Applied patches to drm-misc-next On 10/19/26 09:47, David Zhang wrote:
This patch series extends the amdxdna accelerator driver to support AMD AIE4 (NPU3) platforms, including classic and PF/VF for SR-IOV with kernel-mode command submission (KMQ), system power management, and runtime power management for the classic device.Changes in v5: - Rebase on the uninterruptible HMM invalidate wait fix, and move aie2_hmm_invalidate() into common GEM code for AIE2 and AIE4. - Rebase on the umap range fault serialization fix, and move aie2_populate_range() into common GEM code for AIE2 and AIE4. - Move runtime PM reference calls to the last patch, so each patch is bisectable on its own. Changes in v4: - Return -ENODEV instead of -EAGAIN from aie4_cmd_wait() when the context is in error, so waiters fail after a failed resume instead of retrying forever. - Take a runtime PM reference in aie4_hwctx_init() so context creation does not send mailbox messages to a suspended device. - Make the HMM invalidate wait uninterruptible, so a pending signal does not let pages be freed while the NPU may still access them. - Record whether a command wait ended on completion or disconnect, so a context unlinked and relinked during the wait cannot report a false completion, and a torn read_index cannot stop the worker. - Resume the device in the wait ioctl when the context is suspended by power management instead of returning -EAGAIN, and return 0 for commands completed before suspend. - Add a job timeout that stops the hardware context when a job does not complete, so job fences published to dma_resv are always signaled in finite time. Changes in v3 (addresses v2 review comments): - Lock, check and fence chained sub-command BOs with the job BOs, since the NPU writes completion to command BO headers. - Use the validated payload pointer in submit_one_cmd() to fix a TOCTOU on the user-writable command header. - Drop the last_read_index cache; get_read_index() returns 0 when read_index is still invalid after a re-sample, so callers treat the command as incomplete and keep waiting. - Align the runtime config message buffer to 4 bytes. - Move the HMM invalidate wait and populate_range into common GEM code, and repopulate invalidated user mappings in AIE4 command submission instead of failing the submit. - Drop the patch that finalized runtime PM before taking dev_lock on removal; the PCI core keeps the device active during remove, so there is no deadlock. - Use a stateless PF/VF device link, folded into the suspend and resume patch, so it survives VF driver rebind such as vfio-pci. - Return -EBUSY from runtime suspend when a device has no runtime_suspend callback, so the PM core keeps it active. Changes in v2: - Drop the redundant cert_feature_table configuration. - Call aie4_partition_fini() when aie4_restore_power_mode() fails, so the partition is not leaked on hardware start failure. - Return -ETIME from best-effort firmware configuration steps, so a mailbox timeout fails the probe instead of letting it succeed with unresponsive firmware. - Fix a synchronize_srcu() deadlock in context teardown by stopping the hardware context and waking all waiters before destroying it. - Use READ_ONCE() for read_index and for fields read from user-shared command buffers, so checked values cannot change after validation. - Check that cert_comp is still linked in all command wait paths, so -EAGAIN is only returned on a real disconnect. - Check that trace buffer addresses fit within 48 bits, as required by the current CERT design, instead of silently truncating them. - Hold references to chained sub-command BOs until job release, so userspace cannot free a BO while hardware may still access it. - Document that job fences are signaled when submission fails or the hardware context is stopped, destroyed or suspended, and that job timeout detection and recovery (TDR) will follow in a separate change to guarantee the fence signaling in finite time. - Address use-after-free comments on fence dev_name(). - Abort hardware contexts when resume fails to start the hardware, so submitters waiting for a context to reconnect fail instead of blocking forever. - Document that VF driver rebind requires re-enabling SR-IOV via sriov_numvfs, which recreates the PF/VF PM device links. - Add pm_ptr() wrapper to fix the build with CONFIG_PM=n. - Remove hwctx_config for AIE4; use-after-free cases in the debug buffer handling need more work. - Detect VF passthrough by comparing against the PF driver instead of the module owner, which is NULL for built-in drivers. Changes in v1: - Firmware version 6.0: return IOMMU_PASID_INVALID from aie4_msg_pasid(). - Power mode: consolidate cached power mode override restoration on hardware start directly into the power mode introduction patch, and respect user buffer size. - Transport hooks: fix header include to use <drm/amdxdna_accel.h>. - Command submission & fencing: fold fence timeline naming and unique timeline context allocation into the command submission patch. Series structure: The patch layout is the same as v0 except for the v3 and v4 changes above; see the v0 cover letter for a per-patch overview. Testing: - Tested on AMD AIE4/NPU3 hardware in both classic and SR-IOV (PF/VF) modes. - Verified kernel-mode command submission with direct and indirect execution packets under concurrent workloads. - Verified system suspend/resume (S2idle/S3) during idle and active command submission, and runtime autosuspend on the classic device. - Verified SR-IOV VF binding and execution. - Confirmed no regression on existing AIE2 devices (NPU1/NPU4). David Zhang (19): accel/amdxdna: Rename NPU3 firmware files accel/amdxdna: Remove mmap for doorbell accel/amdxdna: Add CERT firmware version support accel/amdxdna: Support firmware 6.0 accel/amdxdna: Add NPU3 classic device support accel/amdxdna: Add AIE version query to aie4_get_info accel/amdxdna: Add get and set power_mode for AIE4 accel/amdxdna: Add clock, DPM frequency, and resource info queries for AIE4 accel/amdxdna: Add context switch hysteresis with debugfs control accel/amdxdna: Refactor AIE4 hardware initialization sequence accel/amdxdna: Decouple AIE4 doorbell and MSI-X notify transport hooks accel/amdxdna: Implement AIE4 kernel queue lifecycle and memory layout accel/amdxdna: Prepare for AIE4 command submission accel/amdxdna: Move HMM invalidate wait into common GEM code accel/amdxdna: Make populate_range common for AIE2 and AIE4 accel/amdxdna: Implement AIE4 command packet building and submission accel/amdxdna: Enable AIE4 firmware logging to DRAM accel/amdxdna: Implement AIE4 suspend and resume accel/amdxdna: Implement runtime suspend and resume support drivers/accel/amdxdna/aie.c | 33 +- drivers/accel/amdxdna/aie.h | 47 +- drivers/accel/amdxdna/aie2_ctx.c | 91 +- drivers/accel/amdxdna/aie2_message.c | 4 +- drivers/accel/amdxdna/aie2_pci.c | 67 +- drivers/accel/amdxdna/aie2_pci.h | 40 +- drivers/accel/amdxdna/aie2_pm.c | 10 +- drivers/accel/amdxdna/aie4_ctx.c | 1129 +++++++++++++++++++++-- drivers/accel/amdxdna/aie4_host_queue.h | 72 +- drivers/accel/amdxdna/aie4_message.c | 245 +++++ drivers/accel/amdxdna/aie4_msg_priv.h | 149 ++- drivers/accel/amdxdna/aie4_pci.c | 818 +++++++++++++++- drivers/accel/amdxdna/aie4_pci.h | 117 ++- drivers/accel/amdxdna/aie4_sriov.c | 66 +- drivers/accel/amdxdna/amdxdna_ctx.c | 72 +- drivers/accel/amdxdna/amdxdna_ctx.h | 28 +- drivers/accel/amdxdna/amdxdna_debugfs.c | 3 + drivers/accel/amdxdna/amdxdna_gem.c | 87 +- drivers/accel/amdxdna/amdxdna_gem.h | 1 + drivers/accel/amdxdna/amdxdna_pci_drv.c | 46 +- drivers/accel/amdxdna/amdxdna_pci_drv.h | 20 +- drivers/accel/amdxdna/amdxdna_pm.c | 26 + drivers/accel/amdxdna/amdxdna_pm.h | 2 + drivers/accel/amdxdna/amdxdna_sysfs.c | 2 +- drivers/accel/amdxdna/npu1_regs.c | 19 +- drivers/accel/amdxdna/npu3_regs.c | 105 ++- drivers/accel/amdxdna/npu4_regs.c | 30 +- 27 files changed, 2922 insertions(+), 407 deletions(-)
