I/O virtual memory (IOMMU) support

CHANGELOG.md

@@ -7,6 +7,7 @@
 - \[[#311](https://github.com/rust-vmm/vm-memory/pull/311)\] Allow compiling without the ReadVolatile and WriteVolatile implementations
 - \[[#312](https://github.com/rust-vmm/vm-memory/pull/312)\] `GuestRegionContainer`, a generic container of `GuestMemoryRegion`s, generalizing `GuestMemoryMmap` (which 
   is now a type alias for `GuestRegionContainer<GuestRegionMmap>`).
+- \[[#327](https://github.com/rust-vmm/vm-memory/pull/327)\] I/O virtual memory support via `IoMemory`, `IommuMemory`, and `Iommu`/`Iotlb`
 
 ### Changed
 

Cargo.toml

@@ -16,13 +16,15 @@ default = ["rawfd"]
 backend-bitmap = ["dep:libc"]
 backend-mmap = ["dep:libc"]
 backend-atomic = ["arc-swap"]
+iommu = ["dep:rangemap"]
 rawfd = ["dep:libc"]
 xen = ["backend-mmap", "bitflags", "vmm-sys-util"]
 
 [dependencies]
 libc = { version = "0.2.39", optional = true }
 arc-swap = { version = "1.0.0", optional = true }
 bitflags = { version = "2.4.0", optional = true }
+rangemap = { version = "1.5.1", optional = true }
 thiserror = "1.0.40"
 vmm-sys-util = { version = ">=0.12.1,<=0.14.0", optional = true }
 

DESIGN.md

@@ -2,8 +2,8 @@
 
 ## Objectives
 
-- Provide a set of traits for accessing and configuring the physical memory of
-  a virtual machine.
+- Provide a set of traits for accessing and configuring the physical and/or
+  I/O virtual memory of a virtual machine.
 - Provide a clean abstraction of the VM memory such that rust-vmm components
   can use it without depending on the implementation details specific to
   different VMMs.
@@ -122,6 +122,29 @@ let buf = &mut [0u8; 5];
 let result = guest_memory_mmap.write(buf, addr);
 ```
 
+### I/O Virtual Address Space
+
+When using an IOMMU, there no longer is direct access to the guest (physical)
+address space, but instead only to I/O virtual address space.  In this case:
+
+- `IoMemory` replaces `GuestMemory`: It requires specifying the required access
+  permissions (which are relevant for virtual memory).  It also removes
+  interfaces that imply a mostly linear memory layout, because virtual memory is
+  fragmented into many pages instead of few (large) memory regions.
+  - Any `IoMemory` still has a `GuestMemory` inside as the underlying address
+    space, but if an IOMMU is used, that will generally not be guest physical
+    address space.  With vhost-user, for example, it will be the VMM’s user
+    address space instead.
+  - `IommuMemory` as our only actually IOMMU-supporting `IoMemory`
+    implementation uses an `Iommu` object to translate I/O virtual addresses
+    (IOVAs) into VMM user addresses (VUAs), which are then passed to the inner
+    `GuestMemory` implementation (like `GuestMemoryMmap`).
+- `GuestAddress` (for compatibility) refers to an address in any of these
+  address spaces:
+  - Guest physical addresses (GPAs) when no IOMMU is used,
+  - I/O virtual addresses (IOVAs),
+  - VMM user addresses (VUAs).
+
 ### Utilities and Helpers
 
 The following utilities and helper traits/macros are imported from the
@@ -143,7 +166,8 @@ with minor changes:
 - `Address` inherits `AddressValue`
 - `GuestMemoryRegion` inherits `Bytes<MemoryRegionAddress, E = Error>`. The
   `Bytes` trait must be implemented.
-- `GuestMemory` has a generic implementation of `Bytes<GuestAddress>`.
+- `GuestMemory` has a generic implementation of `IoMemory`
+- `IoMemory` has a generic implementation of `Bytes<GuestAddress>`.
 
 **Types**:
 

coverage_config_aarch64.json

@@ -1,5 +1,5 @@
 {
   "coverage_score": 85.2,
   "exclude_path": "mmap/windows.rs",
-  "crate_features": "backend-mmap,backend-atomic,backend-bitmap"
+  "crate_features": "backend-mmap,backend-atomic,backend-bitmap,iommu"
 }

coverage_config_x86_64.json

@@ -1,5 +1,5 @@
 {
-  "coverage_score": 91.78,
+  "coverage_score": 92.68,
   "exclude_path": "mmap_windows.rs",
-  "crate_features": "backend-mmap,backend-atomic,backend-bitmap"
+  "crate_features": "backend-mmap,backend-atomic,backend-bitmap,iommu"
 }

src/atomic.rs

@@ -2,7 +2,7 @@
 // Copyright (C) 2020 Red Hat, Inc. All rights reserved.
 // SPDX-License-Identifier: Apache-2.0
 
-//! A wrapper over an `ArcSwap<GuestMemory>` struct to support RCU-style mutability.
+//! A wrapper over an `ArcSwap<IoMemory>` struct to support RCU-style mutability.
 //!
 //! With the `backend-atomic` feature enabled, simply replacing `GuestMemoryMmap`
 //! with `GuestMemoryAtomic<GuestMemoryMmap>` will enable support for mutable memory maps.
@@ -15,17 +15,17 @@ use arc_swap::{ArcSwap, Guard};
 use std::ops::Deref;
 use std::sync::{Arc, LockResult, Mutex, MutexGuard, PoisonError};
 
-use crate::{GuestAddressSpace, GuestMemory};
+use crate::{GuestAddressSpace, IoMemory};
 
 /// A fast implementation of a mutable collection of memory regions.
 ///
 /// This implementation uses `ArcSwap` to provide RCU-like snapshotting of the memory map:
-/// every update of the memory map creates a completely new `GuestMemory` object, and
+/// every update of the memory map creates a completely new `IoMemory` object, and
 /// readers will not be blocked because the copies they retrieved will be collected once
 /// no one can access them anymore.  Under the assumption that updates to the memory map
 /// are rare, this allows a very efficient implementation of the `memory()` method.
 #[derive(Clone, Debug)]
-pub struct GuestMemoryAtomic<M: GuestMemory> {
+pub struct GuestMemoryAtomic<M: IoMemory> {
     // GuestAddressSpace<M>, which we want to implement, is basically a drop-in
     // replacement for &M.  Therefore, we need to pass to devices the `GuestMemoryAtomic`
     // rather than a reference to it.  To obtain this effect we wrap the actual fields
@@ -34,9 +34,9 @@ pub struct GuestMemoryAtomic<M: GuestMemory> {
     inner: Arc<(ArcSwap<M>, Mutex<()>)>,
 }
 
-impl<M: GuestMemory> From<Arc<M>> for GuestMemoryAtomic<M> {
+impl<M: IoMemory> From<Arc<M>> for GuestMemoryAtomic<M> {
     /// create a new `GuestMemoryAtomic` object whose initial contents come from
-    /// the `map` reference counted `GuestMemory`.
+    /// the `map` reference counted `IoMemory`.
     fn from(map: Arc<M>) -> Self {
         let inner = (ArcSwap::new(map), Mutex::new(()));
         GuestMemoryAtomic {
@@ -45,9 +45,9 @@ impl<M: GuestMemory> From<Arc<M>> for GuestMemoryAtomic<M> {
     }
 }
 
-impl<M: GuestMemory> GuestMemoryAtomic<M> {
+impl<M: IoMemory> GuestMemoryAtomic<M> {
     /// create a new `GuestMemoryAtomic` object whose initial contents come from
-    /// the `map` `GuestMemory`.
+    /// the `map` `IoMemory`.
     pub fn new(map: M) -> Self {
         Arc::new(map).into()
     }
@@ -75,7 +75,7 @@ impl<M: GuestMemory> GuestMemoryAtomic<M> {
     }
 }
 
-impl<M: GuestMemory> GuestAddressSpace for GuestMemoryAtomic<M> {
+impl<M: IoMemory> GuestAddressSpace for GuestMemoryAtomic<M> {
     type T = GuestMemoryLoadGuard<M>;
     type M = M;
 
@@ -86,14 +86,14 @@ impl<M: GuestMemory> GuestAddressSpace for GuestMemoryAtomic<M> {
 
 /// A guard that provides temporary access to a `GuestMemoryAtomic`.  This
 /// object is returned from the `memory()` method.  It dereference to
-/// a snapshot of the `GuestMemory`, so it can be used transparently to
+/// a snapshot of the `IoMemory`, so it can be used transparently to
 /// access memory.
 #[derive(Debug)]
-pub struct GuestMemoryLoadGuard<M: GuestMemory> {
+pub struct GuestMemoryLoadGuard<M: IoMemory> {
     guard: Guard<Arc<M>>,
 }
 
-impl<M: GuestMemory> GuestMemoryLoadGuard<M> {
+impl<M: IoMemory> GuestMemoryLoadGuard<M> {
     /// Make a clone of the held pointer and returns it.  This is more
     /// expensive than just using the snapshot, but it allows to hold on
     /// to the snapshot outside the scope of the guard.  It also allows
@@ -104,15 +104,15 @@ impl<M: GuestMemory> GuestMemoryLoadGuard<M> {
     }
 }
 
-impl<M: GuestMemory> Clone for GuestMemoryLoadGuard<M> {
+impl<M: IoMemory> Clone for GuestMemoryLoadGuard<M> {
     fn clone(&self) -> Self {
         GuestMemoryLoadGuard {
             guard: Guard::from_inner(Arc::clone(&*self.guard)),
         }
     }
 }
 
-impl<M: GuestMemory> Deref for GuestMemoryLoadGuard<M> {
+impl<M: IoMemory> Deref for GuestMemoryLoadGuard<M> {
     type Target = M;
 
     fn deref(&self) -> &Self::Target {
@@ -125,12 +125,12 @@ impl<M: GuestMemory> Deref for GuestMemoryLoadGuard<M> {
 /// possibly after updating the memory map represented by the
 /// `GuestMemoryAtomic` that created the guard.
 #[derive(Debug)]
-pub struct GuestMemoryExclusiveGuard<'a, M: GuestMemory> {
+pub struct GuestMemoryExclusiveGuard<'a, M: IoMemory> {
     parent: &'a GuestMemoryAtomic<M>,
     _guard: MutexGuard<'a, ()>,
 }
 
-impl<M: GuestMemory> GuestMemoryExclusiveGuard<'_, M> {
+impl<M: IoMemory> GuestMemoryExclusiveGuard<'_, M> {
     /// Replace the memory map in the `GuestMemoryAtomic` that created the guard
     /// with the new memory map, `map`.  The lock is then dropped since this
     /// method consumes the guard.
@@ -143,7 +143,7 @@ impl<M: GuestMemory> GuestMemoryExclusiveGuard<'_, M> {
 mod tests {
     use super::*;
     use crate::region::tests::{new_guest_memory_collection_from_regions, Collection, MockRegion};
-    use crate::{GuestAddress, GuestMemory, GuestMemoryRegion, GuestUsize};
+    use crate::{GuestAddress, GuestMemory, GuestMemoryRegion, GuestUsize, IoMemory};
 
     type GuestMemoryMmapAtomic = GuestMemoryAtomic<Collection>;
 
@@ -157,7 +157,8 @@ mod tests {
         let mut iterated_regions = Vec::new();
         let gmm = new_guest_memory_collection_from_regions(&regions).unwrap();
         let gm = GuestMemoryMmapAtomic::new(gmm);
-        let mem = gm.memory();
+        let vmem = gm.memory();
+        let mem = vmem.physical_memory().unwrap();
 
         for region in mem.iter() {
             assert_eq!(region.len(), region_size as GuestUsize);
@@ -176,7 +177,7 @@ mod tests {
             .map(|x| (x.0, x.1))
             .eq(iterated_regions.iter().copied()));
 
-        let mem2 = mem.into_inner();
+        let mem2 = vmem.into_inner();
         for region in mem2.iter() {
             assert_eq!(region.len(), region_size as GuestUsize);
         }