[catpowder] Enhancement: add buffer pooling and multiple ring entries to XDP backend

Cargo.toml

@@ -36,7 +36,7 @@ x86 = "0.52.0"
 yaml-rust = "0.4.5"
 
 # Demikernel crates (published on crates.io).
-demikernel-dpdk-bindings = { version = "1.1.6", optional = true }
+demikernel-dpdk-bindings = { version = "1.1.7", optional = true }
 demikernel-network-simulator = { version = "0.1.0" }
 
 # Windows-specific dependencies.
@@ -53,7 +53,7 @@ windows = { version = "0.57.0", features = [
     "Win32_System_SystemInformation",
     "Win32_System_Threading",
 ] }
-demikernel-xdp-bindings = { version = "1.0.0", optional = true }
+demikernel-xdp-bindings = { version = "1.0.1", optional = true }
 # for interacting with socket2.
 windows-sys = { version = "0.52.0", features = ["Win32_Networking_WinSock"] }
 

scripts/config/azure.yaml

@@ -8,11 +8,21 @@ raw_socket:
   linux_interface_name: "abcde"
   xdp_interface_index: 0
   xdp_cohost_mode: false
+  xdp_always_poke_tx: false
   # Enable the following for XDP cohosting mode, or override in environment:
   # xdp_tcp_ports: [80, 443]
   # xdp_udp_ports: [53]
   # Enable the following line if you have a VF interface
   # xdp_vf_interface_index: 0
+  # The number of buffers to allocate for sending packets. Must be larger than the tx_ring_size.
+  tx_buffer_count: 4096
+  # The number of buffers to allocate for receiving packets for each RSS queue. Must be larger than
+  # the rx_ring_size.
+  rx_buffer_count: 4096
+  # The number of entries in the TX producer/consumer rings; must be a power of 2.
+  tx_ring_size: 128
+  # The number of entries in each RX producer/consumer ring for each RSS queue; must be a power of 2.
+  rx_ring_size: 128
 dpdk:
   eal_init: ["-c", "0xff", "-n", "4", "-a", "WW:WW.W", "--proc-type=auto", "--vdev=net_vdev_netvsc0,iface=abcde"]
 tcp_socket_options:

scripts/config/default.yaml

@@ -8,11 +8,21 @@ raw_socket:
   linux_interface_name: "abcde"
   xdp_interface_index: 0
   xdp_cohost_mode: false
+  xdp_always_poke_tx: false
   # Enable the following for XDP cohosting mode, or override in environment:
   # xdp_tcp_ports: [80, 443]
   # xdp_udp_ports: [53]
   # Enable the following line if you have a VF interface
   # xdp_vf_interface_index: 0
+  # The number of buffers to allocate for sending packets. Must be larger than the tx_ring_size.
+  tx_buffer_count: 4096
+  # The number of buffers to allocate for receiving packets for each RSS queue. Must be larger than
+  # the rx_ring_size.
+  rx_buffer_count: 4096
+  # The number of entries in the TX producer/consumer rings; must be a power of 2.
+  tx_ring_size: 128
+  # The number of entries in each RX producer/consumer ring for each RSS queue; must be a power of 2.
+  rx_ring_size: 128
 dpdk:
   eal_init: ["", "-c", "0xff", "-n", "4", "-a", "WW:WW.W","--proc-type=auto"]
 tcp_socket_options:

src/rust/catpowder/win/api.rs

@@ -66,3 +66,6 @@ impl Drop for XdpApi {
         }
     }
 }
+
+unsafe impl Send for XdpApi {}
+unsafe impl Sync for XdpApi {}

src/rust/catpowder/win/ring/generic.rs

@@ -1,6 +1,8 @@
 // Copyright (c) Microsoft Corporation.
 // Licensed under the MIT license.
 
+use std::{marker::PhantomData, mem::MaybeUninit};
+
 use crate::runtime::libxdp;
 
 //======================================================================================================================
@@ -9,25 +11,31 @@ use crate::runtime::libxdp;
 
 /// A wrapper structure for a XDP ring.
 #[repr(C)]
-pub struct XdpRing(libxdp::XSK_RING);
+pub struct XdpRing<T>(libxdp::XSK_RING, PhantomData<T>);
 
 //======================================================================================================================
 // Implementations
 //======================================================================================================================
 
-impl XdpRing {
+impl<T> XdpRing<T> {
     /// Initializes a XDP ring.
     pub(super) fn new(info: &libxdp::XSK_RING_INFO) -> Self {
-        Self(unsafe {
+        let ring: libxdp::XSK_RING = unsafe {
             let mut ring: libxdp::XSK_RING = std::mem::zeroed();
             libxdp::_XskRingInitialize(&mut ring, info);
             ring
-        })
+        };
+
+        if !ring.SharedElements.cast::<T>().is_aligned() {
+            panic!("XdpRing::new(): ring memory is not aligned for type T");
+        }
+
+        Self(ring, PhantomData)
     }
 
     /// Reserves a consumer slot in the target ring.
-    pub(super) fn consumer_reserve(&mut self, count: u32, idx: *mut u32) -> u32 {
-        unsafe { libxdp::_XskRingConsumerReserve(&mut self.0, count, idx) }
+    pub(super) fn consumer_reserve(&mut self, count: u32, idx: &mut u32) -> u32 {
+        unsafe { libxdp::_XskRingConsumerReserve(&mut self.0, count, idx as *mut u32) }
     }
 
     /// Releases a consumer slot in the target ring.
@@ -36,8 +44,8 @@ impl XdpRing {
     }
 
     /// Reserves a producer slot in the target ring.
-    pub(super) fn producer_reserve(&mut self, count: u32, idx: *mut u32) -> u32 {
-        unsafe { libxdp::_XskRingProducerReserve(&mut self.0, count, idx) }
+    pub(super) fn producer_reserve(&mut self, count: u32, idx: &mut u32) -> u32 {
+        unsafe { libxdp::_XskRingProducerReserve(&mut self.0, count, idx as *mut u32) }
     }
 
     /// Submits a producer slot in the target ring.
@@ -46,7 +54,18 @@ impl XdpRing {
     }
 
     /// Gets the element at the target index.
-    pub(super) fn get_element(&self, idx: u32) -> *mut std::ffi::c_void {
-        unsafe { libxdp::_XskRingGetElement(&self.0, idx) }
+    pub(super) fn get_element(&self, idx: u32) -> &mut MaybeUninit<T> {
+        // Safety: the alignment of ring elements is validated by the constructor. We rely on the XDP runtime to
+        // provide valid memory for the ring.
+        unsafe { &mut *libxdp::_XskRingGetElement(&self.0, idx).cast() }
+    }
+
+    #[allow(dead_code)]
+    pub(super) fn needs_poke(&self) -> bool {
+        unsafe { libxdp::_XskRingProducerNeedPoke(&self.0) != 0 }
+    }
+
+    pub(super) fn has_error(&self) -> bool {
+        unsafe { libxdp::_XskRingError(&self.0) != 0 }
     }
 }

src/rust/catpowder/win/ring/mod.rs

@@ -5,7 +5,6 @@
 // Modules
 //======================================================================================================================
 
-mod buffer;
 mod generic;
 mod rule;
 mod rx_ring;
@@ -16,6 +15,5 @@ mod umemreg;
 // Exports
 //======================================================================================================================
 
-pub use buffer::XdpBuffer;
 pub use rx_ring::RxRing;
 pub use tx_ring::TxRing;