@samitouri / QOSamiQemu / commits / 295ca0e14c

hw/arm/tegra241-cmdqv: Document the CMDQV design and lifecycle

Add an overview describing the Tegra241 CMDQV passthrough model, MMIO layout, guest-driven lifecycle, and per-VM isolation. Tested-by: Nicolin Chen <nicolinc@nvidia.com> Signed-off-by: Shameer Kolothum <skolothumtho@nvidia.com> Tested-by: Eric Auger <eric.auger@redhat.com> Reviewed-by: Eric Auger <eric.auger@redhat.com> Reviewed-by: Nicolin Chen <nicolinc@nvidia.com> Message-id: 20260609112552.378999-31-skolothumtho@nvidia.com Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

Shameer Kolothum committed Jun 9, 2026 at 12:25 UTC 295ca0e14c6169e831c4a277e48df3a4aa0d8606
1 file changed +100
hw/arm/tegra241-cmdqv.c
+100
@@ -7,6 +7,106 @@
7 * SPDX-License-Identifier: GPL-2.0-or-later
8 */
9
10 +/*
11 + * Tegra241 CMDQV - overview
12 + * =========================
13 + *
14 + * NVIDIA Tegra241 extends SMMUv3 with a Command Queue Virtualization (CMDQ-V)
15 + * block. It lets a guest issue SMMU invalidation commands directly to
16 + * dedicated hardware queues (vCMDQs) without trapping into the hypervisor on
17 + * the fast path. vCMDQs are exclusively allocated to Virtual Interfaces
18 + * (VINTFs); the host kernel allocates one VINTF per emulated SMMUv3 instance
19 + * via iommufd. QEMU emulates the CMDQV MMIO region and drives the host kernel
20 + * calls (VIOMMU_ALLOC, HW_QUEUE_ALLOC, mmap); the actual command processing
21 + * happens on real hardware.
22 + *
23 + * A vCMDQ becomes functional only once allocated to the host VINTF; until then
24 + * no command processing happens, and trapped register accesses fall back to a
25 + * QEMU-side cache. After allocation, the cached register state is migrated to
26 + * the hardware and command processing runs on the host; guest accesses to the
27 + * live control/status registers then bypass QEMU and reach the host directly.
28 + *
29 + * MMIO layout (64KB pages, total TEGRA241_CMDQV_IO_LEN)
30 + * -----------------------------------------------------
31 + * 0x00000 CMDQV Config page: QEMU-trapped.
32 + * 0x10000 Direct vCMDQ Page 0 (control/status): QEMU-trapped and routed
33 + * to either the mmap'd host VINTF Page 0 (if the vCMDQ has been
34 + * allocated to a VINTF) or a per-vCMDQ register cache (otherwise).
35 + * 0x20000 Direct vCMDQ Page 1 (BASE / DRAM addresses): QEMU-trapped.
36 + * 0x30000 VINTF Page 0 (per-VINTF control/status): the guest's virtual
37 + * VINTF Page 0 aperture, backed by the host VINTF Page 0 (mmap'd
38 + * via iommufd) and installed into guest MMIO as a RAM-device
39 + * subregion when VINTF is enabled; subsequent accesses bypass QEMU.
40 + * 0x40000 VINTF Page 1 (per-VINTF BASE): QEMU-trapped. Although this is
41 + * a HW alias of the direct Page 1, the kernel only exposes mmap
42 + * for the host VINTF Page 0; the host VINTF Page 1 is not mmap'd
43 + * and stays trapped.
44 + *
45 + * The direct vCMDQ apertures (0x10000/0x20000) are HW aliases of the VINTF
46 + * apertures (0x30000/0x40000); they expose the same per-vCMDQ register slots
47 + * under different addressing.
48 + *
49 + * The direct vCMDQ Page 0 stays trapped rather than aliased to the host VINTF
50 + * Page 0 mmap. The CMDQV architecture allows software to program a vCMDQ
51 + * through the direct aperture before allocating it to a VINTF; aliasing to
52 + * the host VINTF Page 0 mmap would route those accesses into unallocated
53 + * logical slots where the hardware silently drops them, so trapping keeps
54 + * accesses well-defined for an unallocated vCMDQ.
55 + *
56 + * Lifecycle (driven by guest events)
57 + * ----------------------------------
58 + * 1. First vfio-pci device attach (.set_iommu_device) triggers:
59 + * - tegra241_cmdqv_probe(): IOMMU_GET_HW_INFO confirms host CMDQV support.
60 + * - IOMMU_VIOMMU_ALLOC: the kernel allocates and enables a VINTF for this
61 + * VM, configures the VM's VMID (from its stage-2 HWPT) in VINTF_CONFIG,
62 + * forces HYP_OWN=0, and returns the mmap offset/length for the host
63 + * VINTF Page 0, which QEMU then mmap()s.
64 + *
65 + * 2. Guest writes VINTF_CONFIG.ENABLE = 1:
66 + * QEMU installs the mmap'd host VINTF Page 0 into guest MMIO as the guest's
67 + * virtual VINTF Page 0 aperture (a RAM-device subregion) and reports
68 + * STATUS.ENABLE_OK = 1. The aperture is now a direct window onto the host
69 + * page, so accesses no longer trap into QEMU; a vCMDQ within it operates as
70 + * a real command queue only once it has been allocated (step 3).
71 + *
72 + * 3. Guest completes vCMDQ setup (BASE, CMDQ_ALLOC_MAP.ALLOC, CMDQV_EN,
73 + * VINTF.ENABLE, in any order; each precondition write retries the HW queue
74 + * allocation):
75 + * IOMMU_HW_QUEUE_ALLOC grants the guest a new host vCMDQ in this VM's
76 + * VINTF, binding the guest BASE GPA (translated through stage-2 and pinned
77 + * by the kernel) to it.
78 + *
79 + * 4. Guest SMMU driver programs a Stream Table Entry for a passthrough
80 + * device: IOMMU_VDEVICE_ALLOC programs SID_MATCH/SID_REPLACE in this VM's
81 + * VINTF so that the HW translates the device's guest vSID into its host
82 + * pSID. Commands referencing unmapped SIDs are rejected by HW.
83 + *
84 + * This reflects the current accel SMMUv3 design, which allocates the
85 + * vDEVICE when the guest programs the STE.
86 + *
87 + * Per-VM isolation
88 + * ----------------
89 + * - Each VM has its own iommufd FD; all iommufd objects (VINTF, vdevices,
90 + * hw_queues, mmap regions) belong to that FD. Cross-FD lookups fail, so
91 + * one VM cannot reach another VM's IDs.
92 + * - IOMMU_VIOMMU_ALLOC configures the VM's VMID in VINTF_CONFIG; the CMDQV
93 + * hardware substitutes / checks VMID on every command the guest issues.
94 + * - The kernel allocates the VINTF with HYP_OWN = 0, which restricts the
95 + * guest to a safe subset of commands.
96 + * - IOMMU_VDEVICE_ALLOC populates SID_MATCH/SID_REPLACE so invalidations
97 + * only reach the host StreamIDs assigned to this VM (see step 4).
98 + * - IOMMU_HW_QUEUE_ALLOC binds each vCMDQ to a single VINTF, so a guest
99 + * cannot reach a vCMDQ that belongs to another VM.
100 + *
101 + * Limits exposed to the guest
102 + * ---------------------------
103 + * One VINTF per emulated SMMUv3 and two vCMDQs per VINTF. The HW maximum
104 + * vCMDQ size is 8MiB, but the size QEMU exposes to the guest may be smaller.
105 + * The queue must be physically contiguous in host memory, so QEMU caps the
106 + * exposed size to the host memory-backend page size. Use hugepage backing to
107 + * reach the 8MiB maximum.
108 + */
109 +
110 #include "qemu/osdep.h"
111 #include "qemu/log.h"
112