use buddy allocator for tx queue

Author: braindigitalis

include/buddy_allocator.h

@@ -103,6 +103,10 @@ void buddy_init(buddy_allocator_t* alloc, int min_order, int max_order, int grow
  * @param size   Requested size in bytes.
  * @return Pointer to allocated memory, or NULL if out of memory.
  *
+ * @note Alignment of the returned pointer is at least the largest power-of-two
+ *       that divides sizeof(buddy_header_t); effectively 8 bytes with the current
+ *       header layout. The underlying block base is aligned to its block size.
+ *
  * If the active region cannot satisfy the request, a new region is
  * automatically allocated of size (1 << grow_order). Returned blocks
  * are aligned to their block size (power of two).
@@ -138,6 +142,10 @@ void buddy_destroy(buddy_allocator_t *alloc);
  * @param s      NUL-terminated string to duplicate.
  * @return Newly allocated string copy, or NULL on OOM.
  *
+ * @note Alignment of the returned pointer is at least the largest power-of-two
+ *       that divides sizeof(buddy_header_t); effectively 8 bytes with the current
+ *       header layout. The underlying block base is aligned to its block size.
+ *
  * The returned string is allocated from the buddy allocator's
  * private heap and must be freed with @ref buddy_free.
  */
@@ -149,11 +157,39 @@ char* buddy_strdup(buddy_allocator_t* alloc, const char *s);
  * Works like krealloc(): if ptr is NULL, behaves like buddy_malloc().
  * If size == 0, frees the block and returns NULL.
  *
+ * @note Alignment of the returned pointer is at least the largest power-of-two
+ *       that divides sizeof(buddy_header_t); effectively 8 bytes with the current
+ *       header layout. The underlying block base is aligned to its block size.
+ *
  * @param alloc  Buddy allocator to allocate from.
  * @param ptr    Pointer previously returned by buddy_malloc().
  * @param size   New desired size in bytes.
  * @return       Pointer to resized block, or NULL on OOM.
  */
 void* buddy_realloc(buddy_allocator_t* alloc, void *ptr, size_t size);
 
+/**
+ * @brief Allocate zero-initialised memory from the buddy allocator.
+ *
+ * Behaves like standard calloc(): allocates space for @p num elements of
+ * @p size bytes each, rounds up to the allocator’s block size, and fills
+ * the returned block with zeroes.
+ *
+ * @param alloc Buddy allocator to allocate from.
+ * @param num   Number of elements to allocate.
+ * @param size  Size of each element in bytes.
+ * @return      Pointer to zero-initialised memory, or NULL on OOM.
+ *
+ * @note Alignment of the returned pointer is at least the largest power-of-two
+ *       that divides sizeof(buddy_header_t); effectively 8 bytes with the current
+ *       header layout. The underlying block base is aligned to its block size.
+ *
+ * @see buddy_malloc(), buddy_free(), buddy_realloc(), buddy_destroy()
+ */
 void* buddy_calloc(buddy_allocator_t* alloc, size_t num, size_t size);
+
+/**
+ * Ensure that if we make changes here, pointers returned by the buddy allocator
+ * are still aligned to at least 8 bytes, as some systems assume this.
+ */
+_Static_assert((sizeof(buddy_header_t) % 8) == 0, "Alignment too small");

modules/virtio_net/virtio_net.c

@@ -1,6 +1,8 @@
 #include <kernel.h>
 #include "virtio_net.h"
 
+buddy_allocator_t tx_buffer_allocator;
+
 static bool virtio_net_hw_enable(pci_dev_t pdev) {
 	uint32_t bar4 = pci_read(pdev, PCI_BAR4);
 	uint32_t bar5 = pci_read(pdev, PCI_BAR5);
@@ -289,7 +291,7 @@ bool virtio_send_packet(void *data, uint16_t len) {
 
 	memset(hdr, 0, VNET_HDR_SIZE);
 
-	void* pay = kmalloc_aligned(len, 16);
+	void *pay = buddy_malloc(&tx_buffer_allocator, len);
 	if (!pay) {
 		virtq_free_desc(tq, (uint16_t) h);
 		virtq_free_desc(tq, (uint16_t) p);
@@ -324,7 +326,7 @@ static void vnet_tx_complete(void) {
 
 		void* paybuf = (void *) (uintptr_t) tq->desc[pay].addr;
 		if (paybuf) {
-			kfree_aligned(paybuf);
+			buddy_free(&tx_buffer_allocator, paybuf);
 		}
 
 		virtq_free_desc(tq, pay);
@@ -442,6 +444,8 @@ bool EXPORTED MOD_INIT_SYM(KMOD_ABI)(void) {
 		dprintf("virtio-net: no modern virtio-net device\n");
 		return false;
 	}
+	/* We use a buddy allocator for TX buffer as it is faster and aligned as standard */
+	buddy_init(&tx_buffer_allocator, 6, 22, 22);
 	if (!virtio_net_start(&dev)) {
 		return false;
 	}

modules/virtio_net/virtio_net.h

@@ -5,77 +5,77 @@
 
 /**
  * @def VIRTIO_DEVICE_NET_MODERN
- * @brief PCI device ID for modern (non-transitional) virtio-net.
+ * @brief PCI device ID for modern (non-transitional) virtio-net
  *
- * Transitional virtio-net typically reports 0x1000; modern devices use 0x1041.
+ * Transitional virtio-net typically reports 0x1000; modern devices use 0x1041
  */
 #define VIRTIO_DEVICE_NET_MODERN 0x1041
 
 /**
  * @def VIRTIO_PCI_CAP_COMMON_CFG
- * @brief Vendor capability: common configuration MMIO window selector.
+ * @brief Vendor capability: common configuration MMIO window selector
  */
 #define VIRTIO_PCI_CAP_COMMON_CFG  1
 /**
  * @def VIRTIO_PCI_CAP_NOTIFY_CFG
- * @brief Vendor capability: notify (doorbell) MMIO window selector.
+ * @brief Vendor capability: notify (doorbell) MMIO window selector
  */
 #define VIRTIO_PCI_CAP_NOTIFY_CFG  2
 /**
  * @def VIRTIO_PCI_CAP_ISR_CFG
- * @brief Vendor capability: ISR byte MMIO window selector (legacy/INTx paths).
+ * @brief Vendor capability: ISR byte MMIO window selector (legacy/INTx paths)
  */
 #define VIRTIO_PCI_CAP_ISR_CFG     3
 /**
  * @def VIRTIO_PCI_CAP_DEVICE_CFG
- * @brief Vendor capability: device-specific configuration MMIO window selector.
+ * @brief Vendor capability: device-specific configuration MMIO window selector
  */
 #define VIRTIO_PCI_CAP_DEVICE_CFG  4
 
 /**
  * @def VIRTIO_F_VERSION_1
- * @brief Feature bit indicating the device speaks modern virtio 1.x.
+ * @brief Feature bit indicating the device speaks modern virtio 1.x
  *
- * This lives in the upper 32-bit feature word; keep the constant 64-bit.
+ * This lives in the upper 32-bit feature word; keep the constant 64-bit
  */
 #define VIRTIO_F_VERSION_1         (1ULL << 32)
 /**
  * @def VIRTIO_NET_F_MAC
- * @brief Feature bit: device exposes a fixed MAC address in its config space.
+ * @brief Feature bit: device exposes a fixed MAC address in its config space
  */
 #define VIRTIO_NET_F_MAC           (1 << 5)
 
 /**
  * @def VIRTQ_DESC_F_NEXT
- * @brief Virtqueue descriptor flag: this entry chains to `next`.
+ * @brief Virtqueue descriptor flag: this entry chains to `next`
  */
 #define VIRTQ_DESC_F_NEXT          1
 /**
  * @def VIRTQ_DESC_F_WRITE
- * @brief Virtqueue descriptor flag: device may write to this buffer.
+ * @brief Virtqueue descriptor flag: device may write to this buffer
  */
 #define VIRTQ_DESC_F_WRITE         2
 
 /**
  * @def VNET_QSIZE
- * @brief Driver’s requested queue size (clamped to device’s `queue_size`).
+ * @brief Driver’s requested queue size (clamped to device’s `queue_size`)
  */
 #define VNET_QSIZE                 256
 /**
  * @def VNET_RX_BUF_SIZE
- * @brief Size of each posted RX buffer (must cover virtio header + frame).
+ * @brief Size of each posted RX buffer (must cover virtio header + frame)
  */
 #define VNET_RX_BUF_SIZE           2048
 /**
  * @def VNET_HDR_SIZE
- * @brief Size of the mandatory virtio-net header prepended to each packet.
+ * @brief Size of the mandatory virtio-net header prepended to each packet
  */
 #define VNET_HDR_SIZE              (sizeof(virtio_net_hdr_t))
 
 /**
- * @brief Device-specific configuration for virtio-net.
+ * @brief Device-specific configuration for virtio-net
  *
- * Only the fixed MAC and a status word are consumed by this driver.
+ * Only the fixed MAC and a status word are consumed by this driver
  */
 typedef struct {
 	/** Station MAC address provided by the device when @ref VIRTIO_NET_F_MAC is set */
@@ -85,11 +85,11 @@ typedef struct {
 } __attribute__((packed)) virtio_net_config_t;
 
 /**
- * @brief Minimal virtio-net per-packet header (no offloads negotiated).
+ * @brief Minimal virtio-net per-packet header (no offloads negotiated)
  *
- * This header precedes every TX and RX payload. All fields are little-endian.
+ * This header precedes every TX and RX payload. All fields are little-endian
  * The driver zero-initialises it; the device may fill some fields on RX when
- * certain features are negotiated (not used here).
+ * certain features are negotiated (not used here)
  */
 typedef struct {
 	/** Flags (eg. checksum present); unused in this minimal driver */
@@ -109,15 +109,15 @@ typedef struct {
 } __attribute__((packed)) virtio_net_hdr_t;
 
 /**
- * @brief One-byte ISR window (legacy/INTx ack by read-to-clear).
+ * @brief One-byte ISR window (legacy/INTx ack by read-to-clear)
  */
 typedef struct {
 	/** Interrupt cause byte; reading acknowledges in legacy paths */
 	volatile uint8_t isr;
 } virtio_pci_isr_t;
 
 /**
- * @brief In-memory representation of a single virtqueue.
+ * @brief In-memory representation of a single virtqueue
  */
 typedef struct virtq {
 	/** Queue size (power-of-two), negotiated from device `queue_size` */
@@ -139,7 +139,7 @@ typedef struct virtq {
 } virtq_t;
 
 /**
- * @brief Driver static state for a single virtio-net instance.
+ * @brief Driver static state for a single virtio-net instance
  */
 static struct {
 	/** Pointer to the common configuration MMIO window (from caps) */