add function to VmResources for allocating I/O memory

If swiotlb is requested through the API, this function will allocate a
single contiguous region intended for swiotlb use. It needs to be
continguous because we must be able to describe it using a single memory
range in FDT, as devices cannot be assigned to reserved memory
consisting of multiple regions (e.g. FDT assignment is to memory
regions, not to #reserved-memory nodes).

While we're at it, always use anon memory for the "normal" part of guest
memory if swiotlb is enabled, as if swiotlb is enabled we know that
vhost devices will never need to access that part of memory.

Signed-off-by: Patrick Roy <[email protected]>

Author: roypat

src/vmm/src/arch/aarch64/mod.rs

@@ -51,6 +51,14 @@ pub fn arch_memory_regions(offset: usize, size: usize) -> Vec<(GuestAddress, usi
     )]
 }
 
+/// How many bytes of physical guest memory are addressible before the final gap in
+/// the address space on this architecture.
+///
+/// There are no architectural gaps in the physical address space on aarch64, so this is 0
+pub fn bytes_before_last_gap() -> usize {
+    0
+}
+
 /// Configures the system and should be called once per vm before starting vcpu threads.
 /// For aarch64, we only setup the FDT.
 ///

src/vmm/src/arch/mod.rs

@@ -14,9 +14,9 @@ pub mod aarch64;
 
 #[cfg(target_arch = "aarch64")]
 pub use aarch64::{
-    ConfigurationError, MMIO_MEM_SIZE, MMIO_MEM_START, arch_memory_regions, configure_system,
-    get_kernel_start, initrd_load_addr, layout::CMDLINE_MAX_SIZE, layout::IRQ_BASE,
-    layout::IRQ_MAX, layout::SYSTEM_MEM_SIZE, layout::SYSTEM_MEM_START,
+    ConfigurationError, MMIO_MEM_SIZE, MMIO_MEM_START, arch_memory_regions, bytes_before_last_gap,
+    configure_system, get_kernel_start, initrd_load_addr, layout::CMDLINE_MAX_SIZE,
+    layout::IRQ_BASE, layout::IRQ_MAX, layout::SYSTEM_MEM_SIZE, layout::SYSTEM_MEM_START,
 };
 
 /// Module for x86_64 related functionality.
@@ -25,10 +25,10 @@ pub mod x86_64;
 
 #[cfg(target_arch = "x86_64")]
 pub use crate::arch::x86_64::{
-    ConfigurationError, MMIO_MEM_SIZE, MMIO_MEM_START, arch_memory_regions, configure_system,
-    get_kernel_start, initrd_load_addr, layout::APIC_ADDR, layout::CMDLINE_MAX_SIZE,
-    layout::IOAPIC_ADDR, layout::IRQ_BASE, layout::IRQ_MAX, layout::SYSTEM_MEM_SIZE,
-    layout::SYSTEM_MEM_START,
+    ConfigurationError, MMIO_MEM_SIZE, MMIO_MEM_START, arch_memory_regions, bytes_before_last_gap,
+    configure_system, get_kernel_start, initrd_load_addr, layout::APIC_ADDR,
+    layout::CMDLINE_MAX_SIZE, layout::IOAPIC_ADDR, layout::IRQ_BASE, layout::IRQ_MAX,
+    layout::SYSTEM_MEM_SIZE, layout::SYSTEM_MEM_START,
 };
 
 /// Types of devices that can get attached to this platform.

src/vmm/src/arch/x86_64/mod.rs

@@ -97,6 +97,14 @@ pub fn arch_memory_regions(offset: usize, size: usize) -> Vec<(GuestAddress, usi
     }
 }
 
+/// How many bytes of physical guest memory are addressible before the final gap in
+/// the address space on this architecture.
+///
+/// On x86_64, this is the number of bytes that fit before the MMIO gap.
+pub fn bytes_before_last_gap() -> usize {
+    u64_to_usize(MMIO_MEM_START)
+}
+
 /// Returns the memory address where the kernel could be loaded.
 pub fn get_kernel_start() -> u64 {
     layout::HIMEM_START

src/vmm/src/resources.rs

@@ -437,43 +437,92 @@ impl VmResources {
         Ok(())
     }
 
+    /// Returns true if any vhost user devices are configured int his [`VmResources`] object
+    pub fn vhost_user_devices_used(&self) -> bool {
+        self.block
+            .devices
+            .iter()
+            .any(|b| b.lock().expect("Poisoned lock").is_vhost_user())
+    }
+
+    /// The size of the swiotlb region requested, in MiB
+    #[cfg(target_arch = "aarch64")]
+    pub fn swiotlb_size_mib(&self) -> usize {
+        self.machine_config.mem_config.initial_swiotlb_size
+    }
+
+    /// The size of the swiotlb region requested, in MiB
+    #[cfg(target_arch = "x86_64")]
+    pub fn swiotlb_size_mib(&self) -> usize {
+        0
+    }
+
+    /// Whether the use of swiotlb was requested
+    pub fn swiotlb_used(&self) -> bool {
+        self.swiotlb_size_mib() > 0
+    }
+
+    fn allocate_memory(
+        &self,
+        offset: usize,
+        size: usize,
+        vhost_accessible: bool,
+    ) -> Result<Vec<GuestRegionMmap>, MemoryError> {
+        let regions = crate::arch::arch_memory_regions(offset, size);
+        if vhost_accessible {
+            memory::memfd_backed(
+                regions.as_ref(),
+                self.machine_config.track_dirty_pages,
+                self.machine_config.huge_pages,
+            )
+        } else {
+            memory::anonymous(
+                regions.into_iter(),
+                self.machine_config.track_dirty_pages,
+                self.machine_config.huge_pages,
+            )
+        }
+    }
+
     /// Allocates guest memory in a configuration most appropriate for these [`VmResources`].
     ///
     /// If vhost-user-blk devices are in use, allocates memfd-backed shared memory, otherwise
     /// prefers anonymous memory for performance reasons.
     pub fn allocate_guest_memory(&self) -> Result<Vec<GuestRegionMmap>, MemoryError> {
-        let vhost_user_device_used = self
-            .block
-            .devices
-            .iter()
-            .any(|b| b.lock().expect("Poisoned lock").is_vhost_user());
-
         // Page faults are more expensive for shared memory mapping, including  memfd.
         // For this reason, we only back guest memory with a memfd
         // if a vhost-user-blk device is configured in the VM, otherwise we fall back to
         // an anonymous private memory.
         //
+        // Note that if a swiotlb region is used, no I/O will go through the "regular"
+        // memory regions, and we can back them with anon memory regardless.
+        //
         // The vhost-user-blk branch is not currently covered by integration tests in Rust,
         // because that would require running a backend process. If in the future we converge to
         // a single way of backing guest memory for vhost-user and non-vhost-user cases,
         // that would not be worth the effort.
-        let regions = crate::arch::arch_memory_regions(
+        self.allocate_memory(
             0,
-            self.machine_config.mem_size_mib << MIB_TO_BYTES_SHIFT,
-        );
-        if vhost_user_device_used {
-            memory::memfd_backed(
-                regions.as_ref(),
-                self.machine_config.track_dirty_pages,
-                self.machine_config.huge_pages,
-            )
-        } else {
-            memory::anonymous(
-                regions.into_iter(),
-                self.machine_config.track_dirty_pages,
-                self.machine_config.huge_pages,
-            )
+            (self.machine_config.mem_size_mib - self.swiotlb_size_mib()) << MIB_TO_BYTES_SHIFT,
+            self.vhost_user_devices_used() && !self.swiotlb_used(),
+        )
+    }
+
+    /// Allocates the dedicated I/O region for swiotlb use, if one was requested.
+    pub fn allocate_io_memory(&self) -> Result<Option<GuestRegionMmap>, MemoryError> {
+        if !self.swiotlb_used() {
+            return Ok(None);
         }
+
+        let swiotlb_size = self.swiotlb_size_mib() << MIB_TO_BYTES_SHIFT;
+        let start = (self.machine_config.mem_size_mib << MIB_TO_BYTES_SHIFT) - swiotlb_size;
+        let start = start.max(crate::arch::bytes_before_last_gap());
+
+        let mut mem = self.allocate_memory(start, swiotlb_size, self.vhost_user_devices_used())?;
+
+        assert_eq!(mem.len(), 1);
+
+        Ok(Some(mem.remove(0)))
     }
 }