sockets: use new iovec interface

kernel-src/fs/sockfs.c

@@ -26,38 +26,12 @@ int sockfs_close(vnode_t *node, int flags, cred_t *cred) {
 
 int sockfs_read(vnode_t *node, void *buffer, size_t size, uintmax_t offset, int flags, size_t *readc, cred_t *cred) {
 	socket_t *socket = SOCKFS_SOCKET_FROM_NODE(node);
-	sockdesc_t desc = {
-		.addr = NULL,
-		.buffer = buffer,
-		.count = size,
-		.flags = flags,
-		.donecount = 0,
-		.ctrl = NULL,
-		.ctrllen = 0
-	};
-
-	int e = socket->ops->recv(socket, &desc);
-
-	*readc = desc.donecount;
-	return e;
+	return socket_read(socket, buffer, size, flags, readc);
 }
 
 int sockfs_write(vnode_t *node, void *buffer, size_t size, uintmax_t offset, int flags, size_t *writec, cred_t *cred) {
 	socket_t *socket = SOCKFS_SOCKET_FROM_NODE(node);
-	sockdesc_t desc = {
-		.addr = NULL,
-		.buffer = buffer,
-		.count = size,
-		.flags = flags,
-		.donecount = 0,
-		.ctrl = NULL,
-		.ctrllen = 0
-	};
-
-	int e = socket->ops->send(socket, &desc);
-
-	*writec = desc.donecount;
-	return e;
+	return socket_write(socket, buffer, size, flags, writec);
 }
 
 int sockfs_poll(vnode_t *node, polldata_t *data, int events) {

kernel-src/include/kernel/iovec.h

@@ -47,12 +47,12 @@ int iovec_iterator_copy_from_buffer(iovec_iterator_t *iovec_iterator, void *buff
 // copies byte_count bytes from the iovec_iterator into the ringbuffer
 // returns the number of bytes copied
 // NOTE: the iovec can have userspace addresses. in this case, the copy can possibly fail by returning RINGBUFFER_USER_COPY_FAILED
-size_t iovec_iterator_copy_to_ringbuffer(iovec_iterator_t *iovec_iterator, ringbuffer_t *ringbuffer, size_t byte_count);
+size_t iovec_iterator_write_to_ringbuffer(iovec_iterator_t *iovec_iterator, ringbuffer_t *ringbuffer, size_t byte_count);
 
 // copies byte_count bytes from the ringbuffer into the iovec_iterator
 // returns the number of bytes copied
 // NOTE: the iovec can have userspace addresses. in this case, the copy can possibly fail with RINGBUFFER_USER_COPY_FAILED
-size_t iovec_iterator_copy_from_ringbuffer(iovec_iterator_t *iovec_iterator, ringbuffer_t *ringbuffer, size_t byte_count);
+size_t iovec_iterator_read_from_ringbuffer(iovec_iterator_t *iovec_iterator, ringbuffer_t *ringbuffer, size_t byte_count);
 
 // copies byte_count bytes from the ringbuffer into the iovec_iterator, without moving the ringbuffer read pointer
 // returns the number of bytes copied

kernel-src/include/kernel/net.h

@@ -78,7 +78,7 @@ void udp_process(netdev_t *netdev, void *buffer, uint32_t ip);
 void arp_process(netdev_t *netdev, void *buffer);
 void tcp_process(netdev_t *netdev, void *buffer, ipv4frame_t *ipv4frame);
 int arp_lookup(netdev_t *netdev, uint32_t ip, mac_t *mac);
-int udp_sendpacket(void *buffer, size_t packetsize, uint32_t ip, uint16_t srcport, uint16_t dstport, netdev_t *broadcastdev);
+int udp_sendpacket(iovec_iterator_t *iovec_iterator, size_t packetsize, uint32_t ip, uint16_t srcport, uint16_t dstport, netdev_t *broadcastdev);
 int ipv4_sendpacket(void *buffer, size_t packetsize, uint32_t ip, int proto, netdev_t *broadcastdev);
 size_t ipv4_getmtu(uint32_t ip);
 uint32_t ipv4_getnetdevip(uint32_t ip);

kernel-src/include/kernel/sock.h

@@ -39,7 +39,7 @@ typedef struct {
 
 typedef struct {
 	sockaddr_t *addr;
-	void *buffer;
+	iovec_iterator_t iovec_iterator;
 	size_t count;
 	uintmax_t flags;
 	size_t donecount;
@@ -79,6 +79,53 @@ typedef struct {
 #define SOCKET_TYPE_TCP 2
 #define SOCKFS_SOCKET_FROM_NODE(nodep) (((socketnode_t *)(nodep))->socket)
 
+static inline int socket_read(socket_t *socket, void *buffer, size_t size, uintmax_t flags, size_t *bytes_read) {
+	sockdesc_t desc = {
+		.addr = NULL,
+		.count = size,
+		.flags = flags,
+		.donecount = 0,
+		.ctrl = NULL,
+		.ctrllen = 0
+	};
+
+	iovec_t iovec = {
+		.addr = buffer,
+		.len = size
+	};
+
+	iovec_iterator_init(&desc.iovec_iterator, &iovec, 1);
+
+	int e = socket->ops->recv(socket, &desc);
+
+	*bytes_read = desc.donecount;
+
+	return e;
+}
+
+static inline int socket_write(socket_t *socket, void *buffer, size_t size, uintmax_t flags, size_t *bytes_written) {
+	sockdesc_t desc = {
+		.addr = NULL,
+		.count = size,
+		.flags = flags,
+		.donecount = 0,
+		.ctrl = NULL,
+		.ctrllen = 0
+	};
+
+	iovec_t iovec = {
+		.addr = buffer,
+		.len = size
+	};
+
+	iovec_iterator_init(&desc.iovec_iterator, &iovec, 1);
+
+	int e = socket->ops->send(socket, &desc);
+
+	*bytes_written = desc.donecount;
+	return e;
+}
+
 static inline int sock_convertaddress(sockaddr_t *sockaddr, abisockaddr_t *abisockaddr) {
 	switch (abisockaddr->type) {
 		case AF_INET:

kernel-src/io/iovec.c

@@ -119,7 +119,7 @@ int iovec_iterator_copy_from_buffer(iovec_iterator_t *iovec_iterator, void *buff
 	return error;
 }
 
-size_t iovec_iterator_copy_to_ringbuffer(iovec_iterator_t *iovec_iterator, ringbuffer_t *ringbuffer, size_t byte_count) {
+size_t iovec_iterator_write_to_ringbuffer(iovec_iterator_t *iovec_iterator, ringbuffer_t *ringbuffer, size_t byte_count) {
 	size_t iterator_offset_save = iovec_iterator->total_offset;
 	size_t remaining_total = min(byte_count, iovec_iterator->total_size - iovec_iterator->total_offset);
 	size_t total_done = 0;
@@ -140,6 +140,7 @@ size_t iovec_iterator_copy_to_ringbuffer(iovec_iterator_t *iovec_iterator, ringb
 			break;
 
 		iovec_iterator_skip(iovec_iterator, copy_current);
+		remaining_total -= current_done;
 		total_done += current_done;
 
 		if (total_done == byte_count || iovec_iterator_finished(iovec_iterator))
@@ -154,7 +155,7 @@ size_t iovec_iterator_copy_to_ringbuffer(iovec_iterator_t *iovec_iterator, ringb
 	return total_done;
 }
 
-size_t iovec_iterator_peek_from_ringbuffer(iovec_iterator_t *iovec_iterator, ringbuffer_t *ringbuffer, size_t byte_count, uintmax_t ringbuffer_offset) {
+size_t iovec_iterator_peek_from_ringbuffer(iovec_iterator_t *iovec_iterator, ringbuffer_t *ringbuffer, size_t offset, size_t byte_count) {
 	size_t iterator_offset_save = iovec_iterator->total_offset;
 	size_t remaining_total = min(byte_count, iovec_iterator->total_size - iovec_iterator->total_offset);
 	size_t total_done = 0;
@@ -164,12 +165,13 @@ size_t iovec_iterator_peek_from_ringbuffer(iovec_iterator_t *iovec_iterator, rin
 		size_t remaining_current = iovec_iterator->current->len - iovec_iterator->current_offset;
 		size_t copy_current = min(remaining_current, remaining_total);
 
-		current_done = ringbuffer_peek(ringbuffer, (void *)((uintptr_t)iovec_iterator->current->addr + iovec_iterator->current_offset), total_done, copy_current);
+		current_done = ringbuffer_peek(ringbuffer, (void *)((uintptr_t)iovec_iterator->current->addr + iovec_iterator->current_offset), offset + total_done, copy_current);
 		if (current_done == 0 || current_done == RINGBUFFER_USER_COPY_FAILED)
 			break;
 
 		iovec_iterator_skip(iovec_iterator, copy_current);
 		total_done += current_done;
+		remaining_total -= current_done;
 
 		if (total_done == byte_count || iovec_iterator_finished(iovec_iterator))
 			break;
@@ -183,8 +185,8 @@ size_t iovec_iterator_peek_from_ringbuffer(iovec_iterator_t *iovec_iterator, rin
 	return total_done;
 }
 
-size_t iovec_iterator_copy_from_ringbuffer(iovec_iterator_t *iovec_iterator, ringbuffer_t *ringbuffer, size_t byte_count) {
-	size_t done = iovec_iterator_peek_from_ringbuffer(iovec_iterator, ringbuffer, byte_count, 0);
+size_t iovec_iterator_read_from_ringbuffer(iovec_iterator_t *iovec_iterator, ringbuffer_t *ringbuffer, size_t byte_count) {
+	size_t done = iovec_iterator_peek_from_ringbuffer(iovec_iterator, ringbuffer, 0, byte_count);
 	if (done == RINGBUFFER_USER_COPY_FAILED)
 		return done;
 

kernel-src/io/net/localsocket.c

@@ -363,7 +363,7 @@ static int localsock_send(socket_t *socket, sockdesc_t *sockdesc) {
 		goto leave;
 	}
 
-	sockdesc->donecount = ringbuffer_write(&peer->ringbuffer, sockdesc->buffer, sockdesc->count);
+	sockdesc->donecount = iovec_iterator_write_to_ringbuffer(&sockdesc->iovec_iterator, &peer->ringbuffer, sockdesc->count);
 	if (sockdesc->donecount == RINGBUFFER_USER_COPY_FAILED) {
 		error = EFAULT;
 		goto leave;
@@ -467,13 +467,13 @@ static int localsock_recv(socket_t *socket, sockdesc_t *sockdesc) {
 	}
 
 	if (flags & SOCKET_RECV_FLAGS_PEEK) {
-		sockdesc->donecount = ringbuffer_peek(&localsocket->ringbuffer, sockdesc->buffer, 0, recvcount);
+		sockdesc->donecount = iovec_iterator_peek_from_ringbuffer(&sockdesc->iovec_iterator, &localsocket->ringbuffer, 0, recvcount);
 		if (sockdesc->donecount == RINGBUFFER_USER_COPY_FAILED) {
 			error = EFAULT;
 			goto leave;
 		}
 	} else {
-		sockdesc->donecount = ringbuffer_read(&localsocket->ringbuffer, sockdesc->buffer, recvcount);
+		sockdesc->donecount = iovec_iterator_read_from_ringbuffer(&sockdesc->iovec_iterator, &localsocket->ringbuffer, recvcount);
 		if (sockdesc->donecount == RINGBUFFER_USER_COPY_FAILED) {
 			error = EPIPE;
 			goto leave;

kernel-src/io/net/tcp.c

@@ -1305,7 +1305,7 @@ static int tcp_send(socket_t *socket, sockdesc_t *sockdesc) {
 	size_t writesize = min(TCB_RINGBUFFER_SIZE - RINGBUFFER_DATACOUNT(&tcb->transmitbuffer), sockdesc->count);
 	__assert(writesize);
 
-	sockdesc->donecount = ringbuffer_write(&tcb->transmitbuffer, sockdesc->buffer, writesize);
+	sockdesc->donecount = iovec_iterator_write_to_ringbuffer(&sockdesc->iovec_iterator, &tcb->transmitbuffer, writesize);
 	if (sockdesc->donecount == RINGBUFFER_USER_COPY_FAILED) {
 		error = EFAULT;
 		goto leave;
@@ -1386,13 +1386,13 @@ static int tcp_recv(socket_t *socket, sockdesc_t *sockdesc) {
 
 	size_t copycount = min(RINGBUFFER_DATACOUNT(&tcb->receivebuffer), sockdesc->count);
 	if (flags & SOCKET_RECV_FLAGS_PEEK) {
-		sockdesc->donecount = ringbuffer_peek(&tcb->receivebuffer, sockdesc->buffer, 0, copycount);
+		sockdesc->donecount = iovec_iterator_peek_from_ringbuffer(&sockdesc->iovec_iterator, &tcb->receivebuffer, 0, copycount);
 		if (sockdesc->donecount == RINGBUFFER_USER_COPY_FAILED) {
 			error = EFAULT;
 			goto cleanup;
 		}
 	} else {
-		sockdesc->donecount = ringbuffer_read(&tcb->receivebuffer, sockdesc->buffer, copycount);
+		sockdesc->donecount = iovec_iterator_read_from_ringbuffer(&sockdesc->iovec_iterator, &tcb->receivebuffer, copycount);
 		if (sockdesc->donecount == RINGBUFFER_USER_COPY_FAILED) {
 			error = EFAULT;
 			goto cleanup;

kernel-src/io/net/udp.c

@@ -32,7 +32,7 @@ typedef struct {
 static socket_t *ports[65536];
 static spinlock_t portlock;
 
-int udp_sendpacket(void *buffer, size_t size, uint32_t ip, uint16_t srcport, uint16_t dstport, netdev_t *broadcastnetdev) {
+int udp_sendpacket(iovec_iterator_t *iovec_iterator, size_t size, uint32_t ip, uint16_t srcport, uint16_t dstport, netdev_t *broadcastnetdev) {
 	void *newbuff = alloc(size + sizeof(udpframe_t));
 	if (newbuff == NULL)
 		return ENOMEM;
@@ -45,7 +45,7 @@ int udp_sendpacket(void *buffer, size_t size, uint32_t ip, uint16_t srcport, uin
 	};
 
 	memcpy(newbuff, &frame, sizeof(udpframe_t));
-	int error = USERCOPY_POSSIBLY_FROM_USER((void *)((uintptr_t)newbuff + sizeof(udpframe_t)), buffer, size);
+	int error = iovec_iterator_copy_to_buffer(iovec_iterator, (void *)((uintptr_t)newbuff + sizeof(udpframe_t)), size);
 	if (error)
 		goto leave;
 
@@ -203,7 +203,7 @@ static int udp_send(socket_t *socket, sockdesc_t *sockdesc) {
 		socket->state = SOCKET_STATE_BOUND;
 	}
 
-	e = udp_sendpacket(sockdesc->buffer, sockdesc->count, sockdesc->addr ? sockdesc->addr->ipv4addr.addr : udpsocket->peeraddress, 
+	e = udp_sendpacket(&sockdesc->iovec_iterator, sockdesc->count, sockdesc->addr ? sockdesc->addr->ipv4addr.addr : udpsocket->peeraddress,
 			udpsocket->port, sockdesc->addr ? sockdesc->addr->ipv4addr.port : udpsocket->peerport, udpsocket->socket.netdev);
 
 	sockdesc->donecount = sockdesc->count;
@@ -261,16 +261,20 @@ static int udp_recv(socket_t *socket, sockdesc_t *sockdesc) {
 
 	// this abuses an asterisk on the ringbuffer implementation that makes it not need locking when
 	// something is reading and something is writing simultaneously.
-	// this is done to not need to disable interrupts as the ringbuffer_peek could take a fault
+	// this is done to not need to disable interrupts as the iovec_peek_from_ringbuffer could take a fault
 	// multiple readers cannot happen because of the socket mutex and the write on the side of the
-	// NIC is protected by the ringbuffer lock
+	// NIC is protected by the ringbuffer lock. The ringbuffer lock is taken when reading the desc to ensure that
+	// the whole packet has reached the ringbuffer
 	size_t copycount;
 	dataheader_t header;
 
+	long ipl = spinlock_acquireraiseipl(&udpsocket->ringbufferlock, IPL_DPC);
 	__assert(ringbuffer_peek(&udpsocket->ringbuffer, &header, 0, sizeof(header)) == sizeof(header));
+	spinlock_releaseloweripl(&udpsocket->ringbufferlock, ipl);
+
 	copycount = min(header.length, sockdesc->count);
 
-	sockdesc->donecount = ringbuffer_peek(&udpsocket->ringbuffer, sockdesc->buffer, sizeof(header), copycount);
+	sockdesc->donecount = iovec_iterator_peek_from_ringbuffer(&sockdesc->iovec_iterator, &udpsocket->ringbuffer, sizeof(header), copycount);
 	if (sockdesc->donecount == RINGBUFFER_USER_COPY_FAILED) {
 		e = EFAULT;
 		goto cleanup;
@@ -279,11 +283,8 @@ static int udp_recv(socket_t *socket, sockdesc_t *sockdesc) {
 	__assert(sockdesc->donecount == copycount);
 
 	// remove the message if not asked to peek
-	if ((flags & SOCKET_RECV_FLAGS_PEEK) == 0) {
-		bool status = spinlock_acquireirqclear(&udpsocket->ringbufferlock);
+	if ((flags & SOCKET_RECV_FLAGS_PEEK) == 0)
 		ringbuffer_truncate(&udpsocket->ringbuffer, sizeof(header) + header.length);
-		spinlock_releaseirqrestore(&udpsocket->ringbufferlock, status);
-	}
 
 	sockdesc->donecount = copycount;
 	sockdesc->addr->ipv4addr.addr = header.peer;

kernel-src/sys/syscalls/recvmsg.c

@@ -28,10 +28,10 @@ syscallret_t syscall_recvmsg(context_t *, int fd, msghdr_t *umsghdr, int flags)
 		return ret;
 	}
 
-	void *buffer = vmm_map(NULL, buffersize, VMM_FLAGS_ALLOCATE, ARCH_MMU_FLAGS_READ | ARCH_MMU_FLAGS_WRITE | ARCH_MMU_FLAGS_NOEXEC, NULL);
-	if (buffer == NULL) {
-		ret.errno = ENOMEM;
-		goto cleanup;
+	if (iovec_user_check(msghdr.iov, msghdr.iovcount) == false) {
+		ret.errno = EFAULT;
+		sock_freemsghdr(&msghdr);
+		return ret;
 	}
 
 	file = fd_get(fd);
@@ -57,7 +57,6 @@ syscallret_t syscall_recvmsg(context_t *, int fd, msghdr_t *umsghdr, int flags)
 
 	sockdesc_t desc = {
 		.addr = &sockaddr,
-		.buffer = buffer,
 		.count = buffersize,
 		.flags = fileflagstovnodeflags(file->flags) | recvflags,
 		.donecount = 0,
@@ -66,18 +65,12 @@ syscallret_t syscall_recvmsg(context_t *, int fd, msghdr_t *umsghdr, int flags)
 		.ctrldone = 0
 	};
 
+	iovec_iterator_init(&desc.iovec_iterator, msghdr.iov, msghdr.iovcount);
+
 	ret.errno = socket->ops->recv(socket, &desc);
 	if (ret.errno)
 		goto cleanup;
 
-	uintmax_t iovoffset = 0;
-	for (int i = 0; i < msghdr.iovcount; ++i) {
-		ret.errno = usercopy_touser(msghdr.iov[i].addr, (void *)((uintptr_t)buffer + iovoffset), msghdr.iov[i].len);
-		if (ret.errno)
-			goto cleanup;
-		iovoffset += msghdr.iov[i].len;
-	}
-
 	if (msghdr.addr) {
 		abisockaddr_t abisockaddr;
 		ret.errno = sock_addrtoabiaddr(socket->type, &sockaddr, &abisockaddr);
@@ -120,9 +113,6 @@ syscallret_t syscall_recvmsg(context_t *, int fd, msghdr_t *umsghdr, int flags)
 	if (file)
 		fd_release(file);
 
-	if (buffer)
-		vmm_unmap(buffer, buffersize, 0);
-
 	sock_freemsghdr(&msghdr);
 	return ret;
 }

kernel-src/sys/syscalls/sendmsg.c

@@ -23,20 +23,10 @@ syscallret_t syscall_sendmsg(context_t *, int fd, msghdr_t *umsghdr, int flags)
 		return ret;
 	}
 
-	void *buffer = vmm_map(NULL, buffersize, VMM_FLAGS_ALLOCATE, ARCH_MMU_FLAGS_READ | ARCH_MMU_FLAGS_WRITE | ARCH_MMU_FLAGS_NOEXEC, NULL);
-	if (buffer == NULL) {
-		ret.errno = ENOMEM;
-		goto cleanup;
-	}
-
-	uintmax_t iovoffset = 0;
-
-	for (int i = 0; i < msghdr.iovcount; ++i) {
-		ret.errno = usercopy_fromuser((void *)((uintptr_t)buffer + iovoffset), msghdr.iov[i].addr, msghdr.iov[i].len);
-		if (ret.errno)
-			goto cleanup;
-
-		iovoffset += msghdr.iov[i].len;
+	if (iovec_user_check(msghdr.iov, msghdr.iovcount) == false) {
+		ret.errno = EFAULT;
+		sock_freemsghdr(&msghdr);
+		return ret;
 	}
 
 	file = fd_get(fd);
@@ -61,7 +51,6 @@ syscallret_t syscall_sendmsg(context_t *, int fd, msghdr_t *umsghdr, int flags)
 
 	sockdesc_t desc = {
 		.addr = msghdr.addr ? &sockaddr : NULL,
-		.buffer = buffer,
 		.count = buffersize,
 		.flags = fileflagstovnodeflags(file->flags) | ((flags & MSG_NOSIGNAL) ? SOCKET_SEND_FLAGS_NOSIGNAL : 0),
 		.donecount = 0,
@@ -70,6 +59,8 @@ syscallret_t syscall_sendmsg(context_t *, int fd, msghdr_t *umsghdr, int flags)
 		.ctrldone = 0
 	};
 
+	iovec_iterator_init(&desc.iovec_iterator, msghdr.iov, msghdr.iovcount);
+
 	ret.errno = socket->ops->send(socket, &desc);
 	ret.ret = ret.errno ? -1 : desc.donecount;
 
@@ -77,9 +68,6 @@ syscallret_t syscall_sendmsg(context_t *, int fd, msghdr_t *umsghdr, int flags)
 	if (file)
 		fd_release(file);
 
-	if (buffer)
-		vmm_unmap(buffer, buffersize, 0);
-
 	sock_freemsghdr(&msghdr);
 	return ret;
 }