System vulnerabilities

In the Linux kernel, the following vulnerability has been resolved: crypto: loongson - Remove broken and unused loongson-rng The loongson-rng rng_alg has several vulnerabilities, including not providing forward security, and a use-after-free bug due to the use of wait_for_completion_interruptible(). Meanwhile, the rng_alg framework doesn't really have any purpose in the first place other than to access the software algorithms crypto/drbg.c and crypto/jitterentropy.c. Hardware-specific rng_algs have no in-kernel user, and unlike hwrng there's no feed into the actual Linux RNG. As such, there's really no point to this code. There are of course other rng_alg drivers that are similarly unused, but they're similarly in the process of being phased out, e.g. https://lore.kernel.org/r/20260529193648.18172-1-ebiggers@kernel.org and https://lore.kernel.org/r/20260529220430.34135-1-ebiggers@kernel.org Given that, there's no point in fixing forward these vulnerabilities, and it makes much more sense to simply roll back the addition of this driver. If this platform provides TRNG (not PRNG) functionality, it could make sense to add a hwrng driver, but it would be quite different.

In the Linux kernel, the following vulnerability has been resolved: crypto: ccp - Do not initialize SNP for SEV ioctls Sashiko notes: > if SEV initialization fails and KVM is actively running normal VMs, could a > userspace process trigger this code path via /dev/sev ioctls (e.g., > SEV_PDH_GEN) and zero out MSR_VM_HSAVE_PA globally? Would the next VMRUN > execution for an active VM trigger a general protection fault and crash the > host? sev_move_to_init_state() is called for ioctls requiring only SEV firmware: SEV_PEK_GEN, SEV_PDH_GEN, SEV_PEK_CSR, SEV_PEK_CERT_IMPORT, and SEV_PDH_CERT_EXPORT. After the firmware command, it does SEV_SHUTDOWN on the SEV firmware. Since these commands do not require SNP to be initialized, skip it by calling __sev_platform_init_locked() which only initializes the SEV firmware. This way SNP is not Initialized at all, and HSAVE_PA is not cleared. The previous code saved any SEV initialization firmware error to init_args.error and then threw it away and hardcoded the return value of INVALID_PLATFORM_STATE regardless of the real firmware error. This patch changes it to surface the underlying error, which is hopefully both more useful and doesn't cause any problems. Note that it is still safe to call __sev_firmware_shutdown() directly: it calls __sev_snp_shutdown_locked(), which skips SNP shutdown if SNP was not initialized.

In the Linux kernel, the following vulnerability has been resolved: crypto: ccp - Do not initialize SNP for ioctl(SNP_COMMIT) Sashiko notes: > if SEV initialization fails and KVM is actively running normal VMs, could a > userspace process trigger this code path via /dev/sev ioctls (e.g., > SEV_PDH_GEN) and zero out MSR_VM_HSAVE_PA globally? Would the next VMRUN > execution for an active VM trigger a general protection fault and crash the > host? The SNP_COMMIT command does not require the firmware to be in any particular state. Skip initializing it if it was previously uninitialized. The SEV-SNP firmware specification doc 56860 does not mention SNP_COMMIT in Table 5 as a command that is allowed in the UNINIT state, but it is in fact allowed and a future documentation update will reflect that.

In the Linux kernel, the following vulnerability has been resolved: crypto: ccp - Do not initialize SNP for ioctl(SNP_VLEK_LOAD) Sashiko notes: > if SEV initialization fails and KVM is actively running normal VMs, could a > userspace process trigger this code path via /dev/sev ioctls (e.g., > SEV_PDH_GEN) and zero out MSR_VM_HSAVE_PA globally? Would the next VMRUN > execution for an active VM trigger a general protection fault and crash the > host? The SEV firmware docs for SNP_VLEK_LOAD note: > On SNP_SHUTDOWN, the VLEK is deleted. That is, the initialization/shutdown wrapper here is pointless, because the firmware immediately throws away the key anyway. Instead, refuse to do anything if SNP has not been previously initialized. This is an ABI break: before, this was a no-op and almost certainly a mistake by userspace, and now it returns -ENODEV. ABI compatibility could be maintained here by simply returning 0 in the check instead.

In the Linux kernel, the following vulnerability has been resolved: crypto: ccp - Do not initialize SNP for ioctl(SNP_CONFIG) Sashiko notes: > if SEV initialization fails and KVM is actively running normal VMs, could a > userspace process trigger this code path via /dev/sev ioctls (e.g., > SEV_PDH_GEN) and zero out MSR_VM_HSAVE_PA globally? Would the next VMRUN > execution for an active VM trigger a general protection fault and crash the > host? Refuse to re-try initialization if SNP is not already initialized for SNP_CONFIG. This is technically an ABI break: before if SNP initialization failed it could be transparently retriggered by this ioctl, and if no VMs were running, everything worked fine. Hopefully this is enough of a corner case that nobody will notice, but someone does, there are a few options: * do something like symbol_get() for kvm and refuse to initialize if KVM is loaded * check each cpu's HSAVE_PA for non-zero data before re-initializing * once initialization has failed, continue to refuse to initialize until the ccp module is unloaded

In the Linux kernel, the following vulnerability has been resolved: crypto: drbg - Fix returning success on failure in CTR_DRBG drbg_ctr_generate() sometimes returns success when it fails, leaving the output buffer uninitialized. Fix it.

In the Linux kernel, the following vulnerability has been resolved: crypto: qat - protect service table iterations with service_lock The service_table list is protected by service_lock when entries are added or removed (in adf_service_add() and adf_service_remove()), but several functions iterate over the list without holding this lock. A concurrent adf_service_register() or adf_service_unregister() call could modify the list during traversal, leading to list corruption or a use-after-free. Fix this by holding service_lock across all list_for_each_entry() iterations of service_table in adf_dev_init(), adf_dev_start(), adf_dev_stop(), adf_dev_shutdown(), adf_dev_restarting_notify(), adf_dev_restarted_notify(), and adf_error_notifier(). The lock ordering is safe: callers of the static helpers (adf_dev_up() and adf_dev_down()) acquire state_lock before service_lock, and no event_hld callback or service_lock holder ever acquires state_lock in the reverse order.

In the Linux kernel, the following vulnerability has been resolved: crypto: qat - validate RSA CRT component lengths The generic RSA key parser (rsa_helper.c) bounds each CRT component (p, q, dp, dq, qinv) by the modulus size n_sz, but qat_rsa_setkey_crt() allocates half-size DMA buffers (key_sz / 2) and right-aligns each component with: memcpy(dst + half_key_sz - len, src, len) When a CRT component is larger than half_key_sz the subtraction underflows and memcpy writes past the DMA buffer, causing memory corruption. Add a len > half_key_sz check next to the existing !len check for each of the five CRT components so the driver falls back to the non-CRT path instead of writing out of bounds.

In the Linux kernel, the following vulnerability has been resolved: spi: fsl-lpspi: terminate the RX channel on TX prepare failure path When dmaengine_prep_slave_sg() fails for the TX channel, the error path terminates the TX DMA channel but leaves the RX channel running. Since the RX channel was already submitted and issued prior to preparing the TX descriptor, returning -EINVAL causes the SPI core to unmap the DMA buffers while the RX DMA engine continues writing to them, leading to potential memory corruption or use-after-free. Terminate the RX channel before returning on the TX prepare failure path.

In the Linux kernel, the following vulnerability has been resolved: x86/mm: Fix freeing of PMD-sized vmemmap pages Commit bf9e4e30f353 ("x86/mm: use pagetable_free()"), switched from freeing non-boot page tables through __free_pages() to pagetable_free(). However, the function is also called to free vmemmap pages. Given that vmemmap pages are not page tables, already the page_ptdesc(page) is wrong. But worse, pagetable_free() calls: __free_pages(page, compound_order(page)); Since vmemmap pages are not compound pages (see vmemmap_alloc_block()) -- except for HVO, which doesn't apply here -- only first page of a PMD-sized vmemmap page is freed, leaking the other ones. Fix it by properly decoupling pagetable and vmemmap freeing. free_pagetable() no longer has to mess with SECTION_INFO, as only the vmemmap is marked like that in register_page_bootmem_memmap(). The indentation in remove_pmd_table() is messed up. Fix that while touching it. Bootmem info handling will soon be fixed up. For now, handle it similar to free_pagetable(), just avoiding the ifdef. [ dhansen: changelog munging. More imperative voice ]

In the Linux kernel, the following vulnerability has been resolved: regulator: scmi: fix of_node refcount leak in scmi_regulator_probe() scmi_regulator_probe() calls of_find_node_by_name() which takes a reference on the returned device node. On the error path where process_scmi_regulator_of_node() fails, the function returns without calling of_node_put() on the child node, leaking the reference. Add of_node_put(np) on the error path to properly release the reference.

In the Linux kernel, the following vulnerability has been resolved: perf/aux: Fix page UAF in map_range() map_range() reads rb->aux_pages[], rb->aux_nr_pages and rb->aux_pgoff via perf_mmap_to_page() while holding only event->mmap_mutex. Those fields are serialized by rb->aux_mutex, and mmap_mutex is per event. Thus, two events sharing one rb via PERF_EVENT_IOC_SET_OUTPUT can race rb_alloc_aux() with map_range(), leading to a page-UAF scenario as follows: CPU 0 CPU 1 ===== ===== rb_alloc_aux() map_range() [1]: allocate rb->aux_pages[0] [2]: rb->aux_nr_pages++ [3]: perf_mmap_to_page() returns rb->aux_pages[0] [4]: map it as VM_PFNMAP [5]: rb->aux_pgoff = 1 munmap the page [6]: free rb->aux_pages[0] Pages mapped as VM_PFNMAP have no refcount protection, so CPU 1 holds a mapping to a freed physical frame. Fix this by taking rb->aux_mutex across the page walk in map_range().

In the Linux kernel, the following vulnerability has been resolved: tracing: Prevent out-of-bounds read in glob matching String event fields are not necessarily NUL-terminated, so the filter predicate functions (filter_pred_string(), filter_pred_strloc() and filter_pred_strrelloc()) pass the field length to the regex match callbacks, and the length-aware matchers honour it. regex_match_glob() was the exception: it ignored the length and called glob_match(), which scans the string until it hits a NUL byte. Some string fields are not NUL-terminated. One example is the dynamic char array of the xfs_* namespace tracepoints, which is copied without a trailing NUL. For such a field, glob matching reads past the end of the event field, causing a KASAN slab-out-of-bounds read in glob_match(), reached via regex_match_glob() and filter_match_preds() from the xfs_lookup tracepoint. Add a length-bounded glob_match_len() and use it from regex_match_glob() so glob matching always stops at the field boundary. The matching loop is factored into a shared helper so glob_match() keeps its behaviour.

In the Linux kernel, the following vulnerability has been resolved: NFSv4: include MAY_WRITE in open permission mask for O_TRUNC POSIX requires write permission to truncate a file, so an open() that specifies O_TRUNC must be authorized for write access regardless of the O_ACCMODE access mode. nfs_open_permission_mask() builds the access mask passed to nfs_may_open(), which is the local authorization gate for OPENs the client serves itself from a cached write delegation via the can_open_delegated() path in nfs4_try_open_cached(). The mask is derived from O_ACCMODE alone, so an open(O_RDONLY | O_TRUNC) against a file the caller cannot write requests only MAY_READ and passes the local check. The OPEN is then satisfied locally and the truncation is issued to the server as a SETATTR(size=0) over the delegation stateid, which the server accepts under standard write-delegation semantics. POSIX requires that this open fail with EACCES. Include MAY_WRITE in the mask whenever O_TRUNC is set so the local check matches the access the server would have enforced.

In the Linux kernel, the following vulnerability has been resolved: module: decompress: check return value of module_extend_max_pages() module_extend_max_pages() calls kvrealloc() internally and returns -ENOMEM on allocation failure. The return value is never checked. If the initial allocation fails, info->pages remains NULL and info->max_pages remains 0. Subsequent calls to module_get_next_page() will attempt to dynamically grow the array by calling module_extend_max_pages(info, 0) since info->used_pages is 0. This results in kvrealloc(NULL, 0) returning ZERO_SIZE_PTR, which is treated as a success, leading to a dereference of ZERO_SIZE_PTR and a kernel oops. Fix: add the missing error check after module_extend_max_pages() and return immediately on failure. This matches the pattern used by every other kvrealloc() caller in the module loading path. [Sami: Corrected the analysis in the commit message.]

In the Linux kernel, the following vulnerability has been resolved: exfat: bound uniname advance in exfat_find_dir_entry() In exfat_find_dir_entry(), each TYPE_EXTEND (file name) entry advances the output pointer by a fixed amount while the loop guard only tracks the accumulated name length: if (++order == 2) uniname = p_uniname->name; else uniname += EXFAT_FILE_NAME_LEN; len = exfat_extract_uni_name(ep, entry_uniname); name_len += len; unichar = *(uniname+len); *(uniname+len) = 0x0; uniname grows by EXFAT_FILE_NAME_LEN (15) per name entry, but name_len grows only by the actual extracted length, which is shorter when a name fragment contains an early NUL. The only guard is `name_len >= MAX_NAME_LENGTH`, so a crafted directory with many short name fragments lets uniname run far past the p_uniname->name[MAX_NAME_LENGTH + 3] buffer while name_len stays small, causing an out-of-bounds read and write at *(uniname+len). The sibling extractor exfat_get_uniname_from_ext_entry() already stops on a short fragment (the lockstep `len != EXFAT_FILE_NAME_LEN` guard added in commit d42334578eba ("exfat: check if filename entries exceeds max filename length")); exfat_find_dir_entry() never got the equivalent. Track the per-entry write offset as a count and reject a fragment once the offset, or the offset plus the extracted length, would exceed MAX_NAME_LENGTH, before forming the output pointer.

In the Linux kernel, the following vulnerability has been resolved: mm: page_ext: add count limit to page_ext_iter_next to prevent invalid PFN access The page_ext iteration API does not validate if the PFN still belongs to a valid section while advancing the iterator. When dynamically adding memory in the hotplug path, it can lead to a NULL pointer dereference during page_ext_lookup at the boundary of the last valid section when iterator count equals __pgcount. The for_each_page_ext() macro calls page_ext_iter_next() as its loop increment. for_each_page_ext() does a "__page_ext = page_ext_iter_next(&__iter)" at the end. This causes page_ext_iter_next() to increment iter->index past __pgcount and call page_ext_lookup(start_pfn + __pgcount). During memory hotplug (online), the PFN at start_pfn + __pgcount may belong to a section that has not yet been initialized, causing page_ext_lookup() to trigger a NULL pointer dereference. [ 14.555124][ T846] Call trace: [ 14.555125][ T846] lookup_page_ext+0x6c/0x108 (P) [ 14.555127][ T846] page_ext_lookup+0x30/0x3c [ 14.555129][ T846] __reset_page_owner+0x11c/0x260 [ 14.571201][ T846] __free_pages_ok+0x5e8/0x8e0 [ 14.571204][ T846] __free_pages_core+0x78/0xf0 [ 14.571206][ T846] generic_online_page+0x14/0x24 [ 14.597782][ T846] online_pages+0x178/0x30c [ 14.597784][ T846] memory_block_change_state+0x284/0x32c [ 14.597787][ T846] memory_subsys_online+0x4c/0x64 [ 14.597789][ T846] device_online+0x88/0xb0 [ 14.597791][ T846] online_memory_block+0x30/0x40 [ 14.597793][ T846] walk_memory_blocks+0xac/0xe8 [ 14.597794][ T846] add_memory_resource+0x280/0x298 [ 14.656161][ T846] add_memory+0x60/0x98 Move the iteration boundary enforcement inside the iterator functions, so callers cannot inadvertently access beyond the requested range.

In the Linux kernel, the following vulnerability has been resolved: mm: do file ownership checks with the proper mount idmap Ever since idmapped mounts were introduced, inode ownership checks (for side-channel protection) in mincore() and madvise(MADV_PAGEOUT) were done against the nop_mnt_idmap, which completely ignores the file's mount's idmap. This results in odd edgecases like: 1) mount/bind-mount with an idmap userA:userB:1 2) userB runs an owner_or_capable() check on file that is owned by userA on-disk/in-memory, but owned by userB after idmap translation 3) owner_or_capable() mysteriously fails as the correct idmap wasn't supplied In the case of mincore/madvise MADV_PAGEOUT, this is usually benign, because file_permission(file, MAY_WRITE) will probably succeed, as it uses the proper idmap internally, but it does not need to be the case on e.g a 0444 file where even the owner itself doesn't have permissions to write to it. Since this is clearly not trivial to get right, introduce a file_owner_or_capable() that can carry the correct semantics, and switch the various users in mm to it. The issue was found by manual code inspection & an off-list discussion with Jan Kara.

In the Linux kernel, the following vulnerability has been resolved: iommufd: Use sizeof(*hdr) instead of sizeof(hdr) in veventq read The bound-check in iommufd_veventq_fops_read() for the normal vEVENT path uses sizeof(hdr) where the surrounding code uses sizeof(*hdr): if (!vevent_for_lost_events_header(cur) && sizeof(hdr) + cur->data_len > count - done) { hdr is declared as struct iommufd_vevent_header *, so sizeof(hdr) evaluates to the size of the pointer. Surrounding code uses sizeof(*hdr) consistently: if (done >= count || sizeof(*hdr) > count - done) { ... if (copy_to_user(buf + done, hdr, sizeof(*hdr))) { ... done += sizeof(*hdr); struct iommufd_vevent_header is currently 8 bytes (two __u32 fields, flags and sequence), so on 64-bit (sizeof(void *) == 8) the two expressions happen to be equal and the check works as intended. On 32-bit (sizeof(void *) == 4) the check under-counts the header by 4 bytes: a vEVENT whose data_len causes 8 + cur->data_len to exceed count - done while 4 + cur->data_len does not will pass the check, then the loop will copy_to_user 8 bytes of header followed by data_len bytes of payload, writing past the user-supplied buffer. It is also a latent bug for any future expansion of struct iommufd_vevent_header beyond sizeof(void *) on 64-bit; the check should not depend on the type happening to match the host pointer width. Use sizeof(*hdr) to match the rest of the function and the actual amount that will be copied.

In the Linux kernel, the following vulnerability has been resolved: iommufd: Move vevent memory allocation outside spinlock The veventq memory allocation happens inside the spinlock. Given its depth is decided by the user space, this leaves a vulnerability, where userspace can allocate large queues to exhaust atomic memory reserves. Move the allocation outside the spinlock and use GFP_NOWAIT, which can fail fast under memory pressure without dipping into the GFP_ATOMIC reserves or direct-reclaiming from the threaded IRQ handler. On allocation failure, queue the lost_events_header (so userspace learns of the drop) and return -ENOMEM so the caller learns of the kernel-side memory pressure. This is intentionally distinct from the queue-overflow path, which also queues the lost_events_header but returns 0: a full queue is an expected userspace-pacing condition rather than a kernel error. A subsequent change will cap the upper bound of the veventq_depth.

In the Linux kernel, the following vulnerability has been resolved: iommufd: Set veventq_depth upper bound iommufd_veventq_alloc() accepts any !0 veventq_depth from userspace, with an upper bound at U32_MAX. This leaves a vulnerability where userspace can allocate excessively large queues to exhaust kernel memory reserves. Cap the veventq_depth (maximum number of entries) to 1 << 19, matching the maximum number of entries in the SMMUv3 EVTQ (the largest use case today).

In the Linux kernel, the following vulnerability has been resolved: iommufd: Break the loop on failure in iommufd_fault_fops_read() On a copy_to_user() failure inside the inner list_for_each_entry, only the inner loop breaks; the outer while re-fetches the just-restored fault group and retries the failing copy_to_user() forever, spinning the reader at 100% CPU with fault->mutex held. Check rc after the inner loop and break the outer while as well.

In the Linux kernel, the following vulnerability has been resolved: iommufd: Set upper bounds on cache invalidation entry_num and entry_len iommufd_hwpt_invalidate() takes a user-controlled entry_num and entry_len, each bounded only by U32_MAX. An entry_len beyond the kernel's struct size makes the copy helper verify the extra bytes are zero, scanning that excess in one uninterruptible pass; a multi-gigabyte value over zeroed user memory trips the soft-lockup watchdog. A large entry_num is the other half, driving the backend invalidation loop with no reschedule. The VT-d nested handler, for one, copies each entry and flushes caches per iteration, pinning the CPU on a non-preemptible kernel. Cap both in the ioctl. entry_len is held under PAGE_SIZE, above any request struct, and entry_num under 1 << 19, the order of a hardware invalidation queue and well beyond any real batch, bounding the per-call loop length.

In the Linux kernel, the following vulnerability has been resolved: KVM: arm64: nv: Avoid dereferencing NULL VNCR pseudo-TLB VNCR TLB invalidation occurs from MMU notifiers or TLBI instructions, and either can race against a vcpu not being onlined yet (no pseudo-TLB allocated). Similarly, the TLB might be invalid, and the invalidation should be skipped in this case. Both kvm_invalidate_vncr_ipa() and kvm_invalidate_vncr_va() are expected to perform the same checks, except that the latter doesn't check for the allocation and blindly dereferences the pointer. Solve this by introducing a new iterator built on top of the usual kvm_for_each_vcpu() that checks for both of the above conditions, and convert the two users to it.

In the Linux kernel, the following vulnerability has been resolved: KVM: arm64: Bound used_lrs when flushing the pKVM hyp vCPU flush_hyp_vcpu() copies the host vGIC state into the hyp's private vCPU on every run. The vGIC list register save and restore use used_lrs as their loop bound and expect it to stay within the number of implemented list registers. While this is generally the case, flush_hyp_vcpu() copies vgic_v3 verbatim and does not enforce this, so a value provided by the host is used at EL2 to index vgic_lr[] and access ICH_LR<n>_EL2 (host -> EL2). Fix by clamping used_lrs to the number of implemented list registers after the copy, as the trusted path already does in vgic_flush_lr_state(). The number of implemented list registers is constant after init, so it is replicated once from kvm_vgic_global_state.nr_lr into hyp_gicv3_nr_lr rather than read on every entry.

In the Linux kernel, the following vulnerability has been resolved: KVM: arm64: Clear __hyp_running_vcpu when flushing the pKVM hyp vCPU flush_hyp_vcpu() copies the host vCPU context into the hyp's private vCPU on every run. ctxt_to_vcpu() expects a guest context to have a NULL __hyp_running_vcpu, which is only ever set on the host context, so that it resolves the vCPU via container_of(). While this is generally the case, flush_hyp_vcpu() copies the context verbatim and does not enforce this, so a value provided by the host is dereferenced at EL2 (host -> EL2). Fix by clearing __hyp_running_vcpu after the copy.

In the Linux kernel, the following vulnerability has been resolved: KVM: SEV: Pin source page for write when adding CPUID data for SNP guest When populating a guest_memfd instance with the initial CPUID data for an SNP guest, acquire a writable pin on the source page as KVM will write back the "correct" CPUID information if the userspace provided data is rejected by trusted firmware. Because KVM writes to the source page using a kernel mapping, pinning for read could result in KVM clobbering read-only memory. Note, well-behaved VMMs are unlikely to be affected, as CPUID information is almost always dynamically generated by userspace, i.e. it's unlikely for the CPUID information to be backed by a read-only mapping. [sean: rewrite shortlog and changelog, tag for stable@]

In the Linux kernel, the following vulnerability has been resolved: KVM: x86: Ensure vendor's exit handler runs before fastpath userspace exits Move the handling of fastpath userspace exits into vendor code to ensure KVM runs vendor specific operations that need to run before userspace gains control of the vCPU. E.g. for VMX (and soon to be for SVM as well), KVM needs to flush the PML buffer prior to exiting to userspace, otherwise any memory written by the final KVM_RUN might never be flagged as dirty. Note, waiting to snapshot CR0 and CR3 until svm_handle_exit() is flawed in general, as that risks consuming stale state in a fastpath handler. That will be addressed in a future change.

In the Linux kernel, the following vulnerability has been resolved: KVM: guest_memfd: Treat memslot binding offset+size as unsigned values When binding a memslot to a guest_memfd file, treat the offset and size as unsigned values to fix a bug where the sum of the two can result in a false negative when checking for overflow against the size of the file. Passing unsigned values also avoids relying on somewhat obscure checks in other flows for safety, and tracks the offset and size as they are intended to be tracked, as unsigned values. On 64-bit kernels, the number of pages a memslot contains and thus the size (and offset) of its guest_memfd binding are unsigned 64-bit values. Taking the offset+size as an loff_t instead of a uoff_t inadvertently converts the unsigned value to a signed value if the offset and/or size is massive. Locally storing the offset and size as signed values is benign in and of itself (though even that is *extremely* difficult to discern), but operating on their sum is not. For the offset, KVM explicitly checks against a negative value, which might seem like a bug as KVM could incorrectly reject a legitimate binding, but that's not actually the case as KVM_CREATE_GUEST_MEMFD takes a signed value for its size, i.e. a would-be-negative offset is also greater than the maximum possible size of any guest_memfd file. Regarding the size, while KVM lacks an explicit check for a negative value, i.e. seemingly has a flawed overflow check, KVM restricts the number of pages in a single memslot to the largest positive signed 32-bit value: if (id < KVM_USER_MEM_SLOTS && (mem->memory_size >> PAGE_SHIFT) > KVM_MEM_MAX_NR_PAGES) return -EINVAL; and so that maximum "size" will ever be is 0x7fffffff000. The sum of the two is, however, problematic. While the size is restricted by KVM's memslot logic, the offset is not, i.e. the offset is completely unchecked until the "offset + size > i_size_read(inode)" check. If the offset is the (nearly) largest possible _positive_ value, then adding size to the offset can result in a signed, negative 64-bit value. When compared against the size of the file (guaranteed to be positive), the negative sum is always smaller, and KVM incorrectly allows the absurd offset. Opportunistically add missing includes in kvm_mm.h (instead of relying on its parents).

In the Linux kernel, the following vulnerability has been resolved: KVM: arm64: Don't leak PFN when kvm_translate_vncr() races MMU notifier In the case that kvm_translate_vncr() races with an MMU notifier the early return does not release a reference on the faulted in PFN. Add the necessary call to kvm_release_faultin_page() for the unused PFN.

In the Linux kernel, the following vulnerability has been resolved: svcrdma: wake sq waiters when the transport closes Threads parked in svc_rdma_sq_wait() on sc_sq_ticket_wait or sc_send_wait can hang indefinitely in TASK_UNINTERRUPTIBLE state across transport teardown, pinning svc_xprt references and blocking svc_rdma_free(). The close path sets XPT_CLOSE before invoking xpo_detach and both wait_event predicates include an XPT_CLOSE term, but the predicates are re-evaluated only on wakeup. sc_sq_ticket_wait has no completion-driven wake path; it is advanced solely by the chained ticket handoff inside svc_rdma_sq_wait() itself. Without an explicit wake at close, parked threads never observe XPT_CLOSE, hold their svc_xprt_get reference forever, and svc_rdma_free() blocks on xpt_ref dropping to zero. Two close entry points reach this transport. Local teardown runs svc_rdma_detach() from svc_handle_xprt() -> svc_delete_xprt() -> xpo_detach() on a worker thread. A remote disconnect arrives at svc_rdma_cma_handler(), which calls svc_xprt_deferred_close(): that sets XPT_CLOSE and enqueues the transport but does not access either RDMA waitqueue, so a worker already parked in svc_rdma_sq_wait() never re-evaluates its predicate. With every worker parked on this transport, no thread is available to run the local teardown either, and the wake site there is unreachable. Introduce svc_rdma_xprt_deferred_close(), a thin svcrdma wrapper that calls svc_xprt_deferred_close() and then wakes both sc_sq_ticket_wait and sc_send_wait. Convert the svcrdma producers that called svc_xprt_deferred_close() directly: svc_rdma_cma_handler(), qp_event_handler(), svc_rdma_post_send_err(), svc_rdma_wc_send(), the sendto drop path, the rw completion error paths, and the recvfrom flush and read-list error paths. Wake both waitqueues from svc_rdma_detach() as well. The synchronous svc_xprt_close() path (backchannel ENOTCONN, device removal via svc_rdma_xprt_done) reaches detach without flowing through svc_xprt_deferred_close() and therefore does not invoke the new helper. [ cel: add svc_rdma_xprt_deferred_close() to complete the fix ]

In the Linux kernel, the following vulnerability has been resolved: fpga: dfl-afu: validate DMA mapping length in afu_dma_map_region() afu_ioctl_dma_map() accepts a 64-bit length from userspace via DFL_FPGA_PORT_DMA_MAP ioctl without an upper bound check. The value is passed to afu_dma_pin_pages() where npages is derived as length >> PAGE_SHIFT and passed to pin_user_pages_fast() which takes int nr_pages, causing implicit truncation if length is very large. Validate map.length at the ioctl entry point before calling afu_dma_map_region(), rejecting values whose page count exceeds INT_MAX.

In the Linux kernel, the following vulnerability has been resolved: i2c: core: fix adapter deregistration race Adapters can be looked up by their id using i2c_get_adapter() which takes a reference to the embedded struct device. Remove the adapter from the IDR before tearing it down during deregistration (and on registration failure) to make sure its resources are not accessed after having been freed (e.g. the device name).

In the Linux kernel, the following vulnerability has been resolved: i2c: imx-lpi2c: mark I2C adapter when hardware is powered down On some i.MX platforms, certain I2C client drivers keep a periodic workqueue which continues to trigger I2C transfers. During system suspend/resume, there exists a time window between: - suspend_noirq and the system entering suspend - the system starting to resume and resume_noirq In this window, the I2C controller resources such as clock and pinctrl may already be disabled or not yet restored. If a workqueue triggers an I2C transfer in this period, the driver attempts to access I2C registers while the hardware resources are unavailable, which may lead to system hang. Mark the I2C adapter as suspended during noirq suspend and block new transfers until resume, ensuring that I2C transfers are only issued when hardware resources are available.

In the Linux kernel, the following vulnerability has been resolved: Input: synaptics-rmi4 - bound the F3A keymap to the GPIO count rmi_f3a_initialize() takes the GPIO count from the device query register (f3a->gpio_count = buf & RMI_F3A_GPIO_COUNT, range 0..127). rmi_f3a_map_gpios() then allocates gpio_key_map with min(gpio_count, TRACKSTICK_RANGE_END) == at most 6 entries, but rmi_f3a_attention() iterates the full gpio_count and dereferences gpio_key_map[i], and input->keycodemax is set to the full gpio_count while input->keycode points at the 6-entry allocation. A device that reports gpio_count > 6 therefore causes an out-of-bounds read of gpio_key_map[] on every attention interrupt, and out-of-bounds accesses through the input core's default keymap ioctls: EVIOCGKEYCODE reads past the buffer (leaking adjacent slab memory to user space) and EVIOCSKEYCODE writes a caller-controlled value past it, for any process able to open the evdev node, since input_default_getkeycode() and input_default_setkeycode() only bound the index against keycodemax. Size the keymap for the full gpio_count. The mapping loop is unchanged: it still assigns only the first min(gpio_count, TRACKSTICK_RANGE_END) entries; the remaining slots stay KEY_RESERVED (devm_kcalloc zero-fills) and are skipped when reporting.

In the Linux kernel, the following vulnerability has been resolved: Input: synaptics-rmi4 - bound the F30 keymap to the GPIO/LED count rmi_f30_map_gpios() allocates gpioled_key_map with min(gpioled_count, TRACKSTICK_RANGE_END) == at most 6 entries, but rmi_f30_attention() iterates the full f30->gpioled_count (device query register, range 0..31) and dereferences gpioled_key_map[i], and input->keycodemax is set to the full gpioled_count while input->keycode points at the 6-entry allocation. A device that reports gpioled_count > 6 with GPIO support enabled therefore causes an out-of-bounds read on the attention interrupt and out-of-bounds read/write through the EVIOCGKEYCODE/EVIOCSKEYCODE ioctls, which bound the index only against keycodemax. This is the same defect as the F3A handler, which was copied from F30. Size the keymap for the full gpioled_count; the mapping loop still assigns only the first min(gpioled_count, TRACKSTICK_RANGE_END) entries.

In the Linux kernel, the following vulnerability has been resolved: Input: elan_i2c - prevent division by zero and arithmetic underflow The Elan I2C touchpad driver queries the device for its physical dimensions and trace counts to calculate the device resolution and width. However, if the device firmware or device tree provides invalid zero values for x_traces or y_traces, it results in a fatal division-by-zero exception leading to a kernel panic during device probe. Add checks to ensure these parameters are non-zero before performing the division. If invalid trace values are detected, fall back to a safe default of 1. Additionally, prevent an arithmetic underflow in the touch reporting logic. Previously, if the calculated or fallback width was smaller than ETP_FWIDTH_REDUCE (90), the subtraction would underflow, resulting in a massive unsigned integer being reported to userspace. Clamp the adjusted width to a minimum of 0 to safely handle small physical dimensions and fallback scenarios. Completing the probe with safe fallback values ensures the sysfs nodes are created, keeping the firmware update path intact so a recovery firmware can be flashed to the device.

In the Linux kernel, the following vulnerability has been resolved: Input: goodix - clamp the device-reported contact count goodix_ts_read_input_report() copies the number of touch points reported by the device into an on-stack buffer u8 point_data[2 + GOODIX_MAX_CONTACT_SIZE * GOODIX_MAX_CONTACTS]; which is sized for at most GOODIX_MAX_CONTACTS (10) contacts. The only runtime check bounds the per-interrupt count against ts->max_touch_num, but that value is taken verbatim from a 4-bit field of the device configuration block and is never clamped: ts->max_touch_num = ts->config[MAX_CONTACTS_LOC] & 0x0f; The nibble can be 0..15, so a malfunctioning, malicious or counterfeit controller (or an attacker tampering with the I2C bus) can advertise up to 15 contacts. goodix_ts_read_input_report() then accepts a touch_num of up to 15 and the second goodix_i2c_read() writes ts->contact_size * (touch_num - 1) bytes past the one-contact header into point_data - up to 30 bytes (45 with the 9-byte report format) beyond the 92-byte buffer: a stack out-of-bounds write. Clamp max_touch_num to GOODIX_MAX_CONTACTS, the number of contacts point_data[] is sized for, when reading it from the configuration.

In the Linux kernel, the following vulnerability has been resolved: Input: iforce - bound the device-reported force-feedback effect index iforce_process_packet() handles a status report (packet id 0x02) by taking a force-feedback effect index straight from the device wire and using it to address the per-effect state array: i = data[1] & 0x7f; if (data[1] & 0x80) { if (!test_and_set_bit(FF_CORE_IS_PLAYED, iforce->core_effects[i].flags)) ... } else if (test_and_clear_bit(FF_CORE_IS_PLAYED, iforce->core_effects[i].flags)) { ... } The index is masked only with 0x7f, so it ranges 0..127, but core_effects[] holds only IFORCE_EFFECTS_MAX (32) entries. For an index of 32..127 the test_and_set_bit()/test_and_clear_bit() is an out-of-bounds single-bit read-modify-write past the array. core_effects[] is the second-to-last member of struct iforce, so the write lands in the trailing members and beyond the embedding kzalloc()'d iforce_serio / iforce_usb object. data[1] is unvalidated device payload on both transports (the USB interrupt endpoint and serio), and the status path is not gated on force feedback being present, so a malicious or counterfeit device can set or clear a bit at an attacker-chosen offset past the object. Reject an out-of-range index instead of indexing with it. Bound against the array dimension IFORCE_EFFECTS_MAX rather than dev->ff->max_effects so the check guarantees memory safety regardless of how many effects the device registered. A legitimate "effect started/stopped" status always carries an index below IFORCE_EFFECTS_MAX, so well-formed devices are unaffected; the neighbouring mark_core_as_ready() loop is already bounded and is left untouched.

In the Linux kernel, the following vulnerability has been resolved: Input: mms114 - fix touch indexing for MMS134S and MMS136 The MMS134S and MMS136 touch controllers have an event size of 6 bytes rather than 8 bytes. When __mms114_read_reg() reads the touch data packet from the device into the touch buffer, the events are packed tightly at 6-byte intervals. However, the driver iterates through the events using standard C array indexing (touch[index]), where each element is sizeof(struct mms114_touch) (8 bytes) apart. As a result, any touch events beyond the first one are read from incorrect offsets and parsed improperly. Fix this by explicitly calculating the byte offset for each touch event based on the device's specific event size.

In the Linux kernel, the following vulnerability has been resolved: Input: touchwin - reset the packet index on every complete packet tw_interrupt() accumulates each non-zero serial byte into a fixed three-byte buffer with a running index that is only reset once a full packet has been received *and* the device's two Y bytes agree: tw->data[tw->idx++] = data; if (tw->idx == TW_LENGTH && tw->data[1] == tw->data[2]) { ... tw->idx = 0; } The reset is gated on tw->data[1] == tw->data[2], a value the device controls. A malicious, malfunctioning or counterfeit Touchwindow peripheral can stream non-zero bytes whose 2nd and 3rd bytes differ: the index reaches TW_LENGTH without the equality holding, is never reset, and keeps growing, so tw->data[tw->idx++] walks off the end of the three-byte array and the rest of the heap-allocated struct tw, one attacker-chosen byte at a time -- an unbounded, device-driven heap out-of-bounds write. Reset the index on every completed packet and report an event only when the two Y bytes match, like the other serio touchscreen drivers do.

In the Linux kernel, the following vulnerability has been resolved: Input: mms114 - reject an oversized device packet size mms114_interrupt() reads a packet of touch data from the device into a fixed-size on-stack buffer struct mms114_touch touch[MMS114_MAX_TOUCH]; which holds MMS114_MAX_TOUCH (10) events of MMS114_EVENT_SIZE (8) bytes, i.e. 80 bytes. The length of the I2C read into it is taken verbatim from the device: packet_size = mms114_read_reg(data, MMS114_PACKET_SIZE); if (packet_size <= 0) goto out; ... error = __mms114_read_reg(data, MMS114_INFORMATION, packet_size, (u8 *)touch); packet_size is a single device register byte (0x0F) and the only check is the lower bound packet_size <= 0; it is never bounded against the size of touch[]. A malfunctioning, malicious or counterfeit controller (or an attacker tampering with the I2C bus) can report a packet_size of up to 255, so __mms114_read_reg() writes up to 175 bytes past the end of touch[] on the IRQ-thread stack: a stack out-of-bounds write that can overwrite the stack canary, saved registers and the return address. A well-formed device never reports more than the buffer holds, so reject an oversized packet and drop the report, consistent with the handler's other error paths, rather than reading past the buffer.

In the Linux kernel, the following vulnerability has been resolved: RDMA/rtrs-srv: Bound RDMA-Write length to chunk size in rdma_write_sg When the server answers an RTRS READ, rdma_write_sg() builds the source scatter/gather entry for the IB_WR_RDMA_WRITE that returns data to the peer. Its length is taken directly from the wire descriptor: plist->length = le32_to_cpu(id->rd_msg->desc[0].len); rd_msg points into the chunk buffer that the remote peer filled via RDMA-WRITE-WITH-IMM (rtrs_srv_rdma_done() -> process_io_req() -> process_read()), so desc[0].len is attacker-controlled and, before this change, was only rejected when zero. The source address is the fixed chunk start (dma_addr[msg_id]) and the source lkey is the PD-wide local_dma_lkey, which is not tied to the chunk's MR mapping, so the verbs layer does not constrain the transfer length to max_chunk_size. msg_id and off are bounded against queue_depth and max_chunk_size in rtrs_srv_rdma_done(), but desc[0].len is a separate field that was not checked against the chunk size. A peer that advertises desc[0].len larger than max_chunk_size can make the posted RDMA write read past the chunk's mapped region. The resulting behaviour depends on the IOMMU configuration: with no IOMMU or in passthrough mode the read may extend into memory adjacent to the chunk and be returned to the peer, which can disclose host memory; with a translating IOMMU the out-of-range access is expected to fault and abort the connection. In either case the transfer exceeds what the protocol permits and is driven by a remote peer. Reject a descriptor length above max_chunk_size, mirroring the existing off >= max_chunk_size bound in rtrs_srv_rdma_done(). Legitimate clients do not exceed it: the client sets desc[0].len to its MR length, which is capped at the negotiated max_io_size (max_chunk_size - MAX_HDR_SIZE).

In the Linux kernel, the following vulnerability has been resolved: RDMA/siw: bound Read Response placement to the RREAD length In drivers/infiniband/sw/siw/siw_qp_rx.c, siw_proc_rresp() places each inbound Read Response DDP segment at sge->laddr + wqe->processed and then accumulates wqe->processed, but it never checks the running total against the sink buffer length on continuation segments. siw_check_sge() resolves and validates the sink memory only on the first fragment (the if (!*mem) branch), and siw_rresp_check_ntoh() compares the cumulative length against wqe->bytes only on the final segment (the !frx->more_ddp_segs guard). A connected siw peer that answers an outstanding RREAD with Read Response segments that keep the DDP Last flag clear, carrying more total payload than the RREAD requested, drives wqe->processed past the validated sink buffer; the next siw_rx_data() call writes out of bounds at sge->laddr + wqe->processed. siw runs iWARP over ordinary routable TCP, so the peer is the remote end of an established RDMA connection and needs no local privilege. Bound every segment before placement, exactly as siw_proc_send() and siw_proc_write() already do for their tagged and untagged paths, and terminate the connection with a base-or-bounds DDP error when the Read Response would overrun the sink buffer. This is the second receive-path length fix for this file. A separate change rejects an MPA FPDU length that underflows the per-fragment remainder in the header decode; that guard does not cover this case, because here each individual segment length is self-consistent and only the accumulated placement offset overruns the buffer.

In the Linux kernel, the following vulnerability has been resolved: fuse: avoid 32-bit prune notification count wrap FUSE_NOTIFY_PRUNE validates the nodeid payload length with: size - sizeof(outarg) != outarg.count * sizeof(u64) On 32-bit kernels, size_t is also 32 bits, so the daemon-controlled count multiplication can wrap. A prune notification with count 0x20000000 and no nodeid payload passes the check, enters the copy loop, and asks the device copy path to read nodeids that are not present in the userspace write buffer. In QEMU this reaches the fuse_copy_fill() BUG_ON(!err) path. Validate the payload length with array_size() instead. That accepts exactly the same valid messages, but avoids wrapping arithmetic before the copy loop consumes the count.

In the Linux kernel, the following vulnerability has been resolved: fuse: re-lock request before returning from fuse_ref_folio() fuse_ref_folio() unlocks the request but does not re-lock it before returning. fuse_chan_abort() can end the request and the async end callback (eg fuse_writepage_free()) can free the args while the subsequent copy chain logic after fuse_ref_folio() accesses them, leading to use-after-free issues. Fix this by locking the request in fuse_ref_folio() before returning.

In the Linux kernel, the following vulnerability has been resolved: fuse: clear intr_entry in fuse_resend and fuse_remove_pending_req When fuse_resend() moves a request from fpq->processing back to fiq->pending, it sets FR_PENDING and clears FR_SENT but does not remove the requests intr_entry from fiq->interrupts. If the request had FR_INTERRUPTED set from a prior signal, intr_entry remains dangling on fiq->interrupts. When the requesting task then receives a fatal signal, fuse_remove_pending_req() sees FR_PENDING=1, removes the request from fiq->pending and frees it via the refcount path, also without cleaning intr_entry. The stale intr_entry causes use-after-free when fuse_read_interrupt() iterates fiq->interrupts: - list_del_init(&req->intr_entry) -> UAF write on freed slab - req->in.h.unique -> UAF read, data leaked to userspace Remove intr_entry from fiq->interrupts in fuse_resend() for interrupted requests before they are placed back on fiq->pending. Add a WARN_ON if the intr_entry is not empty on request destruction.

In the Linux kernel, the following vulnerability has been resolved: fuse-uring: fix EFAULT clobber in fuse_uring_commit copy_from_user() returns the number of bytes not copied as an unsigned residual on failure (1..sizeof(struct fuse_out_header)). fuse_uring_commit stores that residual in ssize_t err, sets req->out.h.error to -EFAULT, then jumps to out: with err still holding the positive residual. err = copy_from_user(&req->out.h, &ent->headers->in_out, sizeof(req->out.h)); if (err) { req->out.h.error = -EFAULT; goto out; /* err is the positive residual */ } ... out: fuse_uring_req_end(ent, req, err); fuse_uring_req_end() then runs if (error) req->out.h.error = error; which overwrites the just-assigned -EFAULT with the positive residual. FUSE callers such as fuse_simple_request() test err < 0 to detect failure, so the positive value is interpreted as success and the caller proceeds with an uninitialised or partial req->out.args. Fix by assigning err = -EFAULT in the failure branch before jumping to out, so fuse_uring_req_end() receives a negative errno and sets req->out.h.error to -EFAULT.

In the Linux kernel, the following vulnerability has been resolved: fuse-uring: fix moving cancelled entry to ent_in_userspace list fuse_uring_cancel() moves entries that are available (these have no reqs attached) to the ent_in_userspace list. ent_list_request_expired() checks the first entry on ent_in_userspace and dereferences ent->fuse_req unconditionally, which will crash on a cancelled entry that was moved to this list. Fix this by freeing the entry and dropping queue_refs directly in fuse_uring_cancel(). This is safe because cancel is the cancel handler itself - after io_uring_cmd_done(), no more cancels will be dispatched for this command, and teardown serializes with cancel via queue->lock. Since cancel now decrements queue_refs, fuse_uring_abort() must no longer gate fuse_uring_abort_end_requests() on queue_refs > 0, as cancelled entries may have already dropped queue_refs while requests are still queued. Remove the gate so abort always flushes requests and stops queues.

In the Linux kernel, the following vulnerability has been resolved: fuse-uring: end fuse_req on io-uring cancel task work When io_uring delivers task work with tw.cancel set (PF_EXITING, PF_KTHREAD fallback, or percpu_ref_is_dying on the ring context), fuse_uring_send_in_task() takes the cancel branch, assigns -ECANCELED, and falls through to fuse_uring_send(). That path only flips the entry to FRRS_USERSPACE and completes the io_uring cmd; it never discharges the ring entry's owning reference to the fuse_req that fuse_uring_add_req_to_ring_ent() handed it at dispatch time. fuse_uring_send_in_task() tw.cancel == true err = -ECANCELED fuse_uring_send(ent, cmd, err, issue_flags) ent->state = FRRS_USERSPACE list_move(&ent->list, &queue->ent_in_userspace) ent->cmd = NULL io_uring_cmd_done(-ECANCELED) /* ent->fuse_req still set, req still hashed */ The fuse_req stays linked on fpq->processing[hash] and fuse_request_end() is never invoked. The originating syscall thread blocks in D-state in request_wait_answer() until fuse_abort_conn() runs, which can be the entire connection lifetime. For FR_BACKGROUND requests fc->num_background is never decremented either, so repeated cancels inflate the counter until max_background is hit and all later background ops stall. tw.cancel does not imply a connection abort (e.g. a single io_uring worker thread exits while the fuse connection stays up), so this cannot be left for fuse_abort_conn() to clean up. Ending the req but still routing the entry through fuse_uring_send() is not enough: that leaves a req-less entry on ent_in_userspace, and ent_list_request_expired() dereferences ent->fuse_req unconditionally on the head of that list, which would then NULL-deref. Fix the cancel branch to release the entry directly. Remove it from the queue, complete the io_uring cmd, end the fuse_req, free the entry, and drop its queue_refs (waking the teardown waiter if it was the last).