kernel/include, branch linux-6.18.y

io_uring: ensure ctx->rings is stable for task work flags manipulation

2026-03-19T15:08:51Z

Commit 96189080265e6bb5dde3a4afbaf947af493e3f82 upstream. If DEFER_TASKRUN | SETUP_TASKRUN is used and task work is added while the ring is being resized, it's possible for the OR'ing of IORING_SQ_TASKRUN to happen in the small window of swapping into the new rings and the old rings being freed. Prevent this by adding a 2nd ->rings pointer, ->rings_rcu, which is protected by RCU. The task work flags manipulation is inside RCU already, and if the resize ring freeing is done post an RCU synchronize, then there's no need to add locking to the fast path of task work additions. Note: this is only done for DEFER_TASKRUN, as that's the only setup mode that supports ring resizing. If this ever changes, then they too need to use the io_ctx_mark_taskrun() helper. Link: https://lore.kernel.org/io-uring/20260309062759.482210-1-naup96721@gmail.com/ Cc: stable@vger.kernel.org Fixes: 79cfe9e59c2a ("io_uring/register: add IORING_REGISTER_RESIZE_RINGS") Reported-by: Hao-Yu Yang Suggested-by: Pavel Begunkov Signed-off-by: Jens Axboe Signed-off-by: Greg Kroah-Hartman

mm: Fix a hmm_range_fault() livelock / starvation problem

2026-03-19T15:08:49Z

commit b570f37a2ce480be26c665345c5514686a8a0274 upstream. If hmm_range_fault() fails a folio_trylock() in do_swap_page, trying to acquire the lock of a device-private folio for migration, to ram, the function will spin until it succeeds grabbing the lock. However, if the process holding the lock is depending on a work item to be completed, which is scheduled on the same CPU as the spinning hmm_range_fault(), that work item might be starved and we end up in a livelock / starvation situation which is never resolved. This can happen, for example if the process holding the device-private folio lock is stuck in migrate_device_unmap()->lru_add_drain_all() sinc lru_add_drain_all() requires a short work-item to be run on all online cpus to complete. A prerequisite for this to happen is: a) Both zone device and system memory folios are considered in migrate_device_unmap(), so that there is a reason to call lru_add_drain_all() for a system memory folio while a folio lock is held on a zone device folio. b) The zone device folio has an initial mapcount > 1 which causes at least one migration PTE entry insertion to be deferred to try_to_migrate(), which can happen after the call to lru_add_drain_all(). c) No or voluntary only preemption. This all seems pretty unlikely to happen, but indeed is hit by the "xe_exec_system_allocator" igt test. Resolve this by waiting for the folio to be unlocked if the folio_trylock() fails in do_swap_page(). Rename migration_entry_wait_on_locked() to softleaf_entry_wait_unlock() and update its documentation to indicate the new use-case. Future code improvements might consider moving the lru_add_drain_all() call in migrate_device_unmap() to be called *after* all pages have migration entries inserted. That would eliminate also b) above. v2: - Instead of a cond_resched() in hmm_range_fault(), eliminate the problem by waiting for the folio to be unlocked in do_swap_page() (Alistair Popple, Andrew Morton) v3: - Add a stub migration_entry_wait_on_locked() for the !CONFIG_MIGRATION case. (Kernel Test Robot) v4: - Rename migrate_entry_wait_on_locked() to softleaf_entry_wait_on_locked() and update docs (Alistair Popple) v5: - Add a WARN_ON_ONCE() for the !CONFIG_MIGRATION version of softleaf_entry_wait_on_locked(). - Modify wording around function names in the commit message (Andrew Morton) Suggested-by: Alistair Popple Fixes: 1afaeb8293c9 ("mm/migrate: Trylock device page in do_swap_page") Cc: Ralph Campbell Cc: Christoph Hellwig Cc: Jason Gunthorpe Cc: Jason Gunthorpe Cc: Leon Romanovsky Cc: Andrew Morton Cc: Matthew Brost Cc: John Hubbard Cc: Alistair Popple Cc: linux-mm@kvack.org Cc: Signed-off-by: Thomas Hellström Cc: # v6.15+ Reviewed-by: John Hubbard #v3 Reviewed-by: Alistair Popple Link: https://patch.msgid.link/20260210115653.92413-1-thomas.hellstrom@linux.intel.com (cherry picked from commit a69d1ab971a624c6f112cea61536569d579c3215) Signed-off-by: Rodrigo Vivi Signed-off-by: Thomas Hellström Signed-off-by: Greg Kroah-Hartman

qmi_wwan: allow max_mtu above hard_mtu to control rx_urb_size

2026-03-19T15:08:42Z

commit 55f854dd5bdd8e19b936a00ef1f8d776ac32c7b0 upstream. Commit c7159e960f14 ("usbnet: limit max_mtu based on device's hard_mtu") capped net->max_mtu to the device's hard_mtu in usbnet_probe(). While this correctly prevents oversized packets on standard USB network devices, it breaks the qmi_wwan driver. qmi_wwan relies on userspace (e.g. ModemManager) setting a large MTU on the wwan0 interface to configure rx_urb_size via usbnet_change_mtu(). QMI modems negotiate USB transfer sizes of 16,383 or 32,767 bytes, and the USB receive buffers must be sized accordingly. With max_mtu capped to hard_mtu (~1500 bytes), userspace can no longer raise the MTU, the receive buffers remain small, and download speeds drop from >300 Mbps to ~0.8 Mbps. Introduce a FLAG_NOMAXMTU driver flag that allows individual usbnet drivers to opt out of the max_mtu cap. Set this flag in qmi_wwan's driver_info structures to restore the previous behavior for QMI devices, while keeping the safety fix in place for all other usbnet drivers. Fixes: c7159e960f14 ("usbnet: limit max_mtu based on device's hard_mtu") Cc: stable@vger.kernel.org Link: https://lore.kernel.org/lkml/CAPh3n803k8JcBPV5qEzUB-oKzWkAs-D5CU7z=Vd_nLRCr5ZqQg@mail.gmail.com/ Reported-by: Koen Vandeputte Tested-by: Daniele Palmas Signed-off-by: Laurent Vivier Link: https://patch.msgid.link/20260304134338.1785002-1-lvivier@redhat.com Signed-off-by: Jakub Kicinski Signed-off-by: Greg Kroah-Hartman

kunit: irq: Ensure timer doesn't fire too frequently

2026-03-19T15:08:35Z

commit 201ceb94aa1def0024a7c18ce643e5f65026be06 upstream. Fix a bug where kunit_run_irq_test() could hang if the system is too slow. This was noticed with the crypto library tests in certain VMs. Specifically, if kunit_irq_test_timer_func() and the associated hrtimer code took over 5us to run, then the CPU would spend all its time executing that code in hardirq context. As a result, the task executing kunit_run_irq_test() never had a chance to run, exit the loop, and cancel the timer. To fix it, make kunit_irq_test_timer_func() increase the timer interval when the other contexts aren't having a chance to run. Fixes: 950a81224e8b ("lib/crypto: tests: Add hash-test-template.h and gen-hash-testvecs.py") Cc: stable@vger.kernel.org Reviewed-by: David Gow Link: https://lore.kernel.org/r/20260224033751.97615-1-ebiggers@kernel.org Signed-off-by: Eric Biggers Signed-off-by: Greg Kroah-Hartman

irqchip/gic-v3-its: Limit number of per-device MSIs to the range the ITS supports

2026-03-19T15:08:34Z

commit ce9e40a9a5e5cff0b1b0d2fa582b3d71a8ce68e8 upstream. The ITS driver blindly assumes that EventIDs are in abundant supply, to the point where it never checks how many the hardware actually supports. It turns out that some pretty esoteric integrations make it so that only a few bits are available, all the way down to a single bit. Enforce the advertised limitation at the point of allocating the device structure, and hope that the endpoint driver can deal with such limitation. Fixes: 84a6a2e7fc18d ("irqchip: GICv3: ITS: device allocation and configuration") Signed-off-by: Marc Zyngier Signed-off-by: Thomas Gleixner Reviewed-by: Robin Murphy Reviewed-by: Zenghui Yu Cc: stable@vger.kernel.org Link: https://patch.msgid.link/20260206154816.3582887-1-maz@kernel.org Signed-off-by: Greg Kroah-Hartman

kthread: consolidate kthread exit paths to prevent use-after-free

2026-03-19T15:08:32Z

commit 28aaa9c39945b7925a1cc1d513c8f21ed38f5e4f upstream. Guillaume reported crashes via corrupted RCU callback function pointers during KUnit testing. The crash was traced back to the pidfs rhashtable conversion which replaced the 24-byte rb_node with an 8-byte rhash_head in struct pid, shrinking it from 160 to 144 bytes. struct kthread (without CONFIG_BLK_CGROUP) is also 144 bytes. With CONFIG_SLAB_MERGE_DEFAULT and SLAB_HWCACHE_ALIGN both round up to 192 bytes and share the same slab cache. struct pid.rcu.func and struct kthread.affinity_node both sit at offset 0x78. When a kthread exits via make_task_dead() it bypasses kthread_exit() and misses the affinity_node cleanup. free_kthread_struct() frees the memory while the node is still linked into the global kthread_affinity_list. A subsequent list_del() by another kthread writes through dangling list pointers into the freed and reused memory, corrupting the pid's rcu.func pointer. Instead of patching free_kthread_struct() to handle the missed cleanup, consolidate all kthread exit paths. Turn kthread_exit() into a macro that calls do_exit() and add kthread_do_exit() which is called from do_exit() for any task with PF_KTHREAD set. This guarantees that kthread-specific cleanup always happens regardless of the exit path - make_task_dead(), direct do_exit(), or kthread_exit(). Replace __to_kthread() with a new tsk_is_kthread() accessor in the public header. Export do_exit() since module code using the kthread_exit() macro now needs it directly. Reported-by: Guillaume Tucker Tested-by: Guillaume Tucker Tested-by: Mark Brown Tested-by: David Gow Cc: Link: https://lore.kernel.org/all/20260224-mittlerweile-besessen-2738831ae7f6@brauner Co-developed-by: Linus Torvalds Fixes: 4d13f4304fa4 ("kthread: Implement preferred affinity") Signed-off-by: Linus Torvalds Signed-off-by: Christian Brauner Signed-off-by: Greg Kroah-Hartman

Revert "ptdesc: remove references to folios from __pagetable_ctor() and pagetable_dtor()"

2026-03-19T15:08:31Z

commit 2d28ed588f8d7d0d41b0a4fad7f0d05e4bbf1797 upstream. This change swapped out mod_node_page_state for lruvec_stat_add_folio. But, these two APIs are not interchangeable: the lruvec version also increments memcg stats, in addition to "global" pgdat stats. So after this change, the "pagetables" memcg stat in memory.stat always yields "0", which is a userspace visible regression. I tried to look for a refactor where we add a variant of lruvec_stat_mod_folio which takes a pgdat and a memcg instead of a folio, to try to adhere to the spirit of the original patch. But at the end of the day this just means we have to call folio_memcg(ptdesc_folio(ptdesc)) anyway, which doesn't really accomplish much. This regression is visible in master as well as 6.18 stable, so CC stable too. Link: https://lkml.kernel.org/r/20260225002434.2953895-1-axelrasmussen@google.com Fixes: f0c92726e89f ("ptdesc: remove references to folios from __pagetable_ctor() and pagetable_dtor()") Signed-off-by: Axel Rasmussen Acked-by: Shakeel Butt Acked-by: Johannes Weiner Reviewed-by: Vishal Moola (Oracle) Cc: David Hildenbrand Cc: Liam Howlett Cc: Lorenzo Stoakes Cc: Matthew Wilcox (Oracle) Cc: Michal Hocko Cc: Mike Rapoport Cc: Suren Baghdasaryan Cc: Vlastimil Babka Cc: Roman Gushchin Cc: Muchun Song Cc: Signed-off-by: Andrew Morton Signed-off-by: Greg Kroah-Hartman

mmc: core: Avoid bitfield RMW for claim/retune flags

2026-03-19T15:08:29Z

commit 901084c51a0a8fb42a3f37d2e9c62083c495f824 upstream. Move claimed and retune control flags out of the bitfield word to avoid unrelated RMW side effects in asynchronous contexts. The host->claimed bit shared a word with retune flags. Writes to claimed in __mmc_claim_host() or retune_now in mmc_mq_queue_rq() can overwrite other bits when concurrent updates happen in other contexts, triggering spurious WARN_ON(!host->claimed). Convert claimed, can_retune, retune_now and retune_paused to bool to remove shared-word coupling. Fixes: 6c0cedd1ef952 ("mmc: core: Introduce host claiming by context") Fixes: 1e8e55b67030c ("mmc: block: Add CQE support") Cc: stable@vger.kernel.org Suggested-by: Adrian Hunter Signed-off-by: Penghe Geng Acked-by: Adrian Hunter Signed-off-by: Ulf Hansson Signed-off-by: Greg Kroah-Hartman

mm/tracing: rss_stat: ensure curr is false from kthread context

2026-03-19T15:08:28Z

commit 079c24d5690262e83ee476e2a548e416f3237511 upstream. The rss_stat trace event allows userspace tools, like Perfetto [1], to inspect per-process RSS metric changes over time. The curr field was introduced to rss_stat in commit e4dcad204d3a ("rss_stat: add support to detect RSS updates of external mm"). Its intent is to indicate whether the RSS update is for the mm_struct of the current execution context; and is set to false when operating on a remote mm_struct (e.g., via kswapd or a direct reclaimer). However, an issue arises when a kernel thread temporarily adopts a user process's mm_struct. Kernel threads do not have their own mm_struct and normally have current->mm set to NULL. To operate on user memory, they can "borrow" a memory context using kthread_use_mm(), which sets current->mm to the user process's mm. This can be observed, for example, in the USB Function Filesystem (FFS) driver. The ffs_user_copy_worker() handles AIO completions and uses kthread_use_mm() to copy data to a user-space buffer. If a page fault occurs during this copy, the fault handler executes in the kthread's context. At this point, current is the kthread, but current->mm points to the user process's mm. Since the rss_stat event (from the page fault) is for that same mm, the condition current->mm == mm becomes true, causing curr to be incorrectly set to true when the trace event is emitted. This is misleading because it suggests the mm belongs to the kthread, confusing userspace tools that track per-process RSS changes and corrupting their mm_id-to-process association. Fix this by ensuring curr is always false when the trace event is emitted from a kthread context by checking for the PF_KTHREAD flag. Link: https://lkml.kernel.org/r/20260219233708.1971199-1-kaleshsingh@google.com Link: https://perfetto.dev/ [1] Fixes: e4dcad204d3a ("rss_stat: add support to detect RSS updates of external mm") Signed-off-by: Kalesh Singh Acked-by: Zi Yan Acked-by: SeongJae Park Reviewed-by: Pedro Falcato Cc: "David Hildenbrand (Arm)" Cc: Joel Fernandes Cc: Lorenzo Stoakes Cc: Minchan Kim Cc: Steven Rostedt Cc: Suren Baghdasaryan Cc: [5.10+] Signed-off-by: Andrew Morton Signed-off-by: Greg Kroah-Hartman

USB: core: Limit the length of unkillable synchronous timeouts

2026-03-19T15:08:26Z

commit 1015c27a5e1a63efae2b18a9901494474b4d1dc3 upstream. The usb_control_msg(), usb_bulk_msg(), and usb_interrupt_msg() APIs in usbcore allow unlimited timeout durations. And since they use uninterruptible waits, this leaves open the possibility of hanging a task for an indefinitely long time, with no way to kill it short of unplugging the target device. To prevent this sort of problem, enforce a maximum limit on the length of these unkillable timeouts. The limit chosen here, somewhat arbitrarily, is 60 seconds. On many systems (although not all) this is short enough to avoid triggering the kernel's hung-task detector. In addition, clear up the ambiguity of negative timeout values by treating them the same as 0, i.e., using the maximum allowed timeout. Signed-off-by: Alan Stern Link: https://lore.kernel.org/linux-usb/3acfe838-6334-4f6d-be7c-4bb01704b33d@rowland.harvard.edu/ Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") CC: stable@vger.kernel.org Link: https://patch.msgid.link/15fc9773-a007-47b0-a703-df89a8cf83dd@rowland.harvard.edu Signed-off-by: Greg Kroah-Hartman