kernel/linux.git/drivers/iommu, branch v6.12.80

iommu/vt-d: Fix intel iommu iotlb sync hardlockup and retry

2026-03-25T10:08:49+00:00

commit fe89277c9ceb0d6af0aa665bcf24a41d8b1b79cd upstream. During the qi_check_fault process after an IOMMU ITE event, requests at odd-numbered positions in the queue are set to QI_ABORT, only satisfying single-request submissions. However, qi_submit_sync now supports multiple simultaneous submissions, and can't guarantee that the wait_desc will be at an odd-numbered position. Therefore, if an item times out, IOMMU can't re-initiate the request, resulting in an infinite polling wait. This modifies the process by setting the status of all requests already fetched by IOMMU and recorded as QI_IN_USE status (including wait_desc requests) to QI_ABORT, thus enabling multiple requests to be resubmitted. Fixes: 8a1d82462540 ("iommu/vt-d: Multiple descriptors per qi_submit_sync()") Cc: stable@vger.kernel.org Signed-off-by: Guanghui Feng Tested-by: Shuai Xue Reviewed-by: Shuai Xue Reviewed-by: Samiullah Khawaja Link: https://lore.kernel.org/r/20260306101516.3885775-1-guanghuifeng@linux.alibaba.com Signed-off-by: Lu Baolu Fixes: 8a1d82462540 ("iommu/vt-d: Multiple descriptors per qi_submit_sync()") Signed-off-by: Joerg Roedel Signed-off-by: Greg Kroah-Hartman

iommu/vt-d: Skip dev-iotlb flush for inaccessible PCIe device without scalable mode

2026-03-13T16:20:26+00:00

[ Upstream commit 42662d19839f34735b718129ea200e3734b07e50 ] PCIe endpoints with ATS enabled and passed through to userspace (e.g., QEMU, DPDK) can hard-lock the host when their link drops, either by surprise removal or by a link fault. Commit 4fc82cd907ac ("iommu/vt-d: Don't issue ATS Invalidation request when device is disconnected") adds pci_dev_is_disconnected() to devtlb_invalidation_with_pasid() so ATS invalidation is skipped only when the device is being safely removed, but it applies only when Intel IOMMU scalable mode is enabled. With scalable mode disabled or unsupported, a system hard-lock occurs when a PCIe endpoint's link drops because the Intel IOMMU waits indefinitely for an ATS invalidation that cannot complete. Call Trace: qi_submit_sync qi_flush_dev_iotlb __context_flush_dev_iotlb.part.0 domain_context_clear_one_cb pci_for_each_dma_alias device_block_translation blocking_domain_attach_dev iommu_deinit_device __iommu_group_remove_device iommu_release_device iommu_bus_notifier blocking_notifier_call_chain bus_notify device_del pci_remove_bus_device pci_stop_and_remove_bus_device pciehp_unconfigure_device pciehp_disable_slot pciehp_handle_presence_or_link_change pciehp_ist Commit 81e921fd3216 ("iommu/vt-d: Fix NULL domain on device release") adds intel_pasid_teardown_sm_context() to intel_iommu_release_device(), which calls qi_flush_dev_iotlb() and can also hard-lock the system when a PCIe endpoint's link drops. Call Trace: qi_submit_sync qi_flush_dev_iotlb __context_flush_dev_iotlb.part.0 intel_context_flush_no_pasid device_pasid_table_teardown pci_pasid_table_teardown pci_for_each_dma_alias intel_pasid_teardown_sm_context intel_iommu_release_device iommu_deinit_device __iommu_group_remove_device iommu_release_device iommu_bus_notifier blocking_notifier_call_chain bus_notify device_del pci_remove_bus_device pci_stop_and_remove_bus_device pciehp_unconfigure_device pciehp_disable_slot pciehp_handle_presence_or_link_change pciehp_ist Sometimes the endpoint loses connection without a link-down event (e.g., due to a link fault); killing the process (virsh destroy) then hard-locks the host. Call Trace: qi_submit_sync qi_flush_dev_iotlb __context_flush_dev_iotlb.part.0 domain_context_clear_one_cb pci_for_each_dma_alias device_block_translation blocking_domain_attach_dev __iommu_attach_device __iommu_device_set_domain __iommu_group_set_domain_internal iommu_detach_group vfio_iommu_type1_detach_group vfio_group_detach_container vfio_group_fops_release __fput pci_dev_is_disconnected() only covers safe-removal paths; pci_device_is_present() tests accessibility by reading vendor/device IDs and internally calls pci_dev_is_disconnected(). On a ConnectX-5 (8 GT/s, x2) this costs ~70 µs. Since __context_flush_dev_iotlb() is only called on {attach,release}_dev paths (not hot), add pci_device_is_present() there to skip inaccessible devices and avoid the hard-lock. Fixes: 37764b952e1b ("iommu/vt-d: Global devTLB flush when present context entry changed") Fixes: 81e921fd3216 ("iommu/vt-d: Fix NULL domain on device release") Cc: stable@vger.kernel.org Signed-off-by: Jinhui Guo Link: https://lore.kernel.org/r/20251211035946.2071-2-guojinhui.liam@bytedance.com Signed-off-by: Lu Baolu Signed-off-by: Joerg Roedel Signed-off-by: Sasha Levin

iommu/vt-d: Flush dev-IOTLB only when PCIe device is accessible in scalable mode

2026-03-04T12:21:47+00:00

[ Upstream commit 10e60d87813989e20eac1f3eda30b3bae461e7f9 ] Commit 4fc82cd907ac ("iommu/vt-d: Don't issue ATS Invalidation request when device is disconnected") relies on pci_dev_is_disconnected() to skip ATS invalidation for safely-removed devices, but it does not cover link-down caused by faults, which can still hard-lock the system. For example, if a VM fails to connect to the PCIe device, "virsh destroy" is executed to release resources and isolate the fault, but a hard-lockup occurs while releasing the group fd. Call Trace: qi_submit_sync qi_flush_dev_iotlb intel_pasid_tear_down_entry device_block_translation blocking_domain_attach_dev __iommu_attach_device __iommu_device_set_domain __iommu_group_set_domain_internal iommu_detach_group vfio_iommu_type1_detach_group vfio_group_detach_container vfio_group_fops_release __fput Although pci_device_is_present() is slower than pci_dev_is_disconnected(), it still takes only ~70 µs on a ConnectX-5 (8 GT/s, x2) and becomes even faster as PCIe speed and width increase. Besides, devtlb_invalidation_with_pasid() is called only in the paths below, which are far less frequent than memory map/unmap. 1. mm-struct release 2. {attach,release}_dev 3. set/remove PASID 4. dirty-tracking setup The gain in system stability far outweighs the negligible cost of using pci_device_is_present() instead of pci_dev_is_disconnected() to decide when to skip ATS invalidation, especially under GDR high-load conditions. Fixes: 4fc82cd907ac ("iommu/vt-d: Don't issue ATS Invalidation request when device is disconnected") Cc: stable@vger.kernel.org Signed-off-by: Jinhui Guo Link: https://lore.kernel.org/r/20251211035946.2071-3-guojinhui.liam@bytedance.com Signed-off-by: Lu Baolu Signed-off-by: Joerg Roedel Signed-off-by: Sasha Levin

iommu/amd: move wait_on_sem() out of spinlock

2026-03-04T12:21:12+00:00

[ Upstream commit d2a0cac10597068567d336e85fa3cbdbe8ca62bf ] With iommu.strict=1, the existing completion wait path can cause soft lockups under stressed environment, as wait_on_sem() busy-waits under the spinlock with interrupts disabled. Move the completion wait in iommu_completion_wait() out of the spinlock. wait_on_sem() only polls the hardware-updated cmd_sem and does not require iommu->lock, so holding the lock during the busy wait unnecessarily increases contention and extends the time with interrupts disabled. Signed-off-by: Ankit Soni Reviewed-by: Vasant Hegde Signed-off-by: Joerg Roedel Signed-off-by: Sasha Levin

iommu/arm-smmu-v3: Improve CMDQ lock fairness and efficiency

2026-03-04T12:21:12+00:00

[ Upstream commit df180b1a4cc51011c5f8c52c7ec02ad2e42962de ] The SMMU CMDQ lock is highly contentious when there are multiple CPUs issuing commands and the queue is nearly full. The lock has the following states: - 0: Unlocked - >0: Shared lock held with count - INT_MIN+N: Exclusive lock held, where N is the # of shared waiters - INT_MIN: Exclusive lock held, no shared waiters When multiple CPUs are polling for space in the queue, they attempt to grab the exclusive lock to update the cons pointer from the hardware. If they fail to get the lock, they will spin until either the cons pointer is updated by another CPU. The current code allows the possibility of shared lock starvation if there is a constant stream of CPUs trying to grab the exclusive lock. This leads to severe latency issues and soft lockups. Consider the following scenario where CPU1's attempt to acquire the shared lock is starved by CPU2 and CPU0 contending for the exclusive lock. CPU0 (exclusive) | CPU1 (shared) | CPU2 (exclusive) | `cmdq->lock` -------------------------------------------------------------------------- trylock() //takes | | | 0 | shared_lock() | | INT_MIN | fetch_inc() | | INT_MIN | no return | | INT_MIN + 1 | spins // VAL >= 0 | | INT_MIN + 1 unlock() | spins... | | INT_MIN + 1 set_release(0) | spins... | | 0 see[NOTE] (done) | (sees 0) | trylock() // takes | 0 | *exits loop* | cmpxchg(0, INT_MIN) | 0 | | *cuts in* | INT_MIN | cmpxchg(0, 1) | | INT_MIN | fails // != 0 | | INT_MIN | spins // VAL >= 0 | | INT_MIN | *starved* | | INT_MIN [NOTE] The current code resets the exclusive lock to 0 regardless of the state of the lock. This causes two problems: 1. It opens the possibility of back-to-back exclusive locks and the downstream effect of starving shared lock. 2. The count of shared lock waiters are lost. To mitigate this, we release the exclusive lock by only clearing the sign bit while retaining the shared lock waiter count as a way to avoid starving the shared lock waiters. Also deleted cmpxchg loop while trying to acquire the shared lock as it is not needed. The waiters can see the positive lock count and proceed immediately after the exclusive lock is released. Exclusive lock is not starved in that submitters will try exclusive lock first when new spaces become available. Reviewed-by: Mostafa Saleh Reviewed-by: Nicolin Chen Signed-off-by: Alexander Grest Signed-off-by: Jacob Pan Signed-off-by: Will Deacon Signed-off-by: Sasha Levin

iommu/vt-d: Clear Present bit before tearing down PASID entry

2026-03-04T12:20:04+00:00

[ Upstream commit 75ed00055c059dedc47b5daaaa2f8a7a019138ff ] The Intel VT-d Scalable Mode PASID table entry consists of 512 bits (64 bytes). When tearing down an entry, the current implementation zeros the entire 64-byte structure immediately using multiple 64-bit writes. Since the IOMMU hardware may fetch these 64 bytes using multiple internal transactions (e.g., four 128-bit bursts), updating or zeroing the entire entry while it is active (P=1) risks a "torn" read. If a hardware fetch occurs simultaneously with the CPU zeroing the entry, the hardware could observe an inconsistent state, leading to unpredictable behavior or spurious faults. Follow the "Guidance to Software for Invalidations" in the VT-d spec (Section 6.5.3.3) by implementing the recommended ownership handshake: 1. Clear only the 'Present' (P) bit of the PASID entry. 2. Use a dma_wmb() to ensure the cleared bit is visible to hardware before proceeding. 3. Execute the required invalidation sequence (PASID cache, IOTLB, and Device-TLB flush) to ensure the hardware has released all cached references. 4. Only after the flushes are complete, zero out the remaining fields of the PASID entry. Also, add a dma_wmb() in pasid_set_present() to ensure that all other fields of the PASID entry are visible to the hardware before the Present bit is set. Fixes: 0bbeb01a4faf ("iommu/vt-d: Manage scalalble mode PASID tables") Signed-off-by: Lu Baolu Reviewed-by: Dmytro Maluka Reviewed-by: Samiullah Khawaja Reviewed-by: Kevin Tian Link: https://lore.kernel.org/r/20260120061816.2132558-2-baolu.lu@linux.intel.com Signed-off-by: Joerg Roedel Signed-off-by: Sasha Levin

iommu/vt-d: Avoid draining PRQ in sva mm release path

2026-03-04T12:20:04+00:00

[ Upstream commit dda2b8c3c6ccc50deae65cc75f246577348e2ec5 ] When a PASID is used for SVA by a device, it's possible that the PASID entry is cleared before the device flushes all ongoing DMA requests and removes the SVA domain. This can occur when an exception happens and the process terminates before the device driver stops DMA and calls the iommu driver to unbind the PASID. There's no need to drain the PRQ in the mm release path. Instead, the PRQ will be drained in the SVA unbind path. Unfortunately, commit c43e1ccdebf2 ("iommu/vt-d: Drain PRQs when domain removed from RID") changed this behavior by unconditionally draining the PRQ in intel_pasid_tear_down_entry(). This can lead to a potential sleeping-in-atomic-context issue. Smatch static checker warning: drivers/iommu/intel/prq.c:95 intel_iommu_drain_pasid_prq() warn: sleeping in atomic context To avoid this issue, prevent draining the PRQ in the SVA mm release path and restore the previous behavior. Fixes: c43e1ccdebf2 ("iommu/vt-d: Drain PRQs when domain removed from RID") Reported-by: Dan Carpenter Closes: https://lore.kernel.org/linux-iommu/c5187676-2fa2-4e29-94e0-4a279dc88b49@stanley.mountain/ Signed-off-by: Lu Baolu Reviewed-by: Kevin Tian Link: https://lore.kernel.org/r/20241212021529.1104745-1-baolu.lu@linux.intel.com Signed-off-by: Joerg Roedel Stable-dep-of: 75ed00055c05 ("iommu/vt-d: Clear Present bit before tearing down PASID entry") Signed-off-by: Sasha Levin

iommu/vt-d: Drain PRQs when domain removed from RID

2026-03-04T12:20:03+00:00

[ Upstream commit c43e1ccdebf2c950545fdf12c5796ad6f7bad7ee ] As this iommu driver now supports page faults for requests without PASID, page requests should be drained when a domain is removed from the RID2PASID entry. This results in the intel_iommu_drain_pasid_prq() call being moved to intel_pasid_tear_down_entry(). This indicates that when a translation is removed from any PASID entry and the PRI has been enabled on the device, page requests are drained in the domain detachment path. The intel_iommu_drain_pasid_prq() helper has been modified to support sending device TLB invalidation requests for both PASID and non-PASID cases. Signed-off-by: Lu Baolu Reviewed-by: Yi Liu Link: https://lore.kernel.org/r/20241101045543.70086-1-baolu.lu@linux.intel.com Signed-off-by: Joerg Roedel Stable-dep-of: 75ed00055c05 ("iommu/vt-d: Clear Present bit before tearing down PASID entry") Signed-off-by: Sasha Levin

iommu/vt-d: Separate page request queue from SVM

2026-03-04T12:20:03+00:00

[ Upstream commit 4d5440957641fb5652cbef8df6183baf473cab6b ] IO page faults are no longer dependent on CONFIG_INTEL_IOMMU_SVM. Move all Page Request Queue (PRQ) functions that handle prq events to a new file in drivers/iommu/intel/prq.c. The page_req_des struct is now declared in drivers/iommu/intel/prq.c. No functional changes are intended. This is a preparation patch to enable the use of IO page faults outside the SVM/PASID use cases. Signed-off-by: Joel Granados Link: https://lore.kernel.org/r/20241015-jag-iopfv8-v4-1-b696ca89ba29@kernel.org Signed-off-by: Lu Baolu Signed-off-by: Joerg Roedel Stable-dep-of: 75ed00055c05 ("iommu/vt-d: Clear Present bit before tearing down PASID entry") Signed-off-by: Sasha Levin

iommu/vt-d: Flush cache for PASID table before using it

2026-03-04T12:20:03+00:00

[ Upstream commit 22d169bdd2849fe6bd18c2643742e1c02be6451c ] When writing the address of a freshly allocated zero-initialized PASID table to a PASID directory entry, do that after the CPU cache flush for this PASID table, not before it, to avoid the time window when this PASID table may be already used by non-coherent IOMMU hardware while its contents in RAM is still some random old data, not zero-initialized. Fixes: 194b3348bdbb ("iommu/vt-d: Fix PASID directory pointer coherency") Signed-off-by: Dmytro Maluka Reviewed-by: Kevin Tian Link: https://lore.kernel.org/r/20251221123508.37495-1-dmaluka@chromium.org Signed-off-by: Lu Baolu Signed-off-by: Joerg Roedel Signed-off-by: Sasha Levin