping still exists. Fix it by loading ->free_folio while i_lock is still held, similar to what folio_unmap_invalidate() does. Inline filemap_free_folio() into its remaining caller such that delete_from_page_cache_batch() also just loads the callback before dropping the i_lock. The bug was found by code inspection and has not been reproduced. Fix generated by an LLM and sent by Josef Bacik in private. [ david: get rid of filemap_free_folio() entirely; improve patch description ]
-
CVE-2026-99733
In the Linux kernel, the following vulnerability has been resolved: mm/truncate: look again if a folio left the mapping invalidate_inode_pages2_range() skips a folio which no longer belongs to the mapping by the time it has got the folio lock, on the assumption that somebody else has removed it from the page cache. But the folio may have been replaced, and then there is a new folio with the same data at that index. For most cases, like folio migration, that is impossible. However, collapse_file() only checks the refcount of the old folios, under the i_pages lock, and relies on their folio locks from there on. That works for users which repeat the lookup when they find folio->mapping changed. The following scenario is, however, possible: CPU0 pwrite(O_DIRECT) CPU1 collapse_file() kiocb_invalidate_pages() read the folio (old data), lock it refcount check the write completes copy to new_folio kiocb_invalidate_post_direct_write() find_get_entries() -> old folio xas_store(new_folio) folio->mapping = NULL folio_unlock(folio) folio_lock() folio->mapping != mapping -> continue In that scenario, the collapsed folio with the data from before the write stays in the page cache after the write has returned. This used to be impossible, as khugepaged only collapsed files that nobody had open for writing. Restart the lookup at that index, like shmem_undo_range() does. The same problem seems to exist in truncate_inode_pages_range() as well when truncate_inode_folio() fails, so handle it there as well, but slightly differently: simply check once more through find_get_entries().
-
CVE-2026-99734
In the Linux kernel, the following vulnerability has been resolved: mm/shmem: make shmem_file() check the file's inode shmem_file() only checks if file->f_mapping is a shmem mapping. This mapping can belong to another file, however, as with a Coda file backed by a tmpfs container after mmap(). So it does not guarantee that the file inode is an actual shmem inode. Seal handling and memfd preservation use this check before accessing shmem inode state, so they can access an inode of the wrong type. Let's make shmem_file() check the inode's file operations instead, to identify shmem inodes reliably. This will identify regular shmem inodes, including memfds and ordinary tmpfs files, independently of wrapper file operations. Use shmem_mapping() in collapse_file() instead to not change the behavior of collapse_file(): it operates on the page cache and uses mapping->host for shmem inode operations. Focusing on the file does not seem relevant.
-
CVE-2026-99735
In the Linux kernel, the following vulnerability has been resolved: mm/mglru: don't use stack pointers when walking PUD and PMD tables MGLRU passes stack copies of page-table entries in walk_pud_range() and walk_pmd_range() when walking to the next level. These copies can become stale when khugepaged or MADV_DONTNEED (CONFIG_PT_RECLAIM) removes a PTE table concurrently. This pattern dates back to commit bd74fdaea146 ("mm: multi-gen LRU: support page table walks"). Likely, earlier locking masked the latent bug. Later changes exposed it: commit 1d65b771bc08 ("mm/khugepaged: retract_page_tables() without mmap or vma lock") allowed concurrent retraction, and commit cf4a65539c13 ("memcg-v1: no need for memcg locking for MGLRU") dropped the walker's outer RCU protection. The fix is simple, so just fix the original commit so we don't have to speculate. To fix it, pass the actual PMD/PUD pointers. This restores the RCU-protected lookup and validation against the live PMD. Note that the PUD handling is required as well to handle folded page tables correctly in all cases. Fix generated by an LLM and sent by Josef Bacik in private. [ david: handle walk_pud_range() as well, clarify+simplify description, adjust Fixes: ]
-
CVE-2026-99723
In the Linux kernel, the following vulnerability has been resolved: vhost-vdpa: fix pinned page leak in vhost_vdpa_pa_map() error path When pin_user_pages() or vhost_vdpa_map() fails, vhost_vdpa_pa_map() is supposed to unpin the pages pinned after the last successfully mapped chunk, i.e. the range [map_pfn, last_pfn]. The unpin is guarded by nchunks, but nchunks is reset to 0 every time a chunk is mapped from inside the current page list, so a failure after any such boundary leaves that pending range pinned forever. With physically scattered memory nearly every page list contains a boundary, so the leak happens on practically every failed IOTLB update. The pending range is non-empty exactly when a page has been pinned, which last_pfn already tells us, so drop nchunks and test last_pfn.
-
CVE-2026-99726
In the Linux kernel, the following vulnerability has been resolved: drm/v3d: Widen cache_clean_lock over the whole L2TCACTL sequence v3d_clean_caches() and v3d_flush_l2t() both write the single L2TCACTL register and poll its status bits. The mutex cache_clean_lock exists to serialize them, but v3d_clean_caches() only took the lock around its final FLM_CLEAN write. These functions run concurrently: v3d_flush_l2t() is issued from the BIN/RENDER/CSD invalidate path while v3d_clean_caches() runs from the CACHE_CLEAN queue, and each queue's scheduler uses its own ordered workqueue, so their run_job callbacks execute in parallel. Because clean locked only its final write, a concurrent flush can write L2TCACTL during clean's unlocked phase. Both use non read-modify-write writes to the one register, so whichever lands last wins: clean's TMUWCF write can land on the flush's in-flight L2TFLS invalidate, triggering the GFXH-1897 hazard of writing L2TCACTL while a flush is pending. Hold cache_clean_lock across the entire L2TCACTL access sequence so it is fully mutually exclusive with v3d_flush_l2t(), which already takes the lock around its own write.
-
CVE-2026-99727
In the Linux kernel, the following vulnerability has been resolved: mm/huge_memory: properly handle refcounted PUD mappings on fork copy_huge_pud() copies device-DAX entries without taking a folio reference or updating mapcounts and the child's RSS. For example, MADV_GUARD_INSTALL sets VM_MAYBE_GUARD on a device-DAX VMA, forcing copy_page_range() to traverse the whole VMA on fork. Any remaining huge PUDs are then copied without accounting for the new mappings. Unmapping them in the child decrements accounting that was never added. Use vm_normal_page_pud() to identify refcounted mappings. Take a folio reference and add the child's PUD rmap and RSS accounting before installing the copied entry. Pass src and dst VMAs consistently instead of src and dst MM. Move the TODO.
-
CVE-2026-99728
In the Linux kernel, the following vulnerability has been resolved: mm/huge_memory: fix list corruption when deferred_split_scan() races with deferred_split_folio() Since the deferred split queue became a list_lru, the isolate callback moves the folios it took a reference on to an on-stack list, linked through folio->_deferred_list, and deferred_split_scan() takes them off that list with list_del_init() without holding the list_lru lock. deferred_split_folio() does not know about the on-stack list. It only holds the list_lru lock when it looks at the same list_head, and it can run for such a folio: a THP is queued when it is mapped, as a candidate for the underused check, and is queued "again" when a part of it gets unmapped: CPU0 deferred_split_scan() CPU1 deferred_split_folio() list_del_init(&folio->_deferred_list) next->prev = prev prev->next = next entry->next = entry list_lru_lock_irqsave() folio_set_partially_mapped() __list_lru_add() list_empty(item) -> true list_add_tail(item, &l->list) item->next = &l->list item->prev = tail entry->prev = entry The folio now sits on the list_lru with ->prev pointing to itself. The next removal leaves the neighbours pointing to it, and they still do so after the folio has been freed. With CONFIG_DEBUG_LIST: list_del corruption. prev->next should be ffd4000000270090, but was ffd4000000278090. (prev=ffd4000000270090) Fix it by grabbing the lock around list_del_init() so deferred_split_folio() cannot race with it.
-
CVE-2026-99716
In the Linux kernel, the following vulnerability has been resolved: hwmon: (coretemp) Set nr_cores only after core_data allocation succeeds In init_temp_data(), pdata->nr_cores is populated from topology_num_cores_per_package() before allocating pdata->core_data. If that allocation fails, init_temp_data() returns NULL with pdata->core_data still NULL and pdata->nr_cores set to a positive value. coretemp_add_core() logs the error without failing coretemp_cpu_online(), so the CPU still comes online with pdata->hwmon_dev registered. When the CPU is later taken offline (or the module is unloaded), coretemp_cpu_offline() calls get_temp_data(pd, cpu), which loops up to pdata->nr_cores and dereferences the NULL pdata->core_data pointer. Store the core count in a local variable and assign pdata->nr_cores only after pdata->core_data is allocated, preserving the invariant that pdata->nr_cores is zero whenever pdata->core_data is NULL.